caprock

Changelog

What's new in Caprock.

Every release, newest first. The desktop app updates itself: when a release is out, a card in its corner offers Update and restart, and your sessions keep running. The .deb and .rpm packages are updated by downloading the new one.

Upgrade the daemon alone

  1. Homebrew

  2. Scoop

  3. Go

  4. Then restart, so the new binary runs

The daemon's release check is off until you turn it on, and sends nothing about you; it tells you a release is out and never installs it.

v0.169.02026-10-31

ui

  • —At a glance has the share button the other figure panels have: picture without the Charts/Numbers switch or fold toggles, and Copy as text with the Numbers view's tables. So do All time on Home, Last 30 days and Compactions on Cost, Hermes Agent on Cost, Compactions on Week, and Lifetime, Daily cost, Top projects (names hidden by default) and Model mix on Lifetime.
  • —ui: the Agent panel's session chip says uptime 8d instead of started 8d ago, so it stays on the model and context line; the start time is in its tooltip.
  • —ui: the subagent frame in the Agent panel is drawn with the Caprock characters instead of a thin step line. By default the last hour is a chart of faces: twelve five-minute columns, each a stack of faces as tall as the most subagents at once, with the peak and the time it was reached (peak 5 at 09:42); past six at once one face stands for several. Two more looks sit behind three small dots in the frame's corner: Crew, one face per working subagent with the finished faded, and Numbers, the crowd large with N working now. The faces follow the one picked for Claude in Settings, and the choice of look is remembered.
  • —app: the "Caprock vX is available" card on Now updates in place in the desktop app — Update to vX — Restart, a progress bar while it downloads, then the restart — the same one click as the status strip, instead of a Download link to the release page. A browser and installs that cannot update themselves still get the command or the release page.
  • —ui: a share button can no longer leave the sheet stuck on "Drawing…" with every button off. A press on a card whose figures are still being counted waits at most six seconds, saying Counting this month…, then sends the card as shown (its corner still names the reading); the picture embeds only the fonts the page already loaded, from the browser's cache, instead of fetching eight font files from a daemon busy counting the month; and drawing has a 30-second limit, after which the buttons come back with Could not draw the card — try again. Panel pictures get the same limit.
v0.168.02026-10-31

compact

  • —every compaction is now a stat: what it cost and what it has saved since, both estimated from the pricing table and marked ≈. The Agent panel's Context shows the latest one on one line (32h ago · 442k → 19k · cost ≈ $0.09) with a small curve of the saving since and saved ≈ at its end; the ⓘ beside ~15% stays says what /compact does and how each figure is estimated. Cost and Week gain a Compactions panel: the count, cost, saved and net for the period. Every compaction already recorded is filled in once on the first start, its summary's length read back from the transcript. New GET /v1/stats/compactions; GET /v1/sessions/{id}/compaction gains estimate (migration 0077).
  • —pricing: Claude Sonnet 5.5 cache reads are priced at $0.10 per million tokens from 7 October 2026, as Anthropic charges (0.05x the input price); the table had kept $0.20. Turns from that day are repriced once, openly, and the session and daily totals move by the difference. GPT-6.1 Terra gets its own row, so its Codex turns are priced instead of counted as unpriced. Table version 2026-10-31.1.
v0.167.02026-10-30

ui

  • —the Agent panel header shows the session's reasoning effort beside its model, Opus 5.5 · high, as Claude Code's own hooks report it. The daemon now keeps the newest known level per session (low, medium, high, xhigh, max; migration 0076) and returns it as effort; a session whose hooks never named one shows the model alone.
  • —ui: a shared panel's picture is the title and the content alone: fold chevrons, More/Less and Hide toggles, Settings links, a row's ways in, tooltips and hover-only buttons are left out. With Hide project names on, names no longer slip through in part: a folder's last segment, a GitHub slug's owner and repository and a name in another letter case are hidden too, and an activity line such as "editing caprock/internal/api/x.go" reads "editing path 1", its file names and branches replaced.
v0.166.02026-10-29

ui

  • —Files: → expands a folder in place, its contents right below it one level in, as Finder's list view does; → again steps onto its first file, ← folds it or steps back to the folder holding the row. A triangle on each folder row does the same with the mouse; Enter still goes into the folder. Nested rows show git marks, size and date, are tinted in compare mode, stay under What the agent changed, and can be marked, viewed, edited, copied, moved and sent to the Trash.
  • —ui: the top of the Agent panel (⌘I) takes about half the height it did and cuts nothing. One row carries the agent's own mark (a click on it still picks the session's face), the session's name once, the project only when it differs, the branch, the state and the permission mode. Burn rate stays big with tokens and spend on one line under it, beside a new chart of how many subagents worked each minute of the last hour, with how many work now, their types and the hour's peak; with none, the chart says so. The model, the context fill and the start are small chips under them.
  • —api: GET /v1/sessions/{id}/subagents?hour=1 adds the last hour's subagent count a minute (hour: {start, step_ms, counts, peak}).
  • —ui: a folded Plan limits panel is no longer an empty row. Its header keeps one thin bar per window the panel shows — Claude's 5 hours and week, and Codex's when it reports them — each with its short name, the percentage and the reset time, coloured as the gauges are (amber from 60%, red above 85%, grey when the reading is stale). As the row narrows the reset times go first, then the names; the bars and percentages stay, on one line.
v0.165.12026-10-29

app

  • —the window no longer flips to "the daemon stopped" and back to the dashboard on a brief hiccup. A running daemon is kept through 5 s of failed checks (20 s when it is only slow), and a slow /v1/status no longer counts at all; a real stop (its runtime.json gone, or its process exited) is still seen at once. When the app does count the daemon as gone, app.log says why.
v0.165.02026-10-28

ui

  • —Files, a two-panel file manager in the spirit of Norton Commander that knows about your agents (⇧⌘X, Ctrl+Shift+X on Windows and Linux; ⌘K → Open Files; Files in a project's menu). Your checkout on the left, an agent's worktree on the right: Compare tints what differs and what is on one side only, What <agent> changed filters to that worktree's changes, View shows the file or its diff with + and − counts. Copy, move or rename, new folder and delete (F5–F8, or ⌥5–⌥8 on a Mac), each confirmed by name; delete only ever moves to the Trash. Reaches only the project's folder and its worktrees, and is read-only from a phone or another computer. A Norton look paints it blue with cyan frames and the F-key bar.
  • —ui: every figures panel can be shared — Live pulse, Repos, CI/CD, Plan limits, Most expensive, Releases, Plan value and the Cost panels get a share icon in their header: Copy image, Save PNG, Copy as text (Markdown) and Post to X, Threads or LinkedIn through the share card's own sheet. Hide project names, on by default and remembered, turns projects, sessions, branches, folders, pull request titles and logins into "project 1", "session 1"… in the picture and the text alike. Drawn on your machine; only the X, Threads and LinkedIn buttons upload the picture, to caprock.dev, as for the share card.
  • —ci: the Windows Go leg, the slowest job in a CI run, is split into three jobs that run side by side (lint and smoke, the slowest package's tests, the rest of the tests), and a pull request that touches only the changelog, its fragments or .ai/ docs — a release pull request — skips the build and test jobs while the docs checks still run.
  • —ui: the agent cockpit's header names the agent in a small muted label and puts the session's configuration where the big name was: the model and how full its context is (Opus 5.5 · 62% ctx). Under it, the subagents working now are counted by type (3 subagents · 2 Explore, 1 general-purpose), and a type opens the first of its kind in the Subagents section. A figure the daemon does not report is left out.
  • —ui: the tab in front is framed in the accent on every side, the Dashboard tab in front a step heavier, and behind it keeps an accent outline, so it is the one tab the eye finds. The Pipeline and Mail tabs are quiet beside it now, drawn like a session tab until brought forward. The Pipeline tab closes from behind too (its × on hover), and with ⌘W (Ctrl+Shift+W) while it is in front; it stays closed until ⌘P opens it again.
  • —ios: the iPhone app's upload to App Store Connect is no longer refused over export compliance. The build no longer claims an approved compliance code it does not have; App Store Connect asks the encryption questions on the build instead.
  • —ui: the Repos panel's last column (Release) no longer runs past the panel's right edge in a mid-width window: a letter-spaced heading wider than its column no longer widens the column, and the table drops to its narrower layout a little earlier.
v0.164.12026-10-28

ui

  • —the app window uses less memory with terminal tabs left in the background. A tab out of sight for five minutes now gives its terminal's GPU context back, and every closed or replaced terminal gives its context back at once instead of when the page next cleans up. Measured in the app with eight terminal tabs, a long chat and the Dashboard open, after 25 minutes: the window's graphics memory went from 125 to 80 MB and the whole app from 478 to 432 MB. Coming back to a tab looks as it did: the terminal is drawn in the frame it appears in.
v0.164.02026-10-28

ios

  • —Caprock for iPhone, version 1.0, ready for the App Store, in ios-app/: the phone remote in a Capacitor shell, bundled rather than loaded from m.caprock.dev. Pair by scanning the computer's code with the app's own camera, or paste the link; dictate in any language the iPhone recognises on its own, with nothing sent anywhere and a line when a language is not on the device; lock the app behind Face ID; the pairing kept in the Keychain; Try it without a Mac on bundled sample data that connects to nothing; the app's own Settings (paired computer, voice language, Face ID, unpair). Nothing in the app is for sale. A manual workflow (ios-release.yml) archives, signs and uploads a build to App Store Connect; the listing texts are in ios-app/store/. No push notifications in 1.0; they follow in 1.1 (ADR-083).
  • —machines: start an agent on another computer from this one. New agent gains a Machine field when you have linked a computer: pick it, and the folders, agents and models are read from that computer, the agent runs there, and its tab opens here. A computer that cannot take a start is listed with why (offline, waiting for Allow, removed there), and a refusal is said in words, a folder missing there included.
  • —machines: one list across your computers. Sessions on the other computers that need you sort first in the sidebar, Now shows them under On other computers, and the app tells you when one needs you, naming the computer. A computer that goes away is shown as offline with when it last answered, not with the sessions it had. The daemon reads each computer once for every window (GET /v1/machines/sessions) and listens to it for prompts only while a window is open.
  • —machines: a guide for trying it on two computers, docs/machines.md, and an end-to-end test with two real daemons and a local relay in the smoke suite.
v0.163.12026-10-28

app

  • —a daemon that answers slowly on a loaded computer no longer flips the window to "The Caprock daemon stopped". A daemon that takes the connection and answers late is kept for 20 seconds; no connection at all, or silence past that, shows it as stopped.
v0.163.02026-10-27

browser

  • —Open on my phone in phone mode. A local dev server's page (Vite, Next, webpack on localhost) opens on a real phone on the same Wi-Fi: a QR code, scanned with the phone's camera. Caprock shares that one server through a proxy on this computer's Wi-Fi address and a random port, behind a link only the QR carries; live reload works. Sharing stops on Stop sharing, after 30 minutes unused, after 8 hours, or when Caprock quits (ADR-087).
  • —ui: folded, the CI section says "#220 merged" (or "closed") for a pull request that is done, instead of a tally of its old checks.
  • —queue: a message waiting in a session's queue is no longer lost when the session ends. When the process ends, the waiting messages are kept for 7 days, shown under the conversation as Not sent: the session ended, each with Send to a new session (the same agent in the same folder, continuing the conversation when it can; the message then goes in through that session's queue), Copy and Delete, and Send all for several, in order, one per turn. A /clear no longer empties the queue: it moves to the new session and carries on delivering there. Migration 0075 (ADR-053, amendment of 2026-10-27).
  • —ui: a pull request with failed checks no longer says "ready to merge". GitHub calls it clean when no check is required, and Caprock repeated that beside the red checks; it now says "mergeable, checks failed", or "checks running" while they run.
  • —ui: the window holds much less memory over a long day. The Live pulse kept every tool's whole answer (command output, files read in full, an edit's original file) for up to 4,000 events a track; it now keeps what the bars read. An open chat following a busy agent no longer grows for as long as it stays open: past 240 messages it keeps the newest 120 and pages the rest back in when scrolled to. A session's timeline left open keeps 1,000 live events instead of 5,000. The agent cockpit keeps a result's outcome, not its output, and clipped text no longer keeps the full original alive.
v0.162.02026-10-27

share

  • —X, Threads and LinkedIn now put the card you chose in the post itself. Pressing one uploads that exact picture — the picture alone — to caprock.dev and opens the post with its link, whose preview is the card, so there is nothing to paste. The buttons' tooltips and the box after posting say so, and the box can take the picture down again (Remove the shared picture); Settings → Privacy lists every shared picture with Remove. The social site may keep the preview it already fetched. Share…, Copy image, Save image and My Telegram never upload. If the upload fails, the sheet falls back to the clipboard and the ⌘V steps, and says why in one line (ADR-086).
  • —phone: Share your week on the phone home draws the Week card and opens the phone's own share menu with the picture attached — on an iPhone, pick Threads or X. Nothing is uploaded on the phone.
  • —store: an upgrade that found no disk for the safety copy of the database used to say so only in the log. It is now recorded, and /v1/status (upgrade_copy_skipped), caprock status, Now and Status say that no copy was made, how much disk was free and needed, and that older Caprock versions cannot open the upgraded database.
  • —ui: while older history moves in after the 0073 upgrade, History, Memory's search and session pages say so in one quiet line, "Older history is still loading — N% moved", gone once the move ends.
v0.161.02026-10-27

cockpit

  • —Compact context, free. In the Context section, one click types Claude Code's own /compact into a session Caprock started — at once at its prompt, otherwise when its turn ends — and the session carries on in place with a summary; /clear is never typed. The collapsed Context line shows a small icon from 60% full; both are hidden under 100k tokens. Before the press it says what usually stays and where that comes from: your own compactions once five are measured, until then one machine's measurement (70 automatic compactions, 2026-09-29 to 2026-10-25, a median 15% kept). After it, the measured before and after, Compacted: 41.4k → 10.1k tokens — at once where Claude Code writes the size after on its compaction boundary, else with the next reply. A session another terminal holds is never typed into: it shows the command to type there and Continue here. Compaction boundaries are now read from transcripts, including the ones already on disk, once, in the background (GET /v1/sessions/{id}/compaction, POST /v1/agents/{id}/compact).
  • —store: an event names its session by an integer key instead of repeating the session's 36-character id in its row and in most indexes. A 1 GB database came out 128 MiB smaller after a VACUUM (measured 2026-10-19). The upgrade no longer rewrites the events table before the dashboard answers: the first start takes seconds, and the older events move into the new table in the background, newest first, 1 min 52 s for 603,141 of them on a 1.2 GB database (measured 2026-10-23). Until they are in, the Now screen and caprock status say so, older history is incomplete, the other agents' importers wait, and removing a session asks to try again later. A query on caprock.db joins sessions for a session's id (docs/schema.md). No earlier release opens the upgraded database, so the upgrade first keeps a copy of it as caprock.pre-0073.db in the data directory for 30 days: to go back, stop Caprock, put the copy in place of caprock.db, remove caprock.db-wal and caprock.db-shm, and install the older release. caprock up and the desktop app wait for the copy and say the database is upgrading.
v0.160.02026-10-26

ui

  • —the share sheet writes the post from your figures — "My agents shipped 48 PRs this week.", the other counts and which agents did the work — in place of one line with no number; the dollars stay on the card. After X, Threads or LinkedIn opens, a box lists the steps: the card is on your clipboard, the post is open, and pressing ⌘V in the post adds the card.
v0.159.12026-10-25

share

  • —X, Threads and LinkedIn get the card you chose, dark or paper: the post no longer carries a caprock.dev link, which made them show the site's dark preview instead, and a large line says to press ⌘V (Ctrl+V) in the post to add the card from the clipboard.
v0.159.02026-10-25

cost

  • —every cost at API list price now also says what the work was of your subscription, on a muted line under it: plan dollars and the share of the weekly limit ($50 at API · ≈$1.65 of plan · ≈2.3% of a week). On the Cost screen, the session card, a session's page, the Pipeline's rows, its Releases and both Most expensive lists. Each sign-in is priced by its own plan over the last 30 days and calibrated on its own last full week; hover for how, and why use outside Caprock makes the figures run a little high. Nothing is shown on a metered plan, with no plan stated, or before a weekly window has been seen.
  • —share: share from wherever you work: a small share icon on the sidebar's Today card, File → Share Your Figures… in the macOS app, and the share toast now shows over session tabs too.
  • —share: your week, every week: when a week closes (Sunday 18:00, or the first time you open Caprock in the two days after) and it merged at least one PR, one quiet toast offers the card in its own figures ("Your week is done. 33× your plan this week · 48 PRs merged") and opens the sheet on this week. Once per week, whichever way you answer.
  • —share: Threads joins X and LinkedIn as a place to post the card.
  • —resources: one Free up memory button, in Caprock resources and in the memory popover alike, replaces Clean up and End N idle agents. It opens a list in two groups — leftovers and idle agents — each line with its own checkbox and size, the total following the ticks, and says first what happens: conversations are kept, and anything an idle agent started in the background stops with it.
  • —resources: it is stricter about what is safe. A test copy of Caprock is offered only once the agent that started it is gone; one whose agent still runs is kept, with the reason on its row. An idle agent is offered only after 15 minutes idle, with nothing but its own helpers running under it, and never the tab in front; the rest it leaves running are listed with why, and the daemon checks each again at the click.
  • —share: the share card tells the truth. Pull requests come from GitHub's record where Caprock holds one (the lists the Pipeline reads), not only from the gh pr merge commands agents ran — a batch or script merge counts now. Sessions started at / or in your home folder (scheduled scripts) are left out of every figure, and the card says "hand-run sessions". The cost per merged PR is the Pipeline's own per-PR figure, averaged, so the two never disagree.
  • —share: the card's pull request count is GitHub's whole period. When the lists Caprock kept do not reach back to the period's start, the missing older pull requests are read from GitHub once, in the background, and kept; the next card counts them (it read 48 merged where GitHub said 59 before).
  • —share: one card, two formats. Landscape and portrait lead with the same two big figures — your plan's multiple (or the dollars) and merged PRs — instead of a Figures card and a Story card that said different things.
  • —share: a calmer sheet: Share…, Copy and Save in one row; My Telegram, X, Threads and LinkedIn in another; the post text behind Edit post text; one line on what leaves the machine. Send to someone in Telegram is gone — Copy image does that.
  • —ui: a row's ⋯ menu on Chats, Now and the home is no longer cut off by the list's rounded box on a wide window: it is drawn above everything, under the ⋯, and opens upwards when the window has no room below (#816).
  • —terminal: a tab whose agent Free up memory ended shows its conversation with Continue here at once, instead of Can't connect · next try in 10 s over a process Caprock itself ended.
  • —resources: Caprock resources no longer offers to stop Caprock Mail. The separate Mail app was taken for a Caprock command left behind because its process name starts with caprock; an app in its own bundle is never listed now, only the caprock binary itself.
  • —store: a pull request a script opened or merged now gets the agent cost of the session that ran the script. Before, only a gh pr command the session typed itself counted, so on the PRs & CI board three releases made by a release script showed "–". Now GitHub's own opening and merge times are matched to the one session in the repository that had a Bash call running at that moment. With no such session, or more than one, the pull request stays unattributed. Nothing is stored, and every session is still charged once.
  • —share: the week card's footer keeps caprock.dev · What's yours? on one line in the 1200 × 630 picture; its note reads PRs: GitHub + agents.
  • —share: a pull request merged after cd ~/dev/<repo> from another repository's session is no longer counted twice on the share card.
v0.158.02026-10-24

brain

  • —Second Brain no longer reads every session. A session a scheduled job starts can say so (POST /v1/agents {"automation": true}) and is never read or handed memories; Settings → Second Brain → Never read these folders skips every session in a folder (recorded as skipped, with the folder) and keeps those projects out of the graph, search and what a new session is handed unless you open that project. Nothing already kept is deleted. The instructions also tell the model to skip what is true only of one run (what the agent did or saved this time, a dropped connection) and keep what still holds for the product or the codebase.
  • —ui: the Lifetime screen says, under its daily chart, that Claude Code deletes conversations older than its cleanupPeriodDays (30 days unless set), so the history before Caprock's first day was gone before Caprock saw it, and offers Keep them from now on: it writes cleanupPeriodDays: 36500 into ~/.claude/settings.json, that key only and every other byte of the file kept, backed up first, and refuses a file that is not valid JSON. Shown only while the period is unset or 90 days or fewer; nothing leaves the machine (GET/POST /v1/claude/retention).
  • —docs: the README opens the way caprock.dev now does: every coding agent you run, wherever it runs, who is waiting on you, who is stuck and how close the limit is, then each pull request to release with what it cost. Its first sections follow the site's order: pull requests, sessions waiting on you, terminals, cost and limits.
  • —brain: the sidebar card's line at the daily limit reads Limit reached · 1546 instead of being cut mid-word on a 14-inch screen.
  • —brain: Second Brain no longer reads sessions started at the disk root or in the home folder itself, where launchd and cron run scheduled claude -p jobs. They are recorded as skipped with the reason, drop out of the sessions waiting at once and stop using the daily token limit; folders under the home are read as before, and nothing already kept is deleted. Settings → Second Brain says so under Never read these folders.
  • —agents: a session whose program exits a moment after it starts now says why. Its tab shows Claude Code exited at once (code 1) with the last lines it printed, the program that ran, and Try again and Close, instead of Connecting to the session… Can't connect · next try in 5 s over a process that was already gone. The same lines, the command and the binary's version go to service.log at WARN, and GET /v1/agents/{id}/exit serves them (kept in the new session_exits table).
  • —agents: Continue here on a Claude Code session that moved to another folder with cd starts in the folder its conversation began in. Started in the moved-to folder, claude --resume printed "No conversation found" and exited at once.
  • —pipeline: a pull request closed without a merge is no longer drawn in flight as PR open: a session's record of it made the row even though GitHub listed it closed. A row whose only title is its number now reads #796, not #796 #796.
  • —pipeline: one cost per pull request. The in-flight row, the list's Agent cost and the Agents · 14 days figure now read the same recorded spend (the in-flight row used to add up the whole sessions on it), and a pull request closed without a merge in the window counts toward the 14 days: what agents spent on it was spent.
  • —pipeline: with no pull request merged in 14 days, Speed says – and why instead of 0.0/day; with no agent cost recorded on any pull request, Agents says – and why instead of $0.00.
v0.157.02026-10-24

brain

  • —Second Brain no longer reads every session. A session a scheduled job starts can say so (POST /v1/agents {"automation": true}) and is never read or handed memories; Settings → Second Brain → Never read these folders skips every session in a folder (recorded as skipped, with the folder) and keeps those projects out of the graph, search and what a new session is handed unless you open that project. Nothing already kept is deleted. The instructions also tell the model to skip what is true only of one run (what the agent did or saved this time, a dropped connection) and keep what still holds for the product or the codebase.
  • —ui: the Lifetime screen says, under its daily chart, that Claude Code deletes conversations older than its cleanupPeriodDays (30 days unless set), so the history before Caprock's first day was gone before Caprock saw it, and offers Keep them from now on: it writes cleanupPeriodDays: 36500 into ~/.claude/settings.json, that key only and every other byte of the file kept, backed up first, and refuses a file that is not valid JSON. Shown only while the period is unset or 90 days or fewer; nothing leaves the machine (GET/POST /v1/claude/retention).
  • —agents: a session whose program exits a moment after it starts now says why. Its tab shows Claude Code exited at once (code 1) with the last lines it printed, the program that ran, and Try again and Close, instead of Connecting to the session… Can't connect · next try in 5 s over a process that was already gone. The same lines, the command and the binary's version go to service.log at WARN, and GET /v1/agents/{id}/exit serves them (kept in the new session_exits table).
  • —agents: Continue here on a Claude Code session that moved to another folder with cd starts in the folder its conversation began in. Started in the moved-to folder, claude --resume printed "No conversation found" and exited at once.
v0.156.02026-10-24

app

  • —the status strip holds service things only: the session in front's model, context, next-call price and cost, and today's spend, left it — the session's cockpit and the sidebar's Today strip carry them. What the session's processes use (CPU and memory) stays.
  • —app: on a narrower screen the status strip's Update to vX — Restart button and Downloading vX wrapped onto two lines; they stay on one, and a paid plan reads Premium there, with the plan and its end date in its tooltip.
v0.155.02026-10-23

ui

  • —the phone attaches any file Caprock keeps, not only photos: a Markdown note, a PDF, a spreadsheet or a source file, from Choose File on an iPhone, in a session's chat, the Home field and the Save sheet. A document shows as a chip with its name and goes into the session by its path, as a file dropped on the desktop does; one Caprock does not keep, or one over 10 MB, is refused with the reason in a line under the field.
  • —ui: the phone Home's Full dashboard is a row like the others, with what is there under it (New session, all projects and sessions, settings), instead of a faint line.
v0.154.22026-10-23

app

  • —a Page tab opened on an address nobody answers no longer flashes a white rectangle before Nothing answers at …: a new page stays hidden until it has something to show.
  • —app: shrinking the window with a phone preview open no longer draws the page past the window's bottom edge while the tab catches up.
  • —app: a phone page drawn smaller than its true size reads the device's size exactly: an iPhone 16 reads 393×681, where it had read 392 wide side by side and up to three pixels short in height when scaled.
v0.154.12026-10-23

release

  • —a stray ui/node_modules link committed in 0.154.0 stopped that release from building; it is removed, and .gitignore now ignores node_modules whether it is a folder or a link. 0.154.0 was tagged but never published, so this release carries everything listed for it.
v0.154.02026-10-23

ui/brain

  • —Second Brain's Last 14 days now looks like the Dashboard: a row of big figures (tokens over the 14 days, the API-price cost, today against the daily limit, reads), the tokens per day in the Dashboard's chart style with its ruler, still stacked by project with the limit line and the over-limit mark, and a clear top five projects with each one's share and a bar in its chart colour. The rest fold into one and N more line that opens the full table of every project.
  • —pipeline: with no predicted version, the release plan said Next release Next release · 1 merged; the line under the heading no longer repeats it.
  • —brain: a long memory title in the graph panel's Newest list is cut with an ellipsis instead of running past the panel's edge.
  • —ui: the cockpit's CI block, folded, no longer shows an old copy of a pull request's checks as running. Its one line was drawn from what was kept when the block was last unfolded and never read again, so it could say "#801 · 4 running · 11h" of a pull request merged hours before, next to the time of a different answer. The session's pull requests are now read folded too, the line waits for that read and the time beside it is that read's; a kept copy not read again yet no longer has a growing elapsed time.
  • —telegram: a module that answers a button press at once keeps its edit. The press is written to the module's store before its ✓ <label> mark is put under the message, so a module quick enough to replace the text and buttons in that moment could see the late mark put an old button back over its new message. The mark now goes only under the buttons that were pressed, and a module's edit and a press's mark no longer interleave.
  • —app: a Page tab in Phone mode never draws past the tab any more. A device whose page area did not fit kept its true size, and the page's window drew over the status bar, past the app's window and over the Dock; it is now drawn smaller to fit (drawn at 74%) and the page zoomed out to match, so it still lays out at the device's width. The bar's DPR reads 2, not 2.200000047683716.
  • —app: a page where nothing answers says so instead of a blank white page: Nothing answers at 127.0.0.1:22776 (or No address found for a host) with Retry, in a normal tab and inside the phone alike.
v0.153.02026-10-23

pipeline

  • —Most expensive now has a Month · All time switch, the month by default and remembered on this device, and reaches past the newest 50 pull requests GitHub lists: every pull request Caprock knows of in the repository (the ones agents opened or merged, and every one GitHub ever listed to this machine, now kept locally) and every release the checkout's version tags mark, each release costing what the pull requests it shipped cost. All from this machine; nothing more is asked of GitHub.
  • —brain: Second Brain's Last 14 days shows where the tokens went. Each day's bar is stacked by project: the three that spent the most each get a colour, everything else is one grey Other, and hovering a segment names the project, the day and its tokens. A day over the limit keeps its colours and gets a warn dot above it. The table puts the same colour before each project's name and a thin bar after its tokens, sized against the project that spent the most. GET /v1/brain/spend now splits each day by project (days[].projects).
  • —brain: a session opened in / (a background claude run there is labelled /) no longer stalls Second Brain. Its project could not be stored, so every read of it failed whole, was never recorded, and the same chunk was asked again every few seconds: 3,363 times in 13 hours on one machine, its tokens counted toward neither the retry wait nor the daily limit. Such a session is now read in no project; an answer that cannot be stored is recorded as a failed read with its tokens and waits out the retry; and one memory that cannot be stored is left out while the rest of the read is kept, a memory sent with no title taking one from its summary.
  • —brain: the Second Brain strip says a read that did not finish in plain words — The last read didn't finish (Claude's answer could not be saved). Caprock tries that session again by itself in 15–60 min and goes on with the others. — with the raw error only in its tooltip, and so do Settings' Last problem line and the Just read list (given up after 4 tries instead of too short for a session given up on).
v0.152.12026-10-23

brain

  • —a session opened in / (a background claude run there is labelled /) no longer stalls Second Brain. Its project could not be stored, so every read of it failed whole, was never recorded, and the same chunk was asked again every few seconds: 3,363 times in 13 hours on one machine, its tokens counted toward neither the retry wait nor the daily limit. Such a session is now read in no project; an answer that cannot be stored is recorded as a failed read with its tokens and waits out the retry; and one memory that cannot be stored is left out while the rest of the read is kept, a memory sent with no title taking one from its summary.
  • —brain: the Second Brain strip says a read that did not finish in plain words — The last read didn't finish (Claude's answer could not be saved). Caprock tries that session again by itself in 15–60 min and goes on with the others. — with the raw error only in its tooltip, and so do Settings' Last problem line and the Just read list (given up after 4 tries instead of too short for a session given up on).
v0.152.02026-10-23

pipeline

  • —A Most expensive panel under Releases on the Pipeline page: the five pull requests (open, or merged in the last 14 days) and the five releases (of the newest read) with the highest recorded agent cost at API list price, side by side, stacked on a phone. A pull request opens its detail beside the list, a release opens on GitHub; one with no recorded session is left out, and each list says the span it covers.
  • —status bar: the memory item reads "This Mac" instead of "Mac".
  • —mail: Show OAuth failures and expiry in the add-account screen, keep in-progress callbacks pending, and reconnect OAuth accounts from Settings without removing their mail. Explain when browser sign-in is unavailable in a build.
  • —mail: Keep connection timeouts concise without guessing their cause, offer retry and provider help, and keep known server settings folded after a failed account check.
  • —mail: Explain iCloud app-specific passwords step by step and replace raw authentication failures with an actionable error and folded details.
  • —ui: dragging a session row in the sidebar now shows it moving: the row follows the pointer, raised over the others, and they slide apart by its height to show where it will land. Released over the list it lands there; released off the list, or on Escape, it slides back and nothing moves. Before, the row only dimmed and you could not see where it was going. A dragged row or tab now makes way for a neighbour once it is halfway over it, rather than all the way.
  • —ui: dragging a tab in the top strip now shows it moving: the tab follows the pointer, raised over the others, and they slide apart to show where it will land. Released over the strip it lands there; released off the strip, or on Escape, it slides back and nothing moves. Before, the tab only dimmed and you could not see where it was going.
v0.151.02026-10-22

store

  • —the Tools screen's drill-down no longer reads tool payloads: each call's group, each result's failure and each Agent result's subagent are stored beside the event when it is written (migration 0070). On a 2 GB database the Bash drill-down over 30 days went from 4.1 s to 0.8 s. Calls stored before the upgrade are filled in the background, newest first, and read as before until they are, so the upgrade starts as fast as ever and no screen changes meanwhile. The first step towards compressing tool payloads (#672).
v0.150.02026-10-22

app

  • —Phone preview in one click. When a terminal prints a local dev server's address (Vite, Next, webpack, localhost, 127.0.0.1, 0.0.0.0, [::1], *.localhost), Preview on phone appears after it; a click opens the page beside that terminal, split, already in Phone mode on the device used last with Touch on, with no setting to turn on first. ⌘K → Preview on phone… lists the servers your terminals printed, then Preview a URL on phone… for any other address.
  • —app: Phones side by side. Side by side on the phone bar shows the page on an iPhone SE, an iPhone 16 and a Pixel 9 at once, each at its real width (375, 393, 412) and drawn smaller together to fit the pane. Reload and your dev server's hot reload reach all three, each scrolls on its own, and a phone whose page is wider than its screen gets a red cut off. A pane too narrow scrolls the row sideways; a phone's name goes back to that one phone. Off by default and remembered; the two extra phones exist only while it is on.
  • —app: the update card folds into the status strip. A chevron in its corner, or Esc while focus is in it, hides the card, and the download goes on as a short item in the strip: ↓ 0.149.0 · 22% with a thin bar, then 0.149.0 ready · Restart, or the failure in the warning colour. A click on the item brings the card back (on the error, after a failure); the card never comes back by itself, and the fold holds for that version only.
  • —ui: under the Week card, the subagent hint is now three big figures — what the week's subagents cost, about what the same tokens cost on Haiku, and the possible saving — with one button, Start new agents with Haiku subagents, that makes Haiku the default New agent starts subagents on in every project that has not chosen otherwise. The fine print below the card is a row of small chips, each with its full sentence on hover.
  • —pricing: a gateway's free model — an id ending in -free (OpenCode Zen, such as deepseek-v4.1-flash-free) or :free (OpenRouter) — is priced at $0 instead of showing as an unknown model to report. Turns already stored unpriced are priced on the next start. Pricing table 2026-10-22.1. Closes #787.
  • —now: the "sessions were still running when Caprock last stopped" banner no longer offers Continue on claude -p and Agent SDK runs. Those answer one prompt and exit, so a stop that caught one cut nothing off; Caprock now reads how the session was started from its transcript and leaves them out, including from a banner already showing.
  • —loop: a browser or computer-use tool taking the same screenshot, scrolling, waiting, pressing a key or navigating again is no longer flagged as a loop. That is how those tools work. A repeated click, typed text, edit or command still raises the alert.
  • —memory: the memory popover no longer lists an app bigger than the group it sits in (colima at 8.2 GB under Other apps and the system at 7.6 GB). The used figure is RAM, while an app's size counts its compressed and swapped-out memory too, as Activity Monitor does; an idle virtual machine's guest is mostly in swap. The group now takes what its apps hold when that is more, an app running a virtual machine is marked · virtual machine, the card says when the sizes add up past what is in use, and the stacked bar never runs past the used share.
  • —report: the Weekly report panel no longer shows an old "Last send failed" under a fresh "sent — check Telegram". It reads the outcome again after Send one now, and a new bot token or chat forgets the last failure, which was about the bot that is no longer configured.
  • —brain: Second Brain no longer keeps credentials or payment details. Before a memory's title, summary or keywords are saved — by the automatic extraction, by hand, by the re-tag or from an Obsidian note — API keys and tokens (sk-…, GitHub, Slack, AWS, Google, Stripe, Telegram bot tokens, JWTs, bearer values, PEM private keys), card numbers (Luhn-checked), IBANs (checksum-checked), labelled passwords, secrets and bank account or routing numbers become a placeholder such as [redacted: card number]. The prose sent to the model to be read is filtered the same way, and the extraction's instructions say never to record such values. Memories kept before this are scrubbed once on start; the log says how many, by kind. caprock mcp uses the same filter, so its placeholders now name the kind too. The Second Brain settings say so in one line.
v0.149.02026-10-22

ui

  • —the status bar shows the machine's memory: Mac 9.1/16 GB with a small green bar, amber with what the agents hold when the system warns or 85% is in use, and red Memory low when the pressure is critical or swap keeps growing. A click shows who holds it — agents, Caprock, other apps — each agent with its state, the apps using the most, and one button that ends the idle agents Caprock started (their conversations are kept; Continue picks them up). Agents started elsewhere are listed, never touched. Read from the OS's own figures on macOS, Linux and Windows, only while a window is open; a paired phone sees the totals, not other apps' names. An agent whose one process passes 3 GB (or a fifth of the machine), or grows for hours, is flagged in the popover and its inspector with Restart it fresh: the process ends and the same conversation continues in a new one. Memory is each process's physical footprint on macOS, compressed pages included, as Activity Monitor shows. New GET /v1/system/memory, POST /v1/system/memory/end-idle and POST /v1/system/memory/restart.
  • —inspector: every tab's Spend now opens on that agent's own session — its burn and spend — instead of one period shared by every tab, which made a Caprock-wide today show the same chart and dollars in every tab. Today, 7d and 30d across all sessions are still in the period picker, and a pick stays in the tab it was made in. Folded, a chip says whose figures they are: session, or today · all.
  • —app: a terminal tab left out of sight for ten minutes keeps the last 1,000 lines of its scrollback instead of 10,000, and has the full 10,000 again when shown. Every tab ever opened used to hold its whole scrollback: ten full tabs held 103 MB of the window's memory, 24 MB once trimmed (measured 2026-10-22). The tab in front and anything shown in the last ten minutes are unchanged, and nothing is redrawn on the way back.
  • —daemon: cached dashboard answers for a past day or range are dropped once they are too old to show (they were kept for as long as the daemon ran); a session Caprock started that crashes or is killed no longer leaves its last tool inputs in memory; and the daemon sets a soft 256 MiB ceiling on Go's own memory (unless GOMEMLIMIT is set). A heavy synthetic load stayed under 90 MB of it (measured 2026-10-22); near it the daemon collects harder and returns memory to the system instead of growing.
  • —app: dragging a file from iCloud Drive onto the window no longer closes the app. macOS hands over such a file in a way the window's drag code did not expect, and the app quit on the spot (0.141.0 to 0.147.0). Every file is now passed on as it is, whatever its name or where it lives. If the app ever stops on an error again, the reason is written to app.log in Caprock's data folder.
  • —app: the app no longer stops at "Caprock could not start: another Caprock daemon is already running" when its background service is still starting, which happened at login on a busy Mac. It waits for that daemon and connects to it. If the service does not answer, the app restarts it. Try again does the same.
  • —app: dropping a file on the window no longer touches the disk on the thread that draws the app. A file from Finder is handed to the terminal as it is; a screenshot thumbnail or another temporary file is copied on a background thread first. On a Mac short of memory and swapping, that copy could freeze the window for seconds.
  • —hooks: the hook shim in Caprock's data folder is brought up to date every time Caprock starts. The desktop app never replaced it, so a shim an earlier Homebrew install had put there could keep running, months old, under a current Caprock. With no separate shim to copy, the app puts its own caprock there, which works as the shim.
  • —hooks: Caprock's own hook and status line commands in ~/.claude/settings.json are repointed at this install when they run a caprock that is gone or is another copy (an uninstalled Homebrew one, for example). A status line or hook of your own is never touched, and the file is backed up first.
  • —terminal: a terminal tab whose session's process is gone no longer spins on "Reconnecting…" for good. After the background service restarted, tabs for sessions that died with it were refused by the daemon on every try, and a browser is never told why. Now, after three refused tries, the tab asks the daemon and says what is wrong and what to do ("Can't connect" with the reason); a session the daemon still does not hold about 20 s later is ended, and the app shows its conversation with Continue here. A tab restored for a running session whose terminal Caprock lost opens that way at once, and a session the daemon holds is served even when no new agent could be started.
v0.148.02026-10-22

brain

  • —Second Brain now goes back and reads what it skipped. Before this release, a read of a long session kept only its newest part, and the older part was never read. Caprock now finds those skipped parts on its own and reads them oldest first. It does this only while automatic reading is on, only when no new session is waiting, and within the same daily token limit. The Second Brain status shows how far it has got and estimates how many days the rest will take.
  • —brain: Second Brain finds what was decided and knows which decisions still stand. Search matches whole words (PR no longer finds project and prompt: 136 memories instead of 1,097 on the owner's database), any word may match, rarer words and title or keyword hits rank first, and PR, pull request, cost, price, spend, decision and rule find each other. Each new memory gets five to eight English keywords, whatever language its title is in, so an English query finds a memory written in another language (older memories get theirs from the re-tag). A read is shown the project's current decisions and recent memories with their ids: it no longer repeats one, can link to one, and can say a new decision supersedes an older one. A superseded memory is ranked last, dimmed and struck through in the graph, and says what replaced it; its Undo makes it current again. A new session's handoff now leads with the project's current decisions (up to 15), carries one memory per session, and goes even when no recent passage does. caprock mcp gains current_decisions, and search_memories returns up to 30 ranked candidates with keywords and a short summary for the agent to judge.
  • —brain: Second Brain gives English keywords to the memories that have none, the ones kept before keywords existed and the ones typed by hand, so an English search finds a memory titled in another language. It runs on its own while automatic reading is on and no session waits to be read, 40 memories a call, within the same daily token limit; its tokens show in the spend like a read's, in a row of their own (Keywords for older memories), and the strip's second line says Adding keywords to older memories: X / Y while no catch-up is running.
  • —brain: Second Brain says what it spends. A strip on top of the screen reads Today 84k of 300k tokens over a thin bar against the daily limit, with a chip for its share of every Claude token Caprock recorded in the same 24 hours; while a backfill catches up, a second line says X / Y reads, about N days left. A click opens the last 14 days: tokens per day against the limit, and reads, tokens and memories per project. Read from the new GET /v1/brain/spend?days= (open to paired devices); the status gains a generic catch_up field.
  • —mcp: one-click Connect for caprock mcp. Second Brain opens with Let your agents ask Second Brain and a button per installed agent; a click runs claude mcp add -s user caprock -- <path> mcp or codex mcp add … on this machine, Update path fixes a registration that points at an old caprock, and Disconnect removes it. Settings shows the same status above the copy commands. Only on a click, never from a paired phone, and nothing leaves the computer (GET /v1/mcp/agents, POST /v1/mcp/agents/{agent}/connect|disconnect).
  • —modules: a private module can say what waits on you in it, with PUT /v1/x/<id>/attention (a count and up to eight items, each a short label and a #route inside the module). Its sidebar row shows a small 2 waiting; a click unfolds the items, and an item opens the module right there. The row and Second Brain above it fold as an accordion. At 0 the row is unchanged, and a report stops showing a day after the module last sent one.
  • —modules: a module can edit a Telegram message it sent — new text, new buttons, or none (PATCH /v1/x/<id>/telegram/<message_id>) — so a batch of approved drafts shows which cards are done, can thread a new message under one of its own with reply_to_message_id, and can pin one message at a time with pin (silently, replacing only its own previous pin). Presses on edited buttons land in the module's store like the first ones; any message the module did not send is a 404.
  • —tabs: an empty new shell shows one dim line under its prompt, "Run claude, codex or any command · ⌘W closes this tab" (Ctrl+Shift+W on Windows and Linux). It is drawn over the terminal, never typed into the shell, goes at the first key you type, and never shows in a tab that already has output. The tab strip's + now says "New terminal ⌘T" on hover (Ctrl+Shift+T on Windows and Linux).
  • —mcp: memory_topics with a project now counts that project's memories and lists that project alone, instead of the totals for every project beside one project's topics. session_summary says how many files a long session touched beyond the 40 it lists (+N more), counted from the whole session rather than the daemon's first 100.
  • —brain: the screens no longer promise every agent what only Claude Code gets. Any agent's sessions add memories, but only a Claude Code session is handed them at start; the tagline, the agents line (Claude Code and Codex add to it; Claude Code sessions start with it.) and the third step of How it works now say so. Links on the Brain screen counts links from one memory to another only, not each memory's link to its topic or project.
  • —brain: Second Brain reads a long session whole. A read still sends at most 24,000 characters, but a session with more is now read in chunks, oldest first, each read going on where the last one stopped; before, a read kept only the newest part and the older part was never read. A wake still makes at most four calls, so a read costs what it did. The extraction is also told to keep a decision or standing rule you state ("from now on", "rule:") as a decision, and a stretch of mostly your own prompts is read rather than skipped for want of the agent's prose. Machine-made "user" turns (task notifications, reminders, slash-command echoes, messages from other sessions) are no longer read as your words, and a prompt is sent up to 2,000 characters instead of 600.
  • —brain: A short exchange at the end of a session that is still open is no longer marked read and lost. Second Brain waits and reads it together with what the session says next. When the session has ended, the short tail is marked read, so nothing waits forever.
  • —brain: a session that ran in a folder under the system's temp directory no longer files its memories under a throwaway project named after that folder (an archive extracted to ops-seo-site-m143z5ko made one per run). A checkout there with a remote is named after the remote's repository; anything else keeps its memories with their topics and no project.
  • —docs: the guides match the app again. Closing a tab ends an idle agent (its conversation kept) and asks about a working one; the sidebar has no worktree rows or ±N; the phone says No internet on this phone and counts down to its next try; desktop notifications have the one switch When a session needs you; a paired phone starts any agent it lists, not just three; Gemini is read through its telemetry file; and an ended Hermes session is carried on in your own terminal, not in a Caprock tab.
  • —app: dropping a file on the window no longer touches the disk on the thread that draws the app. A file from Finder is handed to the terminal as it is; a screenshot thumbnail or another temporary file is copied on a background thread first. On a Mac short of memory and swapping, that copy could freeze the window for seconds.
  • —hermes: an ended Hermes session can be continued in a Caprock tab again (hermes chat --resume <id>), as Codex and OpenCode ones are; it was refused with "Caprock continues Claude Code, Codex and OpenCode sessions." It is offered once the session has ended, while its folder and Hermes's state.db exist and hermes is installed.
  • —ui: the shortcut list no longer says ⌘W leaves the session running: it closes the tab or pane, an idle agent ends with its conversation kept, and a working one asks Keep it running or Stop it.
  • —phone: Continue on the phone home now lists the sessions open as tabs on your computer, in the tab strip's order and under the same names, and a row no longer jumps around as sessions start and stop working: its status changes in place. Sessions running in the background without a tab stay under All sessions. When the computer's app has not reported its tabs, the live sessions are listed oldest first, in an order activity never changes. The app now reports its tabs in strip order, and the daemon answers GET /v1/tabs (a paired phone may read it).
  • —now: What happened is back on Now, after a merge on 2026-10-14 dropped it an hour after it shipped and brought back the old feed of one line per command. It shows the moments of your sessions again: files edited (grouped by folder), tests and builds with their result, commits, pushes, pull requests, subagents by their task, turn ends, approvals and failed commands, with reads as a quiet +N. All commands brings the full feed back; the panel folds and every row is the same height; the end of a turn no longer reads as "session ended".
v0.147.02026-10-22

modules

  • —a private module can message its owner. POST /v1/x/<id>/telegram sends to the owner's own paired chat through the existing Caprock bot, its first line naming the module, with optional buttons or a reply field; a press or a reply is written back into the module's store (telegram/<message_id>-<n>) and never reaches a session. POST /v1/x/<id>/notify shows a desktop notification from the Caprock app, titled with the module's name, that opens the module when clicked.
  • —ui: a folded inspector section no longer runs past the panel's edge. The Changes row read 511 files +35244 −7781 Review and co, cut mid-word; now each row's summary shortens to fit: the counts compacted (+35.3k −7.8k), Review for Review and commit, then fewer parts, and nothing rather than half a word. The same holds for Subagents and Processes.
  • —ui: the Subagents cost says total: 1 finished $2,549.54 was every subagent's cost over the whole session beside the count of those that finished in the last half hour, and read as one subagent's.
  • —mail: Import everything no longer hides Apple Mail when macOS keeps Caprock out of its account list. iCloud, Google and Exchange accounts in Apple Mail were silently missing because the list sits behind Full Disk Access, which macOS never asks for. The screen now says Caprock can't read Apple Mail's account list, with Open Full Disk Access (opens that list in System Settings and shows Caprock's program in Finder) and Add iCloud with an app password instead. Adding iCloud now walks through getting an app-specific password in three short steps, with a button that opens account.apple.com and one field for the code. An iCloud account found in Apple Mail is marked Needs an app password.
  • —telegram: saving a new bot token, or pairing a new chat, no longer leaves the old Last alert failed under Settings → Telegram until a test is sent. The old outcome is forgotten at once and the line reads Saved. Send a test alert to check it reaches you. With no alert switched on, the section now says the bot is idle: nothing is sent and its messages are not read.
v0.146.02026-10-22

docs

  • —what /clear and a fork do in Caprock: the conversation stays one card linked to its earlier parts, while Claude itself starts with an empty context and gets no handoff note.
  • —brain: memories moving: when a read keeps memories, one silver thread for the whole read drifts from that session's row (else its tab, its project's row or its agent's face) into the sidebar's Second Brain, which ripples and shows +N; when a new session is handed memories, a thread goes the other way into its tab with N memories handed. At most one a minute each way, with counts merged, never while you type, nothing in a hidden window; with reduced motion only the count shows. The setting is now Show memories moving (same switch, one for both directions). The card's agent faces now show only agents whose sessions were read in the last 24 hours, each with how many, and the one being read glows; the orb's background sparks no longer look like memories arriving. The live brain frames gain a read's start and end (phase, run, agent, added) and a handoff frame, and GET /v1/brain/status gains auto.agents_24h.
  • —attention: a session whose turn ended is now either Needs you or Done. Needs you is a permission prompt, a question, a plan to approve (ExitPlanMode, which now says so), a turn that stopped on an error (a StopFailure such as a rate limit, which used to read "working", then "idle"), or a last message that asks you something (a guess, marked with a dashed GUESS chip). Done is a turn whose last message asks nothing: it stays listed, quieter, and is never counted in the badge, the tray head or the phone's amber card. Home, the sidebar, the tray popover and menu (Done in its own submenu) and the phone home all show it; needs me? and not needed move a guess until the session's next turn.
  • —notifications: a turn that ends on a question is Needs your answer instead of Finished, and a Done turn raises no OS notification. Telegram still sends Finished for Done turns, unless the new Also when it asked nothing (Done) switch is turned off.
  • —ui: the session's state under its portrait says Background on one line while background subagents work, instead of wrapping Background agents over two lines in a narrow inspector; the full Background agents working · N stays on hover and is what a screen reader announces.
  • —api: a shell tab's inspector shows its folder's uncommitted changes, instead of asking for a session that does not exist and logging a 404 on every refresh.
  • —tabs: after /clear in an agent tab, the tab carries on with the conversation instead of freezing on the part before it. The terminal stays attached, the chat shows the earlier part above a Context cleared — Claude no longer remembers the part above line with the new part under it, typing works, and the session page no longer says the session has ended while it runs. A fork never takes the original's tab. A Caprock restart finds the newest part of a cleared session again.
v0.145.02026-10-21

app

  • —Phone mode for a browser tab (beta, with the in-app browser on): Phone on the tab's toolbar, or ⌥⌘R (Ctrl+Alt+Shift+D on Windows and Linux), shows the page at iPhone 16, iPhone SE (3rd gen), Pixel 9, Galaxy S24 or a custom size inside a simple device frame. The page itself is resized to the phone's page area, so media queries and innerHeight read the phone's numbers, and it gets the phone browser's user agent. Rotate, browser bars shown or scrolled away, and Touch, which turns the mouse into a finger (drag to scroll with momentum, swipe, pull a sheet). Page heights marked est. are estimates of the bars; the ? lists per system what is real, approximate or not emulated.
  • —ui: The phone home is shorter and says more at a glance. New chat is a filled button in the Chats row. Continue, Releases and Saved fold from their headers (Continue open, the other two folded at first), and a folded section keeps one line: how many sessions wait, work and sit idle; the repository and its next release, with a pull request whose CI failed said first in red; how many notes were saved and the newest. The phone remembers the folds. A new Limits line shows the 5-hour window with its reset and the weekly one, amber from 80% and red from 95% (Codex's when Claude Code has none), and opens Cost at its plan limits. A live Continue row says its context used and its subagents at work.
  • —api: GET /v1/stats/limits answers the plan windows alone (rate_limits, codex_rate_limits), as the summary carries them, so the phone reads two percentages without the whole summary. A paired device may read it.
  • —ui: links in a chat can be tapped, on the phone too. A bare https://… address in an agent's reply was plain text; now every http(s) address in a reply, a prompt, the Answers tab and the last-word popover is a link that opens in a new tab (in the desktop app, your browser), with a sentence's trailing punctuation left out and a long address wrapping instead of widening the screen. Other schemes stay text.
  • —remote: a phone that loses its network no longer offers Pair again. It says No internet on this phone at once, large and full width, with the next try's countdown, and reconnects by itself when the network is back, keeping the last screen dimmed underneath. A relay that cannot be reached and a computer that is not answering each get their own words; Pair again shows only when the computer removed the phone or its keys changed.
  • —api: Claude Code's footer no longer loses characters after you switch tabs or apps or reconnect: no more status line of lone digits or mode line with its words gone. The repaint that redraws a reattached terminal now waits for the session to draw at the narrower size before restoring its width. A busy Claude Code used to get both size changes as one and redraw nothing. A terminal that falls behind mid-session is now repainted too.
v0.144.02026-10-21

modules

  • —a private module opens on a paired phone the owner made a controller, over Wi-Fi or from anywhere: listed in the phone's menu and the header, drawn in the app's theme, with its own files and document store. It runs in a sandboxed frame that reaches only that module's files and store; a phone that only views sees no module, and the admin key and installing a module stay on the computer.
v0.143.02026-10-21

agents

  • —POST /v1/agents takes end_when_done: true for a new Claude Code or Codex session started with a prompt: once its turn finishes (Claude Code's Stop hook, Codex's task_complete) with no permission prompt, question, queued message or working subagent left, Caprock ends it the way End does. A script that starts a one-job session every half hour no longer leaves an idle TUI behind each time — Codex's never exits by itself, and 27 of them had piled up in a day on one machine. Off unless asked for; never applied to a session started without it or not started by Caprock.
  • —agents: a chat or agent Caprock started that has no tab is ended once it has sat idle for 12 hours — nothing typed, drawn or recorded, and no permission prompt, question, queued message or working subagent waiting — so idle sessions no longer pile up in memory. Settings → New sessions → End idle chats after: 2 h, 12 h, 24 h or Never. The conversation stays listed and says Ended after 12 h idle — continue; opening it (or replying to it in Telegram) continues it from where it stopped. Never applied to a session with a tab, a shell, the orchestrator's workers, or a session Caprock did not start. Idle time is measured from the session's own activity, so a daemon restart does not reset it.
  • —ci: the release's desktop app builds start from compiled Rust dependencies instead of from scratch. A new master-only CI job keeps a Rust cache of the macOS universal build (both architectures). CI's Linux app build moves to ubuntu-22.04, the image the release builds the AppImage on, so the two share one cache. The release jobs only read master's caches and no longer write Go caches that no later tag could read. A stale or missing cache costs build time only: signing, notarization and the updater signatures are unchanged.
v0.142.02026-10-21

app

  • —the macOS app is now signed with Caprock's Apple Developer ID and notarized by Apple, the .app and the .dmg stapled, so it opens without "Open Anyway" and the Homebrew cask no longer clears the quarantine flag. macOS keys privacy answers to the app's signature, and this signer is new to it: on the first launch of the first release signed this way Caprock asks once more for the microphone and for each folder it was allowed, and updates keep the answers from then on. Answers given to claude or codex, notifications and keychain passwords are not affected. While the Apple secrets are missing a release falls back to the self-signed certificate, as before.
  • —phone: the phone chat's walkie-talkie bar is gone (owner, 2026-10-21: unused and in the way): the speaker switch, the + sheet, hold-to-talk, and the latest answer's AI;DR, Read aloud, Continue and Stop. The phone chat has the regular field again: Keys, Photo, the dictation language, the microphone and Send. New chat stays on the phone home; Report a bug is in ⌘K on a computer.
  • —store: Caprock no longer opens a database that a newer Caprock has migrated. Going back to an older version used to open it anyway and could lose data without a word; now the daemon stops before writing anything and says, in one line, the database's schema number, the highest this version knows, and to install the newer Caprock. caprock up, the desktop app's "Caprock could not start" screen and caprock doctor show that line, and the database file is left exactly as it was.
v0.141.02026-10-21

sidebar

  • —Recently closed lists the last five tabs you closed, newest first, each with a quiet age (2 min, 1 h), above Running, no tab in the Tabs list and after the projects in the Projects list. A click brings the conversation back in a tab: its own tab while the agent still runs, continued with --resume once it ended. ⇧⌘T reopens the newest, as in a browser (Ctrl+Alt+Shift+T on Windows and Linux). Show all in History opens History; the list is kept across restarts.
  • —sidebar: Running, no tab keeps the sessions you started yourself, titled and active in the last day. Sessions in a temporary folder, ones another tool started with a scripted brief, untitled ones and ones idle for a day fold into one Background · N row at the bottom, opened on a click and remembered.
  • —modules: a private module's key is ⇧⌘1–⇧⌘9 on a Mac (it was ⌥⌘1–⌥⌘9), so the first module opens with ⇧⌘1. Windows and Linux keep Ctrl+Alt+Shift+1–9. macOS keeps ⇧⌘3, ⇧⌘4 and ⇧⌘5 for screenshots, so on a Mac the third to fifth module's row shows no key and says why in its tooltip; the key still works where those screenshot shortcuts are off.
  • —app: the theme picker moved from ⇧⌘T to ⌥⌘P (Ctrl+Alt+Shift+P on Windows and Linux), so ⇧⌘T can reopen the last closed tab.
  • —app: an agent tab's Chat | Terminal pills are one quiet icon at the right of the tab strip. It shows the other view — a chat bubble over the terminal, a terminal glyph over the chat — named Show as chat / Show as terminal, and one click flips it. Each session keeps its own view, and the terminal stays running under the chat as before.
  • —app: the session deck under the agent's terminal (⌘B, the Deck button, Pipeline bar: On/Off in the palette). What it showed is in the inspector and on the Pipeline page: a session's own pull request in the inspector's CI block now offers the one action it needs — Resolve conflict, Update branch, Fix CI or Re-run — asking once in the row, with its agent cost and who is on it, and Log in Pipeline → opens the Pipeline on that pull request with its failed job's log and Investigate. The Pipeline page gains a Default branch panel: each workflow's newest run there, a failed job's log with Investigate, and what GitHub billed for Actions this month. The deck's Branches view is gone. ⌘B is free.
  • —daemon: the dashboard reads far less of the database while agents work. The Work board keeps each live session's newest events and task list until the session writes something, the plan-limit pace is counted again only when a new sample arrives, the status screen's Memory line stops at the first passage it needs instead of parsing a fortnight of replies (it often ran out of time and showed nothing), and a week of spend per subscription is summed once a minute instead of on every request. On a copy of a 1.85 GB database with the Now screen open, the database reads the daemon asked for fell from 2,050 MB to 1,231 MB in three minutes. The Work board also keeps each session's spend today until it changes, and task lists are read from their own index (migration 0066) instead of a walk through every event of the session.
  • —ui: the phone no longer scrolls sideways. On an iPhone, tapping a field smaller than 16px (the phone home's Repository and Project pickers, most Settings fields, Cost's cap) zoomed the page in and left it panning sideways; every field is 16px on a phone or touch screen now. A chat file's Markdown viewer no longer grows to the width of its longest link or path, and a wide table in a reply scrolls in its own frame instead of breaking every cell into two-letter lines. A new e2e test walks every phone screen, sheet and menu at 360, 390 and 430 px in a phone profile and fails on any sideways scroll, anything cut off at the edge or any field under 16px.
  • —smoke: the end-to-end smoke tests start their daemons with the file secret store, so running them on a Mac never writes a test token into the real login keychain; the Telegram smoke had replaced a developer's real bot token there.
v0.140.22026-10-20

app

  • —after an in-app update on macOS, Caprock opens again with its window in front. The updated app was started as a child of the old one, which macOS never brought forward, so it waited behind other windows until a click on its Dock icon; it is now launched the way the Dock launches it.
  • —app: a terminal no longer keeps stale rows after its GPU context is lost (a sleep, a GPU reset, or WebKit letting a background tab's context go): it switches to its plain renderer in that frame instead of drawing on a dead canvas for three seconds, or for good when no loss was reported, and gets a fresh context shortly after. A lost context is also looked for when you come back to the tab or the window. The key hint over Claude Code's input box goes when the window loses focus instead of staying over the quick replies.
  • —app: reopening a terminal or coming back to it after its connection was let go no longer leaves a stale line on its bottom row. Caprock makes the program in it redraw by changing the size for a moment; it used to take one row away, and Claude Code then cleared one row too few. It now takes one column away, which makes it clear the whole screen.
  • —app: a window that goes white comes back by itself instead of staying blank until the app is quit. When macOS ends the WebView's content process (under memory pressure, or a crash), the app reloads that window, waiting longer each time it repeats. A page the app asked for that never finished loading is loaded again while the daemon answers. A page that loaded but drew nothing (a file of it missing while the daemon restarted) shows Reconnecting to Caprock… with Reload and reloads once the daemon answers. Each of these, and the crash screen, writes one line to the app's own log on your Mac (app.log in Caprock's data folder), so a report can say why; nothing is sent anywhere.
v0.140.12026-10-20

deps

  • —routine dependency updates: purego 0.11.2, klauspost/compress 1.20.3, golang.org/x/text 0.43.0 and x/sync 0.24.0; React 19.3.1, Vitest 5.0.5, jsdom 30.1.3 and Playwright 1.64.1 for the dashboard's tests; the goreleaser action 7.3.0 for release builds.
v0.140.02026-10-20

app

  • —an agent tab shows its session as a Chat or as its Terminal, from a switch in the tab strip. Quick chats open on Chat and project sessions on Terminal; the choice is remembered for each session. The chat reads like one: your messages as bubbles, the agent's replies as text, each tool call as one plain line (Searched the web: …), and a question that only the terminal can answer links to Open in Terminal. Switching never restarts the session or types into it, and the hidden terminal does not draw.
  • —perf(api): with the dashboard open and agents working, the daemon no longer counts the whole history again on every 10-second poll. A history that reaches before today (7 days, 30 days, all time) is kept for a minute however fast it was to count; today's keeps its 3-second refresh. A quick count was quick only because the page cache answered it, and on a machine short of memory every count went to disk: one user's daemon read 8.8 GB in an hour. On a preview stand with a 103 MB database and the Now screen open, database reads fell from 193 MB to 71 MB per 3 minutes (2026-10-20). The lifetime and weekly figures can now be up to about two minutes behind.
  • —ui: Settings → Computers says what to do. Connected computers come first — this one, the ones it reaches and the ones that reach it, each with Disconnect — then What is this computer? (The main one or Connects to another) and three numbered steps for that side that tick as they are done: access from anywhere, the other computer's choice, and the link to paste with how long it lasts and Copy. What the relay is, where access from anywhere is switched on and what it sends sit behind the (i) beside Add a computer. Pairing and what it sends are unchanged.
  • —ui: the pull request progress bars end at merged. A merged pull request's bar is now completely filled on Work, the Pipeline, the release plan, the phone's Releases and the right panel's CI block, instead of leaving a grey released block that read as unfinished or closed. Whether it shipped is a quiet word after the bar: in v0.140.0 once released, a muted not released while a repository that cuts releases waits for the next one. A pull request closed without a merge still reads Closed.
  • —ui: Lifetime's Daily cost chart draws every calendar day through today, a day without work as an empty bar, instead of closing the gap and putting a quiet week next to a busy day. Its 30d, 90d and 180d now cover the same days as the range buttons, in local time.
  • —ui: quick chats no longer pile up in the sidebar with no tab. Closing tabs in bulk — a project's Close tabs, its ×, Hide, Remove from list — ends the quick chats in them as ⌘W does (one question for a chat still answering). A quick chat with no tab is listed under Running, no tab only while it works or asks for something, or for 10 minutes after its last answer; it stays on the Chats screen and in the palette. Its row now says what the chat is about, when it started and its model, instead of only Quick chat · <model>, so two chats can be told apart.
  • —ui: on Windows and Linux, hovering a sidebar row (Dashboard, Mail, Settings, New agent and the rest) or a folded footer icon shows one hint, its tooltip with the key in it, instead of a key chip at once and the tooltip a second later. The chips still show under keyboard focus and all together while Ctrl is held; macOS is unchanged.
v0.139.02026-10-19

ui

  • —the right panel's Release section reads at its own width: each pull request's title takes the full row and wraps to two lines instead of being cut beside a progress bar, with one muted line under it in plain words (CI 1/3, CI failed, ready, merged · not released) and its age. Blocked pull requests come first, marked !; a branch a batch pull request already carries shows as that pull request's +N branches; the branches without a pull request fold away by default, one line each with their age; the heading is the version and the last release.
v0.138.02026-10-19

agents

  • —New agent (desktop, and the phone's start sheet) asks which model Claude Code's subagents run on: the session's own (default), Haiku, Sonnet or Opus. It sets CLAUDE_CODE_SUBAGENT_MODEL for the session, is remembered per project, and shows on the session as subagents: Haiku. The Week screen adds a note when subagents ran on a pricier model than Haiku and the difference is a dollar or more: what they cost, and about what the same tokens would cost on Haiku (an estimate at list prices).
  • —perf: a hidden window no longer receives terminal output: after 2 s out of sight each terminal lets its socket go and, when shown, resumes from its last byte, so nothing is lost. A flooding terminal is written at most every 33 ms instead of on every chunk. The status strip's footprint and the shell list arrive as live-socket frames instead of a 10 s poll each. Measured in Chromium against a test daemon, 60 s each (2026-10-19, docs/app-window-footprint-2026-10-19.md): terminal frames to a hidden page 1,218 → 0 and its CPU 1.85% → 0.62%; three tabs with one at 20 lines a second 28.4% → 17.9% browser CPU and 7.3 s → 4.3 s main-thread work; idle Home 28 → 17 requests a minute.
  • —ui: right after a check, the version panel's button is Check again; What's new is a link on the "Checked" line.
  • —brain: keeping a memory while a backfill imports many transcripts no longer fails with SQLITE_BUSY. Each import batch left the database free for a while after it committed, but a batch already waiting was handed the lock at once, so the gap never opened; it now waits the gap out. A write that still finds the database busy is answered 503 with words, and the dashboard tries a drop again rather than showing a bare error.
  • —terminal: a shell terminal's holder stops on SIGTERM within about a second. It passed SIGTERM to the shell, which ignores it, and then waited for good; it now hangs the shell up first (as closing a terminal window does, which also ends the shell's jobs), and kills what is still running 750 ms later.
  • —release: the right panel's Branches pushed, no PR yet no longer lists branches that already shipped. A branch squashed into a batch pull request was still listed once master edited the same files or the release consumed its changelog fragment; it is now found by its commit subjects in the squash's message, or by its changes still standing in master's history. A branch already inside another open pull request's branch (a batch) is left out too. Each branch reads in two lines at 360px — its commit subject, then the branch, its commits and age with ↗ — five at a time behind Show all N, and a branch with no commit in a week waits behind +N older. Pull request titles in the right panel take two lines instead of being cut.
  • —ui: in a short window a sheet uses most of its height, and New agent's First message scrolls into view when focused instead of sitting under the footer.
v0.137.02026-10-19

telegram

  • —forward messages to your paired bot — text, photos, links, files, voice notes, one or several at once — and it asks once Where to?: a running session Caprock started (waiting, working, then recent), a new chat, or Cancel. Everything goes to the agent as one message, your own words first, each item quoted with who wrote it; photos and files are saved where pasted files go and named by path, a voice note is saved and marked not transcribed. A busy session gets it when its turn ends, and the bot says when it went. In Caprock it shows with a small from Telegram. Reply here (off by default, remembered) sends that turn's answer back. Premium, with two-way chat on (ADR-081).
  • —app: on Windows the tab keys read "Ctrl@0" and "Ctrl@1": the chip used the ⇧ glyph, which no font Windows draws it with carries. Off macOS every key now reads as the chord you press, Ctrl+Shift+1. The long Windows and Linux key hints no longer crowd or cut the sidebar's labels ("Settings", the theme's name, "Quick chat") or fill the header: a row shows its key while the pointer is on it, and holding Ctrl shows every key at once. macOS is unchanged.
v0.136.02026-10-19

ui

  • —the line under the next release now says where it stands in one row: what is merged, the open pull request and its CI, agents working and branches ready, and the last release. Each part jumps to its list.
  • —perf(api): the all-time history behind the Now screen and History is no longer counted again on every poll while agents work. It is refreshed at most once per 20 times what counting it took (never past 2 minutes), and its totals reuse the summary's scans instead of making them twice. Today's figures keep their 3-second refresh. Measured 2026-10-18 on an M1 Pro under a synthetic working load (four Claude and two Codex sessions writing, hooks, two dashboard pages): on an 890,000-event history the daemon went from 40% of one core to 11% and the history refresh's allocation from 2,839 MB to 182 MB per 150 s; on a 181,000-event history the refresh's own CPU fell 40% and the daemon's total stayed within noise.
  • —perf(api): the sessions list no longer re-reads every listed session's recent events on every request. Each session's reads are kept until its events change: an insert raises its newest event id, and new database triggers count updates and deletes (migration 0063). Narration, stats, live subagents and the clock are still worked out on every request. Measured 2026-10-18 on an M1 Pro under a synthetic working load (181,000-event history, four Claude and two Codex sessions writing, hooks, two dashboard pages): the list's allocation per request fell from 0.85–0.87 MB to 0.35–0.36 MB and its CPU per request from 1.8–3.2 ms to 0.8–1.1 ms. The daemon went from 13–15% of one core to 11–12%. The cache holds about 9 MB for at most 256 sessions.
  • —perf(rollup): importing history writes many events per transaction instead of one. The Claude transcript reader and the Codex importer commit up to 512 events or 50 ms at a time, with a transcript's read position saved in the same transaction, so a failed batch is read again and nothing is lost or counted twice. Live hook events keep their own transaction and are let in as soon as they wait. Every recorded event also stopped starting four or five short-lived goroutines inside the database layer. Measured 2026-10-18 on an M1 Pro, importing a synthetic 178,000-event history: 149–150 s and 152–160 s of CPU before, 101 s and 83–84 s after; the daemon under a working load went from 12.8% of one core to 11.6–11.9%.
  • —web app: The phone Home is back to what 0.134.0 showed: the Projects list added in 0.135.0 now appears only at tablet width and wider, so long-unused projects no longer fill the phone screen. The chat field is no longer docked at the bottom: it lives in a Chats section, folded by default, with the newest chats and All chats.
  • —web app: The folded Chats row on the phone Home has a + New chat button at its right: one tap opens the chat field with the cursor in it.
  • —ui: Send just sends, on the phone and on the Mac. A message sent while the agent works was already held and typed in by itself when the turn ended, but it was drawn as a dashed "Queued" bubble with three equal buttons (Edit, Cancel, Send now), which read as a choice you still had to make. It is now an ordinary outgoing bubble with one muted line, "Sends when Claude finishes"; Send now, Edit and Cancel are small text links. The box says "Message will send when the agent is free".
  • —chat: a reply now shows the moment it is written. Claude Code writes a reply's thinking first and its words on a later line of the same message; the daemon merged those words into the stored turn without telling the live feed, so a chat showed the turn's cost and no answer until it was opened again. Measured on a preview daemon: never within 30 s before, 108 ms now. The merge is now an event_update frame, and the chat replaces the reply in place.
  • —chat: a chat says when its agent is at work: Claude is thinking… with a clock from your message, the call it runs (Running npm test, Reading main.go), or Waiting for your permission with the Yes / No card under the conversation, on the desktop, in the app's chat tab and on the phone. A quick chat runs in the default permission mode, so a web fetch or a command could stall it on a prompt the chat never showed.
  • —chat: what you send shows at once as your bubble, Sending… then Sent ✓, until the session's own record of it arrives; one the daemon refused reads Not sent with Retry. Before, nothing showed until the transcript caught up, so a message to a busy session looked lost.
  • —chat: text you pasted shows as a quiet Pasted · N lines block, folded past six lines, instead of Claude Code's raw <pasted_content> tags; a <system-reminder> inside a prompt is left out.
  • —codex: a reply no longer shows its text twice when the import was interrupted between saving the turn and saving its words; the import is retried on the next tick, and turns already doubled are corrected the next time their session file is read in full.
  • —ui: AI;DR on a phone reached through access from anywhere no longer answers 403. The relay drops the browser's Origin and fetch metadata, and the tap sent no Content-Type to stand in for them; it now sends an empty JSON body. A 403 on AI;DR says in words that Caprock refused it.
  • —web app: Keep the phone Home and the permission card inside the screen: a long path, tool name or "always" answer in a waiting prompt no longer widens the page, so cards keep their right margin and the header spans the full width.
  • —ui: a dictated message is never silently lost on the phone. When the speech service ends by itself mid-recording (Chrome on Android does after a pause), Caprock listens again and keeps every word; interim words with no final are sent rather than wiped; a later, emptier result never erases what is in the field; Send while listening waits for the last words. A pointer cancel keeps recording instead of dropping it, and sliding off the mic sends nothing but leaves the words in the field. A message the daemon does not take becomes a draft with Retry, a message on its way when the tab is discarded comes back as a draft, and the field's text survives a discarded tab.
  • —ui: the dictation language can be changed from the phone's + sheet, which was the only way on a chat's screen (it was stuck on the phone's own language there), and the recording line names the language it listens in. Speech recognition still hears one language per recording: the browser's API cannot detect a second one, and Caprock has no transcription of its own.
v0.135.02026-10-18

mcp

  • —let your coding agents ask Caprock what earlier sessions learned. caprock mcp is a free, read-only MCP server on stdio: eleven tools for Second Brain, handoff, session history and prose, costs, CI and loop alerts. Settings → Use Caprock from your agents (MCP) has the commands for Claude Code and Codex, including the full binary path for desktop-only installs. Answers enter your agent's conversation and reach its model; credential-shaped text is redacted. No tool starts, stops or types into sessions or changes settings.
  • —brain: import an Obsidian vault into Second Brain. Markdown notes, links, aliases and tags join the graph, with their vault and source files marked; the original files stay unchanged. Saved vaults refresh every ten minutes while Caprock runs, and can be refreshed or removed from the import sheet. Letting agents use the notes is a separate switch per vault, off by default, covering memory handoffs, MCP reads and model topic lists.
  • —Mail: clean Spam into Trash and permanently clean Trash by message age or local spam score, with a count and confirmation scoped to the selected account or all accounts.
  • —Resources: Larger text, shorter explanations and clearer memory and disk figures make the Caprock doctor easier to read.
  • —Second Brain: give the graph more room with a collapsible details panel, clearer connections, a more open layout, zoom buttons and controls for node size, link visibility and label density.
  • —agents: Restore Codex's original glyph and terminal character; show ChatGPT imports with OpenAI's Blossom instead of a generic chat bubble.
  • —perf(api): the all-time history behind the Now screen and History is no longer counted again on every poll while agents work. It is refreshed at most once per 20 times what counting it took (never past 2 minutes), and its totals reuse the summary's scans instead of making them twice. Today's figures keep their 3-second refresh. Measured 2026-10-18 on an M1 Pro under a synthetic working load (four Claude and two Codex sessions writing, hooks, two dashboard pages): on an 890,000-event history the daemon went from 40% of one core to 11% and the history refresh's allocation from 2,839 MB to 182 MB per 150 s; on a 181,000-event history the refresh's own CPU fell 40% and the daemon's total stayed within noise.
  • —perf(api): the sessions list no longer re-reads every listed session's recent events on every request. Each session's reads are kept until its events change: an insert raises its newest event id, and new database triggers count updates and deletes (migration 0063). Narration, stats, live subagents and the clock are still worked out on every request. Measured 2026-10-18 on an M1 Pro under a synthetic working load (181,000-event history, four Claude and two Codex sessions writing, hooks, two dashboard pages): the list's allocation per request fell from 0.85–0.87 MB to 0.35–0.36 MB and its CPU per request from 1.8–3.2 ms to 0.8–1.1 ms. The daemon went from 13–15% of one core to 11–12%. The cache holds about 9 MB for at most 256 sessions.
  • —pipeline: the Flow row is three cards instead of ten equal tiles: Now (All clear, or the most urgent count with the other two beside it, each still a filter of the list), Speed (throughput, with CI, opened to merged and merged to released as one line) and Agents (the cost, the cost per agent-hour and the team now, which still opens the team chart). A sparkline is drawn only where there is history, and what is missing is said once under the row. The cards stack on a narrow window.
  • —ui: the release plan in the session's right panel reads in plain words. Sections say what they hold, with a one-line hint ("Branches pushed, no PR yet", "Will go into the next batch PR"); a branch row says "3 commits · pushed 2h ago" with its branch name under the subject and a "branch" chip instead of a lone bar; a subagent row reads "Agent: <task>" with "working 12m".
  • —web app: Show projects and their sessions near the top of Home on phones and tablets, including saved projects with no sessions.
  • —navigation: Keep the open session address intact when a tablet switches to the phone layout.
  • —chat: a photo attached in the chat — the phone's Photo above all — no longer puts its full saved path in the field. It shows as an Image 1, Image 2 chip with a thumbnail and an × over the field, and your words stay your own. On Send each photo is pasted into the session first, as the desktop terminal pastes a dropped image, so Claude Code shows [Image #1], then the words follow; before, the path went in as text inside the message. A message held while offline keeps its photos.
v0.134.02026-10-18

termbuf

  • —terminal output no longer allocates as it scrolls through the 256 KiB scrollback ring, in the daemon and in the pty-host alike: the ring is one fixed array written in a circle, and the wake-up channel is made only when a terminal is waiting. Measured 2026-10-18 on an M1 Pro under load (BenchmarkRingWrite, full ring, nobody waiting): a 1 KiB write went from 5153 B and 1 allocation to none, a 16 KiB write from 77,938 B and 1 allocation to none.
  • —codex: a live Codex session's transcript is read from where the last poll stopped, and only what the new lines added is offered to the database, instead of parsing the whole file and offering every record again on each append. Measured 2026-10-18 on a synthetic 16,384-record transcript (Apple M1 Pro under load): one poll after a one-record append went from about 201 MB allocated and 1.2 million allocations to about 150 KB and 300. A restart, a rewritten file and a compressed rollout are still read in full once.
  • —docs: document the resource-usage audit, synthetic allocation measurements and prioritized optimization proposals.
  • —ingest: a Claude Code transcript record longer than the 16 MiB line limit is now skipped while it is read, in fragments, instead of being read whole first: one read of a 32 MiB record allocated 67 MB before and 17 MB after, and the same 17 MB for a 64 MiB one. A complete record over the limit is skipped too (before, only an unfinished one was), counted once as a malformed line, and the line after it is read. Skipping an unfinished one no longer leaves the offset inside it, which made the next read start the whole transcript over.
v0.133.02026-10-18

app

  • —the status strip carries Premium, the header's way to buy, which a terminal in front hides; and, while a plan window is 70% spent, pause before the limit, which opens Premium on the plan-window stop. Neither shows to someone who has Premium.
  • —modules: each private module's sidebar row shows a key, ⌥⌘1 to ⌥⌘9 (Ctrl+Alt+Shift+1–9 on Windows and Linux), that opens it from anywhere in the app. The keys exist only while a module is installed.
  • —machines: your other computers in the app (ADR-067). The sidebar lists each linked computer and the sessions running there, tagged by its name, and opens one with its Session view and terminal as on that computer's own screen, through this Caprock. Settings → Computers adds one by the link it shows, renames, forgets and checks them, says when one is offline, waiting for Allow, removed there or refused, lists the steps for the other computer (install, caprock remote on, caprock pair --computer), and shows a link for another computer to reach this one. With none linked, one quiet + Add a machine row in the sidebar.
  • —resources: GET /v1/resources reports the machine as a whole — cores, one-minute load (not on Windows), total and available memory.
  • —bench: bench/peer.mjs measures keystroke-to-echo through a linked computer against the same session typed into locally.
  • —mail: removing an account signed in with Google now also revokes Caprock's access at Google, not only the token on this machine; for a Microsoft account, which offers apps no way to do that, the confirmation says where to remove the access. docs/mail-oauth.md gains the owner's setup checklist and a step-by-step test on a real Gmail and Outlook account, and the Go tests walk a whole account's life against a fake provider and fake mail servers: sign in, sync, read, send after the access token expired, remove.
  • —projects: put several repositories in one project without a folder around them. In New project, Put several repositories in one project turns the folder list into checkboxes: tick repositories in one folder or several, 2 selected · Continue, name it, done. Nothing is made on disk: the first one ticked is the project's home repository, where agents start, and they reach the others. A project's menu has Add a repository… for later, New agent asks which repository to start in, each session in the sidebar names its repository, and the CI/CD section lists every repository of the project.
  • —doctor: the orphan list now covers Codex CLIs whose session ended or whose parent exited, and terminals still running from an older Caprock install (a Homebrew caprock pty-host left from before a move to the app), shown but never ticked in Clean up. Each process shows its CPU and its command line, cut short with tokens and passwords hidden. The status bar says 3 orphans and the doctor Orphans: 3 · 410 MB; caprock doctor prints the same.
  • —release: the right panel's Release section now shows what heads into the next release beyond merged pull requests: pushed branches with no pull request yet (ahead of the default branch, from the last 14 days, titled by their last commit, with their commits ahead and age; ones whose work is already on the default branch by content, as after a squash or batch merge, are left out) and the session's working subagents (working · 12m). The rows stand in sections with counts: In this release, Pull requests, Ready (branches) and In progress (agents).
  • —ui: Settings has a search field at the top of its menu (above the sections on a narrow window). Typing filters the sections live by title and by every label, hint and button inside them, understands common words (token finds Telegram, dark Appearance, phone Use on your phone, shortcut Global shortcut) and marks the matches; Enter or a click jumps to the matching control and lights it. Esc clears; ⌘F (Ctrl+Shift+F, or Ctrl+F, on Windows and Linux) puts the cursor in it while Settings is in front.
  • —app: closing a working agent's tab asks in one short line, and the two choices carry a few words each; Keep it running is drawn filled as the recommended one, Stop it outlined.
  • —ui: Codex is shown by OpenAI's own mark, the Blossom, everywhere it appears — sidebar rows, tabs, Projects, the pulse, Now cards, the tray, At a glance, Second Brain, the Week and share cards. Where the other agents have characters, Codex is the Blossom on a plain white tile instead of the drawn terminal box. Per OpenAI's brand guidelines it is used unchanged, in black or white only, with no face or other additions.
  • —ui: the ⇧⌘-click link chooser (Ctrl+Shift+click off macOS) now offers two choices: Default browser, named by the browser it is, and Caprock browser, always in the desktop app. Picking Caprock browser while the in-app browser is off turns it on, and a toast says where to turn it off. Copy link follows. Other browsers wait behind a quiet Other browser…, which lists real web browsers only: apps that merely handle web links, such as a chat app, a terminal or a bank's secure-browser wrapper, are no longer offered. The hover hint now reads ⇧⌘-click to choose where it opens.
  • —doctor: stopping a listed process sends SIGTERM and then reports whether it is still running, instead of following up with SIGKILL on its own. A process that ignores the request gets Force stop in the panel, or caprock doctor --stop <pid> --force, as a second, separate choice.
  • —settings: Settings → Use on your phone is rebuilt plain. It opens with one sentence that says whether a phone is connected — iPhone · Safari is connected and can control sessions · seen 2 min ago, or No phone connected yet. — with the one fix beside it when something is in the way. One button, Connect a phone, shows the QR with one line (Scan with your phone's camera. Works once, 4:43 left.). The four-step checklist is gone; how it works, the home-screen link, the way back to the page and the off switch are folded under How it works. Each device is one row: its name, Can control or View only, when it was last seen, Allow control or Take control away, and Remove. Devices not seen for 30 days fold into N old devices with Remove all, confirmed once. The phone's own pages now point to Connect a phone and Allow control.
  • —app: the quick replies under a terminal say what they are: Reply to the agent: leads the row, and each one's hover names the message it sends.
  • —ui: setting up or changing the Telegram bot is one card with three numbered steps that tick as they are done — copy the token in @BotFather, paste it here (a Paste button reads the clipboard), send your bot the code with Open @yourbot. Connected, it is one line, Connected to @bot · alerts to <chat>, with real Change bot, Pair again and Disconnect buttons; where the token is kept moved behind the i. A refused token brings the same card back in red.
  • —terminal: The wheel scrolls Codex again. A terminal opened, reloaded or reconnected after Codex's startup had scrolled out of the daemon's 256 KB output buffer came back on the normal screen with the mouse off, so the wheel never reached Codex. The replay now restores the alternate screen and mouse reporting along with the other modes the session set.
  • —pipeline: in a repository that cuts no releases (none published and no version tag in 180 days, such as a site deployed on every merge), merged is now the last step: the CI block's and the Pipeline's bars end at merged and fill, and the release plan says No releases here: work ships when it merges instead of listing every merge as waiting for a release that never comes. Where releases are cut, the empty released step says what it means and what a merge waits for (the next release (v0.133.0), in v0.132.0); where it cannot be read, the step is drawn as not known.
  • —subscriptions: a session started on an added Claude Code or Codex subscription now runs on that subscription's own sign-in even when your shell exports ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, CODEX_API_KEY or another credential the agent would otherwise use first. Settings → Subscriptions names any such variable it sets aside, and sessions on your own sign-in keep them. Removing a subscription now names the macOS Keychain item its sign-in is kept in and offers Sign out first, which types the agent's own sign-out into a shell tab for you to run. Caprock still never touches the Keychain.
  • —ui: the Settings menu draws an icon for Closing a tab, gives Telegram's chat its own icon beside the bot's, and shows a tooltip only on a title it cuts, so it no longer covers the row above.
  • —ui: when Telegram refuses the bot's token (Unauthorized), Settings → Telegram says so in a large box with the two steps to fix it — copy the token from @BotFather, paste it under Change bot — and opens the bot fields.
v0.132.02026-10-17

ui

  • —an agent's "update available" notice in a terminal gets a small bar near the bottom of that terminal — Codex 0.162.0 available · Copy command · Open a terminal · ×. Copy puts the printed command on the clipboard, joined back across wrapped lines and without the "Run … to update." words; Open a terminal opens a plain shell tab in the session's folder and types nothing into it; × hides it for that agent and version. Caprock never runs the command. Codex's banner and Claude Code's update notices are recognised.
  • —now: At a glance shows which way each share is moving. Every legend row, Numbers table row and agent carries a small marker, the change in its share between the last 7 days and the 7 before (↑ 9, ↓ 4, · under a point), and its hover says both shares in words. When a model's share of spend moved five points or more, one sentence under Where the money went says so. Measured shares only, no projection; GET /v1/glance gained trend.
  • —queue: a message Caprock types into a session it started is sent, not left sitting in the prompt. Tell the agent on a loop alert could leave its text typed into a Codex session's composer, unsent, because the agent read the Enter as part of the fast run of text before it. The text now goes as a bracketed paste wherever the session turned paste on (Claude Code and Codex both do), the Enter follows on its own after a pause that grows with the message, and 1.5 s later Caprock reads the screen and presses Enter once more if the prompt still holds the text — never into a menu, and not once the turn has started. The same path serves the chat field, the phone, a delivery at the end of a turn, Send now and Telegram replies. The loop card now says Sent to the agent or Will send when it finishes this turn; Dismiss never sends.
v0.131.02026-10-17

modules

  • —a slot for private modules (ADR-076). A signed admin key pasted into Settings → License names the modules this machine may run; caprock module install <dir> copies one in from a local folder; each installed module gets a sidebar screen in the app's theme and a small JSON document store. Without a key nothing of it is shown.
v0.130.02026-10-17

brain

  • —the Second Brain graph card unfolds into the graph: the dialog opens out of the card's rectangle with the card's picture growing into it, and folds back into the card on close (a full-screen sheet on a phone; a cut under reduced motion). The graph is coloured by project with the projects and the busiest topics named, Find in the graph and a project picker pick nodes out without laying it out again, and a panel beside it says what the hovered or opened node holds — a project's or a topic's memory count, its latest memories, the sessions they came from and Open — else the whole graph's counts, its projects and its newest memories.
  • —brain: thoughts arriving: a memory just kept flies as a spark from its session's sidebar row or tab into the sidebar's Second Brain card, which pulses, and lands in its project's cluster when the graph is drawn. Throttled, never under reduced motion, and off with Settings → Second Brain, Animate thoughts arriving. The live brain frame of an added memory now carries its session and project.
  • —app: the link chooser (⇧⌘-click a terminal link) lists Open in Caprock after the browsers while the in-app browser (beta) is on.
  • —app: the in-app browser is back as a beta, off until you turn it on (Settings → In-app browser). New browser tab (File menu, ⌘K, the + menu) opens a tab after your agents with the cursor in its address box; plain words search DuckDuckGo, or the engine you pick. A page stays exactly where it was while you work elsewhere, a link that opens a new window opens another tab, ⌥← ⌥→ (Ctrl+PageUp/PageDown on Windows and Linux) step between tabs, and when an agent starts waiting on you a strip names it and a playing video pauses. ⌘-click still opens your default browser. Each page now runs in a window of its own rather than inside Caprock's window, which is what crashed the app on macOS 27 and lost drops in 0.124.0; a page is never closed while it is busy.
  • —ui: the highlight where you type can be switched off, or made stronger. Settings → Terminal → Highlight where you type covers both the faint band on Claude Code's input box and the chat field's accent tint, with Subtle or Stronger. ⌥⌘G (Ctrl+Alt+Shift+G on Windows and Linux) flips it in every open terminal at once and remembers it; ⌘K has Toggle input highlight, and pointing at the input box shows the key on its top edge.
  • —ui: a release plan row opens its session in one click, as a Waiting for you row does: an agent's row (in the next release or not in one yet) is itself the button, and a pull request's row with an agent on it gets a small Open session button after its title. In the app it opens the session's tab, in a browser its terminal or timeline, and on the phone an agent's row opens its chat.
  • —ui: the release plan says itself in plain words: Nothing new since v0.129.0, released 9m ago or 3 changes waiting for the next release · last release v0.129.0, 9m ago, with the usual gap on hover only; idle agents sit under Not in a release yet · agents waiting for you, idle 30+ min; an agent's row with no pull request says no PR yet, and every bar segment names the steps in order on hover.
  • —ui: the session panel's Changes section stays inside the panel: a long file path no longer pushes the counts, the bar and Review and commit past its right edge.
  • —ui: the "New" tag on the active top tab (Second Brain) is readable: on the accent fill it takes the label's own ink on a chip instead of faint grey.
v0.129.02026-10-17

chats

  • —import your Claude.ai and ChatGPT conversations into Chats. Ask the service for your data (Claude.ai: Settings → Privacy → Export data; ChatGPT: Settings → Data controls → Export data), then choose the .zip it emails you under Import from Claude.ai or ChatGPT… on the Chats screen. Each conversation becomes a read-only chat marked with its service, searchable, archivable and deletable, and readable from a paired phone. The file is read on your computer and nothing is sent anywhere. Importing a newer export adds only what is new, and a chat you deleted stays deleted. Neither service offers a way for another program to read your history, so this is an import from the export, not a live sync (ADR-074).
  • —app: choose the browser a terminal link opens in. ⌘-click (Ctrl+click on Windows and Linux) still opens the default browser; ⇧⌘-click (Ctrl+Shift+click) opens a small chooser at the pointer with the browsers installed on this machine, the default first, and Copy link. Arrows move, Enter picks, Esc closes. The link's hover hint names both clicks, and both are in the shortcuts list. The daemon finds the browsers from what the OS says handles web links (LaunchServices on macOS, the registry on Windows, desktop entries on Linux) and answers only this machine (GET /v1/browsers, POST /v1/browsers/open).
  • —ui: each repository in the CI/CD section has its own Hide button beside active 10d ago, with an Undo in its place for ten seconds, and the section's last line says N hidden · Show to list them and bring them back, one at a time or all at once. The ⋯ menu that held the only Hide is gone.
  • —ui: a row of the Repos section opens under itself (its chevron, or More for every row): the branch checked out, the newest CI run with its result, open and merged pull requests, red runs, the release, the lines of 7 days, the folder, and git's own error in full when it could not be read.
  • —app: closing the tab of an agent that is still working asks with two equal buttons, each with a line saying what it does: Keep it running (only the tab closes; in focus, so ⌘W then Enter still keeps it) and Stop it (it stops now; continue it later from History or the sidebar). Stopping was a quiet link under one button before. Escape still cancels.
  • —projects: the Repos section no longer says "git could not be read" for a busy repository. Its one git log --all --shortstat over 14 days took 12 s on a checkout with 961 refs and 1,048 recent commits (measured 2026-10-17), past the 5 s git timeout, and the failure was kept for 10 minutes. Commits are now listed without diffs, which stays fast, and each commit's lines are counted once, kept, and diffed in batches under their own budget; when it runs out the row keeps its commits, marks the lines … and goes on counting a minute later.
v0.128.12026-10-17

release

  • —0.128.0 was tagged but never published: a phone Releases test read the "read N ago" line before the stats answer had drawn it, and failed the release run. The test now waits for the line; 0.128.1 carries 0.128.0.
v0.128.02026-10-17

api

  • —AI;DR: POST /v1/events/{id}/tldr squeezes one agent answer into at most two short sentences, in the answer's own language, by a separate call to your own claude (Second Brain's headless runner, Haiku, no session written), only when you tap it. Never typed into the session. The result is kept, so a second tap makes no call; an answer under 280 characters is its own TL;DR. A paired controller phone may use it (ADR-072).
  • —ui: Halloween pumpkins for Claude's portrait: eight of them (a jack-o'-lantern, a witch, a vampire, a ghost, a black cat, a mummy, a monster and a lantern) on a new Pumpkins tab of the face menu, and A pumpkin per session in Settings → Appearance. Nobody's current face changes. Through October the quiet try a duck? line beside an idle portrait says try a pumpkin? instead and opens the pumpkins.
  • —support: Report a bug, from the phone bar's + menu and ⌘K Report a bug…: type or dictate what happened, attach up to three screenshots (made small on the device), add an email if you want a reply, and press Send to support. It goes privately to Caprock support at caprock.dev, never to GitHub, and only on that click; the sheet lists exactly what is sent — your text, the screenshots, the email if given, the Caprock version and the OS (ADR-073).
  • —phone: the chat's bottom bar is one row: a speaker switch that reads new answers aloud (on by default, remembered per device), a + menu with Photo, New chat and Report a bug, a one-line field that Enter sends, and a round hold-to-talk microphone — hold to speak and release to send, tap to start and tap again to send, slide off to cancel. The latest answer carries AI;DR (a two-line summary from your machine), Read aloud, Continue and Stop. Esc, Tab, the arrows, Enter, Ctrl+C and the dictation language moved to the session's ··· menu. The desktop chat is unchanged.
v0.127.02026-10-16

ui

  • —Halloween pumpkins for Claude's portrait: eight of them (a jack-o'-lantern, a witch, a vampire, a ghost, a black cat, a mummy, a monster and a lantern) on a new Pumpkins tab of the face menu, and A pumpkin per session in Settings → Appearance. Nobody's current face changes. Through October the quiet try a duck? line beside an idle portrait says try a pumpkin? instead and opens the pumpkins.
  • —ui: the Pipeline's Next release panel says where the release itself is: the next version's bump and why (minor · a feature since …, patch · fixes only), then its stages as the pull request rows' pills — merged → release PR → CI → tag → builds → published → site — the current one filled, done ones ticked, and any stage Caprock cannot read (the site's deploy) left neutral with a ? rather than guessed. A tag already pushed and still building shows above as releasing, its builds counted as files on the release against the release before it. Under the rows, one line for the last release: when, how many files, and a link.
  • —github: the releases the board lists carry their uploaded file count (assets), from the same request.
  • —phone: a session or chat can be ended from wherever the phone lists it, not only from inside it (owner: trial chats could not be ended). Each live row Caprock started — the home's Chats and Continue, All sessions, the Chats screen — has End session… in its ⋯ (a swipe left opens it too), one confirm, then the polite stop that keeps the conversation; the row then reads ended, and an ended chat stays in Chats to be continued. A chat's own ⋯ has End this chat… first. Where End is not offered the row and the session's sheet say why in one line: a session started in your own terminal is never stopped by Caprock, and a phone that can only watch is told where to allow control (Settings → Use on your phone → Let it control sessions).
  • —ui: the version panel can check again right after a check. When you are on the latest version and the last check is recent, the "Checked 4m ago" line carries a Check again link beside the What's new button.
v0.126.02026-10-16

chats

  • —Chats are their own kind of session. Start one in one tap on the phone home, the new Chats screen, or ⌥⌘N in the app, on the model you used last (a compact model chip changes it). Chats are listed apart from your projects, newest first, with search, rename, archive and delete, and a count; a chat reads as a conversation, never a terminal, and can be continued after it ends. The files a chat writes open in a viewer — Markdown rendered, HTML sandboxed with no scripts, PDF in the browser's viewer, images — and can be downloaded, on a phone too. A paired phone that may control can now start, continue, archive and delete chats. Quick chat with… keeps the model chooser (ADR-071).
  • —app: ⌘W closes the tab in front, not the window. The macOS menu's Close Window held ⌘W, so the key hid the whole window; File → Close Tab holds it now and closes the tab as its × does, asking first about a running agent or shell. With the Dashboard in front it does nothing; Window → Close Window (no key), the red button, ⌘H and ⌘M still put the window away, and in the Mail window ⌘W still closes it. Windows and Linux keep Ctrl+Shift+W, since Ctrl+W is the shell's delete-word.
  • —app: the Premium dialog's Also in Premium chips are bigger, so the row is seen at first glance.
  • —app: the version panel says You're on the latest version in a sentence and shows one button — what's new right after a check, Check now otherwise.
  • —app: a Waiting for you row on Home is one big way back into that session: a click anywhere on it, Enter, or its amber Open terminal → opens the session's tab in the desktop app (opening one if needed), or its terminal in a browser. what did it ask? still opens the answer in place, and the row now names the session after the project, so two waiting rows in one project say which is which.
v0.125.02026-10-16

docs

  • —docs/install-agent.md says what the install does, where it lives and how to remove it; the agent install prompt is now five lines, runs the same script as the site, and leaves Claude Code's settings to the app.
  • —ci: make e2e drives the app's golden paths in Playwright's own headless shell against a sandboxed daemon (temp home and data directory, a free port, the fake claude): the workspace loads with no console errors, sessions open as tabs with a sidebar row in Tree and Flat, typing, paste and a native file drop reach the terminal, and the palette, Settings, Second Brain, Pipeline and the share dialog open. make pr-check runs it when ui/ or internal/api/ changed, and CI runs it on Ubuntu. The app's own tests now check that a drop is heard as both the window's and the webview's event, through one handler (the 0.123.0 bug).
  • —mail: AI sorts the inbox. With AI on, Mail labels your newest 200 conversations, then new mail as it arrives, as Important, Updates, Newsletters or Spam, with a one-line summary in place of the snippet and a Reply mark on mail that waits on your answer. Filters along the top of the inbox — 1 All (everything but spam), 2 Important, 3 Updates, 4 Newsletters, 5 Spam with its count. Nothing moves on the server by itself: ! sends a conversation to Junk, n says it is not spam, and Mail remembers that for the sender too. While it works the list says Sorting 37 conversations…; with AI off one line offers Let AI sort your inbox — Turn on and says where the text goes. Sorting uses the provider you picked in Mail's settings, a batch of conversations per call, in the background. The sample inbox shows the labels without any AI call.
  • —mail: Who does the AI work is named wherever it acts — under the inbox filters, under every summary and on the composer's AI bar — with change opening Settings at the AI section, and the reader shows a conversation's stored one-liner at once. A new Language setting (same as the email, English, Russian, Spanish, German) sets what summaries are written in; a drafted reply keeps the language of the mail it answers.
  • —mail: junk stays out of the Inbox, even when your mail provider leaves it there (owner request, translated: "there's a lot of spam in my inbox — every mail client filters this"). Caprock's own filter runs on the machine, with no AI and nothing sent anywhere, and is on from the start. It reads what the receiving server wrote into each message — SPF, DKIM and DMARC results and the server's own spam verdict — plus a few cheap checks (a brand name on an unrelated domain, replies sent to a free mail address, URL shorteners, image-only mail), and a small learning classifier taught by your Spam folder and by every Spam (!) and Not spam (n, new) you press. Anyone you have written to, or said Not spam for, is never junk. Hidden mail is listed under Caprock thinks these are junk in Spam (g !), whose count the sidebar shows, and the reader says why; nothing is moved or deleted on the server unless you press Spam yourself.
  • —mail: with AI sorting on, Spam is one list and one count whoever found it: the inbox's Spam filter (5) and the Spam folder view show the junk filter's catch first, then the AI's, each with a single Spam label, and n clears both at once — the filter learns, the AI's label is taken back, and the sender is never spam again. Mail the filter caught is never sent to the AI.
  • —mail: Mail from the keyboard, with the key on every action. Each button shows its key as a small chip — e archive, # delete, v move (a numbered folder list), s star, u unread, i summarise, r/a/f reply, reply all, forward, c compose — the folders show g i, g u, g s and the rest, / search, , settings, g r check mail, and ? lists every shortcut. t opens the theme list: j/k show each palette at once, Enter keeps it, Esc goes back, a digit picks — Sunlight outside, Dark Pastel at night. Summarise is there even with AI off: the first press says which provider would answer and where the mail goes, and Turn on and summarise switches it on. In the composer ⌘↩ sends and ⌘J drafts the reply with AI.
  • —app: feedback is one box: type what's wrong, drop or paste a screenshot anywhere on it, and press Send (⌘↩, Ctrl+Enter off macOS). The first line becomes the issue's title. Bug / Idea / Question, the diagnostics, Choose image… and Capture window are under Details; nothing else is required. What you typed is kept until it is sent, so Esc loses nothing. Send still opens a prefilled GitHub issue in your browser for you to submit; Caprock sends nothing itself.
  • —pipeline: The Next release plan uses its width. Its bar grows with the panel, one taller segment per step (agent, PR, CI, merged, released), each named on hover, and the step under way breathes softly. Beside the stage word, as the width allows, each row says what it waits for now (CI on Windows · 6m, needs a rebase, fix: go (windows-latest), a release), the agent cost recorded on it at API list price (— where no session was), a dot per OS its checks ran on, how long its CI took against the usual (14m +27%, 6m / ~11m while running), and how long ago it was opened. In a narrow panel these columns drop first, and the plan never scrolls sideways. On the phone each row also says what it waits for.
  • —brain: Second Brain leads with what it is for. One status line on top (Up to date · 278 memories · 137 new today · 91 topics · $0.15 in 24 h, or the stage and N of M over the progress bar while it builds) replaces the three step blocks and six tiles, with the rest under Details. The middle is What your next session gets — pick a repository and see the up-to-five memories a new session there is handed, in order, and whether one opening now would get them — then Recent memories with search, a project and a span. The graph is a small card on the side; a click opens it in a near-full-screen dialog to zoom, pan and explore (Esc, × or the backdrop close it). The Graph | List switch is gone. New read-only GET /v1/brain/handoff?project= names the memories with the same pick the handoff makes.
  • —app: every dashboard page says where you are and how to get back. Off Home the header reads caprock / Pipeline (the browser tab too: Pipeline · caprock), one item is highlighted — Settings included — and a ← Home button returns to Home in the same tab. In the app the Pipeline and Mail tabs carry ← Dashboard, which brings the Dashboard tab back. Links that leave Caprock are marked ↗ and say Opens in a new tab (or Opens in your browser in the app); links that stay inside use → instead. The version dialog's link to turn release checks on now opens Settings rather than falling through to Home.
  • —app: Mail opens in a window of its own, not as a tab inside Caprock's window with the sidebar beside it. In the desktop app the sidebar's Mail, ⇧⌘M and the palette's Open Mail launch or bring forward the Caprock Mail preview when it is installed (macOS, in Applications); without it they open a second Caprock window, Caprock Mail, with Mail alone, brought to the front when it is already open. In a browser the Mail tab stays, under a banner, Mail works best in its own window, whose button — or ⇧⌘M on the Mail tab — opens Mail alone in a browser window.
  • —app: the Premium dialog shows the whole plan, not only the feature you clicked. Its header counts the plan and states the yearly price ("Premium · 8 features · $30 a year"), and under the feature one row of chips names the other seven; a chip shows that feature in place, and the ones added in the last 30 days are marked New.
  • —app: the right panel's CI block lists only what is in flight: open pull requests, and the ones merged since the latest release. Released and closed ones are gone, so a repository that publishes no releases (a site that pushes straight to master) no longer fills the block with every merge. With nothing in flight it says so in one line, with the last release and how long ago it was, over the three newest merges. Ring and By OS are never blank: without the session deck's own PR they draw the newest pull request in flight, read its checks themselves, and say "Nothing in flight" when nothing is.
  • —ui: a freshly opened window no longer logs a 404 for /v1/projects/dir:…/cleanup: the finished-copies line waits for the daemon's own project list instead of asking about a project the window derived while that list was loading. Found by make e2e.
  • —app: a new release shows within minutes. The release check is due every 15 minutes (it was hourly, so a release could sit unannounced for most of an hour), and when you come back to the window Caprock checks again if the last check is ten minutes old. Never more than one check per ten minutes unless you ask; still one conditional request to GitHub, and only while Settings → Updates allows it.
  • —app: the answer to a check is at the bottom right, beside the version, where you look for it: Checking for updates…, Caprock is up to date for a few seconds, Couldn't check for updates with Retry, or Update to vX — Restart as an amber pill. The banner at the top centre is gone from the app.
v0.124.12026-10-16

brain

  • —the wheel scrolls the page past the Second Brain graph instead of zooming it. Click the graph (or press Enter on it) to zoom and move it; an accent ring shows it is in hand, and Escape, a click outside or the mouse leaving lets it go. A click on a dot still opens it, and Fit works either way.
  • —app: the Page tab is gone for now: with it the app crashed on macOS 27 and drops went missing. ⌘-click opens a link in your browser again (ADR-070, withdrawn).
  • —phone: quick replies under a session's chat take the page's own surface in a light theme (Light, Light Cold, Book and the rest) instead of drawing as dark pills; under an agent's terminal in the app they still sit on the terminal's ground.
  • —app: a file dropped beside a terminal — on the tab strip, a margin, the composer — goes to the terminal you used last instead of vanishing, and a drop with nowhere to go says so.
v0.124.02026-10-16

app

  • —the sidebar's Today card has three states, kept per viewer: the full card, one line ($33.79 · 11 running · 5h 6% · 7d 74%, plus what waits when anything does) by its chevron, or hidden by the × that shows on hover. Right-click offers Full card, One line and Hide; Settings → Appearance → Today card in the sidebar brings it back. Its Share button is gone: sharing stays in ⌘K and on Pipeline.
  • —app: every tab in the strip has a row in the sidebar, in Projects and in Tabs view. A project or session hidden by hand and then opened again (from More…, a link or a restored workspace) had a ⌘N tab and no row; it is now listed while its tab is open, and goes back behind More… when the last tab closes. Hiding a project from its menu also closes its tabs.
  • —ci: a computer that removed this one is said so at once: a second check of a machine, asked while one is in flight, waits for its answer instead of returning the state from before it, so the proxy answers 403 "revoked" right after a revoke rather than passing the other computer's 401 on (the two-daemon smoke test's Windows failure). In New project from a template, a card clicked before the sheet had settled is the one chosen; the first card is chosen by default without an effect that could overwrite the click. make pr-check runs the whole UI suite when ui/ changed, not only the changed test files.
v0.123.12026-10-16

app

  • —a file dragged onto a terminal (a screenshot's thumbnail, a Finder file) types its path again; 0.123.0 heard no drop at all.
v0.123.02026-10-16

app

  • —a Page tab — open a web page inside Caprock, apart from your agents. ⌘-click (Ctrl-click on Windows and Linux) a web link in an agent's terminal or chat, pick Open a page… in ⌘K or the + menu, or type an address such as localhost:5173 into ⌘K. The page gets its own tab after every agent's, with a globe, back, forward, reload, an address field, copy and Open in browser; it takes no ⌘1–⌘9 and no sidebar row. In the desktop app it is the system's own web view with no access to Caprock and no cookies shared with it, loaded only when you ask, and freed after five minutes behind other tabs; in a browser, only a page on this machine opens in a tab and any other goes to the browser.
  • —remote: revoking a device closes its terminal opened through the relay at once, with "this device was removed". A keystroke arriving as the socket closed used to drop the stream without telling the other computer, and its terminal stayed open. After one failed write the stream now drops what is typed next at once instead of writing it into the closing socket.
  • —daemon: stopping the daemon waits for its background work (history backfills, pruning, Second Brain's loop, the projects list and any clone still running) before the database closes, instead of leaving it to read a closed database; a clone in progress is stopped rather than left writing.
  • —ci: make pr-check, the step before a pull request: the changed Go packages with -race, their changed tests three times, a Windows vet and test build, and the changed dashboard tests. Four flaky tests were fixed at the root.
v0.122.02026-10-16

pipeline

  • —what an hour of agent work costs, and how big the agent team is. Two new figures end the Flow row for the picked repository. Cost / agent-hour is the last 24 hours' spend per active agent-hour, at API list price (an estimate). An agent-hour is the burn rate's own rule: one session's clock hour with five or more priced calls. The figure carries a 15-day line, ▲ 35% vs 14-day avg, and tokens per hour and in all. Agent team shows the main sessions and subagents at work in the last ten minutes and today's peak. Its line counts distinct agents a day, captioned with the 14-day average and peak, with each day split by agent kind on hover. Pressed, it opens a 31-day stacked bar chart, like a contributors graph. Both are read from this machine's events by the new GET /v1/pipeline/agents. With too little history they say not enough history.
  • —pipeline: share the figures. A Share pill in the Pipeline's head draws the eight Flow figures, each with its 14-day line, as a 2x card in the screen's theme, with the dates and caprock.dev. Save it, copy it, post it to X or LinkedIn with words you can edit (the card goes on the clipboard to paste in), or send it to Telegram: To me through your own paired bot (POST /v1/telegram/photo, only on that click), or to someone else through Telegram's share link. The card and the words are anonymous by default ("a GitHub repo", aggregates only, no titles, branches, authors or paths); Show repo name is an opt-in for that one share and is never remembered. The dashboard's share sheet gained the same Post and Telegram actions.
  • —machines: one of your computers can work on another's sessions (ADR-067). On the computer to be reached, caprock pair --computer prints a link (turning access from anywhere on if it is off); on the other, caprock machines add '<link>' pairs with it through that computer's own relay, end to end encrypted. Nothing is let in until the first computer's owner allows it by name: a card on its screen, the prompt in the terminal that printed the link, or caprock machines allow <name> over SSH. An allowed computer can start agents in any folder, open shells and run tasks there, but never change that computer's settings, pairing, updates or the doctor's stop. caprock machines revoke <name> or Remove ends it at once, open terminals included, and caprock machines forget <name> on the other side drops it from both. On Linux both caprock service install and caprock pair --computer say when loginctl enable-linger is needed to stay reachable with nobody logged in.
  • —phone: the phone home shows your releases: per repository, the next version and a row per piece of work with its ten-block progress bar and stage word, the release plan the Pipeline and Work → Release draw. A row opens its pull request on GitHub; a chip switches repository. It reads on opening, on release news, on Refresh and on a pull at the top of the home-screen app — never on a timer. A paired phone may now read GET /v1/sessions/{id}/ci/stats (read-only), which the bars come from.
  • —phone: end the session you are in from its bar: a red End on a session Caprock started, one confirm sheet (The conversation is saved and can be continued.), then the polite stop: Claude Code is asked to exit and write its transcript out, and is killed only if it is still running a few seconds later, so the session can be resumed. Sessions started outside Caprock, and phones that may only read, see no End.
  • —phone: the session's bar stays at the top while Timeline, Changes or Answers scroll, over the notch, so the way back is always one tap away.
  • —brain: see Second Brain build. A header strip on its screen says the state (Building · reading session 12 of 48 over a progress bar, Up to date · last built 14:02, Paused, Stopped, Waiting for today's limit), which step a read is at (find sessions, Claude extracts, link into the graph) with the session being read, what it gives (how many new sessions were handed its memories) and counted figures: memories, new today, topics, links, tokens and cost at API list price in the last 24 hours. The folded sidebar card reads Building · 12 of 48 over a thin bar while it builds. GET /v1/brain/status carries stage, the read's phase and these counts; migration 0058 records the memories each new session is handed.
  • —hermes: Caprock sees Hermes Agent. Its sessions, prompts, replies, tool calls, tokens and cost are read from Hermes's own database, read-only, and shown beside every other agent, with a hermes filter and Hermes's own glyph. Cost is Hermes's figure, never re-priced, and the Cost screen says which part the provider reported and which is Hermes's estimate. Cron jobs and gateway chats (Telegram, Slack…) are shown apart as background runs, by source. Delegated subagents sit inside the session that delegated them, counted once. New agent → Hermes starts hermes chat in a terminal tab and links it to its Hermes record; with no model set, hermes model is offered as your own step. See docs/hermes.md.
  • —app: Markdown reads comfortably in a file tab, and goes to another agent in one click. The document sits at a reading measure with a larger heading scale; code blocks have a Copy button, wide tables scroll on their own, task lists show real check boxes, headings take anchors that in-page links jump to, and a relative picture is read from the project through the daemon (one on the web stays its alt text, so nothing is fetched from outside the machine). Contents sit beside a long document on a wide window and in a dropdown on a narrow one, with a thin reading-progress line. The tab's header has a quiet toolbar: Copy path, Copy for an agent (@/absolute/path, which Claude Code, Codex and the others read as a file reference), Copy Markdown, Download, and Send to a session, which types that reference into a running agent Caprock started without sending it, then brings that agent forward for you to add your words. A .md path an agent prints in its chat or terminal now opens the file in a tab (⌘-click in the terminal, Ctrl+click elsewhere), and Open file… lists Markdown files first. Every "link" is a path on this machine; nothing is uploaded.
  • —ci: a pull request no longer carries the built dashboard (internal/api/dist). Every two PRs that touched ui/ used to conflict there, because the bundle's file names are content hashes; now a PR that changes it fails CI's no dist in the PR check (and make check) with the command that drops it (make dist-reset), and a new dist workflow rebuilds and commits it on master after each merge that touches ui/. go install …@latest still embeds the tag's dashboard: release verify and make release-ready check that the tagged commit's copy is its own ui/ built, so a release is tagged after that rebuild lands (ADR-068).
  • —brain: the Second Brain graph looks like a knowledge graph: one round disc of small dots that fills the panel, with topics and projects as little bursts of their memories. Dots are in the text colour (dark in a light theme) with a faint tint for topics, edges are hairlines, and the one accent marks the node you hover or open. No names at rest: the hubs' names fade in as you zoom, then every memory's, and hovering names a node and its neighbours while the rest fades. It settles before it is shown and then stays still (no frames at rest); drag a dot and its neighbours follow, scroll or pinch to zoom, and Fit brings the whole disc back. The sidebar's turning brain uses the same small dots and hairlines.
  • —pipeline: the In flight, Needs a fix and Ready counts open the pull requests behind them. Hovering or focusing one lists up to five with why each is there (CI red, Conflict, Changes requested), and Needs a fix says how many the page saw fixed in the last 24 hours. A click filters the list to them across the whole repository, even from This session, with a Needs a fix · 4 × chip to clear it, and scrolls to the list. A red pull request's CI pill names the failed checks on hover.
  • —app: "Caprock on this computer" (the popover behind caprock 1.6 GB · 1.8 GB in the status bar) says what is wrong in one line — All good — Caprock uses 1.6 GB, or Caprock is using more than it should with one Clean up (frees ~X MB). Below it, one row per kind of problem in plain words, each with a line of why, its size and a small Stop: Test copies of Caprock (left by agent work in temporary folders and worktrees, not your app), Leftover helpers (servers finished agent sessions left running), Disk reads (normal or more than usual). Pids, paths and the long explanations moved under each row's Details; the memory and disk breakdown is behind What the 1.6 GB is spent on. Every stop shows the list first, one checkbox per process with the safe ones ticked, and stops nothing until you click Stop. The Advanced checkbox is gone.
  • —remote: turning access from anywhere on in the first moments after Caprock started could leave it connected but reporting itself off, so the pairing link came back empty until the next restart. The settings are now read once, under the same lock as the switch.
  • —remote: a request cancelled by the phone or another computer is no longer ended as if its answer were complete, so a cut-short answer is never read as whole.
  • —app: only the tab you are in looks selected (a user could not tell where she was: the tabs waiting on her wore the same accent bar along their top as the tab in front, so four tabs looked selected). The tab in front alone has the 3px accent edge, the panel's fill and a bold title, and room for about 16 characters of it — it read ca…; there its project's name gives way first. A tab waiting on you shows an amber dot before its title, pulsing slowly; one looping or failed a red dot. Never a bar.
  • —app: closing an agent's tab is one decision at most, and always says the conversation is saved. An idle agent Caprock started ends with its tab, no question, and a toast says Closed. The conversation is saved — reopen it from History or the sidebar. One still working or waiting gets one sentence, Close tab (it keeps working in the background) and a quiet Stop it instead; no checkbox. Settings → Closing a tab still makes it always keep or always stop. Ending now asks the agent to stop (SIGTERM, then a kill after 5 seconds) instead of killing it at once, so Claude Code writes its transcript out and the session can be continued with --resume; the agent-signal endpoint takes end for this.
  • —tests: the template sheet test reads the plan once the parent folder has loaded, so it no longer fails on a slow runner.
  • —ui: Live Pulse puts the hour's cost in one right-aligned column: every row's chart, money and agents tracks are fixed widths, so $28.98 and $0.03 no longer sit at different places depending on the model name beside them.
  • —ui: Telegram is always a choice on both share sheets (Share your figures and the Pipeline's), in a row right under Share… / Copy / Save. Send to my Telegram sends the card through your own bot to your own chat (or reads Set up Telegram until one is connected); Send to someone in Telegram copies the card and opens Telegram's chat picker — the Telegram app itself in the desktop app when it is installed, else t.me — saying "Image copied — paste it in the chat".
v0.121.02026-10-15

brain

  • —Second Brain says where it is kept — in its Settings and under Details in How it works: the real path of caprock.db as the daemon resolved it (with Copy path), the tables brain_nodes, brain_edges and brain_runs, and that the session text goes to Anthropic through your own Claude to be read while nothing goes to Caprock. GET /v1/brain/status carries the path as db_path.
  • —remote: the groundwork for reaching your other computers through the relay (ADR-067). A Caprock daemon can now be a client of another machine's daemon over the same end-to-end encrypted channel a phone uses, pair with it by its link and keep the token in machines.json (0600), opening the channel only while something needs it. Nothing to see yet: the routes, the target's computer pairing and the app follow.
  • —app: the sidebar's session rows are bigger and easy to read (a user could not read them, nor tell which tab she was in). Rows are 38px with the title at 14px; each says its state as a dot and a word — green and breathing while it works, amber while it waits, faint when idle — and a tab's key is the footer's readable chip. The flat list no longer draws the project's name on each row; the tooltip and screen readers still say it. Only the tab in front looks selected now (an accent tint, a 3px accent edge and a bold title): a row waiting on you is marked by its amber dot and word, no longer by a fill and edge of its own, and the project in front lost its edge. The Today card folds to one line by its chevron and can be hidden from its right-click menu; Settings → Appearance shows it again, and while it is hidden the list's heading counts what waits on you.
  • —app: the sidebar's Second Brain card is quiet by default: folded to one muted row, Second Brain · Idle · 140 memories or Indexing 3 sessions…, whose small glyph pulses slowly only while a read runs. It lifts a pixel on hover, a click opens Second Brain, and the chevron unfolds a calmer orb (smaller, slower, fainter, fewer sparks). A fold you saved is kept. The card now sits under the list and above the footer instead of at the end of the list's scroll, so its status line is never cut off by the footer; when room is short the orb shrinks (and in a window under 640px tall is not drawn). The orb's sparks also follow the sessions working now.
  • —app: the right panel's CI pull-request bars say what their stages are: one faint line, open · CI · merged · released, over the bars' columns, and each segment's tooltip names its step and state (done, running, failed, waiting) with the landing-queue label on the step a PR is on. The rows are as tall as before.
  • —app: on macOS the first click on a Caprock window that is not in front reaches the page: a tab's × in the title-bar strip closes the tab at once instead of the first click only bringing the window forward. Dragging the window by its title bar works as before.
  • —loop: a subagent's edits no longer raise a loop alert on its parent. A subagent's tool calls arrive under the parent's session id, and the detector counted them in the parent's window: a parent and two subagents making the same edit, three each, read as nine repeats of one call. Each thread now counts its own — the main thread, and every subagent apart — and a subagent that really repeats itself still raises the alert on the session it works for. A hook fired inside a subagent with no agent_id is told apart by its transcript path (…/subagents/agent-<id>.jsonl). The Week card's longest loop counts the same way.
v0.120.02026-10-15

brain

  • —the Second Brain banner and How it works show the agents you actually use — up to five of Claude Code, Codex, Gemini CLI, OpenCode and DeepSeek, from your last 30 days, drawn as Caprock's own characters (Claude in the face you picked, duck included) — around the brain, each line carrying a pulse into it: One memory for all your agents, with a line naming them (Claude Code, Codex and Gemini CLI remember together.). With no usage yet it is Claude Code alone. Both now say Free in a quiet accent pill, and the dialog adds Free — runs on your own Claude, on this computer. The pulses hold still with reduced motion.
  • —app: a release plan with progress bars — on the Pipeline, in the right panel and on Work. The next version your changelog entries call for (a feature makes a minor release, fixes a patch; hover it for how it was counted), then a ten-block bar per pull request merged since the last tag, open, or agent still at work, with its step in a word — a landing-queue label such as landing: queued when it has one, the CI step red when a check failed — and drafts and idle agents under Later. It is a Next release panel under the Pipeline's figures, a Release section in the right panel and a third view on Work beside List and Kanban. The right panel's sections now move by a small grip at their top left: drag it, press ⌥⇧↑ ⌥⇧↓ on it (Ctrl+Shift+↑ ↓ on Windows and Linux), or click it for Move up, Move down, Move to top and Reset order; the order is kept for you.
  • —app: the release check runs hourly instead of every six hours, and within five minutes of the computer waking; it is the same conditional request to GitHub, and nothing new is sent. Settings has its own Updates section: the switch, when Caprock last checked, and Check now. In the desktop app it adds a choice for a new version: Ask me (the card, as before) or Update automatically, which downloads the update in the background and restarts into it only when no agent is working or waiting on a permission prompt, you have left the app alone for two minutes, and nothing typed is waiting to be sent — with ten seconds' notice and Not now, and Updated to vX · What's new afterwards. Sessions and tabs come back as after any update. The update card also offers Always update automatically.
  • —app: a check for updates you ask for always answers. Nothing newer: a banner says You're up to date — Caprock X is the latest · checked just now, with Close, and goes after 8 seconds unless the pointer is on it. A failed check shows the error with Retry. The hourly background check stays silent when there is nothing new.
  • —brain: Second Brain starts in one click and you can watch it being built. The empty screen is one picture, Give your agents a memory, Free and Start building, with one line on what it does (your own Claude reads your last 30 days on this computer, up to the daily limit, nothing goes to Caprock) and the rest under Details. The click turns it on and reads the past month. While it builds, the screen shows the session it is reading, 12 of 41 sessions with a bar, memories found counting up, today's tokens, where it all comes from per project, and the last reads, with Pause and the one fix for a problem (Check again, Raise to 1M a day). A build now reads session after session instead of four every five minutes.
  • —brain: the living brain: a slowly turning cloud of your memories with your agents at its rim and sparks flying into it while it reads, in the sidebar card and on the screen. At rest it added about 0.6% of the page's main thread (measured in a headless browser, 2026-10-15), and nothing while out of view.
  • —app: the status bar's agent cpu … ram … opens a popover on hover, focus or click: CPU and memory over the last 15 minutes or hour for the agent in front, all sessions and Caprock itself, each with its figure now, peak and average, and the sessions using the most right now, a click on one bringing its tab forward. The daemon keeps the samples it already takes in a fixed one-hour ring (GET /v1/resources/history): 157 KB for ten sessions, no new sampling.
  • —app: quick replies: under the terminal of every agent Caprock started, and above the chat's field, a row of small buttons sends Continue, Shorter, Explain, Run tests or Commit in one click, the way the composer sends. Edit the list in Settings → Terminal → Quick replies. Sessions started in another terminal never get the row.
  • —changelog: a pull request no longer edits CHANGELOG.md. It adds one file, changes/unreleased/<slug>.md, holding its entry under a kind: line, and the release writes them all into the new version's section with scripts/changelog.py assemble; scripts/changelog.py preview shows that section before then, and make docs-check rejects a malformed file. With many pull requests in flight, editing one shared section made git conflict and, three times on 2026-10-15, put entries silently into releases that had already shipped. The Work page's next-release list reads these files too, from any repository that keeps them.
  • —brain: the sidebar's Second Brain card is one click target into the screen, folds to one row that says what it is doing (Extracting from acme-api, Indexing 3 sessions, Idle, Paused) with the count, and its Start building opens the screen instead of a wall of consent text.
  • —brain: the graph reads at a glance: memories cluster around their project, a dot's size is its links and its colour its kind, only projects and the busiest topics are named, and hovering a node lights its neighbourhood and dims the rest. Settings on the screen are one gear with three choices; How it works says its details in three short lines.
  • —ui: the Take Caprock with you strip has two buttons instead of five: Set up opens a small list of Phone and Telegram, each with its state, its own action and Not needed, and Later hides the strip for three days. The × folds it to a 0 of 2 chip in the status bar that opens it again.
  • —ui: a GitHub read that fails on the network or with a 5xx now says Couldn't reach GitHub. Showing data from 6 min ago. with Retry, the raw error and the exact time in the tooltip, and then what GitHub's status page reports: an outage with a link to it, or all systems operational — likely your connection. Watch GitHub Status in that line, off by default, shows a dot in the status bar while GitHub reports an incident. The daemon reads githubstatus.com only while GitHub is connected, with nothing about you.
  • —store: four indexes that later ones had replaced are dropped, giving back 125 MiB of a 1041 MiB database of 519,000 events, with every query as fast as before. The space returns as the database's idle vacuum runs.
  • —app: terminals reconnect on their own after the network drops. A terminal whose connection was lost retried on a single timer, and the only thing watching it stopped with the socket: when the browser refused to start a new socket once (as a network stack being swapped under it can), or the timer was lost, the tab said Reconnecting… for good and nothing typed reached the session until the tab was closed (user report: every tab stuck after a Wi-Fi drop). A terminal between sockets is now checked every second and always has a retry on its way and on time; a pause for its own backlog resumes after 15 s even if the terminal never reports catching up; and the window getting focus checks the connection at once. The badge says the attempt and the next try, and after 3 failed tries or 10 s down offers Reconnect now, which reconnects the same session — no new process. The dashboard's live connection got the same watchdog and button.
  • —brain: search finds any language's letters in any case, several words, and small typos. Second Brain's search matched one phrase, and ignored case for Latin letters only, so a memory in Greek or with accents was missed when typed in another case. Now every word has to match, in any order, in a title, summary, project or linked topic, with case and accents ignored in every script; when little matches exactly, a word a typo or an ending away counts too, said as including close matches. Results come best first — title above summary above a close match — and the List view marks the matched words. The search box says what it searches.
  • —app: the terminal in front no longer turns to garbage after a long day of tabs. xterm's WebGL renderer shares one glyph atlas among every terminal, merges its pages when it fills, and after the second merge kept drawing from the old page's texture (the new merged page carried the same per-page version number as the old one): letters clipped and overlapped over the whole screen until font size + then −. Every terminal now re-sends its atlas to the GPU and redraws when a merge happens, and the window back in front after half a minute away does the same for the tab shown.
  • —app: the sidebar footer is a band of its own under the list, on a solid fill with a hairline on top, so the edge between the list and the tools reads at a glance in every palette; the list scrolls above it and its last line (Second Brain's Local · your own Claude · free) is whole when scrolled to the end. Folded, a key above an icon is part of that icon's button: a click on ⌘D, ⌘P, ⇧⌘M and the rest does what the icon does, with its tooltip.
  • —pipeline: the repository picker lists only git repositories on GitHub. A folder outside git (a copy in Downloads, a quick chat's folder, the paste folder), a repository with no remote and one on another host are no longer offered, so picking an entry never ends on "This folder is not a git repository with a remote". Two clones of one GitHub repository are one entry, and two entries that would read the same word show owner/name, else their parent folder. The N more… list follows the same rules, and Home's Pipeline card picks from the same list. New read: GET /v1/repos/remotes.
  • —ui: the theme picker is walked with the arrow keys, each palette shown as the focus lands on it; Enter or Space keeps it.
  • —agents: text copied from an agent (its pbcopy) keeps non-Latin letters: every process Caprock starts gets a UTF-8 locale when it has none, not only agent tabs, and the daemon sets one on itself at start.
  • —ui: the UI tests that turned PRs red at random pass every time. The chat test's live store dialled a socket nobody answered and flipped the link to reconnecting mid-test; the cockpit tests compared times drawn from the clock with fixtures stamped when the file loaded; the app workspace tests handed the dashboard a summary without savings, so the Now screen crashed whenever it drew one; and the first test to open the lazily loaded dashboard paid for transforming it inside its own 5 s, then asked for its links before they had drawn. Tests only; nothing in the product changed.
v0.119.02026-10-15

mail

  • —the AI assistant works with the subscriptions you already have. Summarise and Draft reply can go through your own logged-in Claude Code, Codex or Gemini CLI (run headless, with no tools, hooks or session file), through an API key for Anthropic, OpenAI, OpenRouter or DeepSeek (kept in your keychain, never shown again), or through a local model server as before. Automatic, the default, uses the first of claude, codex, gemini and a local server this computer has, so with Claude Code installed turning AI on is the only step. Settings show which provider answers and where the mail text goes; it is sent only when you click an AI action, and the thread reaches the model fenced as data it must not follow.
  • —app: New project from a template — the dialog. From a template… in New project's footer or ⌘K: pick a saved template or look one up (owner/name, its link, or a folder), name the project, paste or drop the brief, and the box under it says what Create will run and send (the name and visibility, to GitHub, through your own gh; nothing without it). Progress shows in the footer; Esc hides it and the agent still starts in its own tab when the project is ready. A failed clone offers Try the clone again and Open on GitHub. Settings → Project templates edits and removes them. In the app a dropped .md or .txt becomes the brief (read_dropped_text, only from the latest drop).
  • —projects: start a project from a template — the daemon side. Add your own template (a GitHub template repository, a folder, or both) with where its brief goes and the agent's first prompt; Caprock reads its TEMPLATE.md to fill those in. Making a project from it runs your own gh on your click (gh repo create --template, then gh repo clone), or copies the folder's last commit into a new repository when you pick the local copy, then writes your brief and commits it — never pushed. A failure takes away only the folder Caprock made and never deletes a repository on GitHub; Retry picks up where it stopped (ADR-065).
  • —pipeline: every figure on the Pipeline page — and the five on Home — shows its last 14 days as a small line, with today set against them underneath: usual, above usual +38%, below usual −30%. In flight, throughput, the CI median, opened → merged, merged → released and the agent cost each get one; green where today is the better side (a shorter CI run or merge, more merges), amber the worse, muted for the open count and the cost. Running totals compare today so far with earlier days at the same hour; a day with nothing to measure is a gap, not a zero; with fewer than three earlier days of data it says not enough history. Hover a day for its value, or the word for what it is compared with. Counted from what the Pipeline already reads: no new GitHub request.
  • —app: the Pipeline has its own tab again, pinned with the system tabs right after the Dashboard tab and before Mail, drawn like them: its glyph and ⌘P (Ctrl⇧P on Windows and Linux), its × while it is in front. ⌘P, Pipeline at the foot of the sidebar and Open Pipeline in the palette bring it to the front, and again go back to where you were; ⇧⌘B still works and ⌘B stays the Pipeline bar. It stays in the Dashboard too: Pipeline in its menu and Open Pipeline → on Home take you to that tab, so the page is drawn once and, behind other tabs, reads nothing. The tab is kept across restarts.
  • —phone: Settings → Use on your phone tells you how to get back to a phone you already paired. Each paired phone says Already paired — just open m.caprock.dev on that phone, and Open on the phone again shows a QR code of that address (the address only: no code and no key, so it pairs nothing) and the Add to Home Screen tip. With a Telegram bot set up, Send the link to my phone sends the address to your Telegram chat through your own bot, and says so before you press it.
v0.118.02026-10-15

app

  • —a live session's row on Home takes you into the running session. Click anywhere on the row (or Enter on it) and the session's tab comes to the front, as its ⌘N does, or opens if it has none; Go to session ↗ with the tab's key shows on hover, and details still opens the session page. A session started in your own terminal says in your terminal and offers Continue in Caprock: a new process on a copy of the conversation, in a new tab, leaving the one in your terminal untouched. An idle session Caprock started gets a small × End, which asks first only when messages wait in its queue; a working session and one your terminal started never do. In a browser the row keeps its open link.
  • —work: the Work page — every piece of work on the machine on its way to a release, one row per branch at its furthest stage, grouped Next release, In review, In work, Planned and Released, or as Kanban columns. Each row has a bar along agent → PR → CI → merged → released filled to its stage (red when CI failed or it conflicts), who works on it (the agent and model, started by Caprock or your terminal), elapsed time and agent cost at list price. Estimates appear only with a measured basis: a turn's rough range, CI against its median, merge to release against the median lag, and the gap since the last tag against the usual gap. Rows that need you come first; agents idle for over 30 minutes fold into one line. It is second in the Dashboard's menu (⇧⌘W; Ctrl+Alt+Shift+W on Windows and Linux, and again back to where you were), and Home shows its headline and first rows above the Live pulse with Open Work →. A session row brings that session's tab to the front; a pull request's title opens it in the Pipeline with its detail open.
  • —work: The Work board's daemon answer, GET /v1/work (ADR-064): every live session on the machine — started by Caprock or by hand — with its state, its running turn's rough range (Claude Code only, never while it waits on you), its own task list, what it cost today and the pull requests it recorded; the task runner's tasks and the queued messages; and each repository's [Unreleased] CHANGELOG entries, newest tags and commits past the last one, read from local git as of your last fetch. Nothing is written, nothing is polled, GitHub is not asked and git never fetches. The screen comes next.
  • —pipeline: The PRs & CI board's flow strip carries how often the repository ships: the median gap between its releases already read (release_gap_median_ms, with three gaps or more), never a date for the next one.
  • —app: a Sunlight theme for working in bright light. It is black text on pure white with no faint greys: every text colour, the faintest included, is 7:1 or more against its background, measured on every surface. Text is one weight heavier, borders and hairlines are dark enough to see in glare, and the sidebar is opaque. The status colours and the accent are saturated and stay apart. Its terminal keeps all 16 ANSI colours at 5.3:1 or more on white (the normal eight 7:1), the yellow and cyan that usually wash out included. Pick it under Light in the theme picker or Settings → Appearance.
  • —app: where you type is easier to find, quietly. The Chat view's field carries a hint of the accent at rest — a faint tint and a hairline edge, a little stronger with the focus — on every theme and terminal palette. In a Claude Code terminal, the rows of its input box (between the two rules, at the ❯ prompt) sit on a faint band of the terminal's own background, lifted a step: it follows the box as it grows or moves, leaves any cell Claude Code coloured alone, reads only the bottom rows of the screen, and draws nothing for other agents, shells, dialogs and menus.
  • —app: keys named after what they open. ⌘D opens the Dashboard (⌘0 still does), ⌘P the Pipeline (⇧⌘B still works), ⇧⌘T the theme picker, and ⇧⌘W Work; ⌘F stays Find and ⌘T New shell. On Windows and Linux they are Ctrl+Shift+D, Ctrl+Shift+P, Ctrl+Alt+Shift+T and Ctrl+Alt+Shift+W — never Ctrl+D, which stays the shell's end of input. The sidebar's footer shows Dashboard ⌘D, Pipeline ⌘P and the theme's ⇧⌘T, folded or not, and the palette, the shortcuts list and the tooltips say the new keys.
  • —dashboard: The Pipeline is a page of the Dashboard now, beside Week in its menu, instead of a tab of its own: ⌘P and Pipeline in the sidebar open the Dashboard tab there (Ctrl+Shift+P on Windows and Linux), and again go back to where you were. It shows one repository — the agent in front's, else the one you looked at last — with buttons beside its title to switch. It is drawn as the Dashboard is: its figures are the Dashboard's tiles, larger and readable, each explained in a card under it on hover rather than a tooltip over the number; every pull request's whole title; and its four steps tell done (ticked), current (filled, ✗ when red, a turning ring while CI runs) and still ahead (outlined) apart by shape as well as colour. A pull request's detail opens beside the list. Home's CI/CD section shows the Pipeline's first five figures at its top — in flight, needing a fix, ready, throughput, CI median — with Open Pipeline →. A Pipeline tab saved by an earlier version opens as the page.
  • —ui: Get premium in the header is a quiet tinted button, not a bright block: the same height, radius and type as Feedback and Settings beside it, a faint premium tint with a hairline border and premium-coloured text, the price a step fainter, a little stronger on hover. Its text reads at 4.5:1 or more in every palette.
  • —brain: Second Brain looks like what it is: a brain made of a graph. Nodes and links in the side view of a brain, in the memory kinds' colours, softly pulsing, with a new node dropping in now and then and linking up (still with reduced motion; CSS only, nothing runs in the background). How it works is big and short — Your agents' memory, one line on what it does, three one-line steps, Over time: your personal Jarvis, one line on privacy and one button; the defaults (Haiku, 300k tokens a day, up to five memories) moved into a folded Details. The empty sidebar banner leads with the same brain, Give your agents a memory, Start building and How it works.
  • —picker: an arrow key pressed in the folder picker the moment a folder's entries appear is no longer lost; the selection used to be cleared just after it was made.
  • —app: the theme picker opens whole, above everything. From the sidebar's footer its palettes were cut off at the sidebar's edge and the terminal drew over them; it now opens over the whole window, kept inside it on every edge at any sidebar width and with the footer folded, and scrolls in a window shorter than itself. Opening it puts the focus on the palette in use; Esc, or tabbing past the last palette, closes it and gives the focus back to the button.
  • —app: the sidebar footer's shortcuts are readable. Folded, the keys above the icons were small faint 10px chips; now every key in the footer is the New agent chip's size, folded or not. Where one has no room — beside a row's words, or over a folded icon — it is left out whole rather than drawn smaller, and the key stays in the tooltip: a row measures itself and brings the chip back when it widens. Off macOS the folded keys still show on hover, now at full size.
  • —mail: the Caprock Mail preview app reloads onto Caprock's new page after Caprock updates (it used to keep showing the page from before the update when the port stayed the same), and a click on its Dock icon brings the window back.
v0.117.52026-10-15

brain

  • —Second Brain's status says when the reader wakes next (next_wake), so a screen can tell you when the first memories may come instead of leaving you at 0. Turning building on still wakes the reader at once, and a double click or a second yes never starts a second read beside the first.
  • —mail: Mail on its own. <daemon>/?mail=1 is the inbox and nothing else of Caprock — no sidebar, no tabs — titled Caprock Mail. app/mail-mac builds a Mac preview app around it: Caprock's icon with an envelope in place of the dot, one window on the running Caprock, so the accounts and mail are the same as in the Mail tab. It waits for Caprock when it isn't running and follows it to a new port.
  • —brain: the sidebar's Second Brain block is a small banner while it is empty, instead of a bare 0 and two faint lines: a faint graph breathing (still with reduced motion), Remembers what your agents learn — so the next one starts smarter., Start building and How it works, and Local · your own Claude · free. Start building asks in place — your own claude and its model, on this computer, the daily token limit, where the prose goes, and that Settings turns it off — and only Turn on switches it on and starts on your last 30 days of sessions, within that limit; without Claude Code it says so and offers no button. Once on and still empty it says what it is doing: Reading your recent sessions…, or how many sessions are waiting, and the last problem if there was one. The same line replaces your personal Jarvis under the Second Brain screen's title.
  • —brain: Second Brain's How it works now explains it in ten seconds. Remembers what your agents learn — so the next one starts smarter: it keeps the decisions, findings, ideas and open questions your sessions arrive at and hands the relevant ones to the next agent you start — over time, your personal Jarvis, an assistant that knows your projects. Three steps say how (a session ends → your own claude reads it on this computer, Haiku and 300k tokens a day by default → up to five memories go into the next session's handoff), and one line says that nothing leaves the computer except that call and where to turn it off.
  • —app: typing a name in New project's folder field now searches the folder it is in. ~/dev/cap used to empty the list; now it lists the folders of ~/dev holding cap — the ones starting with it first, the letters marked — and every key narrows it at once. The first match is highlighted: ↑↓ move, Enter picks it, Tab completes to it and goes inside. No match says No folder starting with 'cap' in ~/dev. Works with ~, backslashes and drive letters on Windows.
  • —windows: Caprock no longer flashes console windows that steal the keyboard focus. The daemon ran without a console, so every git, gh or claude -p it started opened a window of its own for a fraction of a second, every few seconds, and typing in other windows came out ragged (user report on Windows 10). Every child now starts with a hidden console, and a test keeps it that way.
  • —mail: adding an account can no longer sit on Checking… forever. A server that goes quiet while Mail connects or signs in — IMAP had no limit and SMTP waited five minutes per command — now ends the check within 30 seconds with what happened: which server stopped answering, and that another mail app open on the same account can count against the server's connection limit. A synced mailbox still waits in IDLE as long as it needs; the limit is lifted once signed in.
v0.117.42026-10-15

alerts

  • —a loop alert is now one sentence and three actions. It says what happened — caprock ran the same command 5 times in 3 minutes (python3 …/p2e.py), the whole command on hover — with the costs quiet underneath. Tell the agent shows a message marked [Caprock] and signed as sent by Caprock's loop alert, which you can edit, then puts it in that session's queue so the agent reads it when its turn ends; it is offered only on sessions Caprock started, and a session you started yourself gets Open session instead. Dismiss has a menu with Don't warn about this command again in this session and the one remaining cap link, Set a spending cap…. The same card is on Now, in the agent cockpit and on the session page.
  • —brain: it is called Second Brain everywhere it is named — the sidebar block (SECOND BRAIN), the dashboard's screen row, the screen's title, ⌘K (Second Brain, Add to Second Brain…) and Drop into Second Brain — with a brain outline as its mark (it was a drop falling into a bowl, which did not read as a brain) beside the name, and one faint line under it in the sidebar and on the screen: All your agents build your personal Jarvis over time.
  • —app: the chat button in the tab strip says what it is: a speech bubble and the word Chat, as Deck has its word, lit while the chat is shown. Its tooltip says what the chat is: this session as messages — your prompts, the agent's replies, one line per tool call — and a field that types into it, while the terminal keeps running.
  • —app: terminals no longer show rows of wrong letters ("ι ɤr ·r") after another tab redraws its glyphs. All terminals share one set of drawn glyphs; when one tab rebuilt it (back after five minutes, after the Mac woke, on a display change) the others kept pointing at places in it that now held other letters, until their rows were drawn again — for a couple of seconds after closing a tab or leaving the chat, or for good in a tab that printed nothing. Every terminal now redraws from the new set at once, and a tab coming back into view sends its glyphs to the GPU again. Under the chat (and under Changes or a comparison) the terminal is put away as a tab behind another is: what the agent prints meanwhile is drawn at once when you come back.
  • —app: a terminal no longer opens with its letters spread far apart after the app restarts (an update): a letter cell measured before the terminal's font settled was kept until you changed the font size. A cell no monospace font could have is now measured again on its own, when the fonts are in, when the window comes back and when the tab is shown.
v0.117.32026-10-14

pipeline

  • —⇧⌘B (Ctrl+Alt+Shift+B on Windows and Linux) opens the Pipeline in a tab, or brings it to the front; pressed again there, it goes back to the tab you were on. Open Pipeline is in ⌘K too.
  • —cockpit: the CI block's PRs view shows the repository's pull requests as thin progress bars toward a release — Open → CI → Merged → Released, CI in its check colour — with each one's number, title and time on its step. It scrolls; Expand makes it taller and is remembered; Pipeline ↗ opens the page.
  • —brain: Second Brain is free, all of it. There is no cap on the memories you drop in by hand (it was twenty without a licence), and the automatic build from finished sessions works for everyone who says yes — it runs on your own claude. The Brain screen asks everyone once; the sidebar's Brain block starts unfolded, with no lock, and How it works explains it and opens the Brain screen. A Premium licence bought for Second Brain stays exactly as it was, with every other Premium feature. GET /v1/brain/status no longer carries licensed, free_max or free_left, and no Second Brain route answers 402.
  • —pipeline: PRs & CI is now Pipeline, everywhere: the tab, the sidebar footer, the palette, the deck. Tabs you had open come back under the new name. The page is one list of pull requests instead of a board and a list: each row its number, its whole title, its labels, and the steps Open → CI → Merged → Released as pills, the one it is on highlighted, then its agent cost. In flight first, then merged, then released (ten, then Show more). Click a row for its detail over the page; × or Esc closes it. The red-runs and Actions-minutes figures went with the old list.
  • —app: the sidebar footer's button is Pipeline: a click opens the page or goes back from it. Showing or hiding the bar under the agent is now on the bar itself (Pipeline and ×), ⌘B, and Pipeline bar: On/Off in ⌘K.
v0.117.22026-10-14

mail

  • —Import everything. Mail lists the mail accounts it can read on the computer — Thunderbird, and Apple Mail including the Google, iCloud and Exchange accounts in System Settings › Internet Accounts — each address once, under the app it came from and when that profile was last used (accounts from a profile unused for 90 days start unticked). An app it cannot read is not shown. Tick them and press Import: accounts with a saved Thunderbird password connect by themselves, then each other account asks for its one sign-in in turn (Google or Microsoft in the browser, an app-specific password for iCloud, otherwise the password), with a progress list of connected, waiting and failed accounts, Retry, Skip for now and Come back. A build without Google or Microsoft sign-in says so and asks for an app password instead of showing a button that cannot work.
  • —release: Mail's Google and Microsoft client IDs are compiled into releases from the CAPROCK_MAIL_GOOGLE_CLIENT_ID, CAPROCK_MAIL_GOOGLE_CLIENT_SECRET and CAPROCK_MAIL_MICROSOFT_CLIENT_ID GitHub secrets, the desktop app's daemon included; unset, a release still builds.
  • —now: each figure in Today has a small line of your last 14 days and one word against your own median: usual, above usual +38%, below usual −20%. Running totals (cost, sessions, turns) are compared by the same time of day, so a morning no longer reads low against whole days; Burn now is set against your typical 10-minute window while working. Cache hit gets the line only. Hover says what it is compared with; under five days of history it says not enough history. It follows the agent filter. New endpoint GET /v1/stats/trends.
  • —brain: Second Brain builds itself, once you say yes. When a session has been quiet for ten minutes, your own claude (Haiku, no tools, no transcript, none of your hooks) picks out what it decided and found, and each memory links back to the passage it came from. A daily token limit (100k, 300k or 1M), short sessions skipped, failures retried later; Build from my last 30 days shows an estimate first. Settings → Second Brain shows what it used and read. A new session's handoff now carries up to five of the repository's memories.
  • —app: Caprock may also ask about a feature it is considering — running Hugging Face models on your computer, and one of watching short ads to earn money or YouTube while agents work. At most one strip a day; each question is two clicks and an optional email. Nothing leaves the computer until you press Send, which sends those answers to caprock.dev and nothing else; Not interested and × send nothing, and an answered strip never returns.
  • —brain: Second Brain drawn. The Brain screen opens on a graph of your memories and their topics and projects — drag, zoom, click a dot to open it, List for rows — smooth with thousands of memories. The app's sidebar has a Brain block below the projects: the newest thirty as a small graph where each new memory drops in, and the three newest titles. The strip that offered Second Brain ahead of its arrival no longer shows.
  • —prs: the PRs & CI board follows merges and releases instead of freezing. With a dozen open pull requests whose checks were running, the daemon read each one's state one after another, ran past its deadline on every refresh, and the board kept showing its afternoon answer for hours — merged pull requests still under CI running, new releases missing — behind a bare updating…. The open pull requests are now read four at a time within a budget, the board's answer always comes back (a card whose checks could not be read again in time is drawn faint), and releases are read every 2 minutes instead of 10. The header always says how current the data is: Updated just now, Saved 5 h ago · refreshing… while a kept copy is read again, or Couldn't refresh: <why> · showing data from 15:02 · Retry; the CI/CD section on Now says the same. A refresh that keeps failing is no longer retried on every poll.
  • —ui: selected text stays readable in every theme. In a terminal, Claude Code's own selection (fullscreen mode) painted its highlight in the palette's blue, which in a dark palette is light, under light text — in Dark Pastel 1.6:1, nearly invisible. Every terminal palette now keeps text at 4.5:1 or more against whatever is behind it, as the light ones already did. Selected text on the page sits on a lighter wash of the accent (20%, was 35%), written out per palette: body text 8:1 or more, muted text 4.3:1 or more.
  • —phone: a paired phone stays paired. A dropped connection, a computer or relay restarting, a phone asleep or offline, a computer still on an older Caprock, or a QR opened twice no longer send the phone back to Settings → Use on your phone. Three causes: a phone hanging up in the middle of an answer took the computer off the relay for every phone; the relay could hand the next phone a stray message meant for the one before, which the page read as "not your computer"; and the page believed a single bad answer. Now the page asks to pair again only when the computer has turned the phone away on two separate connections, or refused it three times in a row, and it still checks once a minute and comes back by itself if that was wrong. Needs the new page at m.caprock.dev and the relay redeployed; the computer's half comes with the next release.
  • —status: an agent you set going no longer reads waiting for you for seconds after. Answering a permission prompt — a key in the terminal, the card's buttons, the phone, Telegram — shows it working at once (approved — running …), instead of when the allowed command finished; answering No shows it waiting for you. Enter at its prompt shows starting on your message straight away, and goes back if nothing was sent. Every screen (sidebar, tabs, cockpit, tray, phone) now updates a session's status as soon as it changes rather than up to two seconds later, and a Codex or Gemini session you type into is read at once instead of on its next poll. Measured with a test agent: a prompt to working about 1 s → under 10 ms; an approved 15 s command 16 s → under 10 ms.
v0.117.12026-10-14

now

  • —the tokens beside each pull request's cost in the CI/CD section (9.4M tok · $2.12); the hover splits input, output, cache read and cache write.
  • —brain: Second Brain. Keep the answers worth keeping as memories — an idea, a decision, a finding, an open question — each linked to its project, its topics and the moment in the session it came from. Drop into brain sits under every answer on Memory, in a session's Answers and in the chat view; ⇧⌘H (Ctrl+Shift+H on Windows and Linux) or ⌘K → Add to brain… takes a line of your own. Brain in the screen row searches them by words, project and time, and opens one to edit, pin, link to a topic or delete. Twenty memories are free; Premium keeps every one. Kept in Caprock's database on this machine.
  • —app (macOS): macOS no longer asks for the microphone (and for Desktop, Documents or Downloads) again after every Caprock update. The app and its daemon are now signed with one Caprock certificate that stays the same across releases, instead of an ad-hoc signature macOS saw as a new program each time. The first update to this release asks one last time; later updates keep your answers. Opening the app the first time is unchanged: it is still not signed with an Apple Developer ID.
  • —phone: a session's header is one 44px bar: back, the title, a dot for its state, and Chat | Terminal with ··· for the other views. The project, permission mode, model, controls and details are a tap on the title away. In the chat the bar is drawn over the conversation and slides away while you scroll down, back on scroll up or a tap at the top, so the text never jumps. On the relay page at 390×844 the conversation began 178px down the screen (the computer bar and three header rows); now 44px, and 0 while the bar is away (owner, 2026-10-14, translated: "I want the header to take minimum space because I need to read this chat").
  • —phone: the relay page's computer bar ("DimaS-MBP ▾") is gone while the connection is fine and one computer is paired. It comes back while connecting or offline, and with several computers paired it shows outside a session; inside one, the switch is in the title's sheet.
  • —chat: the message field glows softly in the accent while it has the focus, a little more once there is text in it; on the phone and the desktop alike.
  • —now: Live activity is now What happened: the moments of your sessions instead of every command. Files edited (grouped by folder), tests and checks with their result (passed, failed (exit 2), or no result recorded when there is none), builds, commits with their message, pushes, pull requests opened or merged, subagents by their task, a turn ending with the first line of the reply, a session waiting for approval, and a command that failed. Reads, searches and the like are a quiet +N on the moment they led up to. On one day of real use that is 1,974 rows where the old feed had 11,493. All commands brings the old feed back; the panel folds like the others and remembers it; every row is the same height.
  • —app: the sidebar's bottom block reads more clearly. Dashboard ⌘0 has a row of its own again; under it, one row in two halves: PRs & CI named with its mode (CI Off, CI Bar, CI Page; ⇧⌘B in its tooltip) and the theme with its name. On Windows and Linux the two take a row each, and PRs & CI shows its key, Ctrl+Alt+Shift+B.
  • —app: folded, the bottom block's icons are a size bigger, and on a Mac each one has its key drawn small above it (⇧⌘F, ⇧⌘M, ⌘0, ⌘, ⇧⌘B). On Windows and Linux the keys are too long for that and show above the icon you point at.
  • —phone: dictation listened only for the browser's language, so Russian speech came out as English words. A chip beside the microphone (EN, RU…) says the language and changes it: a tap moves through the browser's languages, then opens a list of common ones (English, Russian, Ukrainian, German, French, Spanish, Portuguese, Italian, Polish, Turkish, Japanese, Chinese, Korean, Arabic, Hebrew). Remembered per viewer; shown prominently once when the browser names more than one language.
  • —now: the activity feed no longer says "session ended" at the end of every turn and every subagent's stop; it says finished its turn or subagent stopped, and "session ended" only when one did.
  • —phone: a phone can type the moment it pairs. A newly paired phone could only look until a second click on the computer — impossible once you have left the house. Settings → Use on your phone now has This phone can type and start agents beside Show a code, ticked by default and remembered on the computer; the code carries it, so the phone pairs ready to work. Untick it for a phone that should only look. The panel lists the steps one line each and, while the code is up and after, ticks off what is done and what is left: access from anywhere, the phone paired (by name), whether it can type (with Let it control sessions right there if not) and whether the app is installed on it. The phone opens on a short Ready card — connected to which computer, whether it can type and, if not, exactly where on the computer to fix it, the install button, and Open sessions. m.caprock.dev needs a redeploy for the phone's half.
  • —now: the CI/CD section lists only repositories with a session in the last 14 days, most recent first. Older ones, and any you hide from a block's new ⋯ menu (Hide from CI/CD), fold into one quiet line — 3 older repositories · 1 hidden — show — and are not read from GitHub until you show them; hidden ones come back with Show again.
  • —now: GitHub no longer runs out of time on the CI/CD section: the repositories are read three at a time, each with its own 12 seconds, instead of one after another under a single deadline that left every later one context deadline exceeded. A slow GitHub is one grey line — GitHub didn't answer in time — retrying — once for the section when every repository says the same, and the names show at once while the figures load.
  • —now: a session that drove many pull requests is no longer charged whole to the one whose branch it ended on. It is split between the pull requests it worked on, and such a row says shared; the release pull request #601 read as $1,884.29, the whole orchestrating session, and now reads $2.12 (measured on the owner's database). This also corrects the PRs & CI Board's agent cost.
  • —app: on Windows and Linux the Settings row's words no longer run under phone, Telegram, and Send feedback's key is no longer cut off at the sidebar's edge.
  • —app: the dashboard keeps its place. Switch to a terminal tab and back, or from Home to Cost and back, and each screen is where you left it: scrolled to the same spot, with the Live pulse rows, panels and folds you opened still open. A screen out of sight asks the daemon nothing.
  • —app: ⌘0 (Ctrl+Shift+0 on Windows and Linux) brings the dashboard back where you left it; pressed again, it goes to Home at its top. The sidebar's Dashboard row does the same, and Home in the dashboard's bar scrolls Home to its top.
  • —ui: the main dashboard screen is called Home (it was Now), in the bar, the palette and the back links. #/home opens it; old #/now links still work.
v0.117.02026-10-14

app

  • —closing the tab of an agent Caprock started now asks: Close tab (in focus, so ⌘W then Enter works as before and the agent keeps running) or Close and end the agent, which stops it (asking once more if it is still answering; the conversation is kept). Tick Don't ask again to keep your answer; Settings → Closing a tab changes it back. An agent you started in your own terminal is never ended from Caprock: its tab just closes.
  • —app: in the Tabs list, sessions running with no tab sit under a quiet Running, no tab line, so an agent whose tab you closed no longer looks like an open chat.
  • —app: a Mail button with its own key. Mail ⇧⌘M in the sidebar's bottom block (Ctrl+Shift+M on Windows and Linux) opens Mail in a tab of its own beside the Dashboard tab; pressed again it brings it to the front. With no account yet it opens straight on a sample inbox, with a line on top and Add account (Google, Microsoft, iCloud, Other, import); adding your account leaves the sample. Also in the palette as Open Mail.
  • —app: Send feedback opens with ⇧⌘F (Ctrl+Alt+Shift+F on Windows and Linux) from anywhere in the app, and its row in the sidebar shows the key.
  • —cockpit: put the sections in your own order. Arrange sections under the last one lists them with ↑ and ↓ for each, Reset to default order and Done; nothing moves by dragging a heading, so it never happens by accident, and the order is kept for you.
  • —cockpit: every section folds to one line on its label, like Context and Plan — Now, Subagents, Tasks, Recent tools, Changes, CI and Processes — and stays folded or open as you left it. CI and Processes can be arranged with the rest.
  • —cockpit: rubber ducks. Eight of them sit beside Claude's faces — classic, debugger, on call, party, pirate, hacker, sleepy and rainbow — on a new Ducks tab in the menu a click on the portrait opens. Settings → Appearance → Claude's face can make every session a duck (A duck per session) or mix them in (Faces and ducks); the default stays the eight faces, and every session keeps the face it already had. Now and then an idle Claude asks "try a duck?" beside its portrait: at most once a day, never while you type or a prompt waits, and never again once you have seen the ducks or closed it twice.
  • —app: New project in three big steps. ⌘O, Add project or New project… in New agent: pick a folder (a place such as ~/dev, then a click on the folder; Choose… opens the system's folder picker in the desktop app), see what is there — One repository, 3 repositories as cards all ticked for one multi-repo project (or one each), or a plain folder — and Start an agent here. A folder inside a repository is that repository, with no question. The same step says whether the repository already has a CLAUDE.md or AGENTS.md (Show reads it) and takes optional instructions for every agent there. Clone and a new empty folder are in its footer.
  • —app: what a project's agents are told is in sight. New agent shows a bigger block, Project instructions — Set · 12 lines or None, and the repository's own CLAUDE.md / AGENTS.md (members' too), each read on a click and Caprock's edited in place; the project's page shows the same, and its sidebar row's tooltip says it in words. A multi-repo project's repositories are named under the Project field (3 repos: api, web, infra). The daemon lists the files on each project and returns their text from the folder scan (instruction_files).
  • —now: an opened Live pulse row is calmer and says more. Terminal, Open repo and Open session are quiet links at the end of the project's line instead of a row of bright buttons, and the empty space beside each live session now shows its cost, its last hour, how long it has run, its context, its branch with lines added and removed, and its pull request with CI. The legend can be hidden; the ? in the panel's header brings it back.
  • —app: the PRs & CI view has one button, at the foot of the sidebar beside Dashboard: its picture shows the mode — Hidden, Bar (the session deck under the agent's terminal) or Full page (the PRs & CI tab) — and a click moves to the next. ⌘B and ⇧⌘B work as before and the button follows them. The deck's own Expand and Hide buttons are gone. The theme button moved beside it, out of the Settings row.
  • —cockpit: Recent tools says something. Each call shows what it acted on — a file relative to the project, or a command's program and first words instead of the variables set ahead of it — and how it ended: a quiet ✓, exit 1 or error in red, the time only when it was slow, and how long ago. The same call repeated in a row is one row with ×N. The bar beside each call is gone, and the count in the heading now says what it counts (31,713 this session).
  • —cockpit: CI's OS view is now By OS and shows the newest pull request's checks per runner — Linux ✓ passed 1m 48s, Windows ● running — without opening its row in the deck first; each of the three tabs says what it shows on hover.
  • —phone: one way to pair a phone: from anywhere. Settings → Use on your phone → Show a code now turns on access from anywhere (the encrypted relay) in the same press — a line under the button says so — and the QR works on Wi-Fi and mobile data alike. The Wi-Fi / Tailscale address choice, the Same Wi-Fi step and Keep on after Caprock restarts are gone from the panel. A phone paired on a home-network address worked on that Wi-Fi only, and nothing said so when it went dead on mobile data. For a machine without internet, caprock up --lan and caprock pair still pair over the Wi-Fi from a terminal.
  • —agents: a permission prompt you answered from a button no longer comes back if the daemon crashes a moment later. The answer cleared the stored prompt in the background, so a crash before that write landed restored buttons for a dialog already answered; the answer now returns only once the stored prompt is gone.
  • —app: the PRs & CI bar's header no longer runs past its right edge when the sidebar and the cockpit are both open on a laptop screen: the least useful words drop first and the rest wraps to a second line.
  • —phone: m.caprock.dev says where you are. The name Caprock and caprock.dev stand beside the mark, a line says the page is the other half of the Caprock app on your computer (with where to get it), and the first step shows a small picture of Settings → Use on your phone → Show a code.
  • —phone: the phone no longer sits on Connecting… with nothing after it. A computer that went to sleep kept its relay connection open on paper, and the relay handed the phone to it; the phone waited forever for an answer. The relay now treats a computer it has not heard from in 75 seconds as offline (caprock remote relay and the hosted relay, which needs a redeploy), and the phone gives every attempt 10 seconds, checks the computer itself, not only the relay, while connected, and after 5 seconds says what is wrong — Your computer is offline or Caprock isn't running, Your computer isn't answering, Can't reach the relay, This phone was removed: pair again — with Retry and Pair again. It retries by itself, and at once when you come back to it.
  • —phone: Caprock installs on an Android phone once, when you press Install the Caprock app on m.caprock.dev. Chrome's own install prompt no longer pops up in a browser tab of an already-installed app, a browser tab says Caprock is on your home screen instead, and the old Wi-Fi link no longer offers a second home-screen icon: it says to use the app from m.caprock.dev. Opening the same pairing QR twice keeps the pairing that worked, and a page left open across a new version of m.caprock.dev reloads instead of failing.
  • —mail: Mail opens instead of failing with "null is not an object (evaluating 't.accounts.length')". A mail client found on the machine with no accounts it could read — Apple Mail before Full Disk Access — came back with a null list, and the first-run screen crashed counting it.
  • —app: the closing questions' buttons read on one line. "Keep it running, close tab" broke over two lines and "Stop and close" spilled out of its button; the sheets are wider, every sheet button keeps its label on one line, and the row wraps as a whole on a narrow window.
  • —app: a tab's project name fits. It was cut at nine characters (uncensor…, creatos_f…) even on a wide window; now it takes its whole width first and the session's name is shortened before it, and tabs grow to 340px when the strip has room. When many tabs leave too little even for that, the name ends in an ellipsis and scrolls into view inside its own space while you point at the tab, so nothing in the strip moves (with reduced motion it stays cut; the tooltip names it in full).
  • —cockpit: moving the mouse over CI's pull requests no longer opens a card over the panel. The list of checks opens only once the pointer rests on a row, beside the panel, and closes as soon as it leaves.
  • —phone: every hint on the phone now says where to go: the Caprock app on your computer, Settings → Use on your phone, and the button by name. "This phone can read but not type. On the computer: Settings…" left out that the setting is in the Caprock app, and two hints still sent you to "Settings → Phone", which does not exist.
  • —doctor: stopping several orphans at once reports every one of them. A process that was no longer on the list and one that could not be stopped were recorded at the same moment, and one of the two could drop out of the result, so a process left running was not reported as failed.
  • —phone: a phone still on a Wi-Fi link now hears about it. Settings → Use on your phone → Paired devices says how each device reaches the computer (anywhere, anywhere with Tailscale, Wi-Fi link), and a Wi-Fi link gets one line with Show the code to pair it once more through the relay. On the phone, a strip says: "This link works only on your home Wi-Fi. To use Caprock anywhere: in the Caprock app on your computer, Settings → Use on your phone → Show a code, and scan it."
  • —phone: the pairing steps on the phone's own page no longer send you to "Settings → Phone → From anywhere → Turn on", which does not exist; they name the Caprock app on your computer and Settings → Use on your phone. The refusals a phone gets (not paired, view only) say the same path.
v0.116.12026-10-14

now

  • —an opened Live pulse row is calmer and says more. Terminal, Open repo and Open session are quiet links at the end of the project's line instead of a row of bright buttons, and the empty space beside each live session now shows its cost, its last hour, how long it has run, its context, its branch with lines added and removed, and its pull request with CI. The legend can be hidden; the ? in the panel's header brings it back.
  • —app: closing the tab of an agent Caprock started now asks: Close tab (in focus, so ⌘W then Enter works as before and the agent keeps running) or Close and end the agent, which stops it (asking once more if it is still answering; the conversation is kept). Tick Don't ask again to keep your answer; Settings → Closing a tab changes it back. An agent you started in your own terminal is never ended from Caprock: its tab just closes.
  • —app: in the Tabs list, sessions running with no tab sit under a quiet Running, no tab line, so an agent whose tab you closed no longer looks like an open chat.
  • —app: the sidebar's heading has a worded switch, Projects | Tabs, in place of the small folder icon: Projects lists the sessions under their projects, Tabs in one list in the tabs' order. Settings → Appearance → Sidebar uses the same words.
  • —app: updating no longer moves you off the sidebar you had. 0.116.0 switched everyone who had never touched the sidebar toggle to the flat list; if you used Caprock before and never chose, the sidebar lists the sessions under their projects again. A first visit still starts on the flat list.
  • —agents: a permission prompt you answered from a button no longer comes back if the daemon crashes a moment later. The answer cleared the stored prompt in the background, so a crash before that write landed restored buttons for a dialog already answered; the answer now returns only once the stored prompt is gone.
  • —cockpit: the spend chart no longer jumps when you hover it. At rest the line above the bars stood 23px tall (the ▾ that picks the metric took the panel's larger font) and under the pointer 19px, so the bars and everything below them moved 4px on every hover; the line is now a fixed 18px either way.
v0.116.02026-10-14

deck

  • —the PRs & CI tab opens on a Board: the repository's pull requests as cards in six columns — Draft, CI running, Needs a fix, Ready, Merged, Released — each with its author's face and name, how long it has sat in that column, its CI, the agents on it now and its agent cost (what Caprock recorded agents spending on it, at API list price). Above it, the flow: merges a day, opened → merged, merged → released (an estimate), the CI median. Below it, each recent release with the pull requests it shipped and what the agents spent on them. A card opens the pull request in the List, the old view.
  • —now: a CI/CD section under Repos: GitHub's own status, and per repository the Actions minutes billed this month, CI time and its trend, red runs, runs stuck past their usual time or waiting for a runner, the newest release, the five pull requests the agents spent most on this week and who opened them. Folded, it reads nothing. With GitHub connected it also reads githubstatus.com and shows avatars from avatars.githubusercontent.com, neither sent your token; Settings → GitHub says so.
  • —now: two live sessions in one checkout are pointed out. When two sessions work in the same folder of a repository (not in separate worktrees), each Live pulse row gets a small chip naming the other; when both edited the same files in the last 30 minutes it turns amber, 2 files touched by both, with the files on hover, since one can write over the other's change. Move to a worktree… beside it opens the New session dialog on a new copy of the folder, named after the newer session; nothing is moved for you. The agent cockpit shows the same chip, one line under the loop warning. New read-only GET /v1/overlap.
  • —mail: Mail no longer opens on a bare form. The first screen offers three ways in: Try it with sample mail (two made-up inboxes on servers inside Caprock, with a slim Sample mail — nothing here is real line to leave or to add your own account; your real mail and keychain are never touched), Import my accounts from Apple Mail or Thunderbird, ticked one by one, with a password field where none was saved and the Full Disk Access steps when macOS hides one, and Add an account as provider tiles — Google, iCloud, Outlook, Yahoo, Fastmail, Other. Google and Outlook sign in through the browser when this build offers it; otherwise, and always for iCloud, the form carries that provider's app-password steps and link. Settings' Add account uses the same tiles, and opening the Mail screen no longer creates an empty mail database.
  • —app: Claude's little character comes in eight faces — the hard hat, a beanie, headphones, a crown, a wizard's hat, a tuft of hair, a bandana and a propeller cap, each in its own pastel and expression. Each session gets one at random from its id and keeps it everywhere it shows; click the portrait in the cockpit to give a session another, and pick the default — a face per session or one for all — in Settings → Appearance. Kept in this browser only.
  • —app: the cockpit's state line and permission mode sit under the agent's portrait, so the figures at the right get the row; started moves under the model. The subagent chips come in groups of five and wrap onto a second line instead of being cut; past two lines they shrink to small squares, and past that a +N says how many more.
  • —mail: when you first open Mail it offers the accounts you already have in another client, so you add them with a click instead of retyping servers and passwords. Thunderbird comes over whole — servers and, when there is no primary password, the passwords too, decrypted on your machine (nothing is sent anywhere and no password ever crosses the dashboard connection). Apple Mail brings its server settings; its passwords are not readable, so you type each once, and on macOS without Full Disk Access Caprock says so and how to grant it. Each account is signed in to and verified before it is saved, and one account that fails never stops the others.
  • —remote: a phone reaches Caprock from anywhere — mobile data, another Wi-Fi — with nothing to install, through a relay that carries only end-to-end ciphertext (ADR-058). Settings → Open Caprock on your phone → From anywhere, or caprock remote on; it stays on across restarts. The machine dials out to the relay and proves its id with its Ed25519 key; the phone pins the machine's X25519 key from the QR and runs Noise NK with it, so the relay forwards bytes it cannot read or change. Pairing, device tokens and the viewer and controller roles are the LAN's own.
  • —remote: the phone's page (ui, vite build --mode remote) keeps several machines and switches between them from a strip along the top that says whether the machine is reachable; + Add a machine takes a pasted pairing link, for an iPhone home-screen app.
  • —cli: caprock remote [on|off|reset], caprock pair, which draws the pairing QR in the terminal, and caprock remote relay, a relay of your own. relay/ is the hosted relay, a Cloudflare Worker with one Durable Object per machine.
  • —mail: adding an account finds the servers for about 40 mail services and the hosts people put their own domain on. The table now knows Yahoo, AOL, AT&T, Verizon, Cox, Xfinity, GMX, WEB.DE, mail.com, Zoho, Mail.ru, Rambler, Seznam, Daum, Orange, Free, SFR, La Poste, T-Online, Libero, Virgilio, mailbox.org, Posteo, Purelymail, Runbox, Mailfence and Migadu as well as Gmail, iCloud, Outlook, Fastmail, Yandex and Proton Bridge, each with what it needs (an app password, IMAP switched on) and its help page. A custom domain is recognised by its MX on Google Workspace, Microsoft 365, iCloud+, Zoho, Rackspace, GoDaddy, Namecheap Private Email, Spacemail, IONOS, OVHcloud, Hostinger, Titan, Infomaniak and STRATO. Any other domain is asked for its RFC 6186 SRV records and its own autoconfig file before Caprock falls back to imap. and smtp.; only the domain itself is asked, no directory. The answer says how it was found (source).
  • —mail: sign in with Google or Microsoft. A Gmail, Google Workspace, Outlook.com or Microsoft 365 mailbox connects with a click and an approval in the browser, no app password: the installed-app OAuth flow with PKCE, back to Caprock on loopback, then IMAP and SMTP over XOAUTH2. The refresh token is kept in the keychain like a password; a revoked or expired sign-in shows as a sync error asking to sign in again. Signing in with an address Mail already has switches that account over and keeps its mail. Nothing is sent to Google or Microsoft until the button is clicked. The buttons need the build's client IDs (docs/mail-oauth.md); iCloud keeps app-specific passwords, having no OAuth for mail.
  • —doctor: an Orphans section lists the processes on the machine that look abandoned, each with why: another Caprock daemon (and whether its data directory still exists), terminal holders no Caprock will reattach to from any install, a claude whose session ended or whose parent exited, MCP servers whose claude is gone, and what a Claude Code session started that outlived it. Anything in a terminal you are in, in a live session, of another user or inside an app is never listed. The list is taken when you open the doctor or run caprock doctor; the 10-minute sweep keeps a count for the status bar.
  • —doctor: an Advanced mode (off by default, on the computer only) adds a Stop to each orphan and Stop selected, each confirmed once more; caprock doctor --stop <pid> does the same from a terminal and asks first (--yes skips). The process gets SIGTERM, then SIGKILL after 5 seconds, and only if it is still the one listed — same pid, same start time. Caprock never stops a process it did not start on its own (ADR-059).
  • —phone: Keep on after Caprock restarts in Settings → Use on your phone: phone access comes back on by itself after an update or a reboot. Off by default; Turn off clears it; only paired phones get in either way (ADR-060).
  • —app: the two dots at the start of every tab are gone. A tab shows state only when it needs you: an amber edge along its top while its session waits on you, a red one when it failed; working and idle show nothing. The project is its short name after the title, in its colour, and the Flat sidebar's rows follow the same rule.
  • —cockpit: the context and the plan limits are back near the top, right under Now and above the spend, and each has two looks. Compact, the default, is one line: 62% · 124.0k / 200.0k with a thin bar, and 5h 41% · 05:40 Week 24% · Mon 13:00 with a tiny bar per window, the reset times cut first in a narrow panel. Click Context or Plan for the full view as it was — the meter, what each call re-reads, the plan and each window's bar and forecast. The choice is remembered on this device.
  • —perf(store): the cockpit's subagent rows and the Cost screen's summary read far less of the database. A refresh of a busy session's subagent rows took 25–37 ms instead of 224–303 ms on a copy of a 2.4 GB database, and today's per-turn attribution 42–59 ms instead of 69–79 ms; the burn tile no longer counts a whole summary of the last ten minutes on every request. A new migration builds three small indexes once, a few seconds on a large database.
  • —phone: the phone panel says first where it works — Works on this Wi-Fi only or Works anywhere (Tailscale) — instead of folding it under Then: …. Without Tailscale it names the one step (Tailscale on this computer and the phone, same account) with the downloads; it notices Tailscale by itself while open, listens on its address without turning access off and on, and a new code opens the Tailscale address. A paired computer is no longer counted as a phone: No phone paired yet says so.
  • —remote: the phone's first screen says what to do in three large steps, keeps the pasted link for a home-screen app behind one tap (shown at once when opened from the home screen), and answers the questions a careful person asks — is it safe, what the relay sees, why not an app, a lost phone, a relay of your own — with links to how it works and the relay's source.
  • —app: a new icon — a bold white C on a near-black tile, with a green dot in its mouth for a session running. The amber triangle got lost among its Dock neighbours; a letter reads at a glance. The menu-bar icon, the dashboard's favicon and the phone's home-screen icon carry the same mark.
  • —cockpit: subagent rows stay where they are. They stand in the order the subagents started, first on top, a new one added at the bottom; working, waiting, getting busier or finishing no longer moves a row, and a finished one stays in its place, dimmed with a ✓. What a subagent is doing now gets the row's whole width and two lines, the full text on hover; its figures moved to the line below, and Show in the Subagents header picks which of them each row carries — Compact (just the phrase), Cost (cost and time, the default) or Full (calls, cost, model and time), or any mix — remembered on this device.
  • —app: tabs are the way around. ⌘1–⌘9 (Ctrl+Shift+1–9 on Windows and Linux) switch to the Nth tab, and the tabs stay in the order you opened them: a new tab goes last with the next number, and the numbers change only when you close or drag a tab; a restart keeps them. Every tab shows its key, the one in front too, and its session's name with its project's dot and short name. The sidebar is Flat by default — the open sessions in the tabs' order, each with the same key and its project, then the ones running without a tab — or Tree, sessions under their projects (Settings → Appearance, or the folder button). Projects have no number keys any more; a click on a project opens its page, and ⌥⌘ arrows still walk sessions and projects. The pinned Dashboard tab shows ⌘0 again (⌘, on Settings) and the inspector's button ⌘I, so every key is on screen. One line says so once, with a way to every key.
  • —deck: a pull request's checks read at a glance. In place of a circle per check — twelve tiny circles in a row that said nothing — each row sums them up in one short group: ✗ 1 failed ● 1 running ✓ 10, the failed first, the running one pulsing, skipped ones a dim count only when there are any, and a finished green PR just ✓ all 12 · 6m. Hovering it lists every check with its OS, state, time and the step a failed one stopped at; a click opens that job. The circles-or-words switch is gone, a row no longer says $0 for a public repository's free runners, and the cockpit's CI block uses the same summary, its view tabs now named PRs · Ring · OS.
  • —deck: Expand is a button you can see — outlined, with its key, ⇧⌘B (Ctrl+Alt+Shift+B on Windows and Linux) — and in the PRs & CI tab the same button in the same corner reads Collapse and takes you back to the agent with its deck; the key toggles both ways. Back to the deck is gone. A pull request's title gets every pixel its row's checks leave, a step larger, cut only when the row is full. The circle per check is back as the default, with the summary one click away (Checks: circles | summary, remembered); hovering the circles shows the same card of every check.
  • —settings: every saved credential (DeepSeek key, Telegram bot token, a pasted GitHub token) is entered in one secure field: a lock inside, masked monospace text, an eye only while typing, and once saved a locked pill with its hint and where it is kept, with Replace and Remove. An (i) explains where it is stored, what it is sent to, how backups treat it and how to revoke it at the provider. Settings says under each field whether it is in the macOS login keychain or in a file, with the file's path.
  • —now: Live pulse rows stay where they are. They were sorted by the newest event, so two busy sessions swapped places every few seconds; now a row keeps the order its session started in (sessions open as tabs follow the tab strip among themselves), a new one lands last, and you can drag a row, or press ⌥⇧↑ ⌥⇧↓ (Ctrl+Shift on Windows and Linux) on it, to set the order, remembered in this browser. Each row now leads with the session's name in bold, with the project, branch and short id under it, so two sessions on the same branch can be told apart.
  • —app: switching tabs no longer redraws the terminal's font for a moment. A tab hidden for 30 s had let its WebGL renderer go, so it came back drawn by the DOM renderer, in other glyphs and another cell height, until WebGL was swapped in 1.5 s later; a tab hidden for less had its glyph atlas rebuilt on every switch. The three tabs hidden last now keep their renderer, a tab past that gets it back before it is first painted, and the atlas is rebuilt only after a wake, a pixel-ratio change or a hide of more than five minutes, so the squashed letters of #541 stay fixed.
  • —app: the version panel in the status bar has a close button; Esc and a click outside still close it.
  • —app: a button's tooltip no longer covers the panel it opened: the Caprock doctor, the version, the theme picker and the composer's menus show their tooltip only while closed.
  • —app: a session no longer jumps into Claude Code's background-agents list on its own. Holding ⌥⌘↓ (Alt+Shift+↓ off macOS) to step through sessions and letting go of the modifiers before the arrow left the arrow repeating into the session it had just opened, and Claude Code read that Down as "go to the agents below the prompt". An arrow pressed as an app chord now stays the app's until it is released, and the sidebar's ⌥⇧↑ ⌥⇧↓ (Ctrl+Shift+↑ ↓ off macOS), which typed a Down too when pressed in a terminal, do nothing there. Plain arrows and ⌥← ⌥→ word jumps are unchanged.
  • —app: the desktop app no longer opens to a black window when the deck's CI panel was left on Release and the newest release run lists no jobs. The daemon sent that run's jobs as null instead of an empty list; it now sends [], and the dashboard reads a missing list as empty, cached answers included.
  • —app: a screen that fails to draw no longer blanks the window. The deck fails on its own and the terminal stays; a failure anywhere else shows Caprock hit a snag with Reload, Reset the view and reload (forgets cached figures and the CI panel's view, keeps tabs and settings), and a report the user sends themselves: a prefilled GitHub issue, an email or the details copied. The report holds the error, where it was thrown, the version and the OS; a home directory is folded to ~, and nothing is sent by Caprock.
  • —app: scrolling up a terminal no longer shows Claude Code's prompt rules and status rows smeared and repeated across the history. A tab restored behind another parsed its replay at xterm's default 80 columns, and told the daemon that size, before it had ever been laid out; it now holds the replay and sends no size until it is shown and fitted.
  • —mail: on macOS, account passwords go to the login keychain through the call that names the keychain file, so saving one can never bring up a "Keychain Not Found" dialog. Passwords saved before are read where they are; Windows and Linux are unchanged.
  • —daemon: a session with no process behind it ends as soon as its transcript file or its folder is deleted, instead of staying in the sidebar for 24 hours. Nine synthetic transcripts a test script wrote and removed left nine such sessions listed as "working". A session is also no longer shown as working because of an event stamped ahead of the clock.
  • —mail: a wrong password on Spacemail and other Dovecot hosts says so. They answer it with "NO [UNAVAILABLE] Temporary authentication failure", which read like an outage; the error now says the password was not accepted, and the other sign-in names are not tried, since more failures only lock sign-in longer.
  • —settings: the DeepSeek API key and the Telegram bot token are no longer kept in config.json. They live where the GitHub token does: the macOS login keychain, or a file only your user can read in the data directory on Windows and Linux (or when the keychain refuses). A key an older release wrote to config.json is moved there on the daemon's next start and removed from the file. A DeepSeek session reads its key from there when it starts, into its own environment only.
v0.115.12026-10-13

mail

  • —iCloud and Gmail accounts connect the first time. The sign-in name is retried in the forms iCloud accepts (the full address or the part before the @; for sending, the @icloud.com address), an iCloud+ custom domain is recognised and asks for the @icloud.com sign-in, and a refused password says which app password to make. Mail is read from each message's whole header, so iCloud's malformed answers no longer stop a folder, and one message the server cannot describe is kept and shown instead of halting sync.
  • —app: a terminal no longer shows squashed, half-cut letters after the Mac sleeps, the window moves to another display or the tab comes back from a long spell hidden. WebKit can drop the WebGL glyph atlas without saying so; the terminal now rasterises its glyphs again whenever it is seen again and whenever the pixel ratio changes.
v0.115.02026-10-13

phone

  • —a home for the phone. Caprock on a phone — on its home screen above all — opens on what waits on you (answer a permission prompt there), the sessions to continue and what you saved, with one field docked at the bottom: pick a project and an agent, type, Start chat, and the chat opens; or Save a note, a link or photos for later. New chat also picks what it runs on, DeepSeek included. The full dashboard is one link down (#/now), and the dashboard's header gives way to one bar — Caprock, its connection state and a ⋯ menu. A one-time hint says how to add Caprock to the Home Screen. The manifest declares New chat and Save shortcuts and a share target; they need Caprock served over https and do nothing yet.
  • —app: what the phone saves shows on the Mac in Quick chat under Saved: start a chat from one (its text and photos become the first message, in its project when it has one) or delete it.
  • —daemon: the Inbox — GET/POST /v1/inbox, DELETE /v1/inbox/{id}, image thumbnails at /v1/inbox/{id}/files/{n} and an inbox live frame (migration 0053). A phone you let control saves and deletes; any paired phone reads.
  • —app: rename a session from its tab or its sidebar row: double-click the name, or right-click the row for Rename…; Enter or clicking away saves, Esc keeps the old name, an empty name gives back the one Caprock made. The new name shows at once in the tab and the sidebar, in every window, and stays after a restart (PUT /v1/sessions/{id}/name, migration 0052). A session you have not named is called by the agent's own title, or else by a few words of your first message (Fix the login bug) rather than the whole sentence.
  • —mail: Caprock Mail, a mail client inside Caprock. Every IMAP account — Gmail, iCloud, Fastmail, a company server — in one inbox at Mail (last in the dashboard's screen row, marked new): conversations, instant search over mail kept on this computer, archive, delete, spam, star, folders, compose, reply, forward and attachments, new mail at once over IMAP IDLE, and a keyboard for all of it. HTML mail never runs scripts or loads trackers until you allow images. Optional summaries and reply drafts from a model on your own computer (Ollama, LM Studio). Off until you add an account; it then talks only to the mail servers you entered, passwords in the system keychain. A one-time strip offers it after two days of use.
  • —app: move between projects and their sessions from the keyboard. ⌘1–⌘9 bring a project to the front by its place in the sidebar (⌘9 the last), its key drawn on its row and on the tab it brings; ⌥⌘↓ / ⌥⌘↑ go to the next / previous session of that project, ⌥⌘→ / ⌥⌘← to the next / previous project, both round the end. On Windows and Linux they are Ctrl+Shift+digit and Alt+Shift+arrows (Ctrl+Shift+← → stay PowerShell's select-by-word). Option+← → stay word jumps in the terminal and ⌘↑ ↓ stay the terminal's; in a text field the arrow chords keep selecting there. The project in front has a faint edge in the sidebar, so you see where you are among the others.
  • —app: Group by project, a folder toggle beside the sidebar's heading: off, the sidebar is one list of every agent session in a kept order, each row with its project's colour dot and short name, and ⌘1–⌘9 pick sessions instead of projects. The app remembers the choice.
  • —app: drag a session row under a project onto another to put it there, or focus it and press ⌥⇧↑ / ⌥⇧↓ (Ctrl+Shift off macOS); the order is kept for that project, also after a restart.
  • —app: the cockpit has a CI block: the session's pull requests as a circle per check, one ring with 10/11 · 5m and the time against the median, or the current PR's jobs by OS — switched in its label, kept per viewer. It draws only what the session deck already read and asks nothing; a click opens the PRs & CI tab.
  • —app: a Repos table on Now, under Live pulse: a row per repository with a session running in it — commits a day over 14 days, today's and the week's, lines added and removed, files not committed and the last commit, from the repository's own git (read at most every 10 minutes); and, while GitHub is connected, open pull requests, red runs with the median CI time and the newest release, only from what the session deck and its PRs & CI tab already read — the table asks GitHub nothing. Folds like the other panels, remembered per viewer; folded, it asks nothing.
  • —api: GET /v1/repos/activity, the Repos table's rows.
  • —app: the cockpit's spend chart can show burn rate — each bar the stretch's cost over the hours actually worked in it (hours with five or more calls), so an idle afternoon does not water it down and a stretch with none has no bar rather than $0/h — next to cost, tokens and turns. A faint dashed pace line marks your usual (for burn rate, your 30-day cost per active hour) or the period's average stretch, labelled at its right end (usual $3.00/h); bars above it stand stronger, amber only in your top fifth, and hovering a bar says 38% above your usual. The choice is remembered, and the folded strip follows it. Each project's spark in GET /v1/stats/summary carries active_hours and active_cost from the same scan, and each of usual's distributions its mean.
  • —app: rows under a project keep their places. Picking a session from the sidebar, opening it as a tab or bringing it to the front no longer moves it up: rows go in the order you set by hand, then in the order the sessions started, and a new one lands last.
  • —app: under each project the sidebar lists every agent session — working, waiting or idle — and nothing hides behind a +N more line. Shells and open files have no rows (they are in the tab strip), an ended session drops off, and rows are 28 pixels with no agent mark (the agent is in the tooltip).
  • —app: one quiet More… row at the end of the sidebar replaces the More projects and Hidden folds: it lists the other projects (click to open) and the hidden projects and session rows (click to show again).
  • —app: ⌘1–⌘9 now pick the sidebar's Nth item rather than a tab by its place in the strip, and the strip follows the sidebar's order, so the numbers on the left and on top agree; only the tab ⌘N brings shows its digit. ⌃Tab and ⇧⌘[ ] still cycle the tabs.
  • —app: tabs are named by their session again, so four or six tabs all keep their names: the project is the coloured dot (and in the tooltip with the agent and the branch), the agent mark is gone from the tab, the dots sit closer, and the ⌘ key is a small key at the right end that the × replaces on hover and on the tab in front. The Dashboard tab is its icon alone, its key in the tooltip.
  • —app: the session deck opens on all of the session's pull requests: no row opens by itself, and the one you open is opened again next time. Each row shows a circle per check again — filled, coloured by state, a running one pulsing — at every deck width, with a switch to the words (running · 10/11), kept per viewer.
  • —api: the first GET /v1/sessions/{id}/prs of a session whose composition has to wait on GitHub answers at once with the recorded pull requests, without their CI, marked refreshing; the next poll has them with it. The deck draws the rows straight away.
  • —ui: Dark Pastel is the default dark palette (owner, 2026-10-13), and first among the dark ones in the picker and the cycle; Book stays the default light one. Whoever never picked a palette — a new install, the phone, the menu bar popover, the app's start-up page — now sees Dark Pastel on a dark system, from the first paint. A palette you picked is kept, Dark Warm included, and an old Slate stays Dark Cold; the old Graphite tone was the default rather than a choice, so it moves to Dark Pastel. The home-screen colours and the screenshot scripts' dark pass follow.
  • —app: Settings → phone and Telegram are where the sidebar says. The words phone and Telegram beside Settings are links, each opening its section unfolded and scrolled into view; phone also shows the pairing QR at once, and the Settings row itself opens on the phone. Use on your phone (the QR, three steps — same Wi-Fi or Tailscale, scan the code, Add to Home Screen — and your paired phones) and Telegram now come first in Settings; #/settings/phone and #/settings/telegram link straight to them.
v0.114.02026-10-13

app

  • —the session deck's pull requests and CI, large. Expand beside Hide ⌘B opens a PRs & CI tab: the repository's last 14 days on top — pull requests merged per day, the median time to merge, the median CI run of its slowest workflow, the share of red runs with re-runs, and the Actions minutes GitHub billed this month — its pull requests on the left (this session's, the repository's newest 50, or the open ones), and the chosen one on the right: state, CI, merge state and review as large pills, the diff's size, the deck's Fix / Re-run / Update actions, its jobs as a grid of job × runner OS against each job's median, the jobs on the run's clock, and a failed job's steps and log with Investigate. Back to the deck returns to the agent. It asks GitHub only while it is the tab in front, through the deck's kept answers, plus one list of the repository's newest 100 runs kept 30 minutes; nothing at all while GitHub is not connected. No new key and no sidebar item.
  • —api: GET /v1/sessions/{id}/ci/stats (the tab's figures and the repository's pull requests); a pull request carries created_at, merged_at and its diff size (additions, deletions, changed_files, commits); …/ci?pr=N and the rerun and update-branch routes also take a pull request of the repository that the tab listed.
  • —app: Settings → Status line → Hidden: no line under Claude Code, so the row goes, while your plan limits keep reaching Caprock's cockpit and Cost screen — which Off stops. Recommended on the page, never chosen for you; Full stays the default.
  • —app: the cockpit's headline says how it compares with your own usual — ▲ 38%, ≈ usual, ▼ 55%, ▲ 4.5× left of burn rate ▾ — against your median active session-hour (burn rate), your median session (cost, tokens) or your 30-day hit rate (cache, in points), all measured on this machine over the last 30 days. Amber only in your top fifth; hover says it in full; nothing at all until there is enough history. The new usual on GET /v1/stats/summary carries the baseline.
  • —app: the cockpit's spend chart switches what its bars measure — cost, tokens or turns — from a cost ▾ after its total, in the hero's own menu style and remembered per viewer. Hovering a bar reads its figure in that metric's own format ($12.40, 950.1k, 1,204).
  • —daemon: each project's spark in GET /v1/stats/summary carries two more arrays over the same buckets, input (input tokens with cache reads and writes) and turns, from the same scan as cost, so the chart can plot today, 7 or 30 days by tokens or turns without another request.
  • —agents: Claude Code on DeepSeek, free. Paste your DeepSeek API key in Settings → Subscriptions and New agent and Quick chat offer DeepSeek · your API key in the Subscription picker: the session is Claude Code in your usual setup, with DeepSeek's documented variables and your key in that one process's environment, so its model calls go to DeepSeek's Anthropic-compatible API and DeepSeek bills your key. Your claude.ai sign-in and its plan limits are not used, and the session is kept out of the Claude plan windows and the plan-window stop. The key is write-only (Settings shows sk-…1a2b), never on a command line, in the database or a log; Settings and the picker say what the session sends to api.deepseek.com and that Caprock sends nothing. Its card says DeepSeek.
  • —app: each subagent row in the cockpit says what the subagent is doing now in words — the description Claude wrote for its latest command ("Rerun the store tests"), else its tool and target ("Editing Cockpit.tsx"), else its task — after its type as a letter chip (G, E, P, the full type on hover), with calls and cost beside it; the dimmer line under it shows the raw command only when the phrase does not already say it. GET /v1/sessions/{id}/subagents carries the phrase as doing and doing_from.
  • —dashboard: the Now screen refreshes once, not once per section. Its sections ask on one shared tick, at most every 2 s (all-time figures every 10 s); sections asking the same question share one request; a folded or scrolled-away section asks nothing until it is seen again; and the live pulse and activity feed no longer read every track's history again each time two busy sessions swap places. Under a steady stream of hook events (five sessions, ten events a second) the open screen went from about 650–770 requests a minute to about 100, and the daemon's CPU for it on a copy of a 2.2 GB database from 36 s to 17 s per minute (measured 2026-10-13). A screen opened after events had arrived no longer asks every question twice.
  • —app: the hot-process flame on a sidebar row or tab draws again only when what it says changes: every row's mark redrew on each 5 s resources sample, hot or not. It stays a static mark — no animation to pause.
  • —daemon: the database gets smaller on its own. While Caprock is idle (nothing recorded for 30 minutes) it replaces the screenshots and files that events stored before 0.113.0 still hold as base64 with the marker new events get, [image image/png 512 KB not stored], a few hundred events at a time; the text of every event is kept whole. It then compacts the database once — only with free disk of about twice its size, rolled back unchanged if anything is recorded meanwhile — and from then on gives free space back to the disk as it appears, including what a retention period deletes. New databases do that from the start. On a copy of a 2.2 GB database (2026-10-13): 4,181 events rewritten, 750 MB of payload removed in 79 s, compaction to 1.4 GB in 21 s; a writer committing every 20 ms throughout never failed. caprock doctor lists each step under done on its own; the Storage panel no longer tells you to stop Caprock and run VACUUM, and GET /v1/storage carries returns_free_space.
  • —app: the sidebar lists only active sessions under a project: what is working, looping or waiting on you, plus the tab in front so it always has its row. Idle sessions, rows you hid, files and quiet shells sit behind one faint +N more (its tooltip says what it holds, hidden rows included); opened, the hidden ones come last with an eye to show them again. The separate N hidden · show line is gone, +N more no longer stays open across restarts, and a project with nothing active is just its own line, its arrow listing what is in it.
  • —app: Spend moves up in the cockpit, from after the plan windows to directly under Now. Folded, it is a strip of the chart's own bars and three figures, the chart's metric first, instead of a faint sentence a narrow cockpit cut off.
  • —pricing: deepseek-flash at DeepSeek's off-peak list price ($0.15 input, $0.003 cache hit, $0.60 output per 1M, read 2026-10-13); turns in DeepSeek's weekday peak hours bill double and read low. Table 2026-10-13.1.
v0.113.02026-10-13

cli

  • —caprock doctor says what Caprock costs this machine — memory and CPU of the daemon, the app, the app's window and the terminal holders; the data directory split into database, pastes, logs and binaries; the daemon's disk reads and writes over the last minute and hour — flags a second daemon, a second app, orphaned terminals, a database over 2 GB and heavy disk reads, and exits 1 when something is flagged. --json prints the new GET /v1/footprint. It works with no daemon running.
  • —app: what Caprock costs this computer, one click from the status bar. caprock 1.1 GB · 2.2 GB at its right is the memory of every Caprock process and the data directory on disk (memory alone in a narrow window), in the warning colour with a dot when something is flagged. A click — or the palette's Caprock doctor — opens the doctor as a panel: the verdict in one line, each warning with the processes behind it, what the memory and the disk are spent on, the daemon's disk reads and writes, what it cleaned up on its own, Check now, Stop N orphans and Copy report, which copies exactly what caprock doctor prints.
  • —daemon: the resource meter counts the app's window: on macOS the WebKit helpers that draw it (hundreds of MB that belonged to no one before), elsewhere the app's child processes.
  • —daemon: any database statement slower than 300 ms is logged with the code that ran it, so a slow screen or a burst of disk reads names its cause.
  • —daemon: one daemon per data directory. A second one waits 5 s for a restarting one, then exits naming the pid of the one that runs.
  • —daemon: the automatic orphan sweep runs every 10 minutes and stops every orphaned terminal holding only an idle shell, whichever Caprock started it — eleven left by deleted preview stands survived every hourly sweep before. A write-ahead log over 256 MB is checkpointed on its own. Both are listed in the footprint's actions with their time.
  • —app: the cockpit's figures beside the agent open on burn rate, big, with tokens and cost under it — what a session is doing now first, what it has added up to under it. A choice someone made stays; one that was only ever the old cost-first default takes the new one.
  • —app: a hidden window does no work, and an idle one almost none. Out of sight (minimized, closed or covered), the page stops its polls and clocks, holds live events and applies them together every 2 s (at once for a notification, a permission prompt or a loop alert, and the moment the window is shown), pauses animations, and the terminal in front neither draws nor blinks; the menu bar and Dock badge stay current. Shown, screens that follow agents at work redraw a batch at a time rather than per event, and clocks tick every second only while something on screen is under a minute old.
  • —store: tool calls are stored the way the screens read them. A hook's tool input and output now go through the same cap as the transcript's: screenshots and other base64 become a marker naming their type and size, text past 32 KB is clipped with a marker saying how much was dropped, and every short field — paths, commands, exit codes — stays. Tool sizes still count what the tool really returned. On a copy of one 2.2 GB database the same history needed 841 MB less.
  • —daemon: the database is read far less. The Storage figures are measured when the panel finds them a day old, or once a day while nothing is running, instead of a full read of the database every 30 minutes; a start no longer pre-computes the share dialog's 30-day and all-time Weeks; and the write-ahead log shrinks back to 64 MB after a big write.
  • —daemon: old leftovers are cleaned up on their own, once a day while idle: pasted and dropped files after 30 days, empty quick-chat folders no session used, Gemini telemetry of sessions ended over 30 days ago, and the unused launchd.log. Each sweep logs what it removed and how many bytes, and is listed in the footprint's actions.
  • —deck: the Pull requests & CI panel opens on what it last knew. Its PRs, Branches, Release and Main views draw the last list at once — when the deck opens, when the app starts, when you switch views — and read again behind it, with updating…, updated 12 s ago or couldn't refresh in the header; the very first read shows three grey rows rather than an empty area. The daemon keeps each composed answer too, so a request whose GitHub reads have expired is answered at once from the last one while one refresh runs behind it, instead of waiting seconds on GitHub. GitHub is asked no more often than before.
  • —app: the spend chart's hovered bar shows its own figure in full. While a bar is hovered (or stepped to with ←/→) its cost takes the big figure's place with a short date beside it ($287.12 Sun 11 Oct · 06–12, of $1,299.44 at the right); leaving the chart puts the total back. In a 330px cockpit the figure used to be the part cut off.
  • —terminal: a phone put away gives the PTY back at once. With a session open on the desktop and a phone, the desktop's terminal shrank to the phone's width and stayed so after the phone was locked or closed, until the session was reopened. A terminal going out of sight now says so and the daemon returns the PTY to the terminal still in sight, at its size; the desktop also re-sends its size whenever it is focused or shown, and focusing the composer counts as using the session.
v0.112.42026-10-13

app

  • —the cockpit puts the model on its own line under the agent's name, so a wide figure at the right no longer cuts it to Opus …; how many subagents work now leads their row (3 working).
v0.112.32026-10-13

app

  • —the cockpit's top block holds one calm layout in every state. The agent, its model and state on the left, the big figure fixed at the right of the same row, the permission mode and the two smaller figures under them, and one line of subagent chips at the bottom; working, waiting on you, idle or ended only changes the colours and the chips, never where anything sits. A long model name is cut with the full name on hover instead of wrapping, and the chips fit as many groups of four as the inspector is wide.
v0.112.22026-10-13

Settings → Status line

  • —Full, Short or Off for the line Caprock puts at the bottom of Claude Code. Short is model, context and the plan windows (the reset time only from 80%), without the cost and counters the cockpit already shows; Off removes Caprock's own entry from Claude Code's settings, never one you set. Full stays the default.
  • —app: the cockpit's subagent row keeps its place: with none running it says no subagents running, so the agent's figures and the sections below no longer jump as subagents start and finish.
  • —app: on Windows, closing a shell tab no longer says it is running csrss (or another stranger) when it is at its prompt. A process whose parent exited long ago still names that parent's pid, and Windows reuses pids; a child now also has to have started after the shell.
v0.112.12026-10-13

app

  • —finished copies clean themselves up, asking first by default. A copy Caprock made that is merged (or its pull request is), pushed, clean and has nothing running in it shows on the project's page and the deck's Branches view as 3 copies done · 2.1 GB · Clean up; one click, confirmed on the same line, removes exactly those, and Review opens the copies list with them ticked. Settings → Worktrees: Off, Ask or Auto (removed a day after they finish). Undo for a day, branches kept 7 days with Restore under Recently removed in the copies list, and a stale tag for a copy with no commit in 14 days and no pull request. The main checkout, your own worktrees and unmerged branches are never touched.
  • —app: a session started in another terminal opens in its own tab. A click on its sidebar row (or in ⌘K, the Inbox, a project's Recent) used to switch to the Dashboard tab on the session's page; it now opens a tab named after the session, with the agent's mark, showing its conversation live and read-only with the inspector beside it. Where the terminal would be, it says Started in another terminal — Caprock watches it, it can't type into it, with Continue here (a new process on a copy of the conversation, in a new tab; the original is never touched) and Open in dashboard. An ended session opens the same way, and Continue here carries it on in the same tab. The tab's ⋯ menu hides the session from the sidebar or closes the tab; a hidden session comes back only when it next needs you.
  • —ui: Lifetime's Top projects donut sits beside its table again instead of drawing over it; a donut without a legend keeps its own width.
  • —app: Hide on a project in the sidebar did nothing when an agent in it had finished its turn, or when it was the project in front — from the row's eye, its ⋯ menu and a right-click alike, Outside projects included. Hidden now wins: the project moves under Hidden (the one in front stays in view under that heading, its tabs open), and comes back out only when an agent there asks or finishes a turn after it was hidden.
  • —deck: every open pull request shows its actions. The ⋯ was a faint dot at the row's end that read as part of the age, and the actions only appeared once a row was opened, so they went unseen. Now the row's one problem carries a direct word beside it — Fix for a conflict, Fix and Re-run for failed CI, Update for a branch behind — that asks on the same line, and the ⋯ is a clear button at every width. Without GitHub connected the actions are there greyed, with Connect GitHub.
  • —deck: a pull request row says whether anything is on it in one word: in work (an agent on its branch working, or a re-run, update or fix Caprock started), waiting (an agent there asks you) or queued (a message queued for it, or GitHub's auto-merge on), with the agent's initial beside it. The row shows one problem chip, the most important, the rest in its tooltip, and at a narrow deck the mini pipeline goes before the title does.
v0.112.02026-10-13

app

  • —an optional composer under an agent's terminal (Settings → Terminal, off by default): a highlighted box to write a message in, sent with Enter through the message queue so several lines arrive as one, with the / and @ pickers, saved prompts, pasted images and ↑ ↓ history; it queues while the agent works and points at the terminal while the agent asks. Only for sessions Caprock started; with it off nothing changes.
  • —ui: a Dark Pastel theme: a neutral graphite ground, a dusty blue accent and soft, low-chroma status colours, with a terminal to match — for whoever finds the other dark themes loud. Body text stays at 7:1 or more, muted text and the accent at 4.5:1 or more.
  • —deck: pull requests show what stands in their way and who is on them, and can be fixed or re-checked from the row. An open PR carries one-word chips beside its pipeline — conflict, behind, review (changes requested), draft — from what GitHub already said, and nothing while GitHub is still computing. A live agent whose checkout is on the PR's branch shows as its initial with fixing, waiting or idle (2 agents for several); a click opens its tab. The open row, or its ⋯, offers Fix conflict with agent, Update branch, Fix CI with agent and Re-run checks; each confirms in one line inside the row, then says what it is doing there (agent fixing…, re-running 5 jobs…, updating branch…) until GitHub shows it done. A fix goes to the Caprock-started agent on the branch, else a new agent starts in the branch's checkout or a new worktree at the PR's head; nothing is typed into a session Caprock did not start. Re-run and Update branch use GitHub's own APIs with your token, and a refusal names the permission the token is missing.
  • —Make any subscription the default. Settings → Subscriptions marks the one new sessions start on when you pick none, and every other one has Make default. Your own sign-in can be renamed and made the default again; nothing about ~/.claude or its sign-in moves. The pickers in New agent, Quick chat and the phone's start form open on it.
  • —A plan per subscription. With more than one Claude Code sign-in, Settings → Plan & licence has a row per account with its own plan; one without a plan uses the main plan, your own sign-in's. A session's cockpit shows its account's plan and windows, and the plan chip there sets that account's plan (PATCH /v1/subscriptions/{id} takes make_default and plan).
  • —Continue on another account. A running session never changes account; the inspector now offers a new session on another one, with a brief of this one, and the session page's Continue in… lists each account. The cockpit and inspector name the account a session runs on, your own sign-in included once there is more than one.
  • —ui: Dark Cold's blue accent is one step softer.
  • —ui: Settings → Projects calls its folder Where you keep your code and says what it does: Add project and Browse open there and a new session suggests the repositories inside it; it limits nothing, every folder you run an agent in or add is a project wherever it is.
  • —ui: the Memory screen opens with one paragraph saying what it is for: every answer Claude wrote you, searchable, linked to the moment it was written, and what a new session in the same folder is handed.
  • —ci: a pull request that does not touch the desktop app (app/, the Makefile, ci.yml) skips the three app builds, and the Go jobs no longer wait for the dashboard job. Go and Rust caches are written once per OS from master and read by every job. A release builds the app while verify runs, as two halves of make check side by side, and attaches it only after goreleaser has published, so a red verify still publishes nothing.
  • —app: each project in the sidebar shows one row: the tab you are on (or its last one), else what waits on you, else the agent that worked last. The rest fold behind one quiet +N more line that opens in place, stays open per project, lists what it holds in its tooltip (2 working · 1 idle · 1 shell) and turns amber when something behind it waits (+4 more · 1 asks). It replaces the Running in other terminals line; ⌘1–⌘9 still reach every tab.
  • —app: three figures beside the agent in the cockpit, each one yours to pick: the headline in large type and two short lines under it ($4.08/h burn ▾, 4.01B tok ▾), from one list — cost, burn rate, tokens, cache hit rate, turns, time working, last call, tool calls — never the same one twice (picking a taken one swaps them), cost, burn rate and tokens until you choose. In a narrow inspector the two lines move under the agent, beside the subagent chips, rather than squeezing the agent's own lines or being cut.
  • —app: the Spend block in the cockpit opens on its chart again, unfolded: the period's spend by time with its total and the stretch under the pointer. Numbers is still a click away on the Chart | Numbers switch. A choice saved while Numbers and folded were the default opens on the chart once; what you pick after that holds.
  • —ui: big figures never run into the next tile or lose their leading digits. A figure keeps its size while it fits and shrinks to its tile when it does not ($130,489.86 beside a seven-figure burn rate, in a narrow window or the app with the sidebar open), and the caption under it wraps instead of being cut off.
v0.111.32026-10-12

app

  • —the version panel's update button reads at a glance: Update to v… on one line, with Restarts the app · sessions keep running in small type under it, instead of one sentence wrapping across the button.
  • —app: the cockpit's headline figure stays on the agent's row, at its right, in a narrow inspector too: it shrinks there instead of dropping to the subagent chips' row.
v0.111.22026-10-12

ui

  • —the At a glance donuts read in full on a 14-inch screen. The bill by token type is labelled cache read, cache write, output and input, the names used everywhere else, with re-reading context and uncached as a quiet second line; the donut centre's caption wraps onto two lines instead of ending in an ellipsis, on every donut in the app. A donut narrower than about 400px puts its legend under the chart at full width, and the three sit side by side only when the panel has room for them. Hovering a slice or a legend row, or focusing a row with the keyboard, lights that slice and shows its share, name and figure in the centre, in place, with no popup.
v0.111.12026-10-12

app

  • —the cockpit's Spend block shows numbers or a chart, your pick. Its header holds the period chip (session · today · 7d · 30d) and a small Numbers | Chart switch in place of 0.111.0's Compact | Full, whose compact spark was a couple of bars. Numbers is three figures in a row; each label is a menu of what sits there (spent, burn rate, tokens, cache hit rate, last call, time working, turns, tool calls), never the same one twice and, for the session, never the figure already beside the agent; the choice is remembered. Chart is the period's spend by time under 96px tall, its total on the top line and the stretch under the pointer at its right (←/→ move it); today, 7 and 30 days come from the summaries already read, and a range with nothing to draw says so instead of drawing an empty line. A stored Full opens as Chart, Compact as Numbers.
  • —deck: an open CI job stays in view and folds back in one move. Its line is pinned at the top of the CI area with Fold Esc, its steps are bounded to the deck's height and scroll on their own, and opening a job closes the one before. Esc, while the focus is in the deck, folds the job, then the open PR; the terminal keeps its own Esc. Every PR row has a chevron, the open PR's line stays pinned, and the current PR folded by hand stays folded while the app runs.
  • —term: one session open on the desktop and a phone no longer leaves the desktop's agent drawn at the phone's width. The terminal in use sizes the PTY (as tmux's latest): a phone that only opens the Terminal tab is drawn at the desktop's size, typing on the phone gives the PTY the phone's size, and focusing, clicking or typing in the desktop's terminal gives it back, with a full repaint. A phone's chat never resizes it.
  • —app: in a narrow inspector the cockpit's headline figure moves to the subagent chips' row (or under the name when there are none) instead of hanging alone on a line of its own.
v0.111.02026-10-12

app

  • —multi-repo projects. Add project on a folder that holds several git repositories asks once: one project of them (the default) or a project each, with the repositories found as checkboxes. A multi-repo project is one sidebar row; its page lists each repository's branch, changes, open pull requests and share of today's spend, and Edit renames it, adds or removes repositories, rescans the folder, merges repository projects in or splits it — never deleting a file. An agent started inside one of its repositories can reach the others (--add-dir for Claude Code and Codex, --include-directories for Gemini).
  • —app: the deck's Pull requests & CI cell shows what CI cost: each job beside its time (macOS 4m 45s · $0.31), each run, the runner total and each PR row (✓ 9m 10s · −6% · $0.42); $0 · public repo when every minute ran free on standard runners in a public repository, unpriced for self-hosted or unknown runners, ≥$0.42 while part of it is unpriced or still running — each an estimate at list rate, the rate table's version and source in the tooltip. The Main view adds what GitHub itself billed for Actions this month (Actions this month: 1,240 min · $7.40), or which token permission it needs, with a link to the token settings.
  • —app: a Branches view in the same cell (PRs | Branches | Release | Main): the session's repository's branches, the current one first, then by last activity — ahead and behind the default branch, the last commit's age and author, the branch's pull request with its mini pipeline, what Caprock's sessions on it cost and how many, and a stale mark (no commit in 30 days and no open PR). Fifteen rows, then Show all; a branch whose PR is the session's opens it in the PRs view. Read from local git, one cached list of the repository's pull requests and an indexed query (GET /v1/sessions/{id}/branches, migration 0051).
  • —deck: the Pull requests & CI cell is dense. A run's jobs flow into as many columns as the deck is wide (three at about 1300px, two at about 900px) instead of one tall scrolling column with its right half empty, each column lined up (OS, time, against the median, cost); an open job takes the whole row and its steps flow the same way. The current pull request opens by itself while it is open and has checks, passed included, so its jobs fill the room the rows used to leave empty. The cell's title no longer truncates, and an open PR's CI no longer scrolls sideways by a few pixels.
  • —app: the cockpit's Spend block opens Compact by default: a period chip (session, today, 7d, 30d), three large figures (Spent, Tokens, Cache hit rate) and a small spark of that period's spend, about half as tall. Today, 7d and 30d are every session's figures, from the existing summary. Compact | Full in its header switches back to the cells and Spend over time; both choices are remembered.
  • —app: closing a quick chat (⌘W, the tab's ×, Close tab) now ends its process, so closed chats no longer pile up as idle sessions under Outside projects. A chat that is still answering asks once first. Closing any other agent's tab still leaves it running.
  • —agents: a Claude Code session started with a first message no longer loses it when no project instructions are set: Caprock's own --add-dir folders came right before the message, and the flag takes every word after it as another folder. Each is now written --add-dir=<folder>.
  • —ui: a session with no cache figures yet (a quick chat that just started) no longer trips the session and Now screens' cache line.
v0.110.02026-10-12

ui

  • —a Queue tab on Tasks (#/queue, also in the palette): what each agent Caprock started is doing now (its task-list item, how long the turn has run), every queued message in order with a rough range (for example "starts in 2–9 min · done in 4–14 min"), and the turns finished in the last three hours. Add to queue sends to a session, or to whichever of a project's sessions is free first. Every estimate is a range from that project's past turns, labelled with its basis, and is left out under five turns.
  • —daemon: GET /v1/queue, and a finished notification as soon as a queued message's turn ends (same switches and limits as the finished alert).
  • —github: what a CI run cost. A new runner rate table, pricing/actions.json (version 2026-10-12.1), holds GitHub's per-minute rates for every hosted runner, larger, arm64 and GPU runners included, copied from GitHub's Actions runner pricing page on 2026-10-12 with the source in the file. Each CI job, run and commit the deck reads now carries cost: the runner's SKU, its minutes rounded up the way GitHub bills them, and the dollars at list rate. Standard runners in a public repository are free; self-hosted and unrecognised runners say they are unpriced rather than guess. GET /v1/github/actions-usage reads the month's Actions usage from GitHub's billing report, only when asked, and names the token permission it lacks when GitHub refuses. Not in the interface yet.
  • —app: the session deck is one cell, Pull requests & CI: the session's pull requests, the current one first and up to three earlier ones, each with its state dot, mini pipeline, CI in words (✓ 9m 10s · −6%, ✗ failed at test (windows), running · 3/5 · 4m 10s), review, conflict and age. A click opens a row in place into its jobs, steps, error lines and Investigate; the current PR opens by itself while its CI runs or failed. The Branch cell is gone: the branch is a chip in the header (feat/x → main, with what is not committed or pushed), and the release chip sits there too. PRs, Release and Main stay behind one switch. The deck lists four PRs (GET /v1/sessions/{id}/prs, was three).
  • —app: CI runs and jobs show their time against the median of the same workflow's last five successful runs: 4m 45s +18% (amber past +20%, green below −10%), ≈ 3m left or +12% over median while running, nothing with fewer than two past runs. Past runs' jobs are read once and kept; only while the deck is open (baseline on runs and jobs).
  • —app: the session deck's CI cell has a switch: PR (as before), Release and Main, remembered per viewer. Release is the repository's newest release build — found from the run a tag push or a release started, for any repository — with its tag, the time so far, about N min left (estimate from the last K runs) from the median of that workflow's past successful runs, and its jobs and steps with errors and Investigate; once done, v1.2.3 · published 3 min ago ✓ with a link to the release. Main is the newest run of each workflow on the default branch, so a red default branch shows at once. While a release builds, a chip Release v1.2.3 · ~6 min shows from the other views (GET /v1/sessions/{id}/ci/release, GET /v1/sessions/{id}/ci/main).
  • —app: a GitLab-style mini pipeline beside each pull request in the deck, beside the release tag and beside each default-branch workflow: a circle per job (✓ passed, ✗ failed, a turning ring running, hollow queued, a slash skipped), joined within a workflow. Hover says the job, its state and its time; a click opens that job's steps and error in the CI cell, turning it to that pull request.
  • —Dark Side is neon purple instead of crimson: a near-black violet ground, a glowing purple accent and cursor, neon cyan for premium, and a red of its own for errors. Body text stays at 5.1:1 or more.
  • —app: ⌥⌘N (Ctrl+Alt+Shift+N on Windows and Linux) now asks: it opens the Quick chat chooser with the agent and model you used last already picked, so Enter starts it and Esc cancels. 1–9 pick the agent there too.
  • —app: a subagent in the cockpit opens in place: click its row (or its chip at the top) to read the task the parent gave it, folded to six lines with Show all, every call it makes as it makes them, with how long each took and the failed ones marked, what it answered, and its model, time, tokens, cost and costliest tools. Finished subagents open too, and +N finished lists them as rows. New GET /v1/sessions/{id}/subagents/{agent}; /subagents takes ?recent= (up to 30) and lists up to 24 working. A new index (migration 0049) finds a subagent's rows without reading its parent's.
  • —app: the cockpit's subagent chips show up to 24, in groups of four across the cockpit's width (twelve to a row), instead of four and +N; N subagents working sits right under the state line.
  • —daemon: a subagent's call count includes the calls only its transcript recorded (a call that failed fires no hook result), so the row and the opened subagent agree.
  • —app: the status bar says whose load it shows: agent cpu 17% ram 2.0 GB is the session's own processes (the agent and what it runs), and a new caprock 160 MB beside it is Caprock itself (daemon, app, terminal holders), with the split in its tooltip.
  • —ui: Settings in the dashboard header is a framed button with a gear, like Feedback beside it, instead of grey text.
  • —app: the agent cockpit puts what people watch first: the agent with a headline figure beside it, then what it and its subagents are doing now, its task list, the context and the plan limits; Spent this session follows, folded to one line (Spent $12.40 · 48 turns · 1h 12m) until you open it, and stays as you left it.
  • —app: the headline figure is your choice: cost, burn rate (dollars an hour over the last hour of activity), tokens, turns, cache hit rate or time working, picked from the small menu under it and remembered. The cells under Spent this session show the others, never the headline again.
  • —app: Spend over time is a row that says what it opens and opens in place, as wide as the cockpit, instead of a card over the block.
  • —app: the permission-mode menu is readable: each mode on two lines, its name large with a small mark and what it does under it, the current one marked with an accent bar, and why a row cannot be picked written out.
  • —app: the cockpit's spend over time stays inside a narrow inspector: a long peak … line no longer pushes it past the panel's edge.
  • —app: leaving a quick chat. Once its agent exited (/exit), the tab became a login shell in Caprock's own data folder, under the same name, so it still looked like the chat and only a reload got you out. Now the tab says The chat ended, with Close tab (Enter) and Start again on the same agent and model. A line above a new chat says ⌘W closes it and /exit ends it (× hides the line for good).
  • —app: a quick chat's tab, its sidebar row and the inspector read Quick chat · Opus 5.5, not Outside projects (the group it sits in).
  • —app: the update card no longer stacks on top of the version panel: while the panel is open (it offers the same update), the card steps aside.
  • —app: the Dock shows the new icon after an in-app update. On macOS the app sets its Dock tile from its own bundle's icon at every launch, and on the first launch of a new version re-registers the bundle with LaunchServices, so the cached old icon is replaced without restarting the Dock.
v0.109.12026-10-12

app

  • —the Dock and taskbar icon is a large filled amber triangle on a dark tile, sized to the macOS icon grid, in place of the outlined mark that filled its square and read small beside its neighbours.
v0.109.02026-10-11

Seven themes, with plain names

  • —Dark Warm (once Graphite), Dark Cold (once Slate), Light (once White) and Book (once Paper), plus three new ones — Light Cold, a blue-grey white with an electric blue accent and no orange; Light Side, pale sand with a blue accent, green for success and a soft glow; and Dark Side, near-black with a glowing crimson accent, where danger turns rose-magenta so it never reads as a button. The picker groups them Dark then Light, the switch-theme command walks all seven, and a theme kept under an old name opens as its new one. Each new theme has a terminal palette of its own, also in Settings → Terminal.
  • —Every theme reads at 4.5:1 or better: faint text was lifted in Dark Warm and Dark Cold and darkened in Light, and Light's accent is a shade deeper, so body and secondary text clear 4.5:1 everywhere.
  • —app: change a running session's permission mode. A chip in the session header and the agent cockpit says the mode a Claude Code session is in; for one Caprock started it is a menu. Default, accept edits, plan, bypass (when the process can reach it) and auto are switched in place by Shift+Tab, typed in and checked on the session's own footer; bypass in a process that cannot reach it, and don't ask, are Restart in …: after a confirm, at the prompt only, the same conversation is resumed in a new process in the same tab, with the one-time bypass consent where it applies. A session started elsewhere shows its mode read-only. New GET /v1/sessions/{id}/permission-mode and POST /v1/agents/{id}/permission-mode (ADR-054).
  • —Slate's terminal is cool too: in Slate, the terminal (app, dashboard and phone) takes a neutral ground, a blue cursor and a cool 16-colour set instead of graphite's warm one, and follows a theme switch while open. It is also a palette of its own in Settings → Terminal.
  • —Send feedback picks Bug, Idea or Question from three large cards with an icon, a line each and their key: press 1, 2 or 3 (any keyboard layout) and type the title straight away. Enter in the title moves to the description, ⌘↩ (Ctrl+Enter on Windows and Linux) creates the issue, and each kind asks its own question ("What broke?").
  • —Slate has no orange left: waiting and idle marks read as a cool yellow there instead of graphite's amber.
  • —agents: a new Claude Code session is started with --allow-dangerously-skip-permissions where the bypass warning was already accepted, so bypass is a Shift+Tab away in it without being its mode (ADR-054). Without that consent nothing changes.
  • —agents: text pasted in Russian, Greek, Japanese or any other non-ASCII script no longer turns to \316\272\316\261… or <00ad> in a session or shell Caprock started. The bytes always arrived intact; the session had no locale, because launchd sets none and a terminal application, not the profile, usually exports LANG. A session without one now gets a UTF-8 LANG (en_US.UTF-8 on macOS, C.UTF-8 on Linux); a locale you set is kept.
  • —app: the sidebar has no Waiting on you section any more — everything in it was already in Projects. A session that waits on you is marked on its own row instead: an amber edge and tint and the word waiting, or asks at a permission prompt. One started in another terminal is a row of its own rather than behind Running in other terminals, and a project folded by hand still shows what waits in it. The Today strip's Waiting count still opens the first one.
  • —app: every row under a project can now be closed or hidden. A row with no tab — an agent whose tab you closed, or a session from another terminal — has a × that hides it from the list (it keeps running, and nothing is ever sent to a session Caprock did not start), next to ■ Stop… for what Caprock started, which still asks first. Hidden rows wait behind N hidden · show under the project, where an eye brings each one back, and one comes back by itself when it next waits on you. The same actions are on a right-click or Shift+F10, a tab's menu adds Close tab and hide from the list, the buttons are bigger (26px), and on a touch screen they are always shown.
v0.108.12026-10-11

ci

  • —the session todos test helper no longer trips golangci-lint (unparam, QF1008), which stopped the 0.108.0 release before it built.
v0.108.02026-10-11

app

  • —the agent cockpit shows the task list the agent keeps for itself — Claude Code's TodoWrite and TaskCreate/TaskUpdate, OpenCode's todos, Codex's plan — one large row per task with a spinner, a check or an empty circle, the task in hand as the agent words it, and 3 of 7 done above. Not there when the session kept none. New GET /v1/sessions/{id}/todos.
  • —Lifetime → Daily cost has its own period (30d, 90d, 180d, all) and a ×2 height toggle; both are remembered.
  • —Four themes, picked by hovering: Graphite, Slate (new — a neutral dark with a soft blue accent), Paper and White. The theme button opens a picker: hover a palette to see it on the whole app, click to keep it. Settings → Appearance shows the same four as large swatches.
  • —app: tabs in the sidebar show their jump key, ⌘1–⌘9, as the tab strip does.
  • —Phone: chat first. Now on a phone starts with the running sessions, one tap each into that session's chat. A session on a phone is the whole screen: the page header steps aside, Chat | Terminal is one large switch, the conversation fills the screen with the field at the bottom (above the keyboard when it is up), and Esc, Tab, the arrows and Ctrl+C fold behind Keys. A session that cannot be typed into from here says why where the field would be — started outside Caprock, ended, or this phone is view-only — with Branch here to type or Continue here when that works.
  • —Voice input in the phone's chat. A microphone button dictates into the field, live, to read and then send; the browser's own speech recognition does it (Safari sends the audio to Apple's service, and the button says so while it listens), Caprock gets the text. Browsers allow the microphone only on https pages, so over Caprock's plain-http Wi-Fi or Tailscale address it says in one line to use the keyboard's dictation key.
  • —Pairing says where the phone will work, beside the code: only on this Wi-Fi, or anywhere while Tailscale is on on both devices.
  • —app: each project in the sidebar draws its last hour as a small line — amber while something in it works — and has a + for a new agent there without hovering. A chevron folds a project by hand, even one with something running; the choice is kept. GET /v1/activity serves the line.
  • —Plan limits asks for your plan until it is set, saying why: Caprock can't detect it, and Cost needs it to weigh usage at API prices against what the plan costs. Set my plan opens the list in place.
  • —app: the session deck's first cell is now CI, the live GitHub Actions of the session's pull request: a row per job with its state (a turning ring while it runs, ✓, ✗), the step it is on or failed at, its OS and its time, ticking while it runs; a job opens into its steps with their times, and a failed step into the lines around the error in its log, with Open on GitHub. At the top, the runner time so far, measured from GitHub's job times and split per OS, never priced. Asked every 12 s while a job runs, else every minute, only while the deck is open (GET /v1/sessions/{id}/ci, GET /v1/sessions/{id}/ci/jobs/{job}/log).
  • —app: Investigate on a failed CI job queues it into the agent in front — the PR, the workflow, job and step, the job's link and the last 40 lines of its log, with "investigate and fix; push to the same branch" — through the message queue, so it types in at the prompt or when the turn ends. Only into a session Caprock started, from a device that may control; otherwise the button says why it is off.
  • —app: the pinned Dashboard or Settings tab is its icon and its key — ⌘0, or ⌘, for Settings — the name in the tooltip.
  • —app: the sidebar's Projects heading counts them — ● 3 of 12 has something running — and Outside projects can be hidden like any project.
  • —app: the inspector's plan windows answer "will it last?" at a glance — a tick on each bar where an even pace would be, the percentage large, and one line: resets in 4 h 56 min · ✓ enough, or ⚠ out at 23:50 when the current pace runs out first. A click shows the details; the plan itself is in the header, to confirm or change.
  • —app: the update card says it in one line, large — Update 0.107.6 → 0.107.7 — with one short line under it and bigger buttons.
  • —app: the version panel is the version, whether it is current, when it was checked, the update button and Release notes →. The release notes themselves left the panel, where they were unreadable.
  • —terminal: an image pasted, dropped or sent from the phone's photo button shows in Claude Code as [Image #1], [Image #2] instead of a long path. Each image goes in as its own bracketed paste of its quoted path — measured against the real claude, the only form it attaches; typed, the path stays text. Other files, and a shell that never asked for bracketed paste, still get the quoted path.
  • —Lifetime: the big figures are short — $16.8k, $159 — with the exact sum on hover. Top projects' rows start with the donut's colour and the share, large.
  • —app: every shortcut hint is drawn as a large key — New agent, Quick chat, Add project, Dashboard, Settings — on macOS, Windows and Linux alike.
  • —A lighter terminal on a phone or tablet: no WebGL, no blinking cursor, 2,000 lines of scrollback, and output handed over at most ten times a second. Under a 4x CPU throttle at 390px, a busy synthetic session took 0.6 s of main-thread time per 10 s, from 3.2 s. A desktop is unchanged.
  • —Plan limits is two matching cards, Claude and Codex, each with its character and both windows side by side — 5 hours and Week — in larger, shorter words. A window an agent did not report is an empty ring that says so, instead of a missing one.
  • —ui: the Tasks screen says what it is and how to start. One sentence and a worked example (you write the task, a worker does it on its own branch, Caprock runs your checks, you merge), then a gallery of templates — fix failing tests, bump a dependency, add tests for a file, fix lint and type errors, write docs for a module, port a script — that open the form prefilled. Their checks are the commands the repository declares (Makefile targets, package.json scripts, go.mod, Cargo.toml, pytest/ruff/mypy when configured), found by the new GET /v1/hive/starter, never guessed. The budget field starts empty with the daemon's own ceiling as the hint.
  • —ui: the board is one list grouped by what each task needs — Needs you, Working, Waiting to start, Finished — instead of six narrow columns, and each row says what the task is doing now, how its checks went, what it cost against its budget, and its one action (Approve/Reject, Review & merge, Open). A queued task with nothing moving says the orchestrator has to be started, with the button.
  • —A new queue directory starts empty. It no longer gets an example task in the inbox, which the board showed as waiting and the orchestrator would assign with checks guessed for the repository; the format is in its README.
  • —app: the permission card's Y, N and A work on any keyboard layout; on a Russian one Y typed a Cyrillic letter and did nothing.
  • —ui: every banner is one design — a small scene drawn from Caprock's own characters on the left, a bold title, one short line, one button and a quiet dismiss, in one row. The team card, the Premium offers, the cap and plan-limit nudges, the phone and star strips, the update and hooks notices, the plan question and the locked panels all use it, in every theme. Figures are short ($16.8k), with the exact sum on hover.
  • —Settings is quick to scan: larger titles, buttons and checkboxes; the explanations are hidden until you turn on Show explanations or press a section's ?; a section's title folds it, and the fold is remembered.
  • —app: Cmd+C in a terminal copies on any keyboard layout and says Copied · N chars; on a Russian layout it copied nothing and macOS beeped.
  • —app: a pull request in the deck with no checks no longer reads "CI no CI": a fresh one says CI not started (GitHub has not reported its checks yet), one older than 10 minutes says no checks, each with a tooltip.
  • —app: the deck's Context cell: the status bar already shows it.
v0.107.72026-10-11

Settings has a sections menu

  • —on a wide window a list of every section sits left of the page, each with a simple icon, the one in view marked; a click scrolls there. Section headings carry the same icons.
  • —app: a tab is named by its project — the repo — alone. A second tab of the same project adds its title, faint; the title is in the tooltip.
  • —Now: At a glance — the donuts — comes before All time, the exact table: the pictures first, the numbers after.
  • —Feedback is an amber bug in the header and the sidebar, in place of a grey speech bubble that went unseen.
  • —Now: the agents folded to icons are twice the size and shrink with the window on one line; hovering one shows a card — its share, spend, turns and cost per turn — instead of a one-line system tooltip.
  • —app: keys drawn as keys (the deck's Review changes ⇧⌘G, the tabs' ⌘1–⌘9) are drawn darker and bolder; the faint chip went unseen.
  • —app: the keep-awake mark in the status bar stays when keep-awake is off — it reads awake off, and a click opens the modes — instead of vanishing with no way back from there.
  • —ci: the PTY resize test waits for the process to report its size instead of sleeping, which flaked on a slow macOS runner.
  • —Windows, Linux: ⇧⌘E (split down) and ⇧⌘0 (reset the interface size) are Ctrl+Alt+Shift+E and Ctrl+Alt+Shift+0 there; as Ctrl+Shift they hit other commands. A test now presses every listed shortcut as each OS spells it.
v0.107.62026-10-11

app

  • —the inspector's spend spark says it opens: a large Spend by time ↗ click label on hover and a small ↗ in its corner at rest.
  • —The theme button walks all three palettes: Graphite → Paper → White, its label naming the next one. It only flipped light and dark before, and Paper or White was a trip to Settings.
v0.107.52026-10-11

app

  • —the session deck's top edge is the faint accent line again, as in 0.107.3: it sets the deck apart without catching the eye.
v0.107.42026-10-11

app

  • —the inspector's spend spark opens, on a click, the session's spend by time instead of a bar per model call: bars follow the clock (5 min to a day, by the session's length), and a stretch reads as when, what it cost, how many calls and which tools. It stays open until closed.
  • —app: the session deck's top edge is a plain hairline; the accent line with its glow was too bright.
  • —app: ⇧⌘G (Ctrl+Shift+G) opens the diff of the session in front — its worktree's Changes view — from anywhere in the workspace; the session deck's Review changes shows the key.
v0.107.32026-10-11

app

  • —the session deck's pull requests line up in columns — CI, then state — and a click on the CI opens the run on GitHub: the failing or running job when GitHub links it, else the PR's Checks tab. The deck's top edge glows less.
v0.107.22026-10-11

Tasks, before the task runner is on, is large and plain

  • —one headline (Hand off a task. Get it back green.), three big numbered steps, and a filled Turn on the task runner button.
v0.107.12026-10-11

app

  • —the session deck is a third lower and keeps only what helps while typing: context, branch, and pull requests with CI — a merged one now says CI ✓ 12 · merged. Now and Spent left it (the sidebar and the inspector carry them), so it no longer follows the session's events. Its top edge is an accent line that frames the prompt above it, and a Hide ⌘B button sits on it.
  • —app: the inspector's content no longer runs past its right edge.
v0.107.02026-10-11

The session deck shows the session's pull requests and their CI

  • —Its latest three PRs with a dot each — passed, running (CI 3/5), failed (CI failed: windows), merged — and a click opens one on GitHub. Without GitHub connected, connect GitHub for CI turns it on through your gh login in one click. Beside it the branch, with how many files are not committed and commits not pushed. The per-minute chart and the plan windows left the deck; the dashboard and the sidebar carry those.
v0.106.12026-10-11

app

  • —on Windows and Linux the desktop app now replaces its own daemon when an update brings a new one, as it already did on macOS. Before, the app updated and its daemon stayed on the old version, so the status strip and the update banner reported a version the app had left behind.
  • —app: when the app and its daemon still run different versions, the version in the status strip gets an amber dot and its panel says which is which and what to do.
  • —app: the Keep awake panel says what is happening in a few large words, has one short line per mode and a close button.
  • —app: the session deck's button says Deck and shows ⌘B, and the deck's per-minute chart is as tall as the figures beside it.
v0.106.02026-10-11

Dashboard panels fold

  • —The arrow on a panel's title folds it to that title and back, remembered; the agents row on Now folds to one line of characters with their shares.
  • —Plan limits and the daily cap lead the Cost screen. They were below the month; they are what the screen is opened to check.
  • —Daily cost has a time axis. Months and Monday ticks run under the bars, and the hovered day reads as a weekday and how many days ago.
  • —Larger figures in the dashboard's stat tiles.
  • —The version in the status bar opens what you need about updates. A click shows whether a newer release is out and when Caprock last checked, the one-click update (or the command for how it was installed), the first lines of what changed and Check now; a dot beside the version says an update is out. It replaces the bare what's new link.
  • —Links to the browser open from the desktop app on Windows. download it in the update banner opened nothing there; every link meant for the browser now goes through the app's own opener.
v0.105.02026-10-11

The session deck under an agent's terminal

  • —⌘B (Ctrl+Shift+B on Windows and Linux), the strip's deck button or the palette opens a band under the terminal of the agent in front. It lifts the terminal's bottom, and Claude Code's prompt with it, off the window's edge, and fills the room with the session's live figures in one row: its state and what it is doing this second, what it has cost and what the last call cost, how full its context is and what the next call pays to re-read it, the plan's 5-hour and weekly windows, and the last hour as one bar per minute. Taller, it adds the last few tool calls, turns and tool calls, and the files changed in its checkout. Drag its top edge (or focus it and use the arrow keys) to resize it, up to 60% of the pane; double-click for the default. It is closed by default, its height and whether it is open are remembered per machine, shell tabs have none, and closed it costs nothing: it reads the figures the inspector already holds, sharing one fetch with it.
  • —⌘0 opens the Dashboard. It sits beside ⌘1–⌘9 like the tab before the first; ⇧⌘D still works, and resetting the interface size moved to ⇧⌘0 (and ⌘K). Keys on the tabs and the Dashboard's key in the sidebar are drawn as keys, larger, and from five open tabs the Dashboard tab folds to its icon and key.
  • —Charts and exact numbers, both drawn. Model mix, What it went on and Per project on the Cost screen each open with a donut, and every table row's share is a bar in its slice's colour. On Lifetime the Charts / Numbers switch is larger, the numbers carry the same bars, and Daily cost sits right under the totals instead of at the bottom.
  • —Keep awake says whether it is on right now. The cup in the status bar is in colour with a green dot while Caprock holds the computer awake and reads may sleep, faint, while it does not; a click explains it and switches the mode there.
  • —The full breakdown is a row of its own. Every tool, model and project under the All time panel on Now is a full-width row instead of a faint link in the corner.
  • —The daemon starts lighter. Shortly after a start Caprock counts a whole-history report for the share dialog, and that count used to read every tool call's full payload — an edit's payload carries the whole file it edited. It now reads only the fields it counts, keeps a command's output only where it can mark a commit or a pull request, and normalises tool inputs in one pass instead of four regular expressions. On a copy of a 1.9 GB database, measured 2026-10-11, the daemon's memory footprint 90 seconds after a start fell from 678 MB to 85 MB, and the report's figures are byte for byte the same.
  • —The dashboard costs less to keep open. Five parts of it ask for your whole history once a minute and the cost strip asks for today's summary every few seconds; each ask counted it all again even when no agent had written anything. Caprock now checks whether anything was written since the last count and reuses the answer when nothing was. Measured 2026-10-11 on a copy of a 1.9 GB database with the Now screen open and no agent working, the daemon used 11.5 s of CPU over two minutes before and 4.7 s after.
  • —The block at the bottom of the sidebar folds. The arrow beside Send feedback collapses feedback, Dashboard, Settings and the theme switch to one row of icons, and the arrow there brings the rows back. The app remembers which way you left it.
v0.104.02026-10-11

Queued messages in the chat

  • —Send while the agent is still working and the message waits under the conversation as a Queued bubble; Caprock types it in when the turn ends — one message per turn, in order, and never into a permission prompt or a question. Each bubble has Edit, Cancel and Send now. The computer holds the queue, so the phone and the Mac show the same one and a phone going to sleep loses nothing; it is dropped when the session ends. For Claude Code sessions Caprock started, with its hooks on; other agents are typed into at once, as before.
  • —/ and @ in the chat field. / at the start lists the agent's commands — Claude Code's own, your project's .claude/commands and skills, and yours in ~/.claude (Codex: its own and ~/.codex/prompts); @ lists the files in the session's folder, filtered as you type. ↑ ↓ Enter and Esc on a keyboard, big rows to tap on a phone. Picking only fills the field.
  • —The desktop app no longer reads its daemon whole to see whether it changed. On every start it compared the bundled daemon and terminal holder with the installed copies by reading each file into memory; it now compares them in small chunks and copies with the file system.
v0.103.02026-10-11

Start an agent from a GitHub issue or pull request

  • —In New agent, for a project on GitHub with GitHub connected in Settings, From a GitHub issue or pull request… lists the open ones — number, title, author, labels — searchable by any of them; ⌘K → New agent from issue… opens it directly. An issue fills in a new copy named after it (issue-123-short-slug) and a first message with its title, link and text. A pull request also starts the copy at its latest commit, fetched with your own git credentials (a fork's too), with a message to review and continue it. Everything stays editable before Start. The list is read only when you open the picker, through the GitHub connection you already have.
  • —Each terminal takes about half the memory. The process that keeps a terminal alive across a Caprock restart (one per shell or agent tab) is now a small program of its own, caprock-pty-host, shipped beside caprock, instead of the whole of Caprock run in a special mode. Measured on an M1 Pro Mac, macOS 27, on 2026-10-11, twelve idle login shells over three runs, each holder process: 11.2 MB of memory before and 4.7 MB after (median physical footprint, as Activity Monitor counts it; resident size 20.3 MB before, 10.9 MB after). Sessions behave exactly as before, and a session started before the upgrade keeps running in the holder it started in. If caprock-pty-host is missing — a go install build, say — terminals are held by caprock itself, as before.
v0.102.02026-10-11

One prompt, several attempts, compared by cost

  • —In New agent, with New copy from a branch, set Attempts to 2, 3 or 4: the same agent gets the same first message in that many separate copies, all from the same commit, each with your .worktreeinclude files and setup command. They open as tabs (attempt 1/3 · …). Compare attempts (⌘K, or Compare in the inspector) sets them side by side: what each is doing, files and lines changed, what it has cost so far and how long it has run, with Open and Review changes on each. Nothing is merged for you; keep the one you want.
  • —Clean up copies. A project's ⋯ menu, or ⌘K → Clean up copies…, lists every git worktree of the project with its branch, last commit, size on disk and whether it is merged, pushed, has uncommitted changes or has an agent running in it, plus the total the copies take. Tick the ones you are done with (Select merged and clean picks those that lose nothing) and remove them, optionally with their merged branches. Only copies Caprock made are removed, never the main checkout; one with work not saved anywhere else asks again and says exactly what would be lost, and one with a session running in it stays.
  • —Saved prompts and commands. Save what you type over and over — Run the tests and fix what fails, or a shell line like pnpm dev — with a name, for every project or for one (Settings → Saved prompts and commands). Press ⌘K, or + in the tab strip → Run saved…: a prompt goes to the agent in front (or opens New agent with it), a command runs in a new shell tab in the folder in front. On your phone, your prompts are chips above the message field: tap one, then Send.
  • —Caprock keeps your computer awake while agents work. A long run no longer stops because the computer idled to sleep. It lets go once nothing is working; the screen can still turn off, and closing a laptop's lid still sleeps it. awake in the status bar says when it is holding. Settings → Keep awake: Off, While agents work (the default), or Always.
  • —Review notes on the diff, sent to the agent. In the Changes view, click a line number to leave a note on it (⇧-click for a range); notes sit under their lines and stay with the worktree. Send to agent types them all, each with its file, lines and code, into the session Caprock started in that worktree as one message — or starts a new agent there with them when none is running.
  • —Fix with agent. A commit or push your git hook refused, files left in conflict, and a pull request's failing checks or conflicts each get a Fix with agent button in the Changes view: the output goes to the agent working in that worktree (or a new one there) with a short instruction to find the cause and fix it.
  • —The Changes view said "A merg is under way" during a merge (and "rebas" during a rebase); it now names the operation, and a hook's coloured output is shown without its escape codes.
  • —New copies are ready to run. A new copy of a branch from New agent (or the sidebar) gets the ignored files your repository lists in .worktreeinclude, such as .env, copied from the project's folder. Give a project a setup command (Setup, under the new copy in New agent), such as npm install, and it runs in the agent's terminal before the agent starts, with its output there; Ctrl+C skips it, and a failure or ten minutes starts the agent anyway.
v0.101.32026-10-11

Hide a project with one click

  • —Hovering a project shows an eye; one click moves it under Hidden, even while something runs in it (it keeps running there, with its dot). It comes back out only when it is the current project or something in it waits on you.
  • —Other folders is now Outside projects, with a tooltip saying what it holds: sessions in folders not added as a project.
v0.101.22026-10-11

Every project where something runs stays open in the sidebar

  • —Its sessions are listed under it whichever project is current, so one click opens any of them and picking one never folds another. Projects with nothing running still fold.
  • —Settings is a row you can find. The sidebar's bottom now has Settings ⌘, by name, with a gear and phone, Telegram beside it, instead of a bare sliders icon, and Dashboard and Settings are outlined so they read as buttons; the palette gains Connect Telegram.
v0.101.12026-10-11

At a glance, Numbers, draws its shares

  • —Each block leads with its largest share in large type, a bar of the whole in the donuts' colours, and a bar under every row the width of its share; two blocks to a row unless the window is very wide.
  • —New agent starts on Claude Code, not on whatever ran last. One Codex session no longer makes every later New agent Codex. To start on another agent, pick it and press Make default beside the Agent label.
  • —The keys on Start and Cancel are big enough to read, in both themes: a larger chip, and on Start a solid one in the button's inverse colours.
  • —Telegram set up from three large steps. Settings → Telegram now leads with Make a bot, Paste its token and Send it a code, an Open @BotFather button and one Connect button; the alert switches appear once the bot is connected, with a note if every alert is off. What a message says is folded away.
v0.101.02026-10-11

The phone section reads at a glance

  • —Three large steps — same Wi-Fi, scan the code, let it work — and one large Show a code button whose text is readable in the dark theme; the rest is folded under Then: …. After a restart, with phones paired, the section leads with Phone access is off and a large Turn on.
  • —Keys on buttons you can read. Start ⌘↩ and Cancel Esc (and the same pair on Add project, Quick chat and Send feedback) carry the key as a bordered chip at full contrast, on taller buttons. The feedback dialog is wider, with larger text throughout.
  • —Take Caprock with you. A slim strip above the status bar offers Open on your phone and Connect Telegram, each opening its Settings section. It ticks an item off by itself once a phone is paired or the bot has a chat, Not needed drops an item for good, × hides it for a week, and once both are done or skipped it never comes back.
  • —Interface size in the app. Settings → Appearance sets 100, 110, 125 or 140%, and ⌘+ / ⌘− / ⌘0 (Ctrl+Shift off macOS) step it. The window zooms natively, terminal included. The app now opens at 110%.
  • —Talk to your agents from Telegram (Premium). Reply to a session's message in Telegram and Caprock types it into that session and presses Enter; the answer comes back in the chat as soon as the turn ends, with code blocks left out unless you switch them on. Approve and Deny answer a question the way the app's card does. /sessions lists what is running, /new <project> <message> starts Claude Code in a project (Bypass asks first), /stop <n> stops one. Only sessions Caprock started are typed into. Off until switched on in Settings → Talk to your agents from Telegram. It runs through your own bot: no server of ours, nothing listening on your computer, and it works on mobile data. See [docs/telegram.md](docs/telegram.md).
  • —Telegram set up in two steps. Settings → Telegram: paste the token from @BotFather (Caprock checks it with Telegram before keeping it), then send your bot the six-digit code shown — one tap from the button. No more copying a chat id out of getUpdates; a chat id typed in before keeps working.
  • —Every Telegram message says whose it is. The first line is now the agent, project and session — 🟠 Claude Code · caprock · Fix the cache test — and what happened moves to the second. Alerts carry Mute this session and, with phone access on, Open on phone; /mute all and /unmute work from the chat. Free, like the alerts.
  • —More than one sign-in. Settings → Subscriptions keeps several Claude Code or Codex accounts side by side, each in a folder of its own. Add subscription types the agent's own sign-in into a shell tab for you to run; Caprock never sees a password or token. Each one shows the account it is signed in to, its 5-hour and weekly windows and its last seven days. New agent and Quick chat get a Subscription picker, remembered per project; the session runs on it and its card says which. Near 90% of a window, the picker says so and offers one with room for the next session — running sessions never move and Caprock never switches on its own. Three per agent are free, your own sign-in included; more are Premium.
  • —See what your sessions cost the machine. While Caprock is open it samples every live session's processes every 5 seconds: the status bar shows the session's CPU and memory, the inspector lists its three busiest processes, and a quiet flame marks the sidebar row and the tab of a session that has used a full core or 2 GB for over a minute (GET /v1/resources, a resources frame on /v1/live). Nothing is sampled while no window is open.
  • —Settings → About shows Caprock itself: the daemon, the app and the terminal holders, with their memory and CPU, and how many holders are orphans, with a button that stops them (POST /v1/resources/reap).
  • —CAPROCK_PPROF=127.0.0.1:<port> serves Go's profiler on a separate loopback listener, off unless set. See .ai/10-infrastructure.md.
  • —Orphaned terminal holders are stopped. A caprock pty-host of this Caprock whose program has exited, or that nothing can reach any more and holds only an idle shell, is stopped five minutes after Caprock starts and hourly after that. Holders left by another data folder (a deleted demo or test install) are stopped from Settings. One that runs an agent or a command is never stopped. On the machine this was measured on, 15 holders left by a deleted demo data folder held about 13 MB each.
  • —The daemon uses less CPU and memory. A file write no longer re-checks every transcript (only the written one; a 2-second sweep covers files written in the last 10 minutes), the storage breakdown reads the events table once instead of sorting it through temporary files, and SQLite's page cache is 16 MiB per connection instead of 64. Measured on a copy of a 1.8 GB database with 1,264 transcripts, two terminals and a steady stream of activity: CPU 3.8% → 3.1%, memory footprint 263 MB → 122 MB, the storage measurement over 40 s → 6.7 s.
v0.100.02026-10-10

Start an agent from a fresh main

  • —New agent's Where leads with New copy from a branch: pick the branch it starts from in a search that puts the main branches first (the default one on top) instead of a list of every branch, and the copy starts from the latest of it on the server — Caprock fetches it first unless you untick Pull the latest. Choosing a new copy is remembered, so the next agent starts the same way.
  • —The phone section reads at a glance. Three large steps — same Wi-Fi, scan the code, let it work — and one large Show a code button whose text is readable in the dark theme; the rest is folded under Then: …. After a restart, with phones paired, the section leads with Phone access is off and a large Turn on.
v0.99.02026-10-10

Find a project by typing a few letters

  • —New agent's Project field is now a search that opens with the cursor in it: type any part of a name ("capr") and pick from the projects you added, the folders you worked in recently and the git repositories under your default folder (two levels deep). The matched letters are bold; ↑↓ and Enter pick, Esc closes the list. The current project is first, so Enter, Enter still starts at once. Picking a repository that is not a project yet adds it and selects it.
  • —Add a project suggests as you type. Typing in the Folder field lists the same matches live; no such folder appears only when nothing matches, and then with the nearest names (Did you mean ~/dev/caprock-web?), one click from being filled in.
  • —Grok. xAI's Grok Build CLI (grok) starts from New agent and Quick chat like the other agents, with its own models, permission modes and your project's instructions. Caprock keeps its sessions in your list but does not read their turns or cost yet.
  • —Agents that are not set up say how to set them up. New agent lists the agents that are not installed after the others, never picked for you; choose one to see its official install command, with Install in a shell tab (the command is typed into a new shell, not run), Copy and its docs. Quick chat names them under its list.
  • —Custom agents (Premium). Settings → Custom agents adds any terminal coding agent by name, command and arguments, with {prompt} and {model} where the first message and the model go. It appears in New agent and Quick chat and opens in a tab.
  • —OpenCode no longer offers an empty model list. It lists the models of the providers you connected (a key in your shell, opencode auth login, or its config); with none, it says so and offers Connect a provider.
  • —Gemini CLI and Codex installed with npm under nvm, Volta or a custom prefix are found even when the login shell's PATH does not carry them. Gemini CLI was never a paid feature: when its binary is not found it is now listed with its install command instead of being hidden.
v0.98.12026-10-10

Voice mode can use the microphone on macOS

  • —Claude Code's /voice in a Caprock terminal recorded silence ("No audio detected from microphone") with no prompt: the app and its daemon were signed with the hardened runtime but without the right to ask for the microphone, so macOS refused every session it held Caprock responsible for (tccd log, 2026-10-10). Both are now signed with com.apple.security.device.audio-input and the app says why it asks, so macOS shows its one-time prompt; after Allow, recording works. Update, open the app once, and start a new session for voice ([docs](docs/app.md#voice-mode-and-the-microphone), ADR-045).
v0.98.02026-10-10

Working from your phone is spelled out where you set it up

  • —The phone section of Settings sits right under Appearance, and its steps now run to the end: pair, press Let it control sessions to work rather than watch, add it to the home screen (an iPhone pairs once more there), carry on a session from its Chat, and turn phone access on again after a restart. The command palette has Open Caprock on your phone, which opens that section.
  • —Each tab shows its jump key. ⌘1–⌘8 by position and ⌘9 on the last tab, as iTerm shows them (Ctrl⇧1… on Windows and Linux); the keys worked before, but only a tooltip said so.
  • —The status bar carries the session in front. With an agent's tab in front, the bar along the bottom shows its model, how full its context is, what the next call costs to re-read it, and what the session has cost.
  • —What's new, in the app. The version in the status bar opens the notes of the release you are running, and the update card's What's new opens the new release's notes in the app instead of a GitHub page.
  • —Windows and Linux see their own shortcuts. Hints, tooltips and the shortcuts sheet showed ⌘ on every platform; off macOS they now read Ctrl+Shift+N, Ctrl+Alt+Shift+N, Ctrl+Enter, F5, the keys that actually work there. The pane keys, which exist only on macOS, are no longer offered elsewhere.
  • —Feedback no longer looks like a volume control. Its icon is a speech bubble rather than a megaphone, and in the desktop app it is a labelled Send feedback row above Dashboard instead of a bare icon beside Settings and the theme switch.
v0.97.02026-10-10

Feedback with a type, a title and screenshots

  • —The Feedback dialog now asks what it is (Bug, Idea or Question, filed with the matching GitHub label), a title that becomes the issue's title, and an optional description. Paste (⌘V), drop or choose up to four screenshots, each with a thumbnail and a ×. Diagnostics (version, OS, screen) sit on one line you can open to read or untick to leave out. Create issue opens the prefilled issue in your browser and puts screenshot 1 on your clipboard; a short Attach your screenshots step says to press ⌘V in the GitHub comment box and Copy next brings up the next one. Caprock still sends nothing: GitHub uploads what you paste.
  • —In the desktop app: Capture window, dropped images, the real clipboard. On macOS and Windows, Capture window attaches a picture of the app's own page (the dialog steps aside for it); it is drawn by the app's webview, so no screen-recording permission is asked. Image files dropped on the window are attached (png, jpg, gif or webp, up to 10 MB). Screenshots go onto the clipboard through the app itself rather than the webview, which not every webview allows.
  • —A share card worth posting. The Figures card says one figure big — the period's cost at API list prices, or, with a flat plan set and a week or a month chosen, the multiple of the plan ("28× my $200/mo Max plan") — with two or three facts under it (sessions, tokens, cache hit), the agents that did the work as marks and shares, the caprock mark, and What's yours? caprock.dev. At API list prices — not a bill stays under the figure. Both styles come in Dark or Paper (caprock.dev's cream), chosen in the sheet. The sheet shows the card large, with one obvious button — Share… where the system has a share menu, Copy image where it does not — and Save image and Post to X beside it. The header's Share is an amber pill rather than a solid block, and the app's Today strip shows its share button in amber. GET /v1/glance takes ?range= for the card's agent mix.
  • —The macOS .dmg opens on a branded window: cream ground, the caprock mark, the app and the Applications folder side by side with an arrow between, and "Drag Caprock to Applications".
  • —The Paper light theme is caprock.dev's palette, all of it. It had the site's cream surfaces but the neutral theme's grey text and light amber, so the app looked paler than its own screenshots. Paper now uses the site's warm dark-brown text, its deep amber for links and accents, and the site's orange button with dark text for filled controls. White is unchanged.
  • —A file dropped on a dialog in the app no longer types into the terminal underneath. The terminal checked only its own box, so a drop on a dialog over it typed the file's path at the prompt.
  • —A long branch no longer hides what a session is. In the sidebar and the tab strip the branch shrinks first and the title keeps its words; the title is cut only once the branch is gone, and the tooltip has both.
  • —Sessions started days ago in another terminal show up under "Running in other terminals". The app asked for the live sessions only after the first live update and dropped any answer a newer update overtook, so beside a busy agent an old session (a "Context recovery" left open in a terminal) never reached the sidebar. It now asks at once and always uses the answer.
v0.96.02026-10-10

A share sheet worth opening

  • —Share shows the card itself, with Today, This week, This month and All time, and three ways out: Copy image, Save image and Post to X (copies the card and opens a post with the words and caprock.dev written; X cannot be handed an image by link). In the app it is in ⌘K as Share my numbers and in the corner of the Today strip. When a week closes or a milestone passes (100 sessions, a billion tokens, a week of 95% cache hits), a small toast offers it, at most once a week.
  • —Star on GitHub, asked once. After three days with sessions and ten sessions, a slim strip at the bottom of the dashboard and the app asks for a star. I starred it says thank you and never asks again; × waits a month.
  • —Premium and Teams, offered when they answer something. A Claude plan window at 80% offers Premium's pause before the limit (under Plan limits and in the app's Today strip); a day in the top quarter of your own days offers the daily cap on Cost; and when two or more people commit to your repositories, Cost and Lifetime offer Caprock for Teams. One offer on screen at a time, never to a licence holder, each dismissible. Dismissals are kept by the daemon (settings.prompts), so the app and a browser agree. New endpoint GET /v1/team-signal counts commit authors from git on this machine; counts only, nothing leaves it.
  • —Every dialog has a way out you can see. Each one now has the same 30px × in its header (44px on a phone), closes on a click on the dimmed backdrop, and closes on Esc: Quick chat, New agent, Add project, the command palette, Open file, the shell-close question, Share, Premium, Teams, Feedback, Version, What's new, New session, Continue in, and the Tasks dialogs. Quick chat had no × at all. Esc closes only the newest dialog, and a text selection dragged out of a dialog no longer closes it.
  • —Settings has its own ×. Beside the Settings heading, and Esc from anywhere on the page but a field: in the app both do what the Settings tab's × does; in a browser they go back to the screen you came from. Closing the Settings or Dashboard tab while it is in front no longer leaves the tab in the strip.
  • —Feedback you can find. The header's 11px grey feedback is now a labelled Feedback button with an icon, and the app's sidebar has a feedback icon beside Settings, on screen whatever tab is in front (a dialog opened from the sidebar is no longer squeezed into it). The prefilled GitHub issue opens in your default browser through the app's own link path, as a terminal link does.
v0.95.02026-10-09

Scrub through a session's spend

  • —Hover the small per-call spark beside Spent this session in the app's cockpit (or focus it, or press it on a phone) and it opens over the block into every priced call of the session. Bars fold into runs once there are more calls than the width holds, but the cursor reads one exact call: Call 412 of 640, when, the model, what it cost, its tokens and the tools it asked for (Bash go test ./...); ←/→ step one call at a time, Esc or leaving closes it. A field the data does not carry is left out, never shown as $0. The series comes from a new lean endpoint, GET /v1/sessions/{id}/calls.
  • —See the subagents at a glance. Under the cockpit's state line, one small tinted avatar per subagent: its type's initial (E for Explore, G for general-purpose, P for Plan), pulsing while it works, accent while it waits on you, muted once finished; at most four and +N. Hover names its type, task and model; a click scrolls to the Subagents list.
  • —Quick chat picks a vendor and a model. Clicking Quick chat (the sidebar, the tab strip's + menu, or Quick chat with… in the palette) shows the agents installed here — Claude Code, Codex, Gemini CLI, OpenCode — and their models; a click on a model starts the chat on it, Enter starts the highlighted one. ⌥⌘N still starts at once, on the agent and model used last.
  • —A default folder for new projects. Settings → Projects, or Set as default beside the Add project sheet's folder field; the sheet starts there (your home folder until set) for an existing folder, a new project, a clone and Browse.
  • —The Add project sheet says what is at the path as you type: will be created, exists, empty — will clone here, exists — not empty, clone will fail, not a folder, already in Caprock. A new daemon endpoint, GET /v1/browse/stat, answers it under the folder browser's rules.
  • —Browse is a folder browser: it opens on the default folder, with a breadcrumb, ↑ to the parent, double-click or Enter to go in, Choose to pick, and arrow keys throughout.
  • —Recent folders leave out scratch: temp folders, Caprock's own data directory (quick chats) and .claude/worktrees agent worktrees.
  • —The Add project sheet has keys on its buttons: ⌘↩ adds, Esc cancels, ⌘1–3 switch Existing folder / New project / Clone, ⌘B switches Recent and Browse (Ctrl off macOS).
  • —A clone may go into an empty folder, as git allows; one with anything in it is still refused.
  • —Tabs are easier to tell apart. The tab in front has an amber top edge; the others are shaded tabs of their own. The strip's + is an amber outline that opens a menu — New agent, Quick chat, New shell, Open file, Add project, each with its key; the separate new-shell button is gone. The pinned Dashboard tab stands apart: an amber grid icon, an amber tint, and a divider before the session tabs.
  • —A session no longer takes a background agent's branch. With an agent working in a git worktree, Claude Code wrote the worktree's branch on the parent's own transcript lines, and the session's header read the agent's feat/… branch over the main checkout's master. A live line now takes the branch its folder's checkout actually has, read from its HEAD.
v0.94.22026-10-09

The app's status bar drops "v2 · 167×36"

  • —The front terminal's protocol and size sat beside the version and meant nothing to a reader; they are now in the version's tooltip (terminal protocol v2 · 167×36), and the bar names the terminal's state only while it is not live (reconnecting).
  • —Each agent is shown by its mark, not by letters. The app's sidebar, Waiting on you, the tab strip, the tray panel and the dashboard's Projects panel, pulse and Now cards named an agent with a text monogram — Cl, Cx, $, oc, cdx. Now they show a small mark in the same space: the Claude spark, the Gemini sparkle, the OpenCode and DeepSeek marks, a C in a rounded square for Codex and a >_ prompt for a shell. Brand colours are toned down to read on graphite, white and paper alike, and each mark says the agent's name on hover and to a screen reader. The paths come from Simple Icons (CC0); nothing is fetched. Row heights are unchanged.
  • —Closing a shell's tab closes the shell. Until now the × of a shell's tab, in the strip or the sidebar, and ⌘W left the shell running, and idle shells piled up as muted Shell rows. Now it works as in iTerm, Terminal or VS Code: an idle shell ends with its tab, without asking; a shell running a program asks first — Shell 1 is running claude. Close and stop it? — with Stop and close, Keep running, close tab and Cancel. The same goes for a shell's pane in a split. Closing an agent's tab still never stops the agent, and the tooltips now say which is which (Close shell / Close tab — the agent keeps running). A shell left running with no tab is still a muted row with ■ to stop it. Close project (a project's ×) now ends that project's shells too, tab or no tab: idle ones at once, and when some run a program one question lists them — alpha: Shell 1 is running claude, Shell 2 is running npm. Stop them and close? — with Stop and close, Keep them running, close tabs and Cancel. Its agents are never stopped.
  • —A shell says what it runs. A shell running a program reads Shell 1 · claude in the tab strip and the sidebar. GET /v1/shells carries a new program field: the shell's foreground program on macOS and Linux, read from the terminal's foreground process group; on Windows its child process, best effort. It is read at most every 2 seconds per shell, only for shells Caprock started, and never signals anything.
  • —No "Waiting on you" block with nothing current in it. When only turns put down more than 12 hours ago are left, the sidebar no longer shows the block with its lone Older line; those turns stay on the Dashboard.
  • —Typing stays quick while other agents work. The app redrew its whole window, every terminal included, on each event an agent sent (about 19 times a second with four busy agents); now terminals are left alone and the sidebar updates at most a few times a second. A tab out of sight no longer draws what its program prints until you bring it forward, and the pulsing "working" dots no longer repaint the window every frame.
v0.94.12026-10-09

Links in the terminal open

  • —Cmd+click (Ctrl+click on Windows and Linux) on a URL in any terminal, or on a link Claude Code prints, opens it in your browser; in the app, the default browser. Hovering underlines it and says how to open it. A plain click still selects text. Only web and mail links open: anything else a program prints is ignored.
  • —Closing things in the sidebar is one ×. Each tab under the project has a × (always on the one in front) that closes it as ⌘W does; the agent or shell keeps running. A project's × closes its tabs and moves it under Hidden, or, while something still runs or waits in it, only closes its tabs and leaves it listed. It replaces the crossed eye. A running agent or shell Caprock started that has no tab gets a ■ on hover, and a tab's right-click menu has Close tab and Stop…; both ask first, as the cockpit does. Sessions started in another terminal have no stop. Delete and Backspace close nothing.
  • —The sidebar says each thing's state, in a fixed order. Every agent under the project ends in one word: working, waiting (amber, the only amber in the list), idle or done. Rows keep the tab strip's order, then the order things started, so a state change never moves one. Sessions started in another terminal fold into one Running in other terminals · N line instead of a row each. The tab in front is marked by a plain edge, no longer amber.
  • —macOS no longer asks "caprock" for what a session did. A find / an agent ran inside a Caprock session asked, in Caprock's name, for the Music library and network volumes, and an ls ~/Desktop for the Desktop: macOS counts everything a program starts as that program, and every new release is a new program to it, so the questions came back after each upgrade (tccd log, 2026-10-09). Sessions, shells, verify commands and the login-shell read now start with their responsibility disclaimed, as iTerm2 and Chromium start theirs: a prompt names Claude Code, Codex or the shell that asked, and the answer outlasts both their updates and Caprock's. The process keeps its pid, terminal and environment (internal/disclaim, pure Go, no cgo).
  • —Background reads stay out of the folders macOS guards. The folder picker and the Projects panel no longer read inside Desktop, Documents, Downloads, iCloud Drive, cloud-storage folders, Music, Movies, Pictures or /Volumes, nor follow a link into one (a link in ~/dev to a share is recognised from the link, without touching the share); the picker still offers those folders, unprobed, and opening one is yours to ask for. Finding the repository a session's folder belongs to still reads up that folder in Desktop, Documents and Downloads — a session there means you work there, and projects in ~/Documents stay grouped by repository — but never in Music, Movies, Pictures, iCloud Drive, cloud storage or /Volumes, where the session keeps its folder's name.
v0.94.02026-10-09

Quick chat in the app

  • —Quick chat under New agent in the sidebar, ⌥⌘N (Ctrl+Alt+Shift+N on Windows and Linux) and the palette start a Claude Code session without picking a folder, in a tab under Other folders. It was only on the browser's Now screen.
  • —Subagents show their model and what they cost. Each row in the cockpit's Subagents section shows the model the subagent runs on (for example Haiku 4.5), how long it has worked and what its own calls have cost. A dot pulses while its call runs. Up to three that finished lately stay listed, muted, with model, duration and cost. The section's header shows what all of the session's subagents cost. A cost Caprock cannot price is left out rather than shown as $0.
  • —The New agent sheet says what Where means. This folder · main is the project's own checkout, Existing copy · <branch> one already made, and New copy on its own branch a separate folder on a new branch, with a one-line explanation and a Branch name field — no git vocabulary needed. Start shows its key (⌘↩, Ctrl+↵ off macOS), which works from every field. Model names fit their box: Codex's default reads gpt-… (default), and the Claude and Gemini labels are shorter (Opus 5.5 · all-rounder).
  • —Tabs never vanish when you pick another project. The tab strip holds every open tab of every project, in the order you opened them, each with its project's name and colour; picking a project shows its tabs in the sidebar or its page, and the strip keeps the rest. ⌘1–9, ⌃Tab and dragging run along the whole strip. Tabs saved by an earlier version come back as they were.
  • —A simpler project list in the app. Only the current project is open in the sidebar, and under it are exactly its tabs — the ones in the tab strip, in the same order and under the same names; shells read Shell 1, Shell 2. After them, muted, whatever of that project still runs with no tab — a session you closed the tab of, a shell left running — and a click opens it as a tab; ended sessions are not listed (they are on the project's page and in ⌘K). No branch rows and no worktree rows. Click another project to make it current; there is nothing to expand or collapse, and the list never reorders when you click: pinned projects first, then by name. The tab you are looking at is the one highlighted row. A project's row shows only what runs in it and what waits. With more than eight projects, the ones with nothing running, no tab open and no pin fold under More projects (it was Quiet, after a week without activity); with eight or fewer every project shows. Waiting on you shows only when something waits.
  • —Dashboard opens as a tab. ⇧⌘D and the sidebar's Dashboard open Now, Cost and the other screens in a Dashboard tab pinned at the left of the strip, with the terminals beside it, one click away; Settings opens there too. In the app, Now leaves out the session cards and the projects panel, which the sidebar already lists; the browser dashboard still shows them.
  • —The terminal matches the app. A new colour choice, Match app, is the default: the dark Caprock palette in the dark theme and Paper on the window's own colour in the light one, so the terminal no longer looks like a dark window laid on a beige one. Caprock's ground is now the dark panel colour, the same as the tab strip. A palette you picked by hand is kept; Caprock saved before this version, when it was the default, reads as Match app — pick it again to keep it in the light theme.
  • —A session whose background agents are still working is no longer "waiting on you". When a Claude Code turn ends with subagents still running, Claude Code resumes the session by itself once they finish. The cockpit used to say "Waiting on you" beside "1 subagent working". It now says Background agents working · N. Such a session gets no waiting dot, is left out of the Waiting on you inbox, ⌘J and the project's waiting count, and sends no "finished" notification until its last subagent stops. A permission prompt still waits, whoever asked.
  • —A subagent no longer moves its session to its own folder and branch. Hooks and transcript lines from a background agent working in a git worktree carry that worktree's folder and branch. They overwrote the session's, so a session in ~/dev/caprock on master showed an agent's feat/… branch and worktree path. A subagent's folder and branch now only fill a session that has none yet. A session already moved goes back with its main thread's next event.
  • —The cockpit header names the project. Under the title it shows the project's name as the sidebar shows it, then the branch only when that is not the project's default. The full path is a tooltip. It used to show a branch beside a folder path cut from the left (…nt-a874ec5240fb37a9d).
  • —Focus no longer jumps while you fill the New agent sheet. Each time the app refreshed behind it, the sheet threw the caret back to Project; now focus moves only when you move it. The same fix covers every sheet in the app.
v0.93.12026-10-09

A Codex call its turn cut off no longer says "running" forever

  • —A call with no output reads "interrupted", muted, once its turn has ended: Codex wrote turn_aborted or task_complete (stored as a tool.post marked interrupted, key codex:interrupted:<line>), the next request started, the person typed the next prompt, or the session ended. Only the call still in its turn says "running". Any other agent's call left without a result reads "interrupted" at its turn's Stop or the session's end — never because a prompt came after it: a prompt queued in Claude Code while a long tool runs leaves that tool running.
  • —Codex's exec reads as the command it ran everywhere, not only in the Chat. The Timeline, the activity feed, the cockpit's tool list and its subagents, the Now phrase, notifications and loop alerts showed the JavaScript an exec call sends; they now show the shell command it ran, apply_patch <file> for an edit, or the tool it called. The daemon reads it the way the Chat does (internal/toolcmd). Memory shows only what the agent wrote, never a tool call, so it had nothing to fix.
  • —OpenCode first run after the daemon started. A daemon that found no OpenCode database at startup read none until it restarted or started an OpenCode session itself, so a session run in a terminal on a machine where OpenCode had just been installed stayed off Now. The daemon now looks for the database every ten seconds until it appears, and follows OpenCode 2's service stream from then on. Found while verifying OpenCode 2 on Linux, where the rest of the support checked out end to end.
v0.93.02026-10-09

OpenCode 2

  • —OpenCode 2 is what opencode.ai, brew install opencode and Arch now install, beside OpenCode 1 under the same command. Caprock reads its sessions (session_v2 and session_message in the same database file) next to OpenCode 1's, with OpenCode's own cost, tokens, model, prompts, replies and tool calls (shell is Bash, subagent is Agent); a database OpenCode 2 migrated from 1, and a session started in 1 and continued in 2, count each message once. The live stream follows OpenCode 2's background service as well. New agent asks opencode --version and starts OpenCode 2 with the flags it has: --standalone --session <id> (linked from the first byte), the model and plan agent as OPENCODE_CONFIG_CONTENT, bypass as --auto, and a first message typed once the TUI is up. /v1/status carries opencode_version.
  • —New agent → OpenCode with OpenCode 2 installed started a TUI with OpenCode 1's --port, -m and --agent, which OpenCode 2 does not have, and on a Mac with /home automounted OpenCode 2.0.26 hung before drawing anything; it now gets the flags its version lists and starts.
v0.92.02026-10-09

Read a project's files in a tab

  • —⌘K → Open file… lists the worktree's files and narrows them as you type; the Changes view (o, or the ↗ on a row) and the cockpit's changed files open one too. The file opens read-only in a tab named after it: Markdown formatted, with a Source switch and links to other files opening their own tabs; any other text with line numbers, wrapping and ⌘F. A binary, or the part of a file past its first megabyte, is left to your editor. The tab reads the file again when you return to it or an agent edits it. Nothing outside the project is ever read, through .. or a symlink. GET /v1/projects/{id}/file, GET /v1/projects/{id}/files.
  • —A stop at the plan limit (Premium). When Claude's 5-hour or weekly window passes the share you pick in Settings → Pause at the plan limit (90% until you change it, or off), Caprock pauses the Claude Code sessions it started, names them on Now and in a desktop notification with the reset time, and resumes them after the window resets. A session you resume by hand is left alone until the next window; a session you started yourself is never touched; figures more than ten minutes old pause nothing, and Settings says when that is the case. Without Premium the 90% alert on Now is unchanged and offers the stop. GET /v1/window-stop, settings.window_stop_pct.
  • —Today, at the top of the sidebar. Under New agent: what every agent has spent today, how many agents are running and how many wait on you, and the Claude plan's 5-hour and weekly windows as thin bars with the share used and when each resets. Click the spend for Cost, a window for the plan limits, Running for Now, Waiting for the session that has waited longest. While the sidebar is open the status strip leaves the windows and the spend to it.
  • —A menu on every project. The ⋯ on a project's row (on hover, and always on the project in front), a right-click, or Shift+F10: hide it from the sidebar or show it again, close all its tabs (the sessions and shells in them keep running), open its folder in an editor, or remove it from Caprock after a confirmation — its files and sessions stay as they are.
  • —Subagents in the agent cockpit. While a session's subagents work, the cockpit lists them under Subagents · N: each one's type and the task it was given, its current call and how long it has run, how many calls it has made, and waiting on you while it has a prompt open; those that finished in the last half hour are one line. The daemon sums them up (GET /v1/sessions/{id}/subagents), so thousands of subagent events never reach the page.
  • —Project rows say what runs in them. Each row shows how many agents are running there and what the project cost today, beside the waiting badge rather than in place of it; a project with neither shows neither. The branch takes only the room left over and steps aside, rather than showing as a sliver, when too little is left.
  • —New agent is outlined, and the sidebar's shortcuts are readable. The New agent button keeps the accent as a 1px border and label on the panel instead of a solid fill, and the key hints beside New agent, Add project and Dashboard are set at the size of the label next to them.
  • —Codex sessions read as a conversation in Chat. Your own prompts now appear beside Codex's replies; before, only the replies did. Each command shows the command it ran, not its JSON or JavaScript wrapper. A finished call no longer reads "running" forever: it shows its output, and a failed one says so, with the exit code where Codex recorded one. Sessions already imported get their prompts and outputs once, from the transcripts still on disk.
  • —The cockpit's tool list was empty on a busy parent. A session whose subagents logged hundreds of calls an hour had none of its own among its newest 400 events, so Recent tools read "No tool calls yet" and the cost spark showed one bar. The cockpit now asks for the main thread's own calls and turns (events?newest=1&main=1&kind=…).
  • —A subagent's permission prompt read as the parent's. The card, the cockpit's Now line, the menu bar popover and the notification now say "Subagent (general-purpose) wants to run Bash", lead with what the command does (rm -f $C/* rather than C=/pri…), and show the whole command wrapped, folded when long.
  • —Approving with several prompts open could answer the wrong one. When subagents ask at once, Caprock cannot tell which dialog the terminal is showing. With two or more outstanding, the card lists them all with who asks and offers Open terminal instead of Yes and No, the daemon refuses to type an answer, and notifications carry no Approve. A prompt is now tied to its own call, so the call finishing clears it and nothing else.
v0.91.02026-10-08

A New agent button you cannot miss

  • —The app's sidebar opens with an accent New agent button (⇧⌘N) and Add project (⌘O) under it; the small Add beside Projects is gone.
  • —A shorter project list. Projects with nothing running and no activity for a week fold under Quiet at the bottom of the sidebar, and any project can be hidden from its row (the crossed eye) into Hidden and brought back from there. Both groups start closed and remember being opened. A project with something running or waiting, or the one in front, always shows.
v0.90.02026-10-08

The agent cockpit

  • —The panel beside an agent's terminal now shows what only Caprock knows, live: the agent's character in its state, what the session has cost and what each of its last model calls cost, the context bar and what every call pays to re-read it, the tool call running this second and for how long, the last seven calls with their durations, the uncommitted changes, the plan windows and a loop warning when there is one. It is open by default; ⌘I closes it, and closed is remembered. An agent that does not report a figure (OpenCode's plan, a context before the first answer) shows nothing there rather than a zero.
v0.89.02026-10-08

Adding a project is easier to find and to finish

  • —The sidebar says Add beside Projects instead of showing a lone folder icon. Pasting a repository URL fills the destination the way git clone names it; a destination whose folders do not exist yet is made rather than refused, and ~ works in any path. A new First task field starts an agent on the project once it is added — the instructions field, which only shapes later sessions, now sits behind a link so it is not mistaken for one.
  • —The Premium dialog opens the right page. "Read more" lands on the feature you were reading about on caprock.dev/premium, and the buy buttons tell the checkout which feature they came from — the feature's name only, nothing about you.
v0.88.12026-10-08

?agent=deepseek on the stats endpoints is no longer a 400

  • —The DeepSeek filter the dashboard offers was refused by the API, which accepted only claude, opencode, gemini and codex.
v0.88.02026-10-08

Claude Haiku 5.5 is priced and can be picked

  • —The pricing table gains it at Anthropic's list price (read 2026-10-08) — its turns were unpriced until now — and New agent offers it in place of Haiku 4.5 after it answered the real claude. Anthropic charges more for a prompt over 100,000 tokens; the table carries the lower tier, so such prompts are under-reported, as the note in the table says.
v0.87.02026-10-08

Keyboard shortcuts in the palette

  • —⌘K → Keyboard shortcuts lists every key the app answers to.
  • —A session put down days ago is no longer "Waiting for you" on Now. Past 12 hours the banner goes, as the sidebar already folds it.
  • —The inspector cuts a long folder path from the left, keeping the end that tells two worktrees apart.
  • —Linux notifications behave like the macOS ones. Checked against the released app on dunst and XFCE: Approve, Deny and a click already worked; three things did not. An approval now stays until it is answered (it vanished after about ten seconds, its buttons with it); one answered elsewhere, or by its own button, is taken down, and none is left up when the app quits (a press there did nothing), without leaving a blank notification in its place; and Open shows once (XFCE drew it twice).
v0.86.12026-10-08

What Claude Code writes to itself no longer shows as your message in…

  • —What Claude Code writes to itself no longer shows as your message in Chat. A background task finishing, a reminder or a slash command's output arrived as a user turn of raw XML, a screen long; it is now one quiet line (the task's own summary), with the full text a click away.
v0.86.02026-10-08

The palette finds past sessions

  • —⌘K searches every session the daemon knows as you type and lists matches under History with project, age and cost; before, it saw only what was running and offered to start a new agent on the words instead.
  • —A past session in the app opens on its chat, with one way back in. Resume in a tab is the page's button, the permission mode beside it; Continue in…, Open in your terminal and the resume command moved under More. It opened on the raw timeline, with five controls in a row.
v0.85.02026-10-08

Recent sessions on a project's empty screen

  • —With no tab open, a project now lists its last six agent sessions with cost and age, and one click on Continue carries an ended one on in a new tab. Before, the sidebar showed only what was running and yesterday's session took the dashboard and a session page to reach.
  • —Settings from anywhere in the app. ⌘, opens it, a button sits beside Dashboard, and the palette finds it by what is in it (permission mode, theme, terminal, notifications, phone), along with the dashboard's screens by name. In the app, Settings leads with how new agents start, appearance, terminal and editor; the phone and Telegram follow.
  • —A worktree started from New agent no longer shows up as a change in the main checkout. .caprock-worktrees/ was listed as untracked there, and Commit all would have committed it.
  • —A new worktree needs no name. Left empty, it is named after the first message; a name with characters the daemon refuses is said beside the field, not in a clipped line in the footer.
  • —Waiting on you is about now. A turn that ended more than 12 hours ago folds under Older and stops counting in the badge, the project's count and ⌘J, so the session that just finished is not behind one from last week.
  • —⌘T opens the shell in the folder of the tab in front — its worktree — not the project's root.
  • —A tab is named after its agent, with +N for its split panes; it read "shell" as soon as a shell was split beside it.
  • —The dashboard in the app drops what the window already shows: the website footer, a second link state, theme switch and version.
  • —"1 changed file", not "1 changed files".
  • —A long branch no longer hides the project's name in the sidebar: the branch gives way first.
v0.84.12026-10-08

Permission card buttons work after a restart and with agents in the…

  • —Permission card buttons work after a restart and with agents in the background. The card read the dialog off a replay of the session's recent output, which after a restart could start mid-screen and hide it; every button then said not on the prompt. It now also reads the dialog from what was drawn after the prompt arrived, and allows the extra rows a status line and background agents add under it.
  • —A terminal reopened after a restart draws its whole screen. Reattaching to a running Claude Code session could leave its status rows as bare numbers, their labels lost from the replay. The first size a terminal sends now makes the session redraw everything.
v0.84.02026-10-08

A new install asks before running commands

  • —With no saved preference, new sessions start in Accept edits · asks first rather than bypass. Bypass is one pick away in the New agent sheet, and the mode you pick there is kept for the next agent. If you relied on the old default, pick bypass once.
v0.83.02026-10-08

Notifications on Linux open the session and answer prompts

  • —A click on a notification used to do nothing there. It now opens the session, and a permission request carries Approve and Deny where the desktop's notification server shows actions.
  • —The dashboard header says only caprock. The mission control tagline from before the desktop app is gone.
  • —Built with the modernc.org/sqlite 1.60 driver, and the dashboard with Vite 8.3.2. No change in behaviour.
v0.82.02026-10-07

Project instructions reach Codex too

  • —A Codex session started in a project gets its instructions as Codex's developer instructions, beside the repository's own AGENTS.md.
  • —Opening the dashboard no longer asks for Documents or Downloads. The per-repository cost panel ran git in the folder of every session with spend, so a session once run in ~/Documents or ~/Downloads made macOS ask for that folder after each update. Folders macOS guards are left alone until you open the session itself.
  • —The menu bar says what its number counts. It showed a bare 2; it now reads 2 waiting. The Dock badge stays a number.
v0.81.02026-10-07

Project instructions

  • —A project can carry instructions for its agents — set them as you add it, or from the New agent sheet — and every Claude Code session started in it gets them appended to its system prompt.
  • —/exit leaves one shell, however many windows show it. When a session's tab was open in the app and in a browser (or on a phone), each window started its own shell in its place, and the project listed two. They now share one.
  • —A screenshot dragged from its thumbnail reaches the session. macOS keeps the corner thumbnail in a folder only the app it is dropped on may read, so Claude answered "operation not permitted". The app now copies such a file into Caprock's paste folder and types that path.
  • —The permission card on a phone names no keys. On the session screen, where there is no terminal, it said Keys work from the terminal and drew Y and N on its buttons. It now shows the buttons alone there.
  • —A permission card whose question is gone can be closed. When Claude Code's request timed out or a check denied it, the card stayed until the turn ended and its buttons said that option is not on the prompt. A button now takes such a card away, and × hides any card without typing anything.
  • —The Dock badge counts everything waiting on you. It counted permission prompts only, so an agent that finished and waited for your answer left the Dock bare. It now shows the popover's Needs you. The menu bar shows the same number beside the icon, instead of an unlabelled 5-hour percentage; a plan window appears there, named, once it is 80% full.
v0.80.02026-10-07

The desktop app updates itself in one click

  • —When a newer Caprock is out, the status strip says Update to vX.Y.Z — Restart: it downloads the new app with a progress bar while you work, checks its signature, installs it and restarts, and everything comes back as it was: every session still running (the app moves its own daemon onto the new version without ending them), the same tabs in the same order, splits and their sizes, the sidebar, the window, each terminal scrolled where you left it and anything half typed into an agent. Nothing installs unless it is signed with Caprock's release key for that exact version. On its first launch the app asks once whether to check for updates automatically (the same release check as Settings → Privacy, off unless you say yes); Check for Updates… in the app and tray menus and in the command palette works either way. The macOS app, the Windows installer and the Linux AppImage update themselves; .deb and .rpm installs are told to use their package manager. Each release now carries latest.json and the signed update files.
  • —A newer version announces itself in the corner. A card at the bottom right of the app says Caprock vX.Y.Z is available once per version, with Update and restart, What's new and Later, and shows the download's progress; the status strip keeps its line.
  • —View → Reload (⌘R) in the Mac app; F5 on Windows and Linux. Ctrl+R is left to the shell's history search, and F5 in a focused terminal still reaches the program in it.
  • —The app no longer keeps the old UI after an upgrade. After brew upgrade --cask, the app loaded its window from the previous daemon, then replaced that daemon with the new one and never reloaded: the window ran the old version's code, and the status strip named it, until the app was quit. The app now decides to replace its daemon before loading any page, and reloads the window and the menu bar popover if a replacement still happens under them.
  • —A page reloads itself when its daemon is updated. The daemon names its version in the page it serves; when the live link comes back to a different one, the page reloads on the same screen — in the app, a browser tab or a paired phone. Typed terminal input is kept (it lives in the session). With text typed into an open sheet it offers Reload — Caprock was updated instead, and it never reloads twice for one version.
  • —A command that finishes fast keeps its last lines. Caprock closed a terminal the moment its process exited, so the final output of a short command could be lost (seen on Linux). It now reads the terminal to the end first, waiting at most two seconds.
v0.79.02026-10-07

One consent for bypass, in Caprock

  • —The first bypass session on a computer used to open on Claude Code's own warning, where Enter means No, exit and drops you into a shell. The New agent sheet and the New session dialog now show what bypass means and start with Accept and start, which answers that warning for good. A bypass start without it is refused by the daemon rather than left on that screen.
  • —Start an agent without the mouse. In the New agent sheet ↑ and ↓ change a choice in place instead of opening its list, ⌘↩ starts from anywhere in the sheet (a list or a button too, not only the message), and the footer names the keys: Tab moves, ↑↓ change, ⌘↩ starts, Esc cancels. The dashboard's New session dialog takes the same arrows.
  • —The permission card is back under the terminal, and answers from it. 0.78.2 hid it while the session's terminal was in front. It shows again, and with focus in that terminal Y, A ("don't ask again") and N answer it without reaching the terminal; Enter and Esc stay with Claude Code's own menu. Each button names its key.
  • —New sessions start on bypass. With no saved preference, the New session dialog and the New agent sheet open on Bypass · never asks instead of Accept edits. Settings → New sessions still overrides it, and a bypass start from a paired phone is still confirmed.
  • —The model list names the model you get. It offered "Opus 5" while Opus 5.5 was out, and picking it started the older model. The list is now Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5, each label with its exact version, opening on Opus 5.5; a newer model in the pricing table that the list does not offer fails the build.
  • —"Bypass · never asks" is spawned the way Caprock's own unattended workers are, with --dangerously-skip-permissions. (Claude Code treats it as the same mode as --permission-mode bypassPermissions; an earlier version of this entry said the latter keeps asking. The questions came from permissions.blockReadsOutsideWorkingDirectories, below.)
  • —Caprock's own files are not a question. Claude Code sessions start with --add-dir for the directories Caprock writes to (pasted screenshots and quick chats), so a user with permissions.blockReadsOutsideWorkingDirectories on is no longer asked about files Caprock put there itself. That setting is a perimeter no permission mode lifts; a run meant to be unattended needs it off.
  • —A session that exits leaves a shell, not a dead tab. After /exit, the tab becomes a shell in the same folder, keeping its place in the tab strip and in a split. Before, the process was gone and the tab kept showing its last screen.
  • —The pricing check against the site no longer breaks when the site's pricing file grows a team price: it reads the personal plans only.
v0.78.22026-10-06

Settings → New sessions

  • —the permission mode new agents start in. The New session dialog and the New agent sheet open on it, and a session started without choosing one (the project terminal's "New session here", a quick chat) starts in it. Unset, nothing changes.
  • —make app-local: try a change to the Mac app without a release. It builds this checkout's app for the Mac's own architecture, installs it over /Applications/Caprock.app and relaunches it; the app then moves the daemon onto the build it carries, and running sessions carry on. Stamped <last tag>-dev+<commit>, which About Caprock shows and the update notice does not mistake for an old release. About half a minute once the build cache is warm. make app-local-revert goes back to the released app.
  • —One permission question in the app, not two. When a session's terminal is in front, its own "Do you want to proceed?" menu is where you answer — press Enter — and Caprock no longer draws its approval card above the status strip for it. The card still shows when the chat or a Changes view covers the terminal, and now says ↵ Enter in the terminal = Yes; other tabs' prompts still reach you through the Inbox, the menu bar, the notification and the phone.
  • —Keys on the permission card. Y or Enter is Yes, A the "don't ask again" option when there is one, N or Esc is No, shown on the buttons. They never fire while you are typing in the terminal or a field.
  • —"Yes, for the rest of this session" no longer rejects the call. The button typed 2, assuming Claude Code's usual menu; in auto mode the dialog is "1. Yes 2. No", so it answered No and the tool call was lost. Every permission button now reads the menu on the session's screen and presses the option whose text matches; when that option is not there, it types nothing and the card says that option is not on the prompt — answer in the terminal.
  • —A permission card answers the dialog it shows. Each new permission request (a subagent's too) replaced the last, so a card could name one request while its key landed in another's dialog. Requests now queue the way Claude Code queues them: the card shows the oldest, says how many more wait, and a later request never overwrites an earlier one.
  • —Upgrading the Mac app now upgrades its daemon too. An app started on a new release kept the daemon of its own from the old one running until that was too old to talk to. It now puts the daemon it carries in place once per launch when the two differ; sessions keep running.
  • —Quitting the app with SIGTERM (kill, a script) now quits it the way Cmd+Q does, keeping the window's size and place.
  • —A live session's Changes tab no longer flashes "loading…" on every event. Each new event re-read the diff as if it were a new question, blanking the list and drawing new rows, so a click on a file or on "expand all" in that moment was lost. The list now stays on screen while it is read again.
  • —App shortcuts act on what is on screen. A key pressed in the instant after the sidebar updated could act on the state before it: ⌘J said nothing was waiting with a waiting session in the list, and ⌘T asked to add a project with one already there.
  • —The app's status strip shows the daemon's real version. It was read once when the page loaded, so after the app swapped its daemon it kept the old number. It is now read again whenever the live link reconnects and when the window regains focus.
  • —Continuing a session keeps the permission mode it was running in. A session run with permissions skipped (--dangerously-skip-permissions) came back from "continue here" or "branch here" in the default mode, asking before every command. Caprock now continues it in the mode its hooks last reported — bypass, accept edits, plan — and says which beside the button, where it can be changed before continuing. A session with no recorded mode uses the setting under Added above.
v0.78.12026-10-06

Download links for the desktop app that never change

  • —Every release now also carries the app under names without the version — Caprock-macOS.dmg, Caprock-Windows-setup.exe, Caprock-Linux.AppImage, Caprock-Linux.deb and Caprock-Linux.rpm — so https://github.com/dspv/caprock/releases/latest/download/Caprock-macOS.dmg always fetches the newest build. The versioned files stay, and checksums.txt lists both. A release is marked Latest only once all of them are attached, so a link never points at a release still being built.
  • —Terminal defaults follow the JetBrains IDE terminal: JetBrains Mono at 13 px, line height 1.2 (was 1.15), Regular and Bold weights. Terminal settings you already saved are kept as they are.
  • —The terminal's font looks like JetBrains Mono again, in every glyph. In the macOS app, digits, capitals and punctuation came out heavier and brighter than the letters beside them, at random, for as long as the terminal was open: the WebGL renderer drew some glyphs outside the page and some inside it, and WebKit rendered the two differently. Every glyph is now drawn the same way. The app terminal also asks for every part of the face (Latin, Cyrillic, Greek, Vietnamese; regular and bold) before it opens, instead of measuring whatever had loaded, and measures and redraws itself when a face finishes loading. Ligatures are off in every terminal, so => and !== show the characters that were typed.
  • —A permission prompt survives a daemon crash, not only a restart. The prompt an owned session was waiting on was stored in the background after the hook that drew it had been answered, so a daemon killed in that moment — a crash, or any stop on Windows, which has no graceful one — came back with the dialog still on the session's screen and no buttons to answer it. The prompt is now stored before the hook is answered.
  • —Dropping a file on a terminal in the desktop app works. The app's native drag-and-drop handler took every drop, so a file dragged from Finder onto a session did nothing. The app now types the file's real path, quoted, into the terminal under the pointer, as a terminal does: nothing is uploaded, any file or folder goes, and Claude works on the original rather than a copy. In a browser a drop still uploads, as before. Tabs in the app's tab strip reorder by dragging again (they could not while the native handler was on).
  • —macOS: one "caprock" in Privacy & Security, not one per release. The Homebrew daemon runs from a new Cellar path every release, and macOS lists every path as another program; agent sessions ask for folder access in the daemon's name. The app now runs its own daemon from one path in the data directory and moves a running Homebrew daemon onto it once (unless the Homebrew one is newer; "own_daemon": false in app.json keeps it). Sessions keep running through the move. Each release still asks once until Caprock is signed with a Developer ID. [ADR-040](.ai/08-decisions.md)
  • —The app no longer aborts when the daemon stops while the menu bar popover is still loading its first page. It asked the WebView for its URL, which a page that never committed does not have, and wry unwraps it on the main thread.
  • —Test, bench and preview daemons stay out of Desktop, Documents and Downloads. A daemon running with a temporary HOME no longer watches or runs git in those folders of the real account, even with a copy of the real database, so it never raises a privacy prompt; the bench stand refuses to live there.
v0.78.02026-10-06

GitHub in the app (F14)

  • —Settings → GitHub connects your account with your GitHub CLI login (read when needed, never stored), a pasted token (kept in the macOS login keychain, else a file only you can read) or, once a Caprock OAuth app is configured, Sign in with GitHub. It shows the account, scopes and a health line (last success, requests left, the last error word for word). ⌘O and the phone's Start work list and search your repositories to clone. A project with no remote can be put on GitHub (private by default). A worktree's Changes view opens a pull request — title and description from the commits or the agent's summary, base the default branch, draft switch, pushed first when needed — and then shows its checks, reviews and mergeability; the sidebar marks worktrees with a pull request, green, amber or red. Updates arrive live; GitHub is read at most once a minute per repository with conditional requests, and the app can notify you when CI fails or a review lands. Phones read all of it; a controller phone can open a pull request. Every GitHub error is shown with what Caprock was doing. See [docs/app.md § GitHub](docs/app.md#github).
  • —A guide to the desktop app. [docs/app.md](docs/app.md) walks through the window — projects, tabs, terminal, chat, inspector, the keys — the menu bar, badge and ⌃⌥⌘C, notifications with Approve and Deny (and when Approve is offered), pairing a phone over Wi-Fi or Tailscale, and starting work from the phone. The README now leads with the app, with a screenshot made from demo data; the daemon, the dashboard and the CLI are documented as before.
  • —Split panes in the app. ⌘E puts a new shell beside the terminal in front, ⇧⌘E below it; ⇧Enter in the palette opens a session there. Each pane has a header with its status, the focused one a ring; ⌘[ and ⌘] move between them, the divider drags or takes the arrow keys, ⌘W closes the pane (the session keeps running). Up to four panes a tab, restored on relaunch.
  • —Jump to what is waiting. ⌘J opens the next session waiting on you, permission prompts first, and wraps; the palette lists them at the top.
  • —A sharper command palette. Results rank by where they match (start of the name, start of a word, anywhere); "New agent in <project>" for every project; text that matches nothing starts an agent on it in a new worktree named after it, with the text as its first message.
  • —Menu bar popover (macOS). A click on Caprock's menu bar icon opens a panel under it: what needs you (approve or deny a prompt shown in full), what is running with its cost, today's spend and the Claude and Codex plan windows with when they reset. It opens over the app you are in without switching to Caprock or raising its window. A right click keeps the menu.
  • —Find in the terminal. ⌘F (Ctrl+Shift+F off macOS) opens a find bar over the focused pane: matches highlighted as you type, "3 of 12", match case, regular expressions, Enter and ⇧Enter for next and previous. Esc closes it and puts you back in the terminal. Works in every pane of a split, in agent and shell tabs alike.
  • —Open in your editor. VS Code, Cursor, Zed or a JetBrains IDE, when installed: right-click a project or worktree in the sidebar, use the palette ("Open <project> in Zed"), or the inspector — the session's folder, or a changed file at its first changed line. Settings → Editor picks the default. Only from this computer: a paired phone can never open an editor on its screen.
  • —Terminal colours and fonts. Settings → Terminal: five palettes (Caprock, a light Paper, Catppuccin Mocha, Tokyo Night, Solarized Dark), JetBrains Mono or any of SF Mono, Menlo, Fira Code and other monospace faces found on the machine, size, line height and cursor shape, with a live preview. Every open terminal changes at once; kept per browser.
  • —Update notice in the app. With release checks on, the status strip says when a newer Caprock is out and opens to the exact command — brew update && brew upgrade --cask caprock-app for the app from Homebrew, and the daemon's own command when the formula installed it. "Not now" hides that version. Nothing installs itself.
  • —Review and commit without a terminal. Click a worktree's ±N in the app's sidebar (or Review and commit in the inspector, or "Review changes" in the palette) for its Changes view: staged, unstaged, new and conflicted files; each file's diff unified or side by side, with large diffs collapsed until asked and a 5,000-line diff scrolling smoothly; stage, unstage and discard files (discarding asks first), stage or unstage single hunks; a commit message box that can start from what the agent said last; Commit, Commit & Push, Push (a new branch is published and tracked; never forced), Pull (fast-forward only) and Fetch. Keys: j/k, s, u, d, v, c, r, Esc; ⌘↵ commits, ⇧⌘↵ commits and pushes. Hooks run, and a hook that refuses shows what it printed; a remote that refuses your credentials says so. On the phone, a session's Changes tab gains the same commit and push for its worktree (tap files to choose them); a view-only phone sees the counts.
  • —Release checks run at most every 6 hours, and a running daemon keeps checking. They used to run at startup and when switched on, throttled to a day, so a daemon left running for a week never heard of a release. Repeats are conditional (If-None-Match), so an unchanged release costs GitHub a 304 and no body. Still off until you turn it on.
  • —The app uses less CPU when idle, and stalls less on the first tab. With its window hidden the app checks the daemon every 5 s and the page stops polling (the live connection stays); the visible, silent window polls less and redraws its clocks only when they change. A terminal warms up once after start, and replays are written in slices. Hidden, the app now takes about 0.3% of a core (was 1%); the first tab of a launch still holds the window for up to 70–130 ms (was 70–150 ms).
  • —A half-open connection is caught sooner. After 20 s without a word from the daemon the page pings it and redials if no answer comes within 2 s (about 22 s, was 25 s).
  • —A project's main checkout keeps its own sidebar row beside a linked worktree. With nothing running in either, the project read as one checkout named after the linked worktree's branch.
  • —The app's status strip rounds plan usage as Now does. It floored, so the same window read 41% there and 42% on Now.
v0.77.02026-10-06

Install the desktop app

  • —Each release now carries the app beside the daemon, built from the same tag and bundling that daemon: a universal .dmg for Apple silicon and Intel Macs and a Homebrew cask (brew install --cask dspv/tap/caprock-app), an installer for Windows, and an AppImage, .deb and .rpm for Linux. The app is not notarized or code-signed yet; the cask opens it without a Gatekeeper prompt, and [docs/install-app.md](docs/install-app.md) has the one-time step for the .dmg and for SmartScreen. On first launch the app starts the caprock formula's daemon when Homebrew installed one, so brew upgrade caprock keeps it current, and its own bundled daemon otherwise. A release can now add the .dmg from a Mac with make app-release TAG=vX.Y.Z when Actions is down.
  • —Start work from the phone. A phone allowed to control sessions gets Start work on Now: clone a repository (an https:// or git@ address, into a folder under your home), make a new project, or make a worktree — then Start an agent here, which opens on the session's chat. A clone runs on the computer and the phone follows it: lose the signal or lock the phone mid-clone and the page catches up when it is back, without cloning twice. A view-only phone is refused each of these, on the screen and by the daemon.
  • —Pair over Tailscale by QR. When the computer is on Wi-Fi and Tailscale, Caprock listens on both, and the pairing panel draws the code for the one you pick — Wi-Fi, Tailscale (works on mobile data) or the Tailscale name. Without Tailscale it says plainly that a phone off the Wi-Fi cannot reach it.
  • —The chat on a phone. Its field now holds a message typed while the connection is down ("Will send when connected") and sends it on reconnect unless a permission prompt waits, as the terminal's keys bar does; if the session ends meanwhile it stays as a draft to edit. Back from a long drop, the chat fetches everything it missed instead of starting over at the newest page, and what you were reading stays where it was. A short conversation sits at the bottom, by the field. On the Chat tab the session header is one line and one row of controls, with the rest under Details, so the conversation starts about 270 px higher at 390 px. Photo puts the picture's path in the field, to send with your words.
  • —The desktop app's terminal no longer cuts off its last rows a few seconds after opening, when it switches to the faster renderer.
  • —The terminal no longer covers its own text field on a phone. Once the faster renderer took over, the terminal grew past its box and over the field under it; it now fits itself again.
v0.76.02026-10-06

The phone never needs a manual reconnect, and says honestly where it…

  • —The phone never needs a manual reconnect, and says honestly where it stands. The live feed and every terminal share one reconnect policy: forever, with full jitter from 0.5 s to a 15 s cap, and at once when the page comes back, the browser goes online, or the network changes; a socket that looks open after a wake must answer within 2 s or it is replaced, and a half-open one is caught within 25 s. The header (and the app's status strip, and a pill over the terminal) says Live only while the daemon answered in the last 25 seconds; otherwise Catching up…, Reconnecting (n) · next try in s, Offline since hh:mm, or Control revoked with the reason. A message sent from the phone's keys bar while offline is held as Will send when connected and sent on reconnect if no permission prompt waits — otherwise it stays as a draft with Send now; Esc, arrows, Ctrl+C and prompt answers are never queued.
  • —Read a session as a chat. A new Chat tab shows the conversation — your prompts, what the agent wrote, each tool call on one line that opens — with a field and keys that type into the session. On a phone a session opens on it; in the app, an agent tab switches between terminal and chat. Messages always appear in the order the session produced them, never twice, however late they arrive; scrolled up, nothing moves while new ones come in, and a "↓ N new" pill takes you back down.
  • —Approve and Deny right in the macOS notification. A permission prompt's notification now carries the buttons: Approve answers that prompt without bringing Caprock forward, Deny refuses it, and a click on the notification opens the session. Approve is offered only when the notification shows the whole request; a longer command offers Open in Caprock and Deny, so you read it on the prompt card first. A prompt answered in the meantime is never typed into: you get "Already answered" instead, and a prompt answered in the terminal or on the card takes its notification away. macOS asks once whether Caprock may send notifications. Windows and Linux keep the click that opens the prompt.
  • —The desktop app tells you when a session needs you. An OS notification names the project and session and what it wants — Bash: go test ./..., the file an edit would change, the question asked — the moment an agent is blocked on a permission prompt. Clicking it opens that session with its Yes and No buttons; a prompt answered meanwhile is refused, never typed into. Nothing for the session you are already looking at. On by default in the app; "finished" notifications are there too, off by default; both switch in Settings → Desktop notifications. Same limits as Telegram alerts, which stay off unless you turn them on.
  • —The live feed replays what you missed. Every /v1/live frame now carries a seq, and the daemon keeps the last 2,000 frames or 10 minutes. A dashboard that drops off — a phone locking, Wi-Fi changing — reconnects with ?since= and receives exactly the frames it missed, in order, so a notification sent while it was offline still arrives; when the gap is older than that, a reset frame makes every screen refetch instead. A socket silent for 25 seconds is taken for dead and reconnected, as the terminal already does. Clients that predate this keep working unchanged.
  • —Menu bar, badge and a global hotkey in the desktop app. The menu bar (a tray on Windows and Linux) shows the 5-hour and 7-day plan usage for Claude and Codex, today's spend and the sessions waiting for your approval; click one to open it. The Dock badge counts them and clears when none wait. ⌃⌥⌘C (Win+Alt+C on Windows and Linux) brings the window up from any app, changed or turned off in Settings → Global shortcut; Wayland does not allow global shortcuts. On macOS closing the window now keeps Caprock in the menu bar.
  • —Paging back through a session no longer skips events. Events sharing a millisecond with the oldest one shown were left out of the page before it.
v0.75.02026-10-05

A terminal-first workspace for the desktop app, at #/app (or ?app=1)

  • —A sidebar with what is waiting on you, then every project with its worktrees, sessions and shells, today's cost and status dots; terminal tabs per project that survive a relaunch; an inspector with cost, context, tokens, the permission prompt and the uncommitted changes; a status strip with the plan limits and today's spend. ⌘T a shell, ⇧⌘N an agent, ⌘O a project, ⌘1–9 tabs, ⌘W closes a tab without stopping its session, ⌘K a command palette. The dashboard opens inside it. Sessions in folders that are not a listed repository sit under Other folders; ⌘O clones with git's progress shown. The browser dashboard and the phone are unchanged.
  • —Projects in the daemon. Caprock keeps a list of the projects you work in — seeded from the repositories your sessions ran in, plus any you add, create (git init optional) or clone — with each one's branch, changed files, ahead/behind, worktrees, live and waiting sessions and today's cost (GET /v1/projects). Git state follows a commit or checkout made anywhere within about a third of a second, by watching .git; idle projects run no git at all. A clone shows git's progress and carries on if the phone that started it drops. Unlisting a project never touches its files. The desktop app's sidebar shows it.
  • —Worktrees on any branch. A worktree for an existing branch, a new one, or a branch only the remote has (created tracking it), under .caprock-worktrees; one with any change is never removed.
  • —Shell tabs, engine side. POST /v1/shells starts your login shell in a project, held like an agent session so it survives a Caprock restart. A shell is in no total, card or count.
  • —From the phone, a controller can add, create and clone a project under home and make or remove a worktree. Shells stay on the machine for now.
  • —api_level in GET /v1/status and runtime.json, an integer raised by every API change a client must know about, so the desktop app (and later a bundled UI) can tell whether a daemon is new enough without parsing its version. runtime.json also records the daemon's executable path.
  • —CAPROCK_SERVICE_LABEL runs caprock service install|uninstall|status under another label, so a second, isolated install (a preview, a test) never touches the real dev.caprock.daemon login agent.
  • —The desktop app's shell in app/ (Tauri v2): it finds or starts the daemon and opens its dashboard in a native window. Built and tested in CI on macOS, Windows and Linux; make app runs it from a checkout.
  • —Phone alerts are off until you switch them on. A Telegram bot saved for the weekly report started sending alerts by itself; now Settings → Phone alerts turns each kind on.
  • —A terminal comes back from a dropped connection exactly where it was. When Caprock restarts, the network drops or a phone wakes, the terminal now picks up from the last byte it had instead of clearing and repainting, so what you scrolled back to stays put, and whatever you typed meanwhile is sent once it reconnects — never lost, never typed twice, even across a Caprock restart. It keeps trying for as long as it takes (it used to give up after a minute and ask for a reload), and a pill over the terminal says Reconnecting — your keys will be sent instead of a line written into it. Up to 4 KB of typing is kept while offline.
  • —Lists no longer jump while you read them. Scrolled down the session timeline or the live activity feed, new rows arriving above keep your place, in Safari as in Chrome, and a pill says how many arrived.
  • —Session cards tell same-named folders apart and say ended when ended. Two folders both called caprock now read Downloads/caprock and dev/caprock — the shortest path that tells them apart — on Now, the pulse and the app sidebar; the card leads with the session's title or first prompt, the short id after it. A folder that is not a repository, or a detached checkout, no longer shows HEAD as its branch. A session the daemon lists as ended never shows as idle.
  • —Gemini CLI is no longer labelled "your own key" in the agent picker. It also signs in with Google, so the note was wrong for most people.
v0.74.02026-10-05

Phone alerts say which session and what it did

  • —The first line, the one a lock screen shows, is what happened and the session's title — ✅ Finished · Fix the cache test, ⏳ Needs approval · …. Then the folder as ~/dev/caprock (not just caprock) and the branch; for a dialog, the tool and the command or file it asks about; for a finished run, how long it worked since your last prompt, what it cost, its tool calls and the files it changed. A turn that stopped on an error (a rate limit, an overload) now alerts too and says why. Settings → Phone alerts has a new switch, Include the last reply's first line, on by default. Telegram can read all of it; code and tool output are never sent (ADR-036, amended).
v0.73.12026-10-05

A tunnel on the Mac no longer hands out the owner's rights

  • —Caprock trusted any request that reached it over 127.0.0.1, so cloudflared, ngrok, Caddy, tailscale serve or ssh -R running on the machine let anyone who reached the tunnel read everything and start commands, with no pairing. A loopback request that carries a proxy header (X-Forwarded-*, Forwarded, Via, X-Real-IP, CF-*, Tailscale-*, Ngrok-*, …) or names a host other than localhost or a loopback address is now treated as a device: it needs a paired device's token and gets that device's role, and without one it gets 401 — on the API and on both WebSockets. The CLI, the hook shim, the statusline, the dashboard and an ssh -L forward are unchanged.
v0.73.02026-10-05

Remove a session from Caprock

  • —Remove from Caprock on a session's page (on the machine, never on a paired phone) takes the session, its events and its cost out of Recent projects, every total, Week, Lifetime and exports, after a confirm that names the cost. caprock sessions rm <id> or --cwd-prefix <dir> does the same in bulk; it only lists what it would remove until you add --yes. A removed session is never read back in from its transcript. A session still running is skipped. It cannot be undone.
  • —Review a session's changes on the phone. The session page's Changes tab has 44px file rows, an open file's name stays pinned at the top while its diff scrolls, next moves to the following file, and lines wrap on a phone (wrap switches it; a desktop scrolls them sideways inside the diff, never the page). Added and removed lines are tinted in both themes, line numbers take one narrow column on a phone and two on a desktop, and a big diff is drawn 400 lines at a time with very long lines cut until asked for. From a phone that controls sessions, tap a line and Ask the agent opens the terminal with "In <file> around line N: " already typed.
  • —A roomier phone terminal input. The field under a phone's terminal is full width and grows with what you type (up to five lines, then scrolls), with Photo and Send beneath it; the Esc/Tab/arrow/Ctrl+C keys sit in two rows of three so their labels fit at 320px.
  • —A permission prompt's buttons survive a Caprock restart. A session Caprock started keeps running through an upgrade, and so does the dialog on its screen; its Yes / No buttons now come back with it instead of waiting for the next prompt. A prompt that was answered meanwhile does not come back.
v0.72.02026-10-05

Answer Claude Code's permission prompts with a button

  • —When a session you started in Caprock asks before running a command or editing a file, its page shows the command or file with Yes, Claude Code's own second option (don't ask again, allow all edits this session) and No — on the phone and on the machine. Claude Code sessions only; Codex and OpenCode prompts are still answered from the keys bar. Uses a new hook, so the dashboard asks once to install it.
  • —Attach a photo from the phone. Photo beside the phone's terminal input takes a picture or picks one from the library, shrinks it to 2048 px and types its path into the session.
  • —Phone alerts through your own Telegram bot, free. Settings → Phone alerts sends a message when a session is waiting for approval (at once) or has finished (after a minute with nothing new) — every Claude Code session on the machine, each kind with its own switch, both on once a bot is set up. A message names the project, what happened and the agent, with a link to the session when phone access is on; never code, prompts, replies, file names or commands. One message per question, at most one of a kind per session every three minutes and 20 an hour. Send a test alert checks the bot. The bot is set up in Settings now, so it no longer needs the weekly report's licence.
  • —Now says when a session is waiting for approval. A session at a permission dialog — any session, not only Caprock's — reads waiting for approval with the call it asks about and the waiting on you badge, instead of looking busy.
  • —Caprock on your phone's home screen. Add it from Safari or Chrome and it opens full-screen with its own icon, in your theme, clear of the notch and the home bar. On an iPhone the home-screen app pairs once more — type the code, since a scan opens Safari — and shows up as "iPhone · home screen".
  • —A paired phone opens Week, At a glance and the tool drill-down. They were left out of what a paired device may read and answered 403; the drill's Premium half stays gated as on the machine.
  • —"Hooks need an update" when only some are missing. After an upgrade adds a hook event, Now said Hooks not installed for one missing event; it now says the hooks need an update, with an Update hooks button.
  • —Codex edits count as edits in What it went on. Codex runs its tools through one exec wrapper, so every patch it applied counted as a command. Each call is now placed by the tool it actually ran — a patch is an edit, a web search is web — including calls already stored. On one machine 30-day editing went from 0.4% to 4.8% of spend. The Tools drill-down for exec groups by the command or tool inside the script instead of showing almost every call as const.
  • —caprock service install no longer leaves the daemon stopped on macOS. It booted the old service out and bootstrapped the new one at once; launchd was still tearing the old one down, the bootstrap failed with "5: Input/output error", and nothing was running. It now waits for the old one to go and retries briefly.
  • —"/" is no longer offered as a recent project. The New session picker leaves out the filesystem root, your home directory and the temp directory itself, so something on the machine running claude from "/" every few minutes no longer sits at the top of the list. Folders inside them are offered as before.
  • —No more "handoff lookup failed … context canceled" in the log. A session opening in "/", your home directory or the temp directory gets no handoff and runs no lookup. The lookup itself now skips other projects' passages on the index instead of reading each one: a project with nothing to hand over took 4-6 s on a 1.2 GB database, longer than Claude Code waits for the hook, and now takes under 0.1 s.
  • —The share dialog's all-time card is ready when you open it. The daemon works out the all-time and 30-day Weeks when it starts, and reads the tool calls behind them in parallel slices: all time went from 5.0 s to 1.9 s on a copy of a 1.2 GB database, with the same figures.
  • —A long command no longer pushes a session page sideways on a phone. *waiting for approval — running …* (or *running …*) with a long unbroken path or file name was wider than a 320–360 px screen, so the whole page scrolled sideways. It now wraps.
v0.71.02026-10-04

Your phone as a controller

  • —In settings, Let it control sessions beside a paired phone lets it start a session in any folder under your home directory — browse to it, or make a new one — picking Claude Code, Codex or OpenCode and any permission mode, type into it, answer approvals and stop it. Bypass asks once on the phone before it starts. The terminal on a phone has a large input field and Esc, Tab, ↑, ↓, Enter and Ctrl+C buttons. Take control away is one button and holds at once, even on a terminal already open. Phones stay view-only until you choose; settings and pairing stay on the machine, and so do custom commands and folders outside home. It works over your Wi-Fi or Tailscale, with no relay. The paired-devices list tells two phones of one model apart by browser, id and when each was paired.
  • —OpenCode tool calls keep their input. A bash call's command, a fetch's URL, a search's pattern and an edit's strings were not stored, so every OpenCode bash call in the Tools drill-down read "(input not recorded)". They are stored now, in Claude Code's spelling, and calls already imported are filled in from OpenCode's own database the first time Caprock starts.
  • —A worker with unread mail is kept working again. The Stop hook's "process your inbox first" answer was nested where Claude Code ignores it, so sessions stopped with mail still waiting. It is now in the shape Claude Code acts on.
  • —Now and a session's page fit a 390px phone. Banners, the Today panel's agent switch, the breakdown bars, the donut legends and the session header ran past the right edge and dragged the whole page sideways.
  • —"What it went on" counts tool calls again. Most tool calls were never matched to the turn that paid for them, so their cost showed as "no tool call" (93% of 30-day spend on the owner's machine). With hooks installed, Claude Code's message id was dropped along with the transcript's copy of a call the hook had already stored. Codex and DeepSeek calls were never linked at all. All three are linked now, and calls already stored are repaired the first time Caprock starts: on a copy of the owner's database, unmatched calls went from 31,275 to 7 and "no tool call" from 93% to 5%. Codex's and DeepSeek's own tool names (exec, js, apply_patch, bash, read, edit…) are now counted as commands, edits and reads instead of "other".
v0.70.12026-10-04

Install hooks from the dashboard

  • —The "Hooks not installed" banner on Now has an Install hooks button that does what caprock hooks install does and then says what changed: sessions started from now on report live; ones already running keep their hooks until restarted. It names the settings file it checked, and it can be dismissed until the missing set changes.
  • —Open a Live pulse row in place. Clicking a project's row shows today's cost and tokens for it, every live session in it with agent, model, state and what it is doing, and the terminal, Open repo and Open session buttons. Open rows are remembered.
  • —Settings is plain now, and opening Caprock on your phone is three steps. The screen behind the header's link (now called settings) is short sections with plain titles: Open Caprock on your phone, Plan & licence, Appearance, Privacy, Memory between sessions, Storage, About this install. Each option says in one line what happens when it is on. - Phone: one button, Show a code, turns network access on and shows a QR code with the six digits in big type beside it. Scan it with the phone's camera and the phone pairs by itself, then appears in the list with a Remove button. The corner says the state at a glance — Off, Waiting for your phone…, 1 phone connected. The QR code is drawn in the browser; nothing is sent anywhere to render it. Cancel withdraws the code on the machine, not just from the screen. Phones paired before stay listed while network access is off and can be removed then too. - Plan: set it right there — the same choices as the header's picker. - Theme: Dark or Light, and Paper or White, from Appearance; the header toggle and this control stay in step. - Memory: the setting says what Claude is handed; the "measure whether it helps" experiment is under Experiments. - Storage and the install's own figures keep their detail one click down, under Details.
  • —The header wraps on a phone. It was one row 882px wide, so every screen on a paired phone scrolled sideways.
  • —Plan limits read at a glance. The ring's filled arc is what is used, coloured by level, on an empty neutral track; a forecast is one plain sentence — "At this pace you'll hit 100% around 19:10 — about 1 h before it resets." — amber unless it is under half an hour away, not a red dashed ring. The explainer is one line, with the details behind "more".
  • —Agents are easier to read. At a glance's agent tiles are bigger and their ring says "of spend"; the Live pulse row shows its state as a coloured pill and its agents as "4 agents" with the model at full contrast.
  • —A session you start in Caprock survives Caprock restarting. Upgrading Caprock, launchctl kickstart, caprock down and up: each used to end the sessions started from the dashboard. Each session's terminal now lives in a small process of its own (caprock pty-host), and the restarted Caprock picks it back up — the terminal shows what was on screen and takes typing again, and stop, pause and kill work as before. caprock down stops Caprock, not your sessions; end a session from its page or by exiting it. Sessions started by an earlier release are still ended by the upgrade that installs this one: their terminal is inside the old Caprock process.
  • —Autostart keeps sessions alive across service restarts. The launchd agent sets AbandonProcessGroup and the systemd unit KillMode=process. An existing autostart keeps its old file until you run caprock service install; caprock service status says when the file differs. On Linux this step is needed for sessions to survive a service restart.
  • —A session whose terminal closed says so plainly. For a session started by an earlier release, the Terminal tab now reads "This session's terminal was closed when Caprock restarted. The conversation is saved — Continue it here resumes it in a new terminal." Continue it here carries the same conversation on (stopping the old terminal-less process first, which Caprock started); open a copy instead leaves that process running.
  • —A carrier-NAT address is no longer offered as Tailscale. Network access admitted any address in 100.64.0.0/10 as a Tailscale one. That block is carrier-grade NAT, which Tailscale borrows: a laptop tethered to a phone or behind an ISP's CGNAT can hold such an address on its Wi-Fi, and binding it invites everyone behind the same carrier. The range now counts only on Tailscale's own interface — named tailscale*, or carrying Tailscale's IPv6 prefix fd7a:115c:a1e0::/48, which is how it looks on macOS.
  • —Open a session in your own terminal. The session's title row (next to Open repo, on every tab) and the Projects terminal menu carry Open in Ghostty ↗ (named for your terminal), which runs the agent's own resume command — claude --resume, codex resume, opencode --session — in the session's folder, in Ghostty, iTerm2, Terminal, WezTerm, kitty or Warp on macOS, your $TERMINAL or a common terminal on Linux, and Windows Terminal, PowerShell or cmd on Windows. Caprock keeps measuring it through hooks and the transcript, as it does any session you start yourself. A session Caprock is running can be moved (stopped here, carried on there) or forked; one running somewhere else can only be forked, because Caprock never stops a process it did not start. Pick another terminal from the ▾ menu and it becomes the default.
v0.70.02026-10-04

Open the repository from a session or a project

  • —The session header links the repository — the branch's page where the host has one — and the pull requests the session opened, newest first, with "merged" only where a merge was recorded. A project row links its repository and its last pull request. The address is read from the local git remote (GitHub, GitLab, Bitbucket, Codeberg and other hosts; ssh or https); nothing is asked of the host.
  • —Screens show their last figures while the new ones load. Now, Lifetime, History and a session's header open with what this browser last saw, marked "updated N min ago · refreshing", instead of dashes. Nothing about the present — health, burn, controls — is drawn from the kept copy.
  • —Start Codex and OpenCode sessions from Caprock. The New session dialog has an agent picker — Claude Code, Codex, OpenCode, Gemini CLI — showing only the ones installed on this machine, and it remembers the one you picked last. Each runs as its own TUI in the terminal tab, with its own model choice (Codex's own model list; OpenCode's provider/model) and the permission choice translated into that agent's flags. Cost, turns and answers land on the same session page as the terminal, and an ended Codex or OpenCode session can be continued with "continue here".
  • —Week: a card of what your agents shipped, to post. A new Week tab draws seven days of this machine's work — pull requests opened and merged, commits, files and ≈lines written, what it cost at list price, ≈cost per merged PR, the share spent re-reading context, who did what (Claude Code, its subagents, Codex, OpenCode, Gemini CLI, DeepSeek), the longest loop and the biggest session — as a 1200×675 or 1080×1350 card in your theme. Download it as a PNG or copy it to paste. Everything is counted on your machine from the agents' own tool calls; Caprock does not ask GitHub, and no repository, path, prompt or session title is ever on the card.
  • —Now: At a glance. A collapsible block after All time draws where the money went, the bill by token type (with the share spent re-reading context in the middle) and the most-used tools as donuts, plus one tile per agent with its share of all-time cost. A Charts | Numbers switch shows the same figures as tables; both choices are remembered.
  • —Live pulse says who is working and on what. Each session row shows the current model by name and, when subagents are running, a crowd icon with "×N".
  • —Share: a Story card, for any period. The share dialog gains a second style beside Figures: the Week card — the headline, the money beside it, who did what, the longest loop and the biggest session — for today, this week, this month or all time, at 1200×675 or 1080×1350.
  • —Click a tool for what it did. On Lifetime, a tool's row opens in place: Bash by the command each call ran ("go test", "git diff"), Read and Edit by file, WebFetch by domain, WebSearch by query, an MCP tool by its action, with calls and share. Premium adds what came back, how often each group failed, a trend per group and hints drawn only from those counts.
  • —A project's terminal is a button you can see. Each Projects row has a full-size button that says what it does: Terminal opens the session Caprock is running there, Pick up in a terminal branches one started elsewhere, New session here starts one. The full menu stays behind ⋯. It replaces a small >_ at the row's edge.
  • —The Teams and Premium dialogs show what you get before they describe it. Each now opens with a small picture in the Week card's style — an example team's week for Teams; the daily cap stopping at its line, a question to Gemini about your own sessions, or the Monday report for Premium — clearly tagged as an example, followed by a few short lines and the two ways forward. Premium's prices are the site's: $30 a year, or $100 once marked Best value, in the product's amber rather than blue. Focus now moves into the dialog and back to what opened it.
  • —Plan limits are gauges, directly under Today. Each window is a ring that turns amber at 60% and red above 85%, with its reset time and a countdown, grouped by agent. A forecast still appears only when your pace would hit the limit before the reset, and is drawn on the ring; Codex is never forecast and says how old its reading is. A stale reading is drawn grey and says so.
  • —Plan limits explain themselves. Each window says, in words, how much is used, when it resets and how long that is (in minutes under an hour), and that Claude Code pauses at 100% until then. One line says these are Anthropic's limits per 5-hour window and per week, not Caprock's, and near the limit a note says to wait for the reset or switch to Codex if it has room. The same panel is on Now and Cost, and the limit alert now opens it directly.
  • —Lifetime reads like the site. The money leads, large, with what it came to per active day and per session beside it, then the cache and the counts. Top projects come first, as a donut beside the table; tool usage and the model mix are donuts with the rest grouped as "other", or tables, by a Charts | Numbers switch.
  • —The share dialog shows a card at once. It draws the last figures for the period while the current ones load, and with nothing kept it shows the card's outline with each range ticking off as it answers, instead of an empty box. Saving or sending always uses the current figures.
  • —Continue a session in another agent. "Continue in…" on a session page starts a new Claude Code, Codex, OpenCode or Gemini CLI session in the same folder, with a summary as its first message: the last thing the agent said, what has changed in the working tree, and the pull requests it opened. You read and edit the summary before anything is sent, and both sessions link to each other. It is a new session with a summary, not the same conversation — which also makes it the way to pick up a session whose transcript is gone.
  • —Opening a session's terminal no longer freezes the page. The terminal set up its GPU renderer before it connected, and connected only after the session's figures had loaded; the figures were then reloaded on every event. It now connects and takes the keyboard first, switches to the GPU renderer once you pause, and reloads the figures every 30 seconds while you type. Keys typed the moment it opens reach the session.
  • —A terminal that is starting says so. A resumed session can take half a minute to print anything; the panel showed black until then. It now counts the seconds, and after 30 seconds with nothing, or if the session closes first, says so and offers Retry.
  • —A session's Answers tab shows what Claude wrote again. Claude Code now writes a thinking line ahead of every reply, and Caprock kept only the first line of each reply — so the prose after it was never stored, and a session of long answers showed one line. The rest of a reply is now kept with it, and replies stored since the change are filled from the transcripts on disk once, in the background, after an upgrade. Memory search and the handoff a new session receives read the same prose and are mended with it.
  • —A long link in an answer no longer widens the page. One unbroken URL in a reply made the Answers tab scroll sideways and pushed the session's buttons off the screen; it now wraps.
  • —The Answers tab opens at once. It read every reply on the machine to find one session's; on a 1 GB database that was 3.3 seconds cold for a session of 21 replies, and is 20 ms now.
  • —The terminal is dark in the light theme. Claude Code's dim text and status line were grey on paper and could not be read. The terminal now keeps its dark palette in both themes, and keeps it when you switch theme with a terminal open — it used to stay in the colours of whichever theme was on when it opened.
  • —Now waits less for its figures. The status the screen polls counted every event and searched a fortnight of prose on each call; both are reused for a minute now.
  • —Assistant turns are no longer lost when the database is busy. Since 0.68.0 the recorder checks whether a forked session's turn was already paid for before storing it, and that read-then-write transaction failed at once with "database is locked" whenever another write was in progress — about 400 times an hour on a busy machine. The transcript reader logged the error and moved past the line, so the turn never arrived. Transactions now take the write lock up front and wait for it, a line that still meets a busy database is read again instead of skipped, and on first start the transcripts written since 3 October are read once more to recover what was dropped (on the owner's machine: 259 turns, $18.52).
  • —Every page loads fast on a large database. On a 1 GB database with sessions running, the main screen's requests took up to 3.6 seconds each. All-time history and the 7-day, 30-day and all-time summaries are now computed in the background and served from memory, refreshed as they are asked for; the whole database file is memory-mapped, so a read after a write no longer starts cold; and the sessions list finds each session's first prompt through its own index. The same requests now answer in under 0.1 seconds, and the share dialog's preview no longer waits on four whole-history scans. Today's figures are still computed on every request.
  • —A session with no activity yet no longer reads "idle 739892d ago".
v0.69.02026-10-04

Gemini 3.8 Flash and Gemini 3.1 Flash-Lite are priced

  • —A Gemini CLI session on either model had its tokens counted but no cost, and the Cost screen named the model as unpriced. Both now carry Google's list prices, read 3 October 2026 (pricing table 2026-10-03.1). Gemini 3.8 Flash is on introductory pricing that doubles on 1 January 2027; the table will be refreshed before then.
  • —Documents dropped into a session's terminal arrive. Dragging a Markdown, CSV, JSON, YAML, log or source file from Finder printed "415 Unsupported Media Type", because the browser gives those files no type and only images, PDF and plain text were accepted by type. Files are now accepted by extension — images, PDF, text and documents, Office files, data and source files — every dropped file is attached, not only the first, and each keeps its own name, so Claude reads contract.pdf rather than a timestamp. A file that is refused says which types are accepted.
  • —The README no longer lists starting a Gemini CLI session as Premium. It never was: the New Session dialog offers Gemini CLI to everyone, like every other agent. Asking Gemini about your numbers is the paid part.
v0.68.02026-10-03

The light theme is paper now, like caprock.dev

  • —A cream ground with panels a shade lighter, instead of white. Prefer white? Settings → Light theme → White; the choice is kept in this browser. The dark theme is unchanged.
  • —Test readings no longer sit in your plan-limit history. A unit test posted a made-up 23.5% five-hour reading, with a reset in 2030, to whatever Caprock daemon was running when the tests ran. Any such reading is now removed, by the same rule the daemon already applies to new ones: a reset more than eight days ahead is not a real window.
  • —A forked session no longer counts its parent's spend a second time. Picking up a running session forks it, and Claude Code copies the conversation since its last compaction into the fork's transcript, usage included. Caprock priced every copied turn again, so the copied spend showed up twice in Cost, the daily totals and caprock report. On the machine this was found on, one fork added $212.93. A copied turn now stays in the fork's timeline with its text but costs nothing, and the copies already stored are corrected once when the daemon starts.
  • —A copied turn with no usage is no longer charged for a cache write. Claude Code blanks the token counts of some messages it carries past a compaction but leaves the 1-hour cache-write figure, which was priced on its own.
v0.67.12026-10-03

Codex sessions older than a week are read again

  • —Codex compresses a session it has not touched for seven days into a .jsonl.zst file and removes the original. Caprock read only the uncompressed files, so a new install or a rebuilt history saw one week of Codex and nothing before it. Compressed sessions are now read like the rest.
  • —The Team seat preset is $25, not $30. Picking "Team seat" as your plan priced it at $30 a month; Anthropic lists a standard seat at $25 billed monthly ($20 annually), so Plan value understated what the usage was worth against it.
v0.67.02026-10-03

What Caprock for Teams is, without leaving the dashboard

  • —The Cost screen has a one-line banner that starts from this machine's spend and project count. It, the Projects team line and the footer link open a dialog that says what the team version is, what a lead gets, and why it is worth paying for, with links to book a demo, the team page and pricing. The Cost screen shows this instead of the Premium banner, which stays on Now and Lifetime. "Not now" hides it for a month.
v0.66.02026-10-03

1-click terminal

  • —Each project on the Projects panel has a >_ button. When one session Caprock started is running in that project, a click opens its terminal. Otherwise you choose: open one of Caprock's sessions, continue an ended one, branch one still running in another terminal, or start a new session in the project's folder. A running session card on Now opens straight in its terminal too. Caprock still never types into a session it did not start; continuing or branching starts a second process.
v0.65.02026-10-02

What you ask Gemini CLI is kept, like every other agent's prompt

  • —Gemini sessions started from Caprock now record the question, so Memory can find a session by what you asked it. Sessions from before keep none.
  • —OpenCode replies can be found by the question you asked. What you type in OpenCode is now stored like a Claude Code prompt, so Memory search finds a reply by your own words, and an untitled session is described by its first question. Sessions read before are filled in on the first start. Text OpenCode or another app adds to a message on its own is left out.
  • —A DeepSeek Harness session's timeline shows what you asked. The prompt rows were blank, because DeepSeek keeps the prompt under a different name.
  • —A session shows the model it runs on, not its subagent's. When a session's subagent ran on a different model — an Opus session sending an Explore task to Haiku — the session took the subagent's model on its card, in the session lists and in the context window it was measured against. The subagent's turns still count toward the session's and the day's cost, by their own model. Sessions already stored with the wrong model are corrected once, on the next start.
  • —Context fill no longer drops while a subagent works. The badge measured the session's last turn, and a subagent's last turn starts from a small prompt of its own, so a nearly full context read nearly empty until the subagent finished. It now measures the session's own last turn.
  • —Memory search finds the reply to a question asked earlier in time. A reply was matched to whichever prompt Caprock happened to store just before it, so a question stored later than its answer — all of OpenCode's history, a few Claude Code exchanges — never found it, and some replies were tied to an older question than the one they answered.
  • —DeepSeek sessions no longer list injected instructions as your prompts. The AGENTS.md, CLAUDE.md and runtime-context blocks stored before this was fixed are removed once, checked against DeepSeek's own transcripts.
v0.64.02026-10-01

The weekly report's bot token no longer reaches the log or the screen

  • —When Telegram could not be reached, the error quoted the request URL, and the URL contains the token; that text went to the daemon log and to the report's last-error line in Settings. The error now names the cause only.
  • —Memory shows what OpenCode wrote. OpenCode's replies are now stored like Claude Code's, so they appear in Memory and its search, history included — the first start fills in sessions imported before. A subagent's replies stay out, as Claude Code's do.
  • —What Codex wrote back is in Memory. Codex replies now appear on the Memory screen, in its search and in a session's Answers tab, the same way Claude Code's do. Replies from sessions Caprock had already read are filled in once, in the background, from the transcripts Codex keeps. Sessions Codex imported from Claude Code are left out, because those answers are already there from Claude Code itself, and so is the chatter of Codex's own subagents.
  • —See how much disk Caprock uses, and on what. The Status screen has a Storage panel: the total on disk, what it is made of (events, their indexes, logs and the rest), which agent and which kind of event recorded the most, roughly how much it grows a day, and what you can do about it — how much a retention setting would delete, and whether compacting the database would give anything back.
  • —The Tasks screen says who runs a task. It read "one Claude session per task"; an orchestrator hands each task to a worker, and the screen now says so. The README's pairing steps now say a code appears after Show a code.
  • —Screenshots no longer print a folder name the scrubber missed. Two folders with the same name were numbered <name>-2, and the numbered one kept its real name; the scrubber now rewrites it and refuses to finish if any real project name is left.
  • —Sessions Codex imported from Claude Code no longer count as Codex spend. Codex Desktop can import a Claude Code session, and Caprock read the copy as Codex work: the same tokens counted twice, the second time at OpenAI's prices. Those threads are now skipped, and the first start takes what was already stored back out of every total. On one machine that was $26.70.
  • —Work done by Codex subagents is no longer lost. A subagent's turns could be mistaken for its parent session's own and dropped, along with their cost; on one machine nearly half of them were missing. They are now kept, shown as part of the session that started them, and the first start brings back the ones dropped before.
  • —Opus 5.5 and Sonnet 5 are priced right, including the turns already stored. Opus 5.5 had no row and was priced as Opus 5 — $5/$25 instead of $4/$20, and cache reads at $0.50 instead of $0.20. Sonnet 5's rise to $3/$15 on 31 August was cancelled, but Caprock priced every Sonnet 5 turn from that day at it. Pricing table 2026-10-01.1 adds Opus 5.5 and Sonnet 5.5, keeps Sonnet 5 at $2/$10, and lists both as corrections: on the first start the daemon reprices those turns once and moves the session and daily totals by the difference. A price that really changed still never restates history.
  • —A DeepSeek Harness reply can be found by the question you asked. Memory search looked for the prompt where Claude Code keeps it, and DeepSeek keeps it elsewhere, so no DeepSeek question matched. The instructions DeepSeek injects into a session are no longer taken for something you typed.
  • —A session whose folder is gone offers a command that works. The copy command began with cd into the missing folder, so it failed before claude started. claude --resume <id> finds the session from any folder, so that is what is offered now, while the transcript is still there.
v0.63.02026-10-01

A paired tablet or phone can read Caprock and no longer control it

  • —The device token was the only check, so a paired device could start a command, type into a session, pause or kill it, change settings and the licence, create and approve tasks, start the orchestrator and turn the task runner on — everything the dashboard on the machine can do, although ADR-029 says a tablet is somewhere to read figures, not a second control room. A paired device may now make only a named list of reads; anything else is refused with 403, including the terminal socket and the folder listings, and a route added later is closed to it until it is named. It no longer receives the licence key with the settings. The tablet's dashboard stops drawing the controls it cannot use. Nothing changes on the machine itself.
  • —Codex's plan limits on the Cost screen. The 5-hour and weekly windows Codex writes into its session transcripts now sit in the Plan limits panel under a Codex heading, beside Claude Code's, with the time Codex wrote them. Only the windows your plan has appear — some ChatGPT plans have no 5-hour window — and they are shown as measured, never forecast. /v1/stats/summary carries them as codex_rate_limits.
  • —Codex sessions outside ~/.codex/sessions are imported. Caprock now reads $CODEX_HOME when it is set, and archived_sessions/ as well as sessions/, so archived threads and a relocated Codex home are no longer missing from the totals. A thread archived after it was imported is not counted twice.
  • —The Cost screen no longer says "At Anthropic list prices" above a model table holding GPT rows; it says "At API list prices". caprock report names every vendor whose prices make up its total ("Anthropic and OpenAI list prices") and stops calling a mixed total "Claude Code".
  • —caprock statusline install no longer sends you in a circle. With a status line of your own already set, it said "run caprock statusline install" — the command you had just run. It now names the settings file and the exact statusLine.command value to set, and still changes nothing.
v0.62.02026-10-01

caprock export writes Caprock's record out as TSV, CSV or JSON lines

  • —events (one row per turn, tool call or lifecycle event) or sessions, with the same columns whichever of the five agents wrote the source. Filters by --since and --agent; --payload adds the raw source record to JSON lines. It reads the database read-only and works with the daemon stopped. The column sets are documented in docs/schema.md and only ever grow at the end (FB-041).
v0.61.02026-09-30

The memory handoff can be measured

  • —Ticking Measure whether it helps holds it back from one new session in four, at random, and the Memory screen compares the two groups: minutes and tool calls to a session's first edit. Off by default. On the owner's machine a session continuing yesterday's work reached its first edit no faster than one starting cold, with the handoff already on — this is how to find out whether it does anything.
v0.60.12026-09-30

Codex sessions on gpt-6-astra and gpt-6.1-sol are priced

  • —Codex sessions on gpt-6-astra and gpt-6.1-sol are priced (OpenAI list prices, read 2026-09-30). They showed as "Partial cost"; the turns already stored are priced on the next start. The gpt-6 rows also carry OpenAI's cache-write rate, which Codex does not report tokens for yet.
v0.60.02026-09-30

After a reboot or an OS update, the Now screen names the sessions…

  • —After a reboot or an OS update, the Now screen names the sessions that were still running when Caprock stopped, each with continue, until dismissed (FB-038). They used to be a few ended cards among a hundred, found by opening each.
  • —A /clear or a fork is one card, not several: the earlier parts of the conversation ride on the latest card as links (FB-039). Only recorded facts join sessions — the same process after /clear, or a fork Caprock started — never the same folder or branch.
  • —An ended card shows when the session was worked in — "worked 29 Sep 14:02–16:40" — not when it was closed. After a reboot every card read the reboot's time (FB-037), and the list is now ordered by the same time.
  • —A session from before a reboot is ended even when its old process number now belongs to another program. It stayed "live" instead, and its terminal tab was an empty screen.
  • —The terminal tab of a session Caprock started before it restarted says the terminal went with that run and offers continue, instead of an empty screen (FB-040).
v0.59.62026-09-30

Hook events are no longer lost when the database is busy

  • —The shim waits a second and hangs up, and an event still waiting for the write lock was abandoned with it — 41 on the owner's machine, each logged as "context canceled" and missing from the record. The write now finishes after the shim has gone.
  • —The log no longer grows without bound. A model missing from the pricing table was warned about on every event, re-reads included — 1.1M identical lines and a 4.6GB caprock.log. It is now said once per model, and the daemon's log is rotated at 64MB with one previous file kept; a log the daemon has stopped writing to is cut to its last 8MB.
  • —Restarting the daemon no longer re-reads every Codex transcript. On the owner's machine that was 1.1GB and 30k duplicate writes on every start, while the dashboard and hooks waited; only transcripts that changed are read now.
  • —CI checks the query plan of every aggregate the dashboard polls and fails when one would read full event rows instead of a covering index — the regression behind 0.59.2 and 0.59.3, which a small test database could never show by timing.
v0.59.52026-09-30

A session started from Caprock right after the daemon starts gets…

  • —A session started from Caprock right after the daemon starts gets your login shell's environment instead of the daemon's bare one. Reading it was cut off after 5 seconds, and at start — while every importer does its first pass — it could take longer; the failure left nothing to fall back on but launchd's PATH, so gcloud went missing again (FB-033). The limit is now 30 seconds, a failed first read retries on its own, and a session started meanwhile waits up to 15 seconds for it.
  • —caprock status says which environment sessions start with — your login shell's, still being read, or the daemon's with the reason — so a session that cannot find a tool a terminal can find is one line to diagnose.
v0.59.42026-09-30

Built with the modernc.org/sqlite 1.59 driver, and the dashboard with…

  • —Built with the modernc.org/sqlite 1.59 driver, and the dashboard with Vite 8.3.1. No change in behaviour; the summary and history timings were re-measured against the same database copy as 0.59.3 and match.
v0.59.32026-09-30

The Lifetime screen and the all-time figures on Now load in about a…

  • —The Lifetime screen and the all-time figures on Now load in about a second instead of nine. The indexes the spend aggregates read from stopped covering them when background usage was split out; migration 0026 rebuilds them (about ten seconds on a large database, once, on the first start). The History totals for today read off the time index like the summary's already do (4.8s to 0.02s on the owner's database).
  • —Rebuilding the Codex daily totals no longer fails on a busy first start with "database is locked"; it used to wait for the next restart.
v0.59.22026-09-30

Today's totals load in a fraction of a second again

  • —Two of the queries behind them walked every event in the database instead of today's, once the background-usage split added a column their indexes did not cover; on a 300k-event database the summary took 3.7s at rest and tens of seconds while a session was busy, and screens sat on "reading your figures…" or "loading…". Measured on a copy of the owner's database: today 3.7s → 0.5s, 7d 1.9s → 0.15s.
v0.59.12026-09-30

The dashboard no longer piles up requests while a session is busy

  • —Every live event made each panel ask again, whether or not its last request had answered, so on a large database today's totals fell 30-60 seconds behind and read "reading your figures…" while the daemon ran at over 100% CPU. A panel now waits for its running request and asks once more after it.
  • —make shots waits until every placeholder on a screen has cleared, and fails rather than capturing a half-loaded one.
v0.59.02026-09-30

Every session card says what the session was about

  • —Claude Code's, Codex's or OpenCode's own title for it (the name their resume pickers show), or else its first prompt that says something. Ended sessions can finally be told apart (FB-035). OpenCode sessions imported before this release get their titles too.
  • —Pricing for GPT-6 Sol and GPT-6 Luna, from OpenAI's published list prices (read 2026-09-29). Turns already stored without a price because their model was missing from the table are priced on the next start, at the rate in force when they ran; turns that were priced keep their figures.
  • —An ended session says whether it can be continued and, when it cannot, why: the transcript was deleted by Claude Code, the folder is gone, or the agent is not one Caprock resumes — with the command to resume it yourself where the agent has one (FB-036).
  • —Ended session cards say when the session ended and offer continue (or why not) right on the card. With ended sessions shown, a search box finds a session by its title, any prompt typed in it, its project or branch, and show more reaches past the first 200.
  • —A session started from Caprock can now reach what a terminal can: gcloud, bq, Homebrew tools and everything the shell profile exports. Sessions and verification commands used to inherit the daemon's environment, which under launchd autostart is a bare PATH=/usr/bin:/bin:/usr/sbin:/sbin; they now start with the user's login-shell environment, resolved in the background at start and refreshed in the background once it is more than a minute old (FB-033). Caprock runs your $SHELL -l -i to read that environment; the shell sees CAPROCK_RESOLVING_ENVIRONMENT=1, so a profile can skip interactive-only work. Windows is unchanged.
  • —A session started from Caprock no longer inherits Claude Code's session markers when the daemon itself was started from inside a Claude Code session.
  • —Pasting into the terminal sends the text once. Cmd+V (Ctrl+Shift+V off macOS) used to paste twice whenever the browser allowed a clipboard read, so short text appeared doubled (FB-034).
  • —A long paste collapses into Claude Code's "[Pasted text]" block again after switching tabs or reloading. A terminal that attached after the session's first 256 KB of output never learned that bracketed paste was on; the daemon now restores the terminal modes the session set before replaying scrollback.
  • —Sessions Caprock started can be continued after they end; the button used to disappear for them, including after they had been continued once.
  • —Continuing a session whose transcript is gone is refused with the reason, instead of opening a terminal that printed "No conversation found" and exited.
  • —Daily totals no longer count some Codex turns twice. Migrations 0022 and 0023 re-imported every Codex event without taking the first import out of the daily table; the Codex rows are rebuilt from events once on start (on the owner's machine, five days $3.68 and 7.9M tokens too high). A Codex turn repriced after its model became readable now moves the session and daily totals with it.
  • —Pasting an image with the keyboard uploads it and types its path; only drag-and-drop worked before, because xterm stopped the paste event from reaching our handler.
v0.58.32026-09-13

Refreshed the shipped UI's React runtime, React types, and Node…

  • —Refreshed the shipped UI's React runtime, React types, and Node types; the embedded production bundle is rebuilt and verified across all CI platforms.
v0.58.22026-09-13

Refreshed the SQLite and Go system dependencies and the pinned GitHub…

  • —Refreshed the SQLite and Go system dependencies and the pinned GitHub Actions used by CI, with the full cross-platform verification matrix still green.
  • —Dependabot updates are grouped by ecosystem and capped at three open PRs; npm major updates remain separate for deliberate review.
  • —The README pulse animation now advances every 60 ms instead of 250 ms, so a complete dark or light cycle takes about 5.5 seconds and reads as live motion.
v0.58.12026-09-12

Checked-in DeepSeek Harness transcript fixtures cover current and…

  • —Checked-in DeepSeek Harness transcript fixtures cover current and legacy JSONL formats without depending on a developer's ~/.dsh directory.
  • —Dependabot now watches GitHub Actions, Go modules, and ui/ npm dependencies.
  • —GitHub Actions are pinned to immutable commit SHAs with readable release comments, reducing supply-chain exposure while keeping updates automated.
  • —The incomplete-cost warning now says Partial cost, offers copy/report/why actions, and explains the lower-bound total only on demand. No manual price entry is required.
v0.58.02026-09-12

DeepSeek Harness is the fifth agent on the same screen

  • —Caprock now reads DSH's own session transcripts (~/.dsh/sessions/…, zstd JSONL) and shows its sessions, turns, tool calls and cost beside Claude Code, OpenCode, Gemini and Codex. The model DSH names — deepseek-v4-pro — was already in the pricing table, so no pricing change was needed. Observation-only, like Codex: Caprock reads the files and does not start, steer or stop DSH sessions.
v0.57.02026-09-11

The default port moved from 4173 to 22776

  • —4173 is Vite Preview's default, so a developer running vite preview beside Caprock hit a bind collision. A fresh install now listens on 22776 (unassigned by IANA, spells CAPRO on a phone keypad); an existing install that never wrote config.json keeps its old 4173 origin, because a port is part of a browser origin and moving it would strand bookmarks and LAN-pairing tokens.
  • —Codex's own approval reviewer was reported as user work. codex-auto-review is Codex's hidden "Automatic approval review model", and it writes a normal transcript that raised an unresolvable unpriced-cost warning (OpenAI publishes no price for the id). It is now classified as background machinery, kept out of every user-work total, and reported beside them as a quiet "background usage" line with no dollar value. A genuinely unknown model still gets the actionable "partial estimate" with a prefilled report issue.
v0.56.02026-09-09

The context tax, in the screens you already look at

  • —Every call re-sends the whole conversation as a cache read before it does anything, and Caprock stored every number needed to price that and showed none of them. It now does, in three places and no new ones: the session card's Context caption says what the next call costs at the current fill ($0.07/call), the lifetime Breakdown carries a Context tax row with its dollars and its share of what was priced, and a loop alert states what the repeated calls paid to re-read the conversation. On the author's own archive that row reads $11,117 of $13,910 — 80%.
  • —The recording stand outlived its terminal, forever. Eight of them were found still running, the oldest two weeks old, each holding a daemon and a temp directory open. sleep 3600 in a hold loop is a separate process, and bash will not run a trap until it returns — so a stand started detached never received the interrupt that would have stopped it. It waits on an interruptible sleep now. Affects make record only, not the daemon.
v0.55.02026-09-06

More words for a cache hit rate, and a shorter cost on the status line

  • —98% and 99% were both "outstanding"; there are now excellent (98–99) and very good (96–98) between it and good, which moves down to 90. And the status line's cost read $169.410 — a trailing zero on a line that reprints after every message. It now shows two decimals, matching the dashboard, and four only below a cent: one decimal was the obvious cut, but 92 of 250 real sessions round to zero at one place, and a session that cost money should never report nothing.
  • —The waiting banner said the same thing twice. It read Waiting on you · caprock · waiting for you — one sentence in two prepositions, the repeat in grey as though it were extra information. Now it says it once, as "Waiting for you", and the "what did it ask?" button beside it is the part that actually tells you something.
v0.54.52026-09-06

Codex figures already on your screen are corrected

  • —0.54.4 fixed how token reports are read and could not reach the turns already stored — those kept showing no tokens and no cost. They are re-read from the transcripts on the next start, which takes seconds and needs nothing from you.
v0.54.42026-09-06

Half of Codex's token reports were being thrown away

  • —About half the turns in a Codex transcript state how many tokens they used without breaking it down by kind, and Caprock read only the breakdown — so those turns were stored as having used nothing at all. On the owner's machine that hid 18.6M tokens and turned $23.71 of usage into $0.53.
v0.54.32026-09-06

A Codex transcript with one unexpected field was skipped entirely

  • —The session record was read all-or-nothing, so a single field of a type we did not expect discarded the session id — and with it every turn in that transcript. Each field is now read on its own: an unfamiliar shape costs that field, never the session. Codex is someone else's format and free to change between releases, so this is the difference between missing a detail and missing a day's work.
v0.54.22026-09-06

Codex sessions were missing more than half their turns

  • —The importer keyed each event on a field present in one transcript out of a hundred; in the rest it read as zero, so every turn of a session shared one key and all but the first were discarded as duplicates. One session kept 1 turn of its 55. On the owner's machine 98 turns were stored where the transcripts hold 218, with 18.5M tokens missing along with them.
v0.54.12026-09-06

Codex sessions now all carry a price

  • —v0.54.0 shipped with 83% of Codex tokens costing nothing, because it read the model from the obvious field — present in 4 transcripts out of 100. The other 96 record it elsewhere, as an explicit model id rather than a hint, so both are read now and every session that used tokens can be priced. A session that names no model anywhere still keeps its tokens and no cost, rather than an invented one.
  • —A test daemon read the machine's real Codex transcripts. The smoke test asserts exact turn and cost figures, and on a developer machine with Codex installed it was importing that person's actual sessions and counting them. It turns the import off, as it already did for OpenCode.
v0.54.02026-09-06

Codex sessions are on the dashboard

  • —A machine running Claude Code, OpenCode and OpenAI Codex had its spend split three ways, each tool knowing only its own share. Codex sessions, turns, tool calls and cost now sit on the same screens as the others, filterable by agent. Nothing to install: Codex writes one transcript per session and Caprock reads it, the way it already reads OpenCode's database.
  • —The share dialog's preview never changed when you switched period — and on a real machine it never appeared at all. Two faults in the same box. The draw released its image on teardown, so changing period released the one still on screen, and a released image leaves the previous picture showing. Under that, anything the drawing threw on killed the draw silently and left "drawing…" on screen for good, which looked exactly like a slow one. The card you saved was always right; the preview was a lie about it. A draw that cannot finish now says so.
  • —The "Lots of turns, few files" alert. It counted files touched through Edit/Write only, so everything changed through Bash — sed -i, a heredoc, a python one-liner — was invisible to it. An agent running with permissions bypassed is told to prefer exactly those, so the rule was blind to an ordinary way of working.
v0.53.22026-09-06

A status line that had silently printed nothing is repaired on start

  • —An older version wrote the registered command with quotes around the whole string rather than just the path ("…/caprock statusline"). The shell reads that as one filename, finds no such file, and prints nothing — so the status line was simply blank, with nothing anywhere saying why. Worse, Caprock did not recognise the entry as its own, so statusline install treated it as somebody else's and refused to touch it: the machine it was found on had been in that state for months. caprock up now recognises the spelling, repairs it, and says so. A statusLine that is not ours is still never touched, and neither is a working one.
v0.53.12026-09-06

Every coloured border in the dashboard was grey

  • —A global border reset sat outside Tailwind's cascade layers, and unlayered CSS beats layered CSS whatever its specificity — so it quietly overrode every border-accent, border-premium, border-ok and border-warn in the app. It went unnoticed because the elements using them mostly have a tinted background too: the "+ New session" button looked amber because of its fill, not its border. Buttons and status pills now show the colour they always asked for.
  • —The two session buttons line up with the strip beside them. Quick chat and + New session were shorter than the ALL TIME row they share, which read as ragged; they now match its height. Quick chat gains a soft amber border.
v0.53.02026-09-06

The status line carries the mark

  • —the amber ⛰ from the favicon, the glyph alone, so the line is identifiably ours instead of an anonymous row of figures. Two columns, and it is never dropped however narrow the terminal gets. This is on every status line, in both modes, and needs nothing turned on: if Caprock's status line is already set up, upgrading is enough.
  • —The session's own numbers in the Claude Code status line — N turns · N steps · cache −N% · in N · out N, from counters only the daemon has, added to the line Caprock already prints.
v0.52.32026-09-03

The loop banner sold a cap that could not act, beside a cost that was…

  • —The loop banner sold a cap that could not act, beside a cost that was not the loop's. Both halves of the row overstated what they knew. The spend cap pauses only sessions Caprock started (rule 7), and 122 of the 127 sessions that did real work on this machine were started by hand — so "a cap that stops this" was, nearly every time it appeared, an offer to stop something the cap cannot reach. It now says "a cap covers sessions Caprock starts" and offers "what a cap does" unless the session is one Caprock owns; unknown ownership counts as not owned, since a promise made on a guess is the thing being fixed. The button stays, because the feature is real and the sessions where it applies are exactly the ones someone would want it for.
v0.52.22026-09-03

Sessions cleared before 0.52.1 still said "compacting context"

  • —0.52.1 gave /clear and Escape their own event kinds, but only for events recorded after it shipped — the rows already in the database kept the compact's kind, and on this machine all 18 of them were the last event of their session, which is the phrase the session card shows. So 17 sessions went on claiming a compaction that never happened, and the /clear that started the whole investigation was still captioned "was compacting context" beside a $1,062 cost and a 95%-full context.
v0.52.12026-09-03

/clear showed one editor as two working sessions

  • —Claude Code does not end a session on /clear — it keeps the process and starts a fresh session id inside it. Both rows therefore shared a live pid, and the sweep that decides a session is over asks whether its process is still running, so the old row could never close. It sat on the dashboard as a second working session, frozen at the moment it was cleared, still presenting a six-day cost and a 95%-full context as current. The replacement is now keyed off the SessionStart that follows, which is the only one of the two hooks that names the session meant to survive.
  • —A cleared session was narrated as "compacting context". /clear was stored as context.compact, which is a different event: a compact summarizes the context and keeps its substance, /clear discards it. The dashboard reported a compaction that never happened, at the one moment the owner is deciding whether to wait. It has its own kind now, and the reasons that leave a session running without touching its context (Escape at the prompt) have theirs.
  • —"unknown model" appeared beside the model's own name. A session that had not answered yet has no prompt to measure, so the Context stat was empty — and captioned unknown model directly beside a Cost stat reading claude-opus-5. The two reasons a context can be missing call for opposite reactions (one clears itself in seconds, the other never will), so the server now says which applies, naming the model id when the pricing table cannot size it.
  • —Fable 5 could not be picked for a new session. The model list offered three ids and omitted the most capable one on the machine, though claude --help names fable alongside opus and sonnet. All four are now offered, ordered by price, each verified by a live answer from the real claude — claude-mythos-5 is in the pricing table and stays out, because the binary rejects it.
v0.52.02026-09-03

A switch for the session memory

  • —"Sometimes I need the context and sometimes I don't" — a fair objection from the first person to live in Caprock, who had been bitten by an assistant that remembered his job title and volunteered it while he was asking about something else.
  • —"Answers" is now "Memory". Two halves of one thing did not know about each other: a new session is handed what the last one left, and you can search everything the agents ever said. Same corpus — one was labelled "memory" in settings, the other "Answers" in the nav, which nobody looking for the first would open. The status row names the screen too.
  • —The Memory screen says how far back it goes — "since July 19", measured from the oldest passage still held rather than from the fortnight a handoff reaches. Those are two different dates, and showing the shorter one understated the corpus by a month.
v0.51.12026-09-03

The token column beside each model counted re-reading, not work

  • —"claude-opus-5 · 12.61B tokens" is arithmetically true and reads as how much went through that model. On a normal workload 98.6% of it is cache read — the same context fetched again, at a tenth of the price — against 14.6M of output. The column measured how many times one conversation was re-read.
v0.51.02026-09-03

A session opening a repository you have worked in before already…

  • —A session opening a repository you have worked in before already knows where you stopped. Open claude, and the agent has been handed the last substantial thing an agent said in that repository — no command, no prompt, nothing to press. Verified against a live session, which named the previous session's work while forbidden to read files or git.
v0.50.12026-09-03

The update instructions were set like captions

  • —The commands read fine; everything explaining them was 11px in the faintest grey — the weight this interface uses for a timestamp, something you are meant to slide past. It is the half of the dialog that says why updating takes three commands and why Caprock will not do it for you. The install-route tabs got the same lift; they are a choice, and they were dimmer than the platform tabs beside them.
v0.50.02026-09-03

Let a tablet in from the dashboard, without a terminal

  • —Reading Caprock from a phone shipped needing caprock up --lan, which requires a terminal on the machine — and the person who wants this is holding the tablet. It is a button on the status screen now, with an off switch beside the device list. The flag still works for a machine administered over SSH. Unchanged: it is off again when Caprock restarts, and paired devices are kept.
  • —Caprock uses a Tailscale address when there is one, and says which kind of address you have. A LAN address works when the other device is on the same wifi and never from mobile data — no setting changes that, because the packets do not arrive. The panel said nothing about it, so the honest conclusion from a tablet on mobile data was that the feature is broken.
v0.50.02026-09-03

Let a tablet in from the dashboard, without a terminal

  • —Reading Caprock from a phone shipped needing caprock up --lan, which requires a terminal on the machine — and the person who wants this is usually holding the tablet, not sitting at the machine. It is a button on the status screen now: press it, and the same panel shows the address and the code. The flag still works for a machine you administer over SSH.
v0.49.32026-09-03

"Burn now" says it is measuring instead of extrapolating

  • —0.49.1 made the rate divide by the time it actually covers, which is correct arithmetic and an alarming number: a daemon twenty seconds old showed $3,215/h, because twenty seconds of a busy minute extrapolates to that. Until the window has been lived through, the tile shows a dash and says it needs ten minutes of running — a figure nobody can act on is worse than no figure.
  • —A test that asserted a promise the code never made. One of the night's new tests demanded that a healthy terminal subscriber receive 300 of 400 chunks while another subscriber sat stalled. The buffer is deliberately lossy — a reader that falls behind legitimately misses some — so the test passed on a fast machine and failed in CI, which is what stopped v0.49.2 from publishing. It asserts what the code actually promises now: the writer finishes and a healthy reader keeps receiving.
v0.49.12026-09-03

"Files touched" counted whole session lifetimes

  • —The range filtered which sessions counted, not which files — so a fortnight-long session that touched 300 files put all 300 into "today" the moment it said anything today. Seven days read 417 against a true 144. The ranged answer was already on disk; the wrong table was being asked.
  • —The daily session count read zero on days with real spend. It counted a session on the day of its first turn ever, so a session begun yesterday added nothing to today. Four of eight days read 0 against 1, 2, 2 and 2 — beside cost and token figures in the same rows that were correct, which is worse than being wrong alone. After the migration, ten recent days all match a direct count over the events.
  • —A replaced transcript of the same size lost everything below the old offset. Truncation was detected by "the file got shorter", and Claude Code rewrites transcripts in place — on a resume, on a compact — where the replacement usually is not shorter. Those lines were never parsed, so nothing deduplicated them either: a session simply arrived missing its first turns.
  • —A half-written Gemini record was skipped permanently. The reader advanced past everything available whenever anything parsed, which is the opposite of what its own comment promised. Since the file is read while Gemini writes it, landing mid-record is the normal case.
  • —Two Gemini events in the same millisecond became one — and worse, every event whose timestamp was missing or in an unexpected format shared a single key, so only the first was ever stored.
  • —The burn rate divided a partial window by its full width. A daemon three minutes old has three minutes of history and was dividing by ten, so the rate read a third of the truth with nothing saying the window was still filling.
  • —Six tables can no longer push the page sideways on a narrow screen. On a tablet a wide table took the whole body with it.
  • —More tests where a failure would cost something: the three shapes a tool error arrives in (only one was exercised, and it decides the red badge), the point where rule 7 is enforced, all five pairing handlers, and the file that holds device tokens. Every important test was checked by reverting the behaviour it asserts.
v0.49.02026-09-03

Read Caprock from a tablet

  • —caprock up --lan opens a second listener on this machine's own private address; the status screen shows what to open on the other device and a six-digit code to type. No account, no tunnel, nothing on anybody's server — the other device is talking to your machine.
  • —Two figures ignored the agent filter. Choosing opencode on the Now screen narrowed most of the Today panel and left "burn now" showing every agent's money, in the same grid row as the cost it no longer matched. The unpriced warning had the same fault, and worse: it told a reader looking at one agent that tokens were missing from a total those tokens were never in.
  • —The share card's breakdown was always the last thirty days. Choosing a period changed the heading and the lit tile and nothing else that mattered, so a card headed "My week on Caprock" carried a month's models beneath it — five of them against the one that week ran, and $5,459 of Opus against $1,576. The four tiles still show every period on purpose; it is the breakdown that belongs to the period the card names.
  • —A range or filter change showed the previous answer under the new heading. Pressing 7d while 30d was up swapped the title at once and left every figure below it answering the old range until the response landed — around half a second on a large database, long enough to read a number and believe it. A refresh still keeps the figures on screen, because blanking them every few seconds would be its own defect.
  • —The calendar put days in the wrong weekday column. A day nobody worked has no row at all, so one missing Sunday slid every later day one column left and a Monday rendered under "S" — in a panel whose whole purpose is making "I do not work Sundays" visible at a glance.
  • —"Ended · 200" counted the page, not the sessions. The list is capped at 200, and the label counted the array it received while the lifetime strip above said otherwise. It reads "200 of 431" when they differ.
  • —The largest figure on the screen called itself a cost. On a flat plan it is not one: it is what the work would have cost through the API, and the qualifier saying so sat underneath in grey at a third the size. A reader on a subscription took the whole product for an accusation of waste. The figure is unchanged and now headed "Worth at API list"; anyone billed per token still sees "Cost", because for them it is the true word.
  • —Every screenshot retaken, and three faults in how they are captured fixed: two first-run banners were in every frame, "set plan" was lit as an unfinished action because the capture runs against a throwaway data directory, and the feature crops had silently been failing to find their panels since the database grew.
  • —The docs say what the product does now. An audit found the release check described as the only outbound call (it is one of three, all opt-in), the premium page selling a shipped feature as unbuilt, a release runbook asking you to verify a hook count that can no longer appear, and Gemini missing from every user-facing list of agents.
v0.48.02026-09-02

Every table says what its columns are

  • —Three figures in a row with nothing over them is a puzzle: $7,587.14 · 12.01B · 58% reads as three unrelated numbers until you work out which is which, and the "by cost" in the corner describes the sort order rather than the columns. Every breakdown on Now, Cost and Lifetime now names them, on one line with the table's own title so it costs no vertical space.
  • —The all-time breakdown moved up, under the live pulse. It was last, below a tall activity feed, so the most legible thing on the screen was the thing you had to scroll to find.
  • —"Nothing measured yet" appeared while the figures were still loading. The first screen could not tell "no answer yet" from "the answer is nothing", so for the second it took the first response to arrive it announced, with confidence, that a machine with thousands of turns on it had never run anything. It says "reading your figures…" until it knows. Measured on a 641 MB database: nothing on this screen takes longer than 0.7s, so the delay was never the problem — the wrong answer during it was.
  • —The answers screen showed raw Markdown. Claude writes bold text, lists and tables; the screen kept the line breaks and showed everything else as the markup it is made of — bold with its asterisks, a table collapsed into | | |. The same renderer the release notes got now covers both, and it learned tables and quotes on the way.
  • —A plan you had already chosen looked unchosen. Once set, the picker's button went grey and lost its border — which is exactly what the buttons beside it look like when they are not set — so every session started with the impression that the plan had been forgotten again. It keeps its border and carries a check.
  • —A Gemini error sent people somewhere they no longer need to go. "Set GEMINI_API_KEY in the daemon's environment and restart it" — but the key has a field on the Cost screen and takes effect at once. A wrong instruction is worse than a long one.
  • —Errors carried a Go package name into the dialog box. agents: "/x" is not an absolute path was written for whoever was debugging it.
  • —The tool table says how much each tool hands back. Calls alone told the wrong story. Bash is 46,343 calls returning 13.4 MB; Read is 11,767 calls returning 58.7 MB — four times fewer calls, four times more context. The loud tool is the cheap one, and nothing in the product said so.
  • —Share a week, a month, or today — not only all time. The picker was removed once, on the reasoning that a card showing every period at once made choosing redundant. It does not: somebody sharing a working week does not want their lifetime total to be the headline, and a card that answers four questions answers none of them loudly. The choice now decides what the heading claims and which figure is lit; every other period stays on the card, because a week means nothing without knowing whether it was a normal one. It opens on the week.
  • —You can see the card before you send it. The buttons said where the picture would go and nothing said what was in it: you pressed Save, opened your downloads, and only then found out. The card is drawn in the sheet now, at the size it will be seen, and it redraws as you change the period — which is what makes the choice worth having.
  • —The Share button can be found. It was an outlined button in a row of 11px text, next to a solid premium button — an outline cannot win an argument with a fill. It is filled, a size up, and carries an icon. It earns that: sharing is the only thing on this dashboard that leaves the machine.
  • —Copy across the product says less. Several tooltips explained the attribution algorithm where the reader had asked what a number meant, and several paragraphs defended design decisions nobody had questioned. The longest of them — the per-directory rule — was four sentences of specification on hover; it is now one. "Everything here is measured — no invented numbers" is gone from the Cost screen: readers assume the figures are real until told otherwise, so the sentence created the doubt it answered.
  • —Messages about a missing claude say what to do. "Spawning is unavailable" named an internal verb and diagnosed without a fix; it now says the binary is not on PATH, that installing Claude Code fixes it, and that watching works regardless.
v0.47.02026-09-02

Pick up a session Caprock did not start

  • —It never types into a process it did not launch — two writers on one terminal interleave characters and ruin both — so a session you started yourself was readable here and nothing else.
v0.46.02026-09-02

Send this week's report now

  • —There was no way to find out whether a bot token and chat id were right except to wait for Monday — and when nothing arrived, that looked exactly like a week where nothing moved. A button beside the fields sends one immediately.
  • —What's new showed raw Markdown. ### Fixed as a literal heading, asterisks around every bold phrase, and paragraphs broken at the eighty columns a changelog is wrapped to for the sake of diffs. It was rendered as preformatted text on the reasoning that a local-first tool should not render remote markup — right instinct, wrong conclusion, in a dialog whose only job is to be read.
  • —The dialog linked to GitHub. Someone reading what they just upgraded into wants the changelog written for readers, not a tag page with build artefacts on it. It goes to caprock.dev/changelog now.
v0.45.22026-09-02

Why a weekly report never arrived is remembered now

  • —The reason a send failed was held in memory, so restarting the daemon erased it — and restarting is exactly what somebody does when a feature seems broken, which made investigating the problem the thing that destroyed the evidence.
v0.45.12026-09-02

The premium lock covered the sentence explaining the feature

  • —The caption and the button were centred on the panel, so they landed exactly across the middle of the text — "Off. Nothing is paused, whatever the day costs" was cut mid-word by the thing selling it. Two earlier attempts tuned the preview's opacity, which could not help: the problem was never how transparent the overlay was, it was where it sat. It is a row underneath now, reading in the order the eye already moves — here is the feature, here is what it costs — and covering nothing.
  • —A gemini filter chip appeared on a machine with no Gemini sessions. Having the binary on PATH is what decides whether the New Session dialog offers Gemini; it says nothing about whether any Gemini session exists. A first run showed all · claude · gemini above an empty screen with nothing to sort by. The chip now appears when there is something to filter.
v0.45.02026-09-02

A session stays open until its process exits

  • —It used to close after a stretch of silence, and every number chosen for "a stretch" was wrong for somebody: twelve hours left the day's work marked live at midnight, and one hour closed a session while its owner was at lunch. Both were guesses about a person's day standing in for a fact about a process — and the fact is available. The shim now reports the pid of the Claude Code that ran it, and Caprock already knew the pid of every session it starts itself, so the sweep asks whether that process is alive instead of watching a clock.
  • —The Now screen shows now. A session that has not made a sound in two days no longer appears there, whatever its status column says — status alone cannot tell a quiet session from somebody's history, because an observed agent's sessions arrive weeks old by definition. Two days rather than one, so a session left on Friday evening is still on screen on Monday morning. Anything actually working shows regardless of age, which is what the screen is for. Older sessions are on Lifetime and under "show ended".
v0.44.52026-09-02

A session you left for lunch was closed while you were out

  • —0.44.3 cut the staleness threshold from twelve hours to one, on the reasoning that an hour outlasts lunch. An hour is lunch: the first person to leave a terminal and come back to it found the session gone from the dashboard. On the machine this was checked against, 44 sessions had paused for over an hour and then carried on working, and 86 of those pauses were between one and three hours — a meeting, a walk, an afternoon on something else.
v0.44.42026-09-02

Caprock now sees what a Gemini session does

  • —It could start one and then observe nothing: a live session sat in the list with zero turns, zero tokens and no cost. Gemini has no hooks and writes no transcript, but it does write OpenTelemetry, and it will write it to a file Caprock names — so a spawned session now reports its turns, its tokens and its cost like any other.
  • —Two of the three Gemini models still could not be used. The list was checked against the installed CLI and the pricing table, and both said these were fine. A real key says otherwise: Google has closed gemini-2.5-flash-lite to new keys, and gemini-3.1-pro-preview is barred on a free key by quota (generate_content_free_tier_input_token_count, limit: 0) — so picking either opened a terminal and then failed.
v0.44.32026-09-02

A session you were still working in could close itself

  • —0.44.0 started listening to Claude Code's SessionEnd hook so a session ends the moment it ends rather than whenever the staleness sweep next runs. But SessionEnd does not only mean "the user left": it also fires on /clear, on Escape at the prompt, and with an unhelpful other. All of them were being treated as the end. One session on the machine this was found on had been running for six days when a /clear retired it in the dashboard.
v0.44.22026-09-02

The terminal flickered on every keystroke

  • —A ResizeObserver called fit() straight from its callback, and fit() writes to the element the observer is watching — so a resize could cause a resize. Idle, that settles after one frame and nobody notices. Under a TUI that repaints its whole box on every key, which Gemini CLI does, it was a visible flicker on every key. Fits are now coalesced into one per frame and skipped entirely when the geometry has not moved. It affected any terminal that repaints often, not only Gemini's.
  • —Six columns of zeros above a session Caprock cannot measure. A Gemini session showed $0.00, 0 tokens, 0 turns, 0% cache, with the reason — "no hooks · no transcript" — in small grey text underneath. Zeros are how that bar says "nothing has happened yet", so it read as broken rather than as out of scope. It now says the one true thing instead: Caprock started this session but does not measure it, and the terminal below is live.
v0.44.12026-09-02

Starting a session with Gemini actually works now

  • —0.44.0 shipped the feature built against a fake binary, on four assumptions about the Gemini CLI that were written from memory rather than from gemini --help. Installing the real thing disproved all four in ten minutes.
  • —A Gemini session is no longer described as a Claude one. The agent was added to the launcher and to nothing else: its sessions arrived in the list wearing no badge, with no filter chip to find them by, and the empty state called them Claude Code — because three separate places asked agent === 'opencode' and treated everything else as Claude. There is now a gemini chip beside claude and opencode, shown only on a machine that has it, and one helper decides the badge and the name so a fourth agent cannot reintroduce this.
v0.44.02026-09-02

Start a session with Gemini CLI

  • —Install it (npm install -g @google/gemini-cli), and it appears next to Claude Code in the New Session dialog — same terminal, same directory picker, same row in the sessions list, same cost stream. Caprock passes the key you already entered into the child process, so it works without exporting anything in the shell you happen to be in.
  • —A debounce timer outlived the screen that set it. Closing a screen inside the 400 ms live-tick window left a timer that woke up and set state on a component that no longer existed. It was visible mainly as a failed release — the timer firing after the test runner had torn the DOM down, with all 483 tests passing — which is how it got fixed twice at the wrong end before the cancel-on-unmount that it actually needed.
v0.43.02026-09-02

Claude Fable 5.1 and Mythos 5.1 in the pricing table, contributed by…

  • —Claude Fable 5.1 and Mythos 5.1 in the pricing table, contributed by [@SergeAx](https://github.com/SergeAx) — the first outside contribution to this repository. Cache reads on those two models cost 0.025× the input price rather than the 0.1× every other Claude row follows, which makes $0.25 look like a typo of Fable 5's $1.00; the PR pins it with a test, so a later tidy cannot quietly charge four times the real rate.
  • —fetched_at in the pricing table names the most recent read, and each note carries the date for the rows it describes. One date on a table assembled from three vendors over three weeks said 18 August while holding prices read on 1 and 2 September — which tells a reader checking a figure nothing about that figure.
  • —The Gemini key goes in the dashboard now. It was read only from GEMINI_API_KEY, on the reasoning that a key Caprock never holds is one it can never leak — but a daemon started at login inherits nothing from a shell profile, so switching on a paid feature meant hand-editing a launchd plist. A key nobody can enter protects nothing; the feature just goes unused. Paste it into the panel instead. The environment variable still works and still wins where both exist, and the stored key is never returned by any endpoint — the same write-only rule the Telegram token follows.
v0.42.12026-09-01

The weekly report, to your own Telegram bot

  • —What moved this week against your usual, sent Monday morning — or the next day you open the lid, because a timer set for Monday 09:00 fires for nobody whose laptop is shut at weekends.
  • —A setting no longer resets when you save a different one. "Set plan keeps resetting" was not a save that failed: the dashboard sent the whole settings object every time, assembled from a cache in the page, so any control could write back a field it had read minutes earlier. Two tabs were enough — change the plan in one, turn on update checks in the other, and the second restated the plan from its own stale copy. Only what changed is sent now.
  • —Live activity showed a fortnight of history as what just happened. Two causes at once: the feed asked for a session's first sixty events rather than its last, and the store ordered "most recent" by insertion rather than by time. Those agree for a session captured live and diverge for one whose transcript was re-read, because a backfill inserts old events with new ids. On a long session every row was equally stale, which is also why it read as impossible to scroll.
  • —The new-session dialog cut off the permission label — "Accept edits · asks before com…" hid the consequence, which is the one thing that label exists to state. Wider dialog, shorter labels.
  • —A stray reconnect no longer crashes a torn-down page. The live socket queues a retry when it cannot connect, and that timer could outlive the document that scheduled it — harmless in a browser, where the page is going away anyway, and enough to fail a release build whose every test had passed. It now gives up quietly instead.
  • —The premium chip was two controls disguised as one. The label went straight to the checkout while a chevron beside it opened the explanation, in one border with no seam between them — so clicking the word "premium" to find out what premium is took you to a card form. One control now: it opens the dialog, and once you have bought it names your plan and when it ends.
v0.41.02026-09-01

Ask Gemini, on your own key

  • —The second paid feature: a second model inside Caprock, billed to you by Google at their prices. Caprock never stores the key — it reads GEMINI_API_KEY from the daemon's environment at the moment of the call, and it is never written to disk, never accepted by the settings endpoint and never returned by one. A key we do not hold is a key we cannot leak; the cost is that you set a variable and restart, which the panel says plainly.
v0.40.02026-08-31

A session now ends when it ends, instead of half a day later

  • —The Now screen counted a whole day's finished sessions as live — "14 sessions, 0 active" — and the live pulse drew a row per known session, so a day in one repository became six identical caprock rows over six flat hairlines.
  • —Upgrading Caprock no longer kills the work it was watching. Restarting the daemon — which every upgrade does — sent SIGKILL to every session Caprock had spawned. They died mid-turn, with no warning and nothing flushed: a tool that watches your work should not be the thing that eats it. Sessions are now asked to stop, waited on for five seconds so Claude Code can write out its transcript, and only then killed if they refuse. The upgrade banner also says how many sessions will close before you copy the command; sessions you started yourself are untouched, as always.
  • —The live pulse shows only tracks that ran. Ended sessions are dropped, and so is any track with no events in the window — an hour of flat hairline says only that a session exists, and several of them read as a broken chart rather than as silence. When nothing ran in the last hour, the panel says that in one line.
  • —A pulse row shows what its hour cost, not the session's lifetime. A long-running session printed $4,053.39 beside an hour of bars and a $101.85 day: two true numbers answering different questions, side by side. The row now sums the window the bars cover, with the lifetime on hover.
  • —The new-session dialog answers its own questions. Model and permission mode both opened on an empty "default" that said nothing about what was about to run — and "default" is not a permission mode Claude Code accepts at all, so the dialog could send the binary a value it rejects. They now start on Opus 5 and "accept edits, asks before commands", with each option labelled by what it does rather than by its flag name. The worktree field and "create the directory" moved under Advanced, and the paragraph advertising OpenCode support left the dialog: every field on screen is a decision asked of someone who wanted to press one button.
  • —The session timeline reads newest-first. It was the only list in Caprock ordered the other way, so the same glance meant two different things on two screens — you arrive at a timeline wanting the last thing that happened, not the first. The follow checkbox and its autoscroll are gone with it: new rows arrive at the top now, so there is nothing to chase. History stays behind "load earlier events" rather than being rendered up front, and each click now pages back from the oldest row on screen — it used to refetch the first thousand events of the session and discard nearly all of them, which on a sixteen-thousand-event session fetched the wrong end of the history.
  • —"Live diff" and "Files" are one Changes tab. They answered one question between them — what did this session change — so reading it meant visiting both tabs and holding two lists in your head, and on most sessions "Files (0)" was empty besides. Files now expand independently rather than one at a time (the old accordion closed the first file when you opened the second, so two changes could never be compared), with a caret per row and expand/collapse all for reading a whole branch. Files touched but unchanged keep their own panel underneath. Old ?tab=diff and ?tab=files links still work.
  • —The live diff measures a branch from where it forked. The base was always HEAD, which answers "what have I not committed yet" — so a branch whose work was committed showed no changes at all while the session had rewritten dozens of files. It is now the merge base with master/main: the branch's own commits plus whatever is uncommitted. The panel names its base ("since master") beside the file count. Untracked files show their contents as added lines instead of "no diff against HEAD".
  • —The Projects list holds its order while you point at it. Rows are ranked by spend and refresh every 30 seconds, so a row could swap places between aiming and clicking and open a repository you never pointed at. The values keep updating live; only the sequence holds still, and only while the pointer is inside the panel.
  • —Pulse rows are told apart by branch and session id. Working all day in one repository drew rows labelled caprock, caprock, caprock, most of them subtitled was responding. The branch and a short id are what actually differ, so they are what the row leads with.
v0.39.22026-08-31

Shift+Enter really does insert a newline now

  • —0.39.1 sent the right bytes and still submitted, because the bytes were only half of it: returning false from the key handler stops xterm interpreting the key, but the browser still delivers it to xterm's hidden textarea, which emits its own carriage return. The socket carried ESC CR and then a bare CR immediately after — [27,13] followed by [13] — and Claude Code submitted on the second one.
v0.39.12026-08-31

Shift+Enter typed \n into the prompt and sent the message

  • —The newline keys sent two printable characters — a backslash and the letter n — so a message arrived as first line\n and the multi-line prompt was unusable. They now send ESC CR, which is Alt+Enter as a terminal encodes it and what Claude Code's own macOS instructions bind Option+Enter to.
v0.39.02026-08-31

The daily spend cap

  • —the first thing Caprock does rather than shows. Everything else here is observation: you look and you learn something. A cap acts while you are asleep, which is when a runaway loop does its damage. Set a number for the day; when the day crosses it, the sessions Caprock started are paused.
v0.38.12026-08-30

The folder picker ran past the edge of its dialog

  • —A long path — the ones under ~/Library/Application Support are the worst — made the list 738px wide inside a 520px panel, so rows spilled over the border. A grid item does not shrink below its own content unless told to, so no amount of clipping inside the picker could fix it: the containers above it had to be allowed to be narrower than what they hold. Paths now truncate with the full one on hover.
  • —The picker looked like a different surface from the field above it. It sat on a transparent background with the lighter border, directly beneath an input that had neither, and the two read as two panels at different opacities. It now matches the input exactly.
v0.38.02026-08-30

Pick a folder instead of typing its path

  • —Starting a session meant typing an absolute path from memory into a dashboard that is already showing the repositories you work in every day.
v0.37.52026-08-30

The button on a session Caprock did not start now says what it does

  • —"Start one here" named an action and left one undefined — a reader could reasonably expect it to attach to, restart, or take over the session in front of them. It does none of those.
v0.37.42026-08-30

"Start a session in Caprock" now starts a session

  • —It was a link to the main screen — navigation dressed as an offer, which moved the reader away from what they were doing and left them to find the real button themselves. It opens the New session dialog in place, already pointed at the repository on screen, so the one thing it promises takes one click. Asking for a working directory on a screen that is displaying it was asking someone to retype what they were looking at.
v0.37.32026-08-30

The terminal panel on a session you started yourself now says what to do

  • —It opened with "This is an externally started session — Caprock observes it but never writes into a terminal it does not own": accurate, and written for the people who built it. A reader saw a wall of text where a terminal should be and could not tell what was being asked of them, or whether they had done something wrong.
v0.37.22026-08-30

Buying takes one click

  • —The header chip opened the dialog and nothing else, so someone who had already decided still had to read a screen and then find a price. It is now the price itself — premium $30/yr, straight to the checkout — with a separate chevron for anyone who wants the explanation first. People who are ready should not be routed through a pitch.
  • —The locked preview really is legible now. Raising its opacity in 0.37.1 changed nothing visible, because a full-panel scrim was laid over the top of it: the content was brighter underneath and the sheet dimmed it straight back. The scrim is gone; the caption and the button carry their own backing instead, so the panel behind them can be read.
  • —A missing plan no longer takes the header down. The chip read .url off the yearly plan without checking it was there — an older daemon, or any response without it, crashed every screen rather than hiding one chip.
v0.37.12026-08-30

There was no way to buy from the main screen

  • —The only entry points to paying were the two panels behind glass, on Cost and on Lifetime — so the screen where people actually sit offered no path at all. premium now sits in the header on every screen: the one element there that is not amber, so the eye finds it when it is looking for the thing about money and not otherwise. It stops selling once a licence is active.
  • —The locked preview was unreadable. At opacity-35 it was the worst of both — occupying the space of a demonstration while demonstrating nothing, so a reader could see that something was under the glass but not what. It is now legible through it, and the caption, which is the sales line on that panel, is larger than the body text it sits over rather than smaller.
v0.37.02026-08-30

The premium dialog, after five readers were asked to buy from it and…

  • —The premium dialog, after five readers were asked to buy from it and none did. Not one of them named the price — three said $30 was nothing to them. The verdict was the same sentence in five voices: "I am paying for delivery of something I already see for free."
v0.36.12026-08-30

The premium dialog says the same thing in half the words

  • —The paragraph above the bullets repeated what the bullets said, and the footer explained the licence twice.
  • —Both prices are buttons now, and the lifetime one is brighter. An outlined second button reads as the lesser option, which is backwards: someone weighing the year should see the lifetime as the step up from it. Two filled buttons of ascending brightness make that legible without a word of copy.
  • —The comparison moved under the buttons, one line each. It was a two-row table above them — "Claude Pro $20/month" over "Caprock Premium $2.50/month" — two unrelated numbers standing where a price belonged. Each button now carries what it is worth in the plan the reader already pays for: about six weeks of Claude Pro for the year, about five months for the lifetime.
v0.36.02026-08-30

Premium is ultramarine now

  • —Every interactive element in this product is the brand amber, which left a subscribe button looking exactly like a range filter. Paid is a different kind of thing, so it gets a different hue — deliberately cool against a warm interface. Reserved strictly for surfaces where money is involved; an ultramarine that spread to ordinary controls would just be a second accent, and then nothing would be marked.
  • —The premium dialog offers a year or a lifetime, and not a month. At $5 the monthly plan is the cheapest way to hold a licence key for a month, which is a worse deal for the buyer than either commitment beside it. It remains on the site for anyone who seeks it out.
  • —The weekly report shows what would arrive, not a description of it. It was three rows reading "Sent / To / Contains", which describes an email without showing one — and behind glass that is a blurred settings screen. It now previews the message itself, with this machine's own repository names in it. Deliberately without figures: those are already free two panels down, and locking them would take something away rather than preview something new.
  • —winget. winget install dspv.caprock, once Microsoft has merged the manifest for a release — the package manager already present on every Windows 10 and 11 machine, with no bucket to add first. Deferred until there was Windows demand ([ADR-014](.ai/08-decisions.md)); the first Windows user to build from source asked for it by name.
  • —make build works on Windows. It did not, and the very first command in CONTRIBUTING.md failed on a fresh Windows clone: make with no sh on PATH falls back to cmd.exe, which understands none of mkdir -p, [ -d ], single-quoted -ldflags, or 2>/dev/null. It now uses the Bash that Git for Windows already installs.
v0.35.12026-08-30

The main screen's heaviest request went from 0.61s to 0.001s

  • —/v1/history answers "everything, ever" — four aggregates over the whole events table — and five components on that one screen ask for it on their own timers: the lifetime strip, the breakdown panel, the share card, the share nudge, and the screen itself. A single open tab produced bursts of identical requests and each one was computed from scratch.
v0.35.02026-08-30

Prices can now carry the dates they applied, and a turn is costed at…

  • —Prices can now carry the dates they applied, and a turn is costed at the price it actually ran under. Sonnet 5 launched at an introductory $2/$10 and reverts to $3/$15 on 2026-08-31. With one row per model the only way to record that is to overwrite the figure — which would have silently restated every August turn at a price nobody was charged, growing a month's reported spend by half overnight. That is rule 6's "no invented numbers" pointed at our own history.
  • —The size of the spend on the main screen, not only how it divides. Each model row carries its token count beside its cost, and the panel gained an input / output / cache read / cache write line. A share says how the money split and nothing about how much there was: "opus-5, 54%" reads the same at forty dollars and at four thousand. Output costs five times input, so the ratio is the explanation of the bill, and it was a screen away on Cost.
v0.34.12026-08-30

The premium dialog no longer disclaims its own product

  • —It carried a line reading "Not built yet" directly above the price, which is a sales screen arguing against itself: the features are being built, the commitment is real, and hedging it sells nothing while protecting nobody.
v0.34.02026-08-29

The premium dialog now measures its price against a Claude subscription

  • —Every statement on it was true and the screen still argued against itself: the loudest element was a bordered box reading "Not built yet", so the first thing the eye landed on was a reason not to buy. Three prices sat where a decision belonged, and Subscribe was styled as one of three equal buttons.
v0.33.02026-08-28

Copy and paste in the terminal

  • —There was none: every key went to the process, so Ctrl+C was always SIGINT and nothing could be copied off the screen. It now follows VS Code's rule, which is the one people already have in their fingers — Ctrl+C copies when something is selected and interrupts when nothing is. On macOS the question never arises, since Cmd+C and Ctrl+C are different keys. Ctrl+Shift+C/V elsewhere.
  • —WebGL rendering, with a fallback to canvas. The canvas renderer repaints the whole grid; WebGL draws from a texture atlas, which is the difference between a build log scrolling smoothly and the tab stuttering. A machine without WebGL, a driver that refuses, and a GPU context lost on sleep all fall back rather than failing — a slower terminal is a cost, a blank one is a broken product.
  • —Paste or drop a file into the terminal. A browser hands over an image's bytes and never a path — there is no path for something copied out of a screenshot tool — while Claude Code reads files by path. The bytes now go to the daemon, which writes them into its own data directory, and the path is typed into the session so Claude can read the file.
  • —Scrollback is 10,000 lines, up from 5,000. A build log passes five thousand easily, and losing the start of what you are reading is where a terminal stops being one you can work in.
v0.32.12026-08-28

The terminal now tells the daemon how big it is

  • —fit() resized the canvas and nothing else, so the PTY kept whatever size it was created with — 120×40 by default — for its whole life. Claude Code lays its menus out to the terminal size, so on any other window it drew an interface for a screen that was not there: arrow keys moved a selection nobody could see, which is what the first user reported as "only Enter works".
  • —Shift+Enter inserts a newline again, on a prompt with text in it. It has been wrong twice, and the second one is the interesting failure: a bare line feed (0x0A) is what Ctrl+J sends and what the documentation says works everywhere — but Claude Code reads a lone line feed as submit. On an empty prompt that looks like a newline, because there is nothing to submit; the moment the prompt has text, the same key sends the message. That is exactly what the user reported.
v0.32.02026-08-28

The cache hit rate now says what it means

  • —It was a bare percentage, amber below 90% and identical above — so 99% and 91% read the same and neither said anything, on a figure that runs from 6% to 99.6% across real sessions. One word sits beside it: outstanding at 99%+, good from 95, ok from 85, low below. The bands are set against the observed spread rather than picked for roundness, so outstanding lands on about one session in nine and is worth reading.
  • —The share card asks the reader what theirs is. It carried a figure and a domain and stopped there — someone seeing another person's total had no reason to think it was a thing they could do too. A question, not an install line: a command on a picture is an advertisement and reads as one, and the domain in the heading is where a person who wonders goes to find out.
  • —Dismissing a banner recorded the wrong time. The premium banner, the premium hint and the share nudge all stamped the dismissal with a fresh Date.now() while every other decision in them used the clock they were handed. The two disagreed by however long the render had been on screen — invisible in production, but it meant the test suite passed for a year and then began failing on the day the fixed test clock and the wall clock crossed, with nobody having touched the files. All three use the clock they are given, and the boundary is now tested to the millisecond.
v0.31.32026-08-27

Every way of asking for a newline now works in the web terminal

  • —This shipped supporting Shift+Enter alone, and the first user to want a second line reported back that Option+Enter was the one he pressed — which is what Claude Code's own macOS documentation tells people to enable. Claude Code accepts four ways of asking for a newline and a person reaches for whichever one they learned elsewhere, so all four are handled, each sending exactly what Claude Code expects over the PTY: Shift+Enter sends CSI u, Option+Enter sends ESC then CR, Ctrl+Enter and Ctrl+J send a line feed, and backslash-Enter needs nothing from us.
  • —A line under the terminal saying how to get a newline. Shift+Enter had worked since v0.30.1 and the user who wanted it still could not find it: he pressed Enter, watched half a thought get submitted, and concluded multi-line prompts were not possible. A feature nobody can discover is not shipped.
v0.31.22026-08-27

Homebrew reported "already installed" for a release that was already out

  • —A tap is not served by the Homebrew API: it is read from a local git clone that brew upgrade refreshes only through auto-update, which runs at most once every 24 hours. A user who ran any brew command earlier the same day was told 0.31.0 already installed hours after 0.31.1 shipped — and the command Caprock itself had given them appeared to be broken. Every place that names the command now says brew update && brew upgrade caprock, with a test pinning the reason so it does not get "simplified" back.
  • —Hovering "Save the image" now shows something. It changed only its border colour — one step of grey against a dark panel — so pointing at it looked identical to not pointing at it. A control the eye cannot confirm it is on reads as disabled. It fills instead.
  • —"Quick chat" and "+ New session" line up. Quick chat sat inside a wrapper div next to a bare button, which made them two different boxes in the same flex row. The error message that wrapper existed for is positioned absolutely now, so a rare failure cannot change the height of the row.
  • —The version chip tells you how to update, in steps. It used to print one line — the command for however the daemon guessed this copy was installed — and stop, which answers only the first of three questions: run what, then what, and how do I know it worked. It now shows tabs for macOS, Linux and Windows with the routes each one has, the way this copy appears to be installed opened first, and three numbered steps with their own copy buttons: update, restart, check. The platform choice is remembered.
  • —"What's new" beside the version. The published release's own notes, in a dialog, without leaving the dashboard. The text arrives with the version in GET /v1/update — the same GitHub response, so no extra request and no further exposure — and it is rendered as text, never parsed as markup.
  • —Release notes come from CHANGELOG.md, not from commit subjects. The dashboard shows that text under "what's new", and a generated list of 1c836d5: fix(ui): … lines is not something anyone reads. A version with no changelog section now fails the release build rather than shipping an empty "what's new".
  • —The share dialog's guarantees are two bullets, not a paragraph. A reader scanning for what does and does not leave their machine should not have to read a sentence to find out. Bigger type, one claim per line — and pinned by a test, because this is the copy that gets reworded whenever the dialog is tightened.
v0.31.12026-08-27

Sharing the card produced two images; saving it produced one

  • —That split was the whole diagnosis — the download path draws once and writes one file, so the duplicate could only come from the native share, which handed the OS {files, text}. Two payloads leave the receiving app to decide what they mean, and macOS's Copy resolved it as two items. The card already carries every figure and its caveat, so the caption was never load-bearing; the file now travels alone.
  • —A second way to get two cards, fixed in the same change. One flag both labelled the share button "Drawing the card…" and disabled it. Clearing it early — so the label would stop lying while the OS share sheet sat open — also re-enabled the button underneath that sheet, and a second press drew a second card. Labelling and locking are now separate states.
  • —The version chip is a button, and it tells you how to update. It used to say a newer release existed and stop there, which answers the half of the question nobody needs help with. It now opens a dialog with the running version, what is published, a check-now, and the exact command for how this copy was installed — brew upgrade caprock, go install …@latest, or the release page for a downloaded binary — with a copy button.
  • —make reload builds the dashboard and binaries, installs them over whatever copy is on PATH, and restarts the daemon. Testing a UI change against a running daemon took four commands and getting one wrong meant looking at a stale build and drawing conclusions from it.
  • —Plan limits moved into the Today row. They had a full-width panel holding two percentages, three rows above the money, which made a reference figure look like a headline — and when both windows were stale, spent a band of the screen explaining that the figures beside it meant nothing. Now one cell beside burn, sessions and cache hit, leading with whichever window is closest to its limit, since that is the one that will stop the work.
  • —The share button in the header is visible. Grey 11px lowercase between "feedback" and the build label — invisible enough that the person who commissioned the feature could not find it on his own dashboard.
  • —The premium banner says what the feature does. It read "Premium stops a day that runs away from you, and alerts before a plan window does" — a metaphor for a mechanism, and the verdict on it was "unclear". A banner has one line to name something the reader can picture, so it now says what the modal behind it goes on to explain: sessions pause when the day crosses a limit you set.
v0.31.02026-08-27

The share card looks like the dashboard, not like an advert

  • —What people post is a screenshot of the thing working; the old card was a poster with one enormous number on it. It now carries eight tiles — today, the week, the month, all time, the daily average, tokens, cost per million and the cache rate — over two breakdowns: where the money went by model and what it went on by kind of work, each with a percentage beside the figure.
  • —The share button can be found. It was the word "share" in 11px grey in the header, between "feedback" and "status". It is now an accent-bordered button in the ALL TIME panel, beside the figures it offers to post, and it is always there — whether someone's numbers are worth posting is their call.
  • —A nudge when there is something worth saying, separate from the button. Three occasions, each reaching a different kind of user: a round number crossed, the first full week of data, and a week clear of every week before it. A record needs to beat the previous best by a fifth — while usage grows, "highest ever" is true nearly every week, and a prompt that fires every week is one people stop reading. At most once a week, dismissible.
  • —Thin bars on the share card were drawn as hooks. A corner radius wider than the bar itself: haiku at $0.59 is two pixels beside opus at $5,338, and 3px rounding turned it into a squiggle.
  • —The card had lost its caveat while being rewritten. A dollar figure posted without "not a bill, and not money saved" reads as a bill somebody paid.
  • —Model ids from a gateway carried a slash — minimax/minimax-m3 — which on an image about to be posted is indistinguishable from a repository path.
v0.30.12026-08-27

Shift+Enter in the terminal, for prompts longer than one line

  • —A terminal cannot tell Shift+Enter from Enter — both are carriage return, ASCII 13, and have been since the teletype — so typing a numbered list submitted the first line and threw the rest away. Terminals that support multi-line prompts send CSI u instead (ESC [ 13 ; 2 u), which is what Claude Code listens for; xterm.js does not send it by default, so Caprock now does. Reported by the user who noticed it works in a normal terminal and not here.
v0.30.02026-08-27

Paid features unlock from a licence key

  • —A string carrying its own expiry, pasted into settings, checked on the machine — no signature, no online validation, no machine binding ([ADR-022](.ai/08-decisions.md)). The binary is Apache-2.0, so any check is deletable in five minutes; what an offline check buys instead is that a paying customer's features cannot fail because of a plane, a proxy, or a server of ours being down. Seven days of grace after expiry, announced in words, because a late renewal is a bank's timing rather than a decision to stop paying.
  • —GET /v1/premium reports the licence state alongside what the plan costs, so every surface that mentions the paid version reads one answer.
  • —A lifetime purchase, $100 once. An ordinary key with a distant date rather than a "never expires" flag: a flag would need a second code path in the daemon, and the design is one path.
  • —caprock license — show, set, clear, and issue keys from the terminal. Issuing exists because the Stripe webhook was the only thing that could mint one, which left no way to serve a customer who paid another way, a refund reissued, or a friend. license set refuses a key that will not work rather than storing it and leaving someone to wonder why nothing happened.
  • —Every paid feature is visible in the product, in the place it will occupy, behind glass, with one click to paying. Paywall.test.tsx enforces the rule that makes this honest rather than hostile: a lock may only cover a feature that does not exist yet, and never a panel showing measured data.
  • —Third-party models are priced for everyone, and Premium lost a feature. DeepSeek, MiniMax and OpenAI arrive through OpenCode as usage nobody could cost — $155 of the author's own spend sat outside his total. Adding the providers' published prices was cheaper than building a paywall around them, so providers is no longer a paid feature: charging for something the free version performs is how a paid tier becomes a hostage.
  • —The unpriced warning fired on turns with nothing to price. It tested for tokens being present rather than greater than zero, so turns recorded with explicit zeroes were reported as usage outside the total. The dashboard was telling users money was missing when none was.
  • —The same model reported by two routes was two rows, one unpriced: minimax/minimax-m3 from OpenRouter and MiniMax-M3 from the direct API. Gateway vendor prefixes are stripped.
  • —A quick chat was labelled 2026-08-26-212735 — the directory name that keeps two chats from colliding, shown as an identity. It reads chat · Aug 26, 21:27 now, and a migration renames the ones already stored.
  • —"Stuck in a loop … open" landed on the session's default view. For a loop three hours old that is the wrong end of a long list. It opens the timeline at the moment the repetition started.
v0.29.02026-08-26

Share from anywhere

  • —A share control in the header, on every screen, offering the week, the month or all time. Where the browser supports it the image goes straight to the operating system's share sheet; where it does not, the card is saved and the post opens with the text already written — because a browser cannot attach a picture to a tweet, and a button that appears to post and does not is worse than one that says so. The card carries caprock.dev, so the link travels with the image.
  • —Paid features are shown where they will live, locked. The daily spend cap occupies its real place on the Cost screen, behind glass, with your own figures under it and one click to paying. Inert and marked: a preview that responds to a click is a preview that lies.
  • —The paid version is mentioned on Now. It was deliberately kept to Cost and Lifetime to avoid interrupting anyone at work, which meant a user with no loop and no runaway session never learned a paid version exists at all.
  • —The ALL TIME panel was visibly slower than everything around it. GET /v1/history?range=all took 0.76–1.17s against 0.15s for the rest of the screen, and ToolDistribution was 60% of it: the index it matched carries (kind, ts) but not tool, so ~80k rows were read from the table for one column. A covering index makes the same plan covering — 139 ms to 46 ms warm, and the cold figure was 1780 ms, which is what a person actually waits through. Migration 0016.
  • —The landing page can be installed from. Its final call to action offered caprock up — the command you run after installing — so a reader who scrolled the whole page to decide yes was handed the one instruction that does nothing on a machine without Caprock on it. It now carries the real install, defaulting to the visitor's platform.
  • —The terminal has a strip of its own. It is the reason the first full-time user could drop the Claude Code IDE, and the page had never mentioned it.
  • —The teams page is three sections instead of ten: what you get, where it runs (your VPC — counters leave a machine, content never does), book the call.
v0.28.02026-08-26

Quick chat

  • —A session with no repository: click and start asking. The spawn dialog demanded an absolute path before it would do anything, which is the right question for work and a wall in front of "look this up for me". Caprock makes a directory per chat under <data_dir>/chats/. One per chat, not one shared folder — Claude Code keys a transcript by working directory, so a shared one would collapse every conversation into a single project row with a single transcript.
  • —create on POST /v1/agents, and a checkbox for it in the spawn dialog: starting a new project no longer means leaving the dashboard to make the folder in a terminal. Opt-in, one level deep, under a parent that already exists — a typo in an absolute path fails loudly instead of materialising a chain of directories somewhere you have never been.
  • —Plan-limit alerts. A warning at 90% of a 5-hour or 7-day window, high severity from 95%, with the reset time. Not at 85%, where the Cost screen turns amber: an alert that fires wherever a colour changes is one people learn to scroll past. A window whose reset clock cannot be believed raises nothing, because an alert built on a stale reading would never clear.
  • —A share offer on a rhythm — weekly and monthly, drawing the same card over those periods. It previously went loud only at a money milestone, so anyone whose spend never lands on a round number was never actually asked. The offer says what the image contains before anyone clicks: totals only, no project names, no paths, no prose, saved locally, uploaded nowhere.
  • —The paid version is mentioned in the dashboard, for the first time: a line on Cost and Lifetime that opens with what a day costs on this machine, and a note beside a loop or a session that spent a lot for nothing. Clicking either opens a dialog about that feature, with the price and both a way to read more and a way to subscribe. Never on Now, never over your work, never on an empty dashboard, and dismissible for a month.
  • —GET /v1/premium — what the plan costs, served by the daemon so no price is hardcoded in the UI. It ships in the binary either way (rule 4 forbids fetching it), but it lives once in Go, and a test reads the site's pricing file and fails when the two disagree.
  • —The terminal rendered every glyph in the fallback font. xterm.js paints to a canvas and was handed the literal string var(--font-mono), which a canvas context cannot resolve — invisible in Latin, unreadable in Cyrillic. Two more defects in the same place: the character cell was measured against the fallback and never re-measured, and no subset was ever fetched, because subsets load when a matching character enters the DOM and canvas text never does. All six subsets JetBrains Mono ships are now requested by name.
  • —One directory counted as two projects. Claude Code records repo_root only when it resolves a checkout, so a session started where git could not answer got its own row — labelled with its full filesystem path, since a label is derived from the path when there is no project name.
  • —Plan limits were on a screen nobody looks at for them, and the panel vanished entirely when there was no data, so the one place that could have explained the absence showed nothing. They are now also a line on Now, and the panel says where the data comes from when it has none.
  • —The control that starts a session sat at the bottom of Now in 11px grey, beside a checkbox and a timestamp. It is at the top, at the size of an action.
  • —Screenshots leaked the machine's directory layout. The scrubber renamed projects but not the directories above them, and the Cost screen prefixes parent segments when two checkouts share a basename.
v0.27.42026-08-26

Hooks never fired on Windows

  • —Claude Code runs command hooks through bash, which reads each backslash as an escape, so C:\Users\…\caprock-hook.exe reached the shell as C:Users…caprock-hook.exe and every hook failed. The dashboard still filled from transcript tailing, so nothing looked broken — the only symptom was a "command not found" line in a Stop hook.
v0.27.32026-08-26

Half of the Windows hook fix, superseded by 0.27.4 an hour later

  • —It quoted the shim path, which stops a space splitting the command but leaves backslashes to be eaten by the shell — so hooks still failed on Windows. Upgrade past it.
v0.27.22026-08-26

The lifetime breakdown is its own panel, between the activity feed…

  • —The lifetime breakdown is its own panel, between the activity feed and the session rows. It spent two releases in the wrong place: hidden inside the all-time line nobody found it, and opening it there pushed Today and the live pulse below the fold. Both were the wrong position rather than the wrong default — these are lifetime figures, not something to check between glances at the pulse, so they sit below the live panels where the screen has room and nothing competes with them. The all-time line is one row again.
v0.27.12026-08-26

The tool and model breakdown was hidden behind something that did not…

  • —The tool and model breakdown was hidden behind something that did not look like a control — a toggle styled like the muted caption beside it, same size, same colour, no border, so it read as text and went unfound.
v0.27.02026-08-26

The tool and model breakdowns open on Now

  • —The all-time line offered them behind a link to another screen, which is a strange thing to do with data already in the response that drew the line. Most-used tools by calls and where the money went by cost, expanded in place; the Lifetime screen keeps the full tables. Closed by default, since the total is the point of the line and this screen is mostly live panels.
  • —A spend cap is offered beside the figure that argues for it. Someone reading their own lifetime total is, at that moment, the person most likely to want a limit on it. It stays a link at the weight of the one next to it — a dashboard is not a checkout, and the price lives on the page it points at.
  • —Sessions are full-width rows rather than tiles. Three to a viewport made a session's own figures reference material, small enough that you read the project name and moved on. A session is the unit this screen is about, so each one now gets the width and its numbers come up to the size of the ones in Today.
v0.26.02026-08-25

The all-time total, on the screen you keep open

  • —A hundred and twenty-nine sessions across fifty-eight days at ten thousand dollars of usage is the figure people repeat, and it lived two clicks away — so someone who opens the dashboard daily could go weeks without meeting it. It now rides above Today as one line: the total, the three figures that make it mean something, and a link to the rest. Deliberately a summary rather than the screen copied upward; the tool table and the model mix are a screen's worth of reading and stay on their own.
  • —What the money went on, without leaving Now. "Which repository" and "what kind of work" are two halves of one question, and the second half only existed on the Cost screen. The projects panel now carries both, on the range already chosen there and from the summary it already fetches — no second request, one control for both answers. It renders nothing when the linkage behind it is too thin to mean anything: on a today range most tool calls have not yet been attached to the turn that paid for them, so a bar drawn from that would describe a gap in the data rather than the work.
  • —The History tab is now Lifetime. It promised a log and held a total — the panel inside it already called itself Lifetime, and the tab now says the same. Renamed in the docs and in the screen name feedback issues carry.
  • —Every screenshot recaptured, showing the calendar view, the header's active-tab pill, the plan-value tiles and the work-mix strip — a day of work none of the published images had.
  • —A machine with one busy session and one idle one wasted a screen of space. Each state opened its own three-column grid, so "Active · 1" took a row and left two thirds of it empty, then "Idle · 1" did the same below. The cards now flow through a single grid with each run's label above its first card; grouping reads the same and stops reserving a row per group.
  • —The screenshot anonymiser was written and never called. Every published image had been anonymised by hand instead — which happened to work, and would have stopped working the first time someone forgot. It now runs before the capture and a failure aborts it. It also rewrites the activity feed's text, which the path rewriting never touched: the feed prints whatever the sessions actually did, and on a working machine that is client material. Only five fields reach the screen (internal/narrate), so those are replaced and the rest of the payload is left alone.
v0.25.12026-08-25

The header said nothing about which screen you were on

  • —The active tab was a bg-panel-2 tint against the header's own bg-panel — a few percent of lightness apart — with its label one step of grey brighter than the rest, so on open there was nothing to read. It is now the filled accent pill the agent filter already uses: a solid block of colour is recognised before any text is. The other labels move up to full strength from the muted grey the whole row had been sitting in.
  • —A colour class on a link was silently ignored. a { color: … } carries the same specificity as a text-* utility and was declared after them, so it won on order — the first attempt at the tab above rendered amber text on an amber pill, and the same trap waited for any link anyone tried to colour. The default is now at zero specificity, so a class always takes over.
  • —The feedback dialog asked three questions before you could type one. "Something is broken / missing / unclear" put three sentences in front of the box, and at a glance the buttons were near-identical because the distinguishing word came last in each. One word each now — bug, feature, unclear, other — with the catch-all there so nobody stalls deciding which of three a thought belongs to.
  • —"One sentence is enough" was set in the same faint grey as the fine print below it, so the line that decides whether anyone writes anything at all read as a footnote. The note under it now says the same thing in one line instead of three.
  • —The footer's premium link was unreadable and unclear. "what should paid add?" was too quiet to see and said nothing about where it led; it reads premium now, at a weight that leaves the team line as the only offer on the row.
  • —An install prompt for the agent you already have open. Everyone installing Caprock has Claude Code running in a terminal, and pasting eleven lines is less work than deciding whether you have Homebrew and what Windows does instead. Not a second install method — underneath it is still brew, scoop or a release binary. It names its sources rather than saying "install caprock", since an unrelated package of the same name sits on PyPI, and it says caprock up --yes: consent for the hook is refused rather than assumed when stdin is not a terminal, which is what an agent's shell provides, so without the flag the install completes and the dashboard never sees a session. In the README, on the site, and in /install.md.
v0.25.02026-08-25

The month reads as a calendar

  • —A week per row, one square per day, shade carrying cost — so the rhythm of how you actually work is visible without reading a figure. Thirty bars in a row answer "how much on the 14th" and hide "I do not work Sundays". Shades are cut on a square-root scale, because a few heavy days set the maximum and linear buckets would leave every ordinary day in the palest step, making a busy month look empty. The bars remain, one click away, for when the amounts are the question.
  • —The daily bars have a scale. They are normalised to the tallest day, so without a labelled line a bar's height meant nothing and a $20 day could not be read against a $200 one. Two or three rules, rounded to a 1/2/5×10ⁿ step: a line labelled $237.83 measures one particular day rather than serving as a ruler for the rest.
  • —The live pulse was empty after a daemon restart, while agents were working. Stopping the daemon ends every live session, and ended was permanent — the session row kept it whatever arrived afterwards — so a working agent stayed marked ended until it happened to start a new session. Anyone upgrading Caprock hit this, on the one screen whose entire job is showing what is happening right now. ended is now sticky only against events no newer than the one already stored, which keeps a finished session's transcript from resurrecting it while treating a session that is still emitting as what it is.
  • —The plan-value panel wasted most of its width and buried its own headline. Two columns on a full-width panel left two thirds of the row empty, and the multiple — the figure the panel exists to deliver — was set as a word inside a sentence at body-text size. Three tiles now read left to right as their own sentence: you pay this, the same work costs that, which is this many times over.
  • —A model's snapshot date broke the column it sat in. claude-haiku-4-5-20251001 was the only name that wrapped to a second line; the trailing date is now elided to … with the full id on the row's tooltip, where someone reconciling against an invoice can still read it.
  • —Tile captions sat at different heights across a row. One caption wrapping to two lines dragged its neighbours off the shared baseline, so they read as text stuck to the underside of a number rather than as a row.
  • —Spend from a filtered-out session appeared under the other agent. The projects list has an "orphan" row for spend whose session was deleted; under a filter, excluded sessions looked deleted, so that row collected the other agent's money and showed it unlabelled.
  • —Submissions could be lost silently. Both /api/waitlist and /api/feedback fired their notification and never read the response: a missing chat id, a network error and a 403 from a revoked token were all indistinguishable from success. Team leads now record the failure onto the stored record; feedback, which has no store behind it, returns 502 so the form can offer the email address instead of thanking someone for a message nobody received.
  • —The activity feed told OpenCode users to start claude when it was empty under an OpenCode filter.
v0.21.02026-08-25

An agent switch on the Now screen

  • —all / claude / opencode, in the middle of the Today header, on a machine that runs both. It applies to the whole screen — today's totals, the live pulse, the activity feed, the projects list and the session cards — because a filtered list beside an unfiltered total is how a reader ends up quoting a number that means something other than what the heading says. Totals filter server-side via GET /v1/stats/summary?agent=; an unrecognised value is a 400 rather than everything. Rows from the second agent carry a small oc mark.
  • —caprock status reports the OpenCode reader. A machine running both shows those sessions mixed together, so "no OpenCode sessions yet" and "OpenCode is not being read at all" looked identical — and someone who upgraded for the feature had no way to confirm it was working.
  • —Per-directory cost was silently empty for OpenCode sessions. touch_dir is derived from the event payload by the store rather than trusted from the caller, so that no writer can supply a hand-made value; the OpenCode ingester emitted its own field names and every tool call was stored unplaced. The payload is now shaped like a Claude Code hook payload, which also makes work-kind classification and narration work unchanged.
  • —Caprock's pricing table was applied to unpriced OpenCode turns. The suppression relied on a cost already being present, so a turn OpenCode had not yet priced acquired a figure from different arithmetic — one column holding two costing methods, with nothing on screen to say which produced a given row. No event sourced from OpenCode is priced by Caprock now.
  • —Caprock now watches OpenCode sessions too, on the same screens. A machine that runs both agents had its spend split across two tools that each saw half of it; sessions from either now share one stream, tagged with which agent produced them, and project cost sums across both. OpenCode keeps its own SQLite database with cost, tokens, directory and model already computed, so the import needs no shim, no settings file to modify and no transcript parsing — and its cost is carried across rather than recomputed, because two arithmetics over the same tokens would disagree about one session's total. The database is polled every five seconds and opened read-only; events are keyed on OpenCode's own identifiers, so re-reading a session stores nothing. Verified against a real installation: 70 sessions, 19,236 events and $156.28, matching the source exactly. Live streaming and session control are not built (.ai/16-opencode.md).
v0.20.12026-08-24

The Projects panel grouped by directory, not by repository

  • —The label was the basename of the session's cwd, so one repository showed up as several rows (caprock and ui), a subdirectory posed as a project (app under amarketer), Caprock's own agent worktrees became projects (worker-1), and two unrelated paths ending in the same segment were silently summed into one row — on the owner's own database, two different testrepos and two different repos. A row is now the repository a session's cwd belongs to, resolved by walking up for .git (following a linked worktree to the repository that owns it) once per directory at ingest, and stored on the session so historical rows keep a stable label even after their directory is deleted. Existing databases are backfilled on first open.
  • —The live pulse's cost tiers were indistinguishable in the light theme. The bar colours were hardcoded rgba lifted from the dark palette, so on a white panel around it and well above it composited to a contrast ratio of 1.05 — the same orange to any eye — and the legend named a distinction a light-theme viewer could not see. Every bar was also nearly invisible there (1.46–1.68 against the panel), and the idle hairline sat at 1.13. The tiers now resolve through the design tokens the rest of the dashboard already uses, each with its own alpha, so each theme supplies its own hue: on dark the tiers climb in brightness, on light they climb in depth, because light's --color-accent-strong is darker than --color-accent rather than lighter. Measured from the rendered canvas in both themes, the worst adjacent-tier separation is now 1.93 in light (from 1.05) and 1.59 in dark, with every tier above 1.9 against its panel. The dark theme keeps its palette and its reading order.
  • —A theme switch left the pulse painted in the old theme's colours. The canvas repaints on data change and on resize, and the colours are read from the stylesheet at paint time, so flipping the theme recoloured every panel except the one whose colours had just changed — until its next event arrived. It now repaints on the theme attribute too.
  • —The pulse legend could drift from the bars it explains. The swatches were a second hardcoded copy of the tier colours, and had already drifted: the idle chip was drawn at alpha 0.35 while the canvas drew the hairline at 0.16. Both now come from one definition, so the key cannot say something the chart does not do.
  • —caprock report prints your measured usage in a form you can publish. The figures behind the launch numbers were hard-coded constants read on one day, which is the one failure mode a page premised on honest numbers cannot survive: they go stale silently, and nothing says so out loud. The command re-reads the same measures from the same capture — total at API list prices, the plan multiple, turns, sessions, projects, the cache hit rate and what the cache cut, the date window with its active-day count, the top projects and models, and the pricing-table version — so a published figure can be regenerated rather than remembered. Three shapes: a block for a post, --markdown for a README or anywhere that renders markdown, and --json shaped to regenerate the site's facts block. It reads the running daemon over the existing GET endpoints and issues no writes; with the daemon down it says so rather than opening the database behind its back with a second copy of the aggregation SQL.
  • —The report carries its caveat inline, on every shape. The number is large and has a dollar sign, and shown without a qualifier it reads as an amount owed — the same misreading CostBasis was built for on the dashboard. The wording is the dashboard's own rather than new phrasing, so the CLI, the dashboard and the site say the same thing: on a flat plan it is what the same work would have cost through the API, not a bill, not a discount received, and not money back; on metered billing it is approximately the actual cost and never a saving; with no plan stated nothing is claimed at all. The caveat is the second line rather than a footnote, so quoting the headline visibly cuts something off.
  • —The multiple is refused rather than guessed when there is no fee to divide by. A flat plan with no monthly price stated has no denominator, and the report says so instead of printing a number. It is also computed against the plan fee prorated to the measured window — the same arithmetic PlanValue uses — and prints that fee alongside it, so the division can be checked rather than taken on trust. Comparing a five-week window against a single month's fee would have understated it by a quarter.
  • —The report carries the raw token counts the cache percentages are computed from — cache-read against fresh input, one line in the human output, their own fields in --json, their own rows in --markdown. On the owner's data that is fifteen billion tokens read from cache against a hundred and twenty-eight thousand fresh, which is the concrete half of the cache story: "99% cache hit" is an abstraction, the two counts side by side are a picture a reader who does not think in percentages can hold. They are published as counts and nothing else — no ratio between them is derived anywhere, because the two percentages already state what the cache did and a second derived number would leave a reader working out which one is the real claim. The caveat is unchanged: this adds detail, not a claim.
  • —The prorated plan fee prints with cents. It is the denominator of the multiple beside it, and the whole reason it is printed is that the arithmetic should check out on sight — but $233 rounded from $233.33 made a reader's own division come out at 41.6 against a printed 41.5. Cents cost two characters; a footnote explaining the rounding would have been more noise than the rounding it explained.
  • —A new or nearly-empty database produces an honest sentence, not a wall of zeroes. Day one has no window to describe and nothing to divide, so the report says nothing has been recorded yet and stops. Figures that cannot be measured are absent from --json rather than sent as 0, because a consumer cannot otherwise tell a real zero from a missing value — a 0% cache hit rate reads as a broken cache, not as an absence of data.
  • —The task runner can be turned on from the dashboard, without restarting the daemon (POST /v1/hive). The Tasks screen's off state used to hand over caprock up --hive ~/caprock-tasks to paste into a terminal, which meant the one control the feature needed did not exist in the product. The button opens and seeds the queue directory, starts the board and wires the orchestrator on the running process, and the board replaces the off state in place. It confirms first — naming the queue directory and the repository, and saying that Caprock will be able to spawn Claude sessions with permission prompts skipped — because a silent one-click that begins spawning agents is not something a local-first tool should ship. GET /v1/status now carries suggested_hive and suggested_repo while the runner is off, so the confirmation names real paths rather than a placeholder.
  • —The empty task board says which button comes first. Six empty columns and two buttons do not, and starting the orchestrator over an empty board does nothing. The line appears only while there is no task at all.
  • —Each project row expands to a per-directory breakdown — what ui cost of caprock's total — because "which part of the monorepo is burning the budget" is the question a per-repo number raises and cannot answer on its own. A repository whose work happened in a single directory has no breakdown, since it would restate the row's own total.
  • —The breakdown is charged by which files Claude touched, not by where a session was started. In a monorepo you open the repository root and let Claude edit across /services/api and /services/web; you do not start a separate session in each. Grouping by the session's directory therefore answered "where was the terminal" — and on a real database only one repository expanded at all. Every repository now expands, with rows like /services/api and /internal/store.
  • —A turn counts toward the directory of the most recent file it touched, and work keeps counting there until it touches a file somewhere else. Work happens in stretches: you say "finish /app", and Claude edits a file, runs the tests, reads the output, greps, edits again. That whole stretch is work on /app. Reading, editing or writing a file counts as touching it; running a command does not — so the commands, tests and searches between two edits count toward the directory being worked on rather than falling out of the answer. Each turn's cost goes whole to one directory, never split between two, so the rows still add up to the repository's own total exactly.
  • —Each directory row shows its tokens, its share, and its cost. The share is of the whole repository, including the two rows that are not directories, so the column adds up to 100% and nothing is hidden in the denominator. Shares round down, and a directory with real but very small spend reads <0.1% rather than 0%.
  • —Work whose files live outside the repository gets its own row, labelled "outside the repository" — Claude's notes on the project, agent scratchpads, test-output directories, or another checkout. It is real work on the project and counts toward its total, but it happened outside the tree, so it is not charged to any directory inside it.
  • —The Tasks off state is three steps and a button, not three paragraphs. It carries the same three facts that decide whether anyone would run this — the work happens in a separate git worktree, Caprock runs the checks rather than the agent, and the queue directory is created new — but as a numbered strip rather than prose, because nobody reads a wall of text to decide whether to try a button.
  • —Per-directory attribution was rebuilt after the first rule proved useless in practice. The original rule charged a directory only when every file a turn touched was in it, which was exact and answered almost nothing: on a real 191k-event database it put 87.6% of one project's $1735 into "repository-wide work". Asking what a service costs and being told "we could not tell you" for seven eighths of the money is not an answer. The rule now carries forward from the last file touched, and the same project reads /app 61% ($2090.67), /.ai 6.8%, /app/tests 2.8%. Nothing is split or estimated — each turn still goes whole to one row — but the row it goes to is now decided by a stated rule rather than by whether the turn happened to contain a file edit.
  • —"Repository-wide work" now means one narrow thing: the opening turns of a session, before Claude has touched any file. It is usually nothing at all, and the row is omitted entirely when it cost nothing rather than showing a puzzling $0.00.
v0.20.02026-08-24

Expanding a repository now gives a tree instead of a flat list of…

  • —Expanding a repository now gives a tree instead of a flat list of full paths. It was 43 rows on a real database with nothing showing that /ui/src/components and /ui/src/lib both live under /ui; it is eight, each opening to what is inside it, three levels deep. Spend below that rolls into the deepest row shown and says so, so the parts still add up to the repository total.
  • —The two rows that are not directories — work outside the repository, and work that belongs to no single one — read as part of the same table now rather than as italics from somewhere else.
  • —Clicking a directory row could land you in an unrelated session. The row was never a link: expanding a repository grew the panel past the activity feed beside it, and the click landed there instead. The tree removes the overshoot, and breakdown rows are now explicitly inert.
  • —Plan value spent the right half of a full-width panel on nothing, and its multiple was aligned to the first line of a two-line sentence so the second dropped below it. Plan limits had no padding and ran into its own border. The feedback dialog was sized like a table of figures rather than a box you type a paragraph into.
v0.19.02026-08-23

The Cost screen shows what the money was spent on

  • —commands, code edits, reading, MCP tools, or turns that called no tool at all — beside the existing splits by model and by project. caprock report carries it too, and withholds it rather than guessing when too many tool calls cannot be linked to the turn that paid for them.
  • —Tool calls that name no file — Bash above all, the most-used tool there is — were never linked to the turn that paid for them, so their spend was reported as "no tool call". On a real database that meant 86.5% of a month attributed to turns that called nothing, where the true figure is 9.4%. The links are recovered on first run, in the background, and verified against the transcripts rather than guessed; a call whose transcript is gone stays unlinked.
v0.18.02026-08-23

A task-runner worker could write files anywhere on your machine

  • —Mailbox delivery built its destination from a field inside a message file without validating it, so a message addressed to ../../../x wrote above the queue directory. The author of those files is a Claude session running with permission prompts skipped, so a confused worker reached this with no attacker involved. Validated on both ends now, contained to the queue directory as a second layer, and a refused message is quarantined rather than dropped.
  • —Task ids reached the filesystem unvalidated, so a hand-written task file could read or write outside the queue directory.
  • —caprock could destroy commits on your own branches. Creating a worker's git worktree force-reset a branch of the same name if one already existed. Worker names are predictable and nothing cleaned up after a run, so a second run dropped the first one's commits off the branch tip. It now reattaches to a worktree it already owns and otherwise refuses, naming the branch.
  • —Granting folder trust rewrote the whole of ~/.claude.json to set one field, losing its key order and rounding any integer past 2^53. It now preserves the file, records only the grants it made, and caprock hooks uninstall revokes them.
  • —The settings.json backup was taken once and never refreshed, and nothing could restore it. It refreshes when the file has genuinely changed, keeps the oldest snapshot plus the newest few, and caprock hooks restore exists.
  • —Retention pruning could have deleted every event if retention were ever set to zero at runtime.
v0.17.02026-08-23

caprock report prints your own measured usage in a form ready to…

  • —caprock report prints your own measured usage in a form ready to publish: what it would have cost at API list prices, against what your plan costs for the same window, with the caveat inline so it cannot be quoted without it. --json and --markdown for the other places you would paste it. The plan fee is prorated to the window and printed with cents, so the division is checkable on sight rather than asserted.
  • —In the light theme the pulse's "around it" and "well above it" tiers rendered at a contrast ratio of 1.05 — the same orange — so the legend explained a distinction a light-theme reader could not see. The tier colours were copied from the dark palette instead of reading theme tokens; the legend swatches were a second copy that had already drifted from the canvas they described.
v0.16.02026-08-22

Any page open in your browser could reach the API while the daemon…

  • —Any page open in your browser could reach the API while the daemon was running. The cross-site check trusted a request that carried no Origin header, and browsers omit it on cross-site simple requests — a form post, or fetch with text/plain. Behind that check, unauthenticated, sat an endpoint that takes a command from the request body and runs it. The check is now layered — Sec-Fetch-Site, Origin, Host, and a JSON content type or the daemon's token — and a missing Origin is never trusted. A second hole found while testing the first: the loopback check matched by prefix, so a hostname like localhost.evil.example passed.
  • —The database was world-readable. It holds your prompts and Claude's replies in the clear and was created 0644 while the config beside it was 0600. It and its WAL and SHM siblings are now 0600 on every open. SECURITY.md now says what is stored and how to delete it.
  • —caprock up crashed with a stack trace when settings.json held {"hooks": []} or {"hooks": null} — the first command a new user runs.
  • —Tokens from a model that is not in the pricing table displayed as $0.00, indistinguishable from free. The unpriced volume is now shown, with the models that caused it.
  • —A fatal ingest failure was logged and swallowed: the daemon reported healthy and the dashboard said "no sessions yet" forever.
  • —A new install opened on a screen of zeroes whose only coloured element was a cache warning about a cache that had never been used.
  • —The explanation for a missing claude binary was unreachable, the Tasks screen showed "Error: 501 Not Implemented" instead of the server's actual message, and caprock statusline install exited silently when it had done nothing.
v0.15.12026-08-22

The task runner turns on from a button on the Tasks screen

  • —no restart, no command to copy into a terminal. It confirms first, naming the queue directory and the repository, because this is the one feature that spawns Claude sessions with permission prompts skipped.
  • —The Tasks screen explains itself in three steps instead of three paragraphs: you write a task with the commands that have to pass, Caprock runs one session per task in its own git worktree, Caprock runs the commands and only green is done. An empty board now says which button comes first.
v0.15.02026-08-22

Orchestration is now presented as what it is

  • —an unattended task runner with a test gate. A task with an assignee shows its branch, and opens to the diff, the checks that passed, and the git command to take the work — none of which the product showed before, which is why the feature was hard to explain. The README documents it, a fresh hive seeds a README and an example task, and caprock status says which hive is active.
  • —caprock task create, and POST /v1/orchestrator/stop to stop the orchestrator and every worker at once.
  • —A task with no done_criteria reached done without being checked, contradicting the promise on the screen above it. Criteria are now required, and a task without them escalates instead of passing.
  • —Three ways to spend money unattended with nothing to stop it: the orchestrator had no forced-continue limit, the wake loop had a throttle but no ceiling, and an over-budget task was parked as a file while its process kept running. All three now stop.
  • —A task whose worktree was missing was verified against the main repository — which is clean, so it passed. Unverifiable is no longer treated as verified, and verification output is kept so a green result can be audited.
v0.14.22026-08-22

The Projects panel opens on the last seven days instead of thirty

  • —A project worked on for two days out of thirty drew a sparkline that was almost entirely idle, which reads as no data rather than as a burst of work; a week is dense enough for the shape to mean something.
v0.14.12026-08-22

The per-directory breakdown now counts a turn toward the directory of…

  • —The per-directory breakdown now counts a turn toward the directory of the most recent file it touched, and keeps counting there until work moves elsewhere — so the commands, tests and searches between two edits count toward the directory being worked on. The previous rule counted only turns whose every touch was in one place, which left 87.6% of a repository in a single row and answered nothing. Cost is still never split between directories, and the rows still sum to the repository total exactly. Turns whose work was outside the repository — notes, scratch files, test output — get their own row rather than being dropped.
v0.14.02026-08-22

Cost and tokens per directory inside a repository, taken from the…

  • —Cost and tokens per directory inside a repository, taken from the files each turn touched rather than from where the session was launched — so a monorepo shows what each service costs without anyone having to run Claude from inside it. Every figure is measured: a turn counts toward a directory only when all of its work was there, and turns are never split between directories. Turns that ran commands, searched, or built are counted as repository-wide work, named as such and explained on hover, because dividing them up would be an estimate.
v0.13.02026-08-22

Each project row carries a sparkline of when its spend happened…

  • —Each project row carries a sparkline of when its spend happened, bucketed to follow the selected range — 30 days by day, today by hour — so the picture is always consistent with the figure beside it. It replaces the share bar, which restated the ranking the sorted numbers already gave.
  • —Rows state both tokens and cost legibly. On a subscription the dollar figure is a proxy for consumption rather than a bill, and the second number used to sit in the faint chrome tone reserved for timestamps.
  • —The breakdown bar inside an expanded repository scaled on the first row, assuming it was the largest. Rows are sorted by cost while the bar measures tokens, so another row could render past its track — reproduced at 900% width.
v0.12.22026-08-22

Projects are grouped by repository instead of by directory name

  • —Work in ~/dev/caprock/ui counted as a project called "ui"; two unrelated paths ending in the same directory name merged into one row and summed together, so the figure was wrong rather than merely oddly labelled. Each repository expands to show where inside it the spend went, as paths — /, /app, /ui/src.
  • —The login agent still throttled the dashboard. 0.12.1 replaced ProcessType=Background with Adaptive, which measured no better: the process still landed at scheduler priority 4 against a normal 20, and requests still took over a second where the same binary from a terminal took 185ms. Any ProcessType puts the job in a managed band, so the key is now omitted entirely. Re-run caprock service install to pick up the corrected agent.
v0.12.12026-08-22

The dashboard took over a second to answer every request when the…

  • —The dashboard took over a second to answer every request when the daemon ran as a login service, which made the range switches on Cost and Projects appear to stick: a click highlighted the new range while the numbers stayed on the old one, because the request had not come back before the next refresh overwrote it. The launchd agent declared ProcessType=Background with LowPriorityIO — correct for a watcher, wrong for a process that also serves a UI, since macOS throttles the I/O of anything it has been told is batch work. The same binary answering the same query measured 1.2s under launchd against 185ms from a terminal; ProcessType=Adaptive closes the gap. Re-run caprock service install to pick up the corrected agent.
v0.12.02026-08-22

caprock service install|uninstall|status

  • —registers the daemon with the OS's own login supervisor so it survives a reboot: a launchd agent on macOS, a systemd user unit on Linux, a Startup-folder script on Windows. All user-level, nothing written outside your home. The service runs with --no-hooks; hook and statusline registration stay interactive consent decisions.
  • —Budget enforcement on the task board. budget_usd was validated, stored and rendered red past the limit while nothing compared cost against it; the reconciler now parks an over-budget task in needs you with the reason attached.
  • —Finished tasks no longer accrue cost forever. The router opened a task's cost window on the session id and verification closed it on the agent id, so the window never closed — and an open window has no upper bound, so a completed task went on absorbing everything that session spent afterwards. A $0.42 task billed $9.42.
  • —Verification can no longer strand a task. Both the success and failure paths guarded their transitions in a way that could silently no-op, leaving a task nobody could move and the next verify erroring outright.
  • —The model mix is answered from an index. idx_events_cost_cover carries model but leads on kind, so a range query filtering only on ts read the table for every matching row — 146ms against 56ms over 30 days on a 190k-event database. Every /v1/stats/summary computes this, and the dashboard asks for one on an interval, so overlapping requests could queue faster than they drained and leave a panel waiting.
  • —The pulse legend names what its colours compare against. The tiers are relative to each session's own median minute, but the labels read "cheap" and "expensive" as though absolute, so a short bar marked expensive looked like a contradiction rather than a quiet minute carrying a lot of context. The hover readout now prints the median beside the minute's own cost.
v0.11.22026-08-21

Answers looked like it had lost your history

  • —The list loaded a fixed 500 notes and stopped, which on a busy machine is half a day — so it showed an answer from 22 hours ago followed immediately by one from 30 days ago, with nothing between. Nothing was lost; the middle had never been fetched. The list now pages backwards on demand, and reaches everything.
  • —An answer did not feel connected to its session. Clicking through opened the session at the top of a timeline holding thousands of events, leaving you to find the passage you had just clicked. It now opens at the moment the answer was written, with those events marked.
v0.11.12026-08-21

"What did it ask?"

  • —a session waiting on you now has a button showing the last thing Claude said, so you do not have to go back to the terminal to find out what it wants. It shows the last complete thought rather than the newest line: roughly a third of sessions end mid-sentence, and "Let me check that…" is not an answer.
v0.11.02026-08-21

A feedback button that files an issue without sending anything

  • —Two clicks and a sentence: pick whether something is broken, missing or unclear, type what you saw, press the button. A prefilled GitHub issue opens in a new tab and you submit it yourself.
v0.10.12026-08-21

Over half the loop alerts were ordinary work

  • —Replayed over 64,733 real tool calls, the detector fired 436 times — and 236 of those were a repeated Read of one file, which is what re-reading after an edit looks like. Tool calls cost nothing, so those alerts could not have been about the budget the feature exists to protect, and noise in an attention surface teaches people to stop reading it.
v0.10.02026-08-21

Live pulse

  • —the shape of the work, one bar per minute. A track per session on Now, showing the last hour. A bar's height is how much happened in that minute and its colour is what it cost, so the shape of a track is the shape of the work: a task that ramped up and finished draws a bell, steady grinding draws a plateau, working in bursts draws a comb.
  • —The pulse said "working" while Claude was waiting for you. It inferred a state from the bars instead of using the one the daemon already knew. The bars describe the past hour; health describes this moment, and they answer different questions.
  • —?newest=1 on the session events endpoint. Paging from the start returns the oldest events, so anything showing recent activity rendered an empty window on a busy session — hours of history, none of it recent, with nothing to say why.
v0.9.92026-08-21

"Active days" was undercounting by a third

  • —The History screen read 21 active days on a database with 32 days of events, because it counted the days sessions were started rather than days with work in them. One session that ran twelve days contributed one. It also got four times faster along the way (0.37s to 1.63s and back down to 0.36s).
  • —A partial settings body wiped the rest. PUT /v1/settings with an empty or short body answered 200 and reset everything: a stated plan reverted to "not stated", and the release-check opt-in switched itself off. The plan decides what every cost figure claims to be, and update checks are the only outbound call Caprock makes — neither may be toggled by omission. Settings are now a patch: a body changes what it names and nothing else.
  • —Tasks were silently lost when created together. Ids were minted from the millisecond alone, so twelve tasks added at once produced four and eight "already exists" rejections. All twelve now succeed.
  • —The tasks endpoint accepted tasks nobody could use — no title, a hundred-thousand-character title, a negative budget, or 1e308. Each is now refused with a reason.
  • —An unknown API path answered 200 with a web page. The dashboard is served from /, so anything unmatched fell through to it — a caller that mistyped an endpoint got a success and HTML, then failed elsewhere parsing it as JSON. Unmatched /v1/ paths return 404 with a JSON body; the dashboard's own deep links still work.
  • —?days= out of range returned the wrong total. Asking /v1/stats/daily for everything clamped to the default of 30 days rather than the ceiling, so the answer was a month with nothing to indicate truncation. On a real database that disagreed with the summary endpoint by $1,603.
  • —A query that stopped early passed for a complete one. Three loops ignored the error a partial scan reports, so a cancelled or failed read returned what it managed to fetch as though it were everything — twice followed by an update that changed every matching row.
  • —"Files touched" is labelled honestly. It sums distinct files per session, so a file edited in three sessions counts three times; the tile says so instead of letting the number pass for a count of distinct files.
  • —Test coverage of internal/ is 80.4%, with the packages where a bug reaches the user's own machine — the hook shim and the process manager — covered first. Two real defects surfaced there: a data race closing a PTY, and a clean exit reported as an error. See .ai/13-testing.md.
v0.9.82026-08-20

Every cost figure now says what it is

  • —Shown to five readers, three took a large dollar number for a bill. The explanation existed on one screen and nowhere else — History said API-equivalent, which is internal jargon, and Cost and Session said nothing at all beside the number.
  • —The dashboard has a visual hierarchy. Cost leads at the size the number deserves; reference figures step down. Previously the money sat fourth of six at the same size as a turn counter.
  • —Percentages round down. A 99.5% cache hit shown as 100% is a perfect score the data does not support.
  • —Cache hit is no longer permanently green. It sits near 99% forever, so an always-on colour pointed attention away from the money. It speaks up below 90%, where a drop means something broke.
  • —An attention row for sessions with many turns and few files. Every other rule skips ended sessions, so this shape had no surface at all. It is worded as the measurement, not a verdict — reading, investigating and designing all look like this from the outside, and only the person who was there can tell them from a session that went nowhere.
  • —Note search matches the prompt that produced an answer, not only the answer's own text — so searching your own words finds what Claude replied.
  • —The live indicator stopped updating while nothing was happening. It computed the age of the last frame only when a frame arrived, and on an idle machine no frame arrives by definition — so it froze at live · now and kept claiming that for hours. A stale liveness indicator is worse than none, because it is the one people trust.
  • —Cost subtitles no longer truncate on the Cost and Session screens.
v0.9.72026-08-20

The Claude desktop app's plan usage, on the status screen

  • —If you also use the desktop app for work that never touches a repo, one line now says how much of your plan went there: 8% of the 5-hour window · 5% of the 7-day.
v0.9.62026-08-20

The dashboard is three to ten times faster on a large history

  • —Measured on a real 184,000-event database:
  • —Panels show placeholders while loading instead of announcing "No history yet" and then replacing it with real numbers. An empty state now means the answer really is empty.
v0.9.52026-08-20

The running version is in the header

  • —"Which build am I on?" was a question you had to open the status page to answer. When release checks are on and a newer version exists, the chip shows v0.9.0 → v0.9.4 and links to the details. A build from source says "dev build" rather than showing a git describe string, and is never told it is out of date — it is not a published release, so there is nothing to upgrade to.
v0.9.42026-08-20

A session that looped twice showed two identical banners

  • —same tool, same count, same cost — which reads as a rendering fault and makes the attention strip look untrustworthy. There is now one banner per session, showing the most recent alert.
  • —The Tasks screen explained itself in terms of our internal build phases. "Phase 2 runs when..." means nothing to anyone outside the repo; it now says what to do.
  • —The favicon and the in-app logo were different marks. Same colour, different shape, so a browser tab did not read as the product it belongs to. Both are now the triangle from the header, on the site as well.
v0.9.32026-08-20

The README and agent docs now describe what the product actually is

  • —They still listed the screens as they stood several releases ago, and never mentioned two of the most useful things in it: per-repository cost on the landing screen, and Answers — the prose Claude wrote, searchable across every session. Nothing in the software changed; the descriptions caught up.
v0.9.22026-08-20

One bad timestamp could make every session invisible, permanently

  • —A transcript line stamped in the year 9999 rolls past year 10000 in any positive UTC offset, and Go refuses to serialize that — which aborts the encoding of the whole list it appears in. The API then returned HTTP 200 with an empty body, so the dashboard showed no sessions at all while the database held them, with nothing in the logs. Because the event persists, restarting did not help. Fixed at all three layers: responses are serialized before the status code is sent so a failure is an honest error, impossible timestamps are rejected on the way in, and negative token counts are clamped.
  • —A malformed live event could silently freeze the dashboard. A tool call is arbitrary JSON, and one with an unexpected shape threw inside the WebSocket handler — outside React, where no error boundary can catch it. That stopped every later update, so the page sat on stale numbers showing no error. Leaf values are now coerced and one failing listener can no longer starve the others.
  • —A single missing daily cost blanked the whole 30-day chart rather than one bar, while the header still showed a total.
  • —Figures that read as broken: an empty range rendered "NaNh NaNm" in the Avg session tile; ratios could print "$∞" or "Infinity%"; the plan progress bar overflowed its card when a session reported more steps done than planned; the loop banner could say "ran the same call undefined× in undefined min".
  • —"Active sessions" was wildly inflated during a first run. A session is marked active on its first event and reaped later, so a backfill counted every historical session at once — a new user's first impression was a count in the dozens that then fell to one. It is now bounded to the last 30 minutes.
  • —The Answers tab and search could crash on a note with no text, and the Now screen could crash on a session missing its stats or activity.
  • —A range like 90d silently meant "today", so a longer range reported fewer sessions than a shorter one. Any <n>d range now works. The dashboard was never affected — it only offers the four presets.
  • —The release gate could not see a stale dashboard. The built UI is committed to the repository so go install works without Node, and a check existed to catch it drifting from the source — but it was never wired into make check, and it used git diff, which does not notice new files. Since a rebuild emits a new filename, it reported a clean tree while an old bundle sat committed beside it. Caught one commit before tagging this release, which would otherwise have shipped a dashboard missing every UI fix listed above.
v0.9.12026-08-20

Numbers that lied

  • —History's "files touched" ignored the selected range, showing the lifetime total under a "today" heading beside five stats that did move with it. "Avg session" reported 32h 59m — it measures first-event to last-event, so a session left open overnight counted its sleeping hours; it is now labelled "Avg session span · first to last event". The plan-value multiple divided all-time usage by a 30-day fee on the "all" range while the caption claimed 30 days. The bar readout announced "0 sessions" next to a $257 day. A session's token subtitle broke out input and output while the total is 99% cache reads.
  • —Asking for more data returned less. SessionNotes, SearchNotes and EventsAfter fell back to a small default when asked for more than their ceiling — notes?limit=5001 returned 200 rows where limit=5000 returned 2372. All now clamp.
  • —Stale plan limits presented as facts. Values relayed by the status line were stored unchecked, so a five-hour window could claim it resets in 2030. Implausible samples are rejected, and a window whose reset has passed is marked stale rather than rendered as a clock.
  • —Active days were counted in UTC while the daily chart uses local time — "21 active days" beside a 31-bar chart.
  • —Content you could not reach. The session timeline showed only the newest events with no way back, so a long session opened as a peephole onto its last few seconds; there is now a "load earlier events" control. Tool output — a failing test tail, a stack trace — was a 160-character stub whose only escape was raw JSON, and now renders as text. The Files tab capped at 100 while the stat above said 132, with nothing admitting the list was cut.
  • —Controls that led nowhere. The spawn dialog was never rendered anywhere, though the Terminal tab told you to use it — so no session could ever be owned. It is now on Now. The settings screen offered nothing to set. The Tasks empty state explained itself in terms of our internal build phases.
  • —Correctness and races. Starting the orchestrator and running verification used the HTTP request's lifetime, so a browser disconnect could leave a real claude process running with nothing recording that Caprock owned it, and could kill a five-minute verification mid-run. Configuration was read without the lock that guards it while the settings endpoint writes it. Two more places sliced UTF-8 by byte, the same fault fixed in the parser last release — activity phrases and loop samples could render as mojibake.
v0.9.02026-08-20

Answers

  • —what Claude actually said. For a large share of sessions the deliverable is not the diff but the conclusion: "this is done, but I could not verify X; check with the team and then we finish it." Caprock always stored that paragraph and never showed it — the timeline rendered a 200-character slice on one line, so it survived only in terminal scrollback.
  • —GET /v1/sessions/{id}/notes and GET /v1/notes?q=.
  • —Assistant prose was clipped on bytes and corrupted. The transcript parser capped text at 2000 bytes and cut at an arbitrary offset, so non-English prose was clipped at roughly half the intended length and about a fifth of clipped rows ended in a corrupted character — landing hardest on closing summaries, the very thing worth keeping. Text is now clipped on character boundaries at a far higher limit, and a daemon started against an older database repairs the damaged rows once from the transcripts still on disk, rewriting only the text and leaving ids, costs and everything else untouched. On the author's machine that took 452 corrupted rows to 3.
  • —Asking the events endpoint for more than its ceiling returned fewer. A limit above the maximum silently fell back to 500, so a caller requesting everything received the start of a session and could mistake an early fragment for its ending. It now clamps to the ceiling.
v0.8.12026-08-20

Optional update notice

  • —When a newer release exists, Now leads with a line naming it and the exact command for how this copy was installed — brew upgrade caprock, scoop update caprock, or go install …/cmd/caprock@latest — one click to copy. When no package manager owns the binary it links to the release page instead.
  • —GET /v1/update and POST /v1/update/check; update_checks added to /v1/settings.
  • —Engineering rule 4 now states the exception honestly — "no outbound calls" became "no outbound calls except the release check the user explicitly turns on", in the rules, the product doc, and the README, rather than leaving a promise the code no longer keeps literally.
v0.8.02026-08-20

Live activity feed on Now

  • —One column of what every session on the machine is doing, newest first, fed by the existing live WebSocket and seeded from recent history so it is never empty on open. A session list says what exists; the feed says what is happening. Only events worth reading become lines — successful tool results, assistant turns, cost ticks and mail are dropped, because a feed of raw event kinds is noise. Long absolute paths in shell commands collapse so the verb stays visible. Pause to read.
  • —Plan value. What your measured usage would have cost at API list price, against what you actually pay. Caprock cannot detect your plan — Claude Code does not report it, and inferring one from usage would be an invented number — so you state it in a header chip that is one click from being changed. On a flat plan (Pro/Max/Team seat) you get a multiple; on metered billing (API key, Bedrock, Vertex, Enterprise at API rates) no multiple is shown at all, because that figure is approximately your real bill. It never says "you saved $X": without the plan you would not have run that much.
  • —Attention strip. Reports a live loop (with the evidence and what that session has spent), a session that errored, and a session that has been waiting on you long enough to cost you time. There is no "all clear" state — it renders nothing when nothing is wrong. Being expensive is never on its own a reason to fire.
  • —GET/PUT /v1/settings for the stated plan, validating rather than coercing so a typo cannot drive a wrong headline figure.
v0.7.02026-08-20

Per-project spend on Now

  • —The landing screen opens with a Projects roll-up: one row per repository with its measured cost, tokens, session count, a bar showing its share of the largest project, and a green dot when a session is live in that repo. A session list could not answer "what does this repo cost and who is working in it" — this does. The range selector (today / 7d / 30d / all) defaults to 30d, because "today" is empty most mornings and an empty panel reads as broken rather than as an honest zero. Every figure is measured from captured events at API list price.
  • —sessions in each projects entry of GET /v1/stats/summary — the count of distinct sessions that touched a project in the range.
  • —Numbers now carry the cards. A stat's value sat at 17px against its own 10px label, close enough in weight that a card had to be read up close rather than scanned; values are larger throughout.
  • —The orchestration graph left the top nav. With no orchestrator running it can only draw session ids around a hub — topology rather than work — so a permanent nav slot bought a screen that says nothing to a solo user. The route stays at #/graph and the Tasks board links to it while any task is assigned and unfinished.
  • —The graph reads at a glance when it is meaningful: a headline of verified / in-flight / worker counts, each worker labelled with the task it is working on and a plain-language status, and larger nodes and gates.
  • —Inline links no longer underline whole cards. A session card is an <a>, so a global hover rule underlined every label and number inside it; underlining is now opt-in for genuine inline links.
v0.6.02026-08-20

Light theme

  • —A header toggle (sun/moon) flips the dashboard between the dark ops-room look and a light theme; the choice is persisted and, when unset, follows your OS preference. Every screen and the live terminal adapt.
  • —go install works with a real UI. go install github.com/dspv/caprock/cmd/caprock@latest now embeds the built dashboard (the UI is committed and a CI check keeps it in sync), so a Go install is a full Caprock, not a placeholder page.
  • —Plan-limit forecast uses the daemon clock (consistent + deterministic), and the status-line command quotes only the binary path — so a caprock installed under a path with spaces registers correctly.
v0.5.12026-08-20

Windows install via Scoop

  • —scoop bucket add dspv https://github.com/dspv/scoop-bucket then scoop install caprock — no more hand-downloading a zip. The manifest is pushed to the bucket on each release, and the README, install guide, and site now show the Windows path alongside Homebrew.
  • —Project map in the README linking the site, the Homebrew tap, and the Scoop bucket, so any repo in the project points to the rest.
v0.5.02026-08-20

Plan limits, set up for you

  • —caprock up now offers to register the caprock statusline command (the 5h/7d plan-limit windows on the Cost screen) under the same consent contract as hooks — a TTY prompt, or --yes for scripts. New caprock statusline install / caprock statusline uninstall subcommands manage it explicitly. It backs up settings.json once and never touches a status line you set yourself. New users get plan limits without hand-editing any file.
  • —CODE_OF_CONDUCT.md (Contributor Covenant 2.1), linked from CONTRIBUTING.
  • —Honest first-run errors. When the daemon can't start (most often the port is already taken), caprock up now surfaces the real cause — e.g. "port 127.0.0.1 is already in use — try caprock status / caprock down, or --port <n>" — instead of a bare "did not report ready" timeout.
  • —Readable MCP tool names. In History's Tool Usage, mcp__server__tool now renders as server·tool with the full name on hover, instead of truncating.
  • —Release can't ship a red commit. The release workflow now runs a verify job (make check + Windows cross-build on the tagged commit) before goreleaser builds or publishes anything, and writes the Homebrew formula to Formula/ (not the repo root, which brew ignored — a stale formula could otherwise be served).
v0.4.12026-08-19

Homebrew install now includes the hook shim

  • —The formula installed only caprock, not caprock-hook, so a brew install fell back to the caprock hook self-command. Both binaries ship now.
  • —Hook status/uninstall recognize the self-hook form. When hooks were registered as …/caprock hook (no sibling shim), caprock status read 0/8 and hooks uninstall silently left them in place (causing duplicates on reinstall). Inspection now matches both the dedicated shim and the self-hook command, so status is honest and uninstall is clean.
v0.4.02026-08-19

Plan-limit windows

  • —caprock statusline (register as Claude Code's statusLine.command) reads Claude Code's rate_limits and shows your 5-hour / 7-day window usage and reset time on the Cost screen, with an honest "at current pace" forecast shown only when your measured usage would reach the limit before the window resets. Pro/Max only; absent otherwise. The command also prints a compact one-line status (model · context% · cost · limits) and can never break or slow the session.
  • —Orchestrator: workers now stop cleanly. A worker's fire-once mail lingered in its inbox after it acted, so the Stop-loop forced continuation forever and the worker was re-kicked into an endless inbox-poll. The router now archives a consumed message to the agent's processed/ dir once its task moves past the state that message was driving — both a picked-up assign and a verify-bounce the worker has since fixed. Live mail (questions, un-acted bounces) is kept.
  • —Orchestrator: Start is idempotent and race-safe. Starting the orchestrator while a session is already live no longer spawns a duplicate (which leaked the first and raced a second router loop on the same hive); it re-kicks the live session so it picks up newly-queued tasks. A starting guard closes the check-then-spawn window so two concurrent starts can't both spawn.
  • —Homebrew formula, not cask. A CLI binary ships as a formula (casks are for GUI apps), so install is now brew install dspv/tap/caprock (no --cask); the formula also works on Linux Homebrew.
v0.3.02026-08-19

On-disk hive (agents / tasks / mailboxes / append-only ledger)…

  • —On-disk hive (agents / tasks / mailboxes / append-only ledger), single writer, atomic writes, dependency-free YAML.
  • —Tasks board (kanban over tasks/*.md), New-task dialog, approvals.
  • —Orchestrator agent — a real claude session with a hive-aware system prompt that spawns and coordinates workers via mailboxes; the router is a reconciler that spawns a worker per assigned task, runs verification, and wakes idle sessions with unread mail.
  • —Stop-loop autonomy — a worker's Stop hook is answered to force it to keep going while its inbox is non-empty, with a hard guard (N=10) that escalates.
  • —Verification before done — a task's done_criteria run in the worker's worktree; only green checks reach done; red bounces the failing output back (R=3 rounds, then escalate). Destructive commands never run unattended (they escalate to needs_you). Cost is attributed to the task.
  • —New endpoints: POST /v1/orchestrator/start, POST /v1/tasks/{id}/verify; --hive / --repo flags.
v0.2.02026-08-19

Spawn real claude sessions from the UI into an optional git…

  • —Spawn real claude sessions from the UI into an optional git worktree, with a live xterm.js terminal (bidirectional) in Session Detail.
  • —Owned-session controls: pause / resume / kill — only for sessions Caprock spawned; externally started sessions stay observe-only.
  • —Opt-in auto-pause of a looping owned session.
  • —History screen — lifetime stats: cost per project / day / model, tool distribution, model mix, top projects.
  • —New endpoints: POST /v1/agents, /v1/agents/{id}/input|signal, WS /v1/agents/{id}/term, GET /v1/history.
v0.1.02026-08-19

Single static Go binary

  • —a loopback daemon (127.0.0.1:4173) that serves the REST API, a /v1/live WebSocket, the hook receiver, and the embedded dashboard.
  • —Hook plane — a tiny caprock-hook shim registered in ~/.claude/settings.json (non-destructive, backed up, cleanly removable) forwards the core Claude Code hook events; a broken or absent daemon never affects the user's session.
  • —Transcript plane — tails ~/.claude/projects/, schema-versioned parser, tolerant of malformed lines and unknown fields; usage counted once per response.
  • —Normalized event stream in SQLite (pure-Go, no CGO); cost from a versioned pricing table (Anthropic list prices, dated) and cache-savings math.
  • —Loop detector — flags a session repeating the same tool with similar input (K=5 in T=3 min), with an alert banner.
  • —Dashboard: Now (per-session narration, health, plan progress, live burn), Session Detail (event timeline, live git diff), Cost (burn, model mix, per-project, 30-day), History (lifetime stats, tool distribution).
  • —CLI: caprock up | down | status | hooks install|uninstall|status | tasks.
  • —Event retention (retention_days, default off) caps database growth.

← back to caprock.dev

See every coding agent you run.

Who is waiting on you, who is stuck, how close the limit is. Free and open source; nothing leaves your machine.

Download for Mac

Notarized by Apple. One file for Apple silicon and Intel, macOS 13 or later.

Or run it yourself:

curl -fsSL https://caprock.dev/mac.sh | sh

Paste in Terminal: it installs Caprock and opens it.Or with Homebrew: brew install --cask dspv/tap/caprock-app