Changelog

Caprock releases, November 2026.

28 releases, v0.170.0 to v0.184.2, newest first. The newest releases and every other month are on the changelog.

v0.184.2 2026-11-19

cockpit

  • — the Release section's stages are five dots on one line under the version, and only the stage the release is at is named, with its detail: Release PR #952 · CI 1/3, Building · 12 of 23 files, Published · 5 min ago; red when that stage failed. The word strip no longer wraps in a narrow panel, and hairlines set the heading, the entries and the lines below apart.
  • — cockpit: the Spend header is one line from 280 to 460px: the period (start · today · 7d · 30d, Since start on hover) and Chart | Numbers. Whose spend moved to the right of the line under it, this session ▾ or all sessions ▾, a menu of the two; in a 330px cockpit the three switches had wrapped into three ragged rows.
  • — chat: scrolling straight to the top of a long chat just after it opens keeps the older messages it reveals; the first page's trim to the newest messages could run a moment later and take them back.
  • — ui: New agent's project list, with nothing typed, lists the projects you worked in last first, not the ones with the shortest names.
v0.184.1 2026-11-19

cockpit

  • — the folded CI line draws its checks as a bar, one segment per check coloured by its state with a failure at the left edge, followed by a count: 1/1 passed, 11/13 · 2 running · 5m, 1 failed · 12/13. A single check now shows how full the run is; too many checks for a segment each fill the bar in proportion.
  • — cockpit: the session panel's actions sit two by two in short words (Dashboard, the editor, Close tab ⌘W, Stop session…) instead of four full-width rows; each keeps its full name on hover and for a screen reader.
v0.184.0 2026-11-19

ui

  • — folded, the CI section's second line starts under the word CI, not under its chevron, so it reads as part of the section.
  • — cockpit: the right panel's Release section reads as release notes: the stages strip always on top, the predicted version with its bump and why, the changelog entries in their own words grouped by kind, the batch pull request as one via line with its checks, and Not in it yet folded for pushed branches with no pull request. The entries the open batch pull request carries count before it merges, read from its branch as this machine last fetched it, so the version is predicted then too. Folded, the line says the next step: 0.183.1 · 2 changes · next: merge #949. The pull request and agent rows are gone from it: the CI and Subagents sections show those.
  • — cockpit: Spend's period chip is split into whose and when: session | all and since start · today · 7d · 30d (since start, from the session's start, for this session only). This session over today, 7 or 30 days is its own calls in that window, by the hour or by the day, so three tabs on 30 days no longer draw the same chart; a window with no calls says so instead of showing zeros. Each tab keeps both picks, and the label always says both (Today · this session, 30 days · all sessions).
  • — cockpit: Spend's Numbers row draws a figure from 1,000 up with a multiplier and two decimals ($7.92k, $1.25M, 13.71B) instead of cutting it off; the exact value is on hover and for a screen reader.
  • — ui: a pull request a session opened in another repository no longer shows as this repository's. The right panel's Release and the Work board list only the repository's own pull requests; the cockpit's CI block and the Pipeline's This session list still show it, named with its repository before the number (caprock-web #247), and the CI line prefers this repository's open pull request over it.
v0.183.0 2026-11-19

ui

  • — add a second Claude Code or Codex subscription right where an agent starts. In New agent and Quick chat the Subscription picker ends with Add subscription…, and with a single sign-in a quiet Add subscription link sits where the picker would be. It opens the same add form as Settings → Subscriptions in the dialog, with the same free number and Premium path; the sign-in is typed into a shell tab, and the new subscription is picked for that start. Adding still needs the computer; the phone's New chat says where it is done.
  • — sidebar: Running, no tab folds to its one line with a click, Background and all, showing how many run and an amber dot when one waits on you; the choice is kept.
  • — phone: the phone's New chat says where a second Claude Code or Codex subscription is added. Under Who works, and in Start in a project under Runs on (or under Agent with a single sign-in), a muted line reads Add subscription — on your Mac: Settings → Subscriptions. It is a hint, not a button: adding signs in through a terminal and a browser on the computer, and a phone cannot add one.
  • — ci: Rerun jobs GitHub never ran, a switch in Settings → GitHub and right beside Rerun on a job GitHub gave no machine (Always rerun these). Off unless you turn it on. While Caprock runs, it then asks GitHub to re-run that run's failed jobs itself — the same request as the Rerun button, sent to api.github.com with your GitHub token, for your repositories only — when nothing else in the run failed, release runs included. Once per run, remembered across restarts: if GitHub gives the re-run no machine either, it is shown and left to you. The run then says re-run by Caprock on the folded CI line and in the list, with when and what on hover.
  • — ui: with more than one Claude Code or Codex sign-in, Plan limits on Now and Cost lists every one, grouped by agent: its plan, the default badge, how old its figures are, what is left of its 5-hour and weekly windows and when each resets ("reopened 18:20" once a reset has passed), seven days of spend at API price, sessions running on it and, while continuing on a limit is automatic, its place in that order. The Subscription select in New agent and Quick chat, and the phone's Who works rows, say how much is left of each option's tightest window ("62% of 5-hour left"). You still choose: nothing is picked for you by these figures.
  • — app: a file dropped on an open chat in the desktop app now attaches exactly as the Attach button's pick does, whatever its type: the app reads its bytes (only a file from that very drop, at most 10 MB), Caprock saves a copy and the chip shows its short name — not the file's full path, which is how a dropped PDF, Markdown or source file used to go. Over 10 MB it gets Attach's line. A folder is still turned away as before; an app built before this keeps attaching a dropped document by its path.
  • — docs: the README's first paragraph says in one sentence what Caprock is and which agents it runs, so a reader, or an AI assistant asked about it, gets the answer before the screenshot.
  • — ui: the right panel's CI line is drawn the moment you switch to a tab, from what was last read for that repository (muted, with its age), while it is read again; it used to wait for the daemon to compose the pull request's checks for that session, seconds and up to 15 s when measured. The daemon answers such a first ask at once from the pull request it already keeps. Folded, the CI line's check summary shortens in a narrow panel instead of running past its edge.
  • — ui: folded, the Release section now says where the release is — Nothing new since v1.4.0, 1.5.0 · 5 changes waiting · PR #40 green, 1.5.0 · release PR #42 · CI running, 1.5.0 · building · 12 of 23 files, v1.5.0 out · 5 min ago — and unfolded it opens on one stages strip, Merged ─ Release PR ─ Tagged ─ Building ─ Published, the current stage marked.
  • — ui: the folded Changes row of a clean worktree says nothing to commit, without a Review and commit action that had nothing to review.
  • — ci: a CI job GitHub never ran — it waited for a machine, then ended with no step run and no runner, as hosted macOS jobs sometimes do — is no longer shown as a failure. The check summary says ⚠ 1 no runner in muted amber, its hover says the code was not tested, and the folded CI line offers Rerun right there (a click and a confirm; it re-runs the failed jobs with the GitHub connection you set up). A pull request whose only non-passing jobs got no runner is not counted as failed and sends no CI failed notification; real failures stay red.
  • — ci: folded, the CI section is two lines: the pull request it summarizes — number and title — with the time it was read, then its checks with the one action that fits (Fix CI on a failure, Rerun on a job GitHub gave no machine, nothing while running or green). It fits from 280 to 460 px without a third line. The folded line and the unfolded views now count the same pull request's checks the same way, and each open pull request's row shows its checks.
  • — cursor: Cursor chats show their Context reading again. Cursor's newer builds (composer-1 and Auto chats) record only how full the window is as a percentage, not in tokens, and Caprock read only the tokens, so those chats said no context reading yet — or kept an older, lower reading. The Context tile, the cockpit's Context line and the Now card now show Cursor's percentage, with Cursor doesn't record this chat's context in tokens where the sizes go, and a live chat's reading follows Cursor's latest measure after each reply instead of staying on the one taken before it. Chats already imported are read again once to pick it up. A cursor-agent (CLI) chat still has no reading: its hooks carry none.
  • — mail: a reply's original stays marked answered, and a message you mark read, unread, flagged or unflagged keeps that mark, when a sync was already under way. That sync read the server's old flags and wrote them back over yours until the next sync, up to five minutes later.
  • — app: an update no longer opens a stray Caprock Mail tab in the browser. The Mail window reloads while the daemon restarts, and with no daemon running the app took its own page for a link and handed it to the browser; a page on the daemon's address is now refused quietly until the daemon is back.
  • — daemon: the status check answers at once while the database is busy. Once a minute it used to work out the memory figures inside the request, which on a 2 GB database often took over a second, so the desktop app's 0.8-second check gave up and logged the daemon as failing, up to 24 times in ten minutes. Those figures now come from a copy refreshed in the background, a few seconds to a minute old: under heavy dashboard reads the slowest status answer went from 2.2 s to 0.16 s on a copy of that database.
  • — store: faster on a large history. Searching Claude's prose for a phrase it rarely wrote read every assistant turn on the machine; a new index (migration 0085) holds what the search tests, and a reply's prompt is looked up only where it can match. The Cursor overhead report and a session's overhead estimate no longer read every event of every session, the cockpit's call list reads only its candidate rows, and the session search no longer reads every prompt a second time for a capitalised spelling SQLite already folds. On a copy of a 2 GB, 1.09M-event database (2026-11-18, warm medians): a rare-phrase prose search 3.1 s to 0.31 s, the overhead report 454 ms to 202 ms, a session's overhead estimate 70 ms to under 1 ms, a session search 360 ms to 197 ms. Building the index takes a few seconds once, on the first start after the update.
  • — ui: Home's Projects no longer goes back to "Reading your projects…" over a list it has already shown. Opened again, the panel keeps the last list it drew until the daemon answers, instead of waiting seconds on a busy one.
  • — ui: Lifetime's Daily cost chart draws its bars on all again over a long history. The 3px gap between bars did not shrink, so a history of 585 days needed more room for its gaps than the panel had and every bar was squeezed to nothing; the gap now shrinks with the bars, on Cost's daily chart too.
v0.182.4 2026-11-18

app

  • — the Updated to X · What's new line after an update closes itself after 10 seconds, a ring round its × showing the time left. Pointing at it holds the count, and so does the window being in the background.
  • — cockpit: the CI section's read time says now instead of just now and stays on one line, where it wrapped in a narrow panel.
v0.182.3 2026-11-17

ui

  • — a file dropped on an open chat now attaches, as the Attach button does: the same Image N or file-name chip over the field, the same line for a file that can't be attached, and the same input to the agent on Send. On the Chat screen, a session's Chat tab and the Full dashboard's Chat view. In a browser a Drop to attach outline shows while a file is over the chat. In the desktop app an image is read and saved as a picked one is, and any other kept file goes by its real path, as a drop on the terminal does; a terminal under the chat no longer also types it.
  • — pricing: Codex sessions on GPT-6.2 Astra and GPT-6.2 Sol now show a cost. OpenAI released both on 2026-11-17 and Codex lists them first in its picker, but the table had no row for them, so their turns were unpriced. Prices are OpenAI's list prices, read 2026-11-17: Astra $8 input, $0.40 cached, $40 output per 1M; Sol $2, $0.10 and $10; both doubled on input and 1.5x on output over 272K input tokens, with a 1.05M context window. Turns already recorded are priced on the next start. Pricing table 2026-11-17.4.
v0.182.2 2026-11-17

session

  • — Cost, cumulative on a session page plots the whole session and ends at the Cost tile's figure. It summed only the 60 events loaded, so a Cursor chat, whose cost sits on one reconciled row that later work had pushed out of that window, read $0.00 flat beside a priced Cost tile.
  • — mail: Caprock Mail's AI on the DeepSeek API now defaults to deepseek-flash. The old default, deepseek-chat, is a legacy name DeepSeek said it would discontinue on 2026-07-24, and it had no price.
  • — pricing: Cursor's Grok 4.7, 4.6 and 4.5 (with Fast and 500k variants) and Muse Wonder now have rows from Cursor's models page (read 2026-11-17), and Cursor's gemini-3.1-pro and the picker's sonnet-4 map to their Gemini 3.1 Pro and Claude Sonnet 4 rows, so the Cursor overhead's "at list price" column has a figure for them. A Cursor chat is still never priced from the table. deepseek-chat, which DeepSeek said it would discontinue on 2026-07-24, gets no row and is an open question (OQ-29). Pricing table 2026-11-17.3.
v0.182.1 2026-11-17

ci

  • — folded, the CI line shows the checks of the repository's open pull request when the session's own is already merged, instead of just "merged". The session's own open pull request still keeps the line, and the line is drawn only from a read made since the block was shown.
v0.182.0 2026-11-17

cursor

  • — a Cursor chat's turns now end the way a Claude Code session's do, so Now and the phone say "done" or "waiting for you" from what Cursor wrote last instead of going by the clock. Each turn's end keeps how long it ran and how long the model thought, where Cursor timed it.
  • — cursor: the summaries Cursor writes when a chat runs out of room are recorded as compactions, with the context before and after where Cursor measured them. They count in the Context look and the typical share a compaction keeps; they carry no cost, which Cursor does not record.
  • — cursor: chats from Cursor's older format (spring 2025) now show their prompts, replies and tool calls, not only their cost.
  • — cursor: the mode Cursor recorded (plan, ask, auto-run) shows on the session's mode chip, and whether you accepted or rejected a call is kept with its result.
  • — cursor: with Cursor's hooks on, a cursor-agent chat also records its start and end, its subagents, each shell command's output (so its pull requests and pushed branches are found), file edits and how long each tool ran. The added hooks only observe; press Turn on Cursor hooks again to add them to an earlier install.
  • — cursor: edits Cursor named by a path starting with the project's own folder name are filed under the right folder, not one level too deep, and Cursor's MCP calls and newer edit tools carry the names Claude Code's do. Chats stored before are corrected once on the next start, with their cost unchanged.
  • — cursor: a Cursor session's Turns tile counts its replies and names the requests Cursor counted, instead of counting Cursor's running totals as turns.
  • — cursor: the estimate of what Cursor adds to each request is recorded again for chats billed to several models, as an unnamed model.
v0.181.0 2026-11-17

faces

  • — the face menu's set tabs show each set's face with a short name, Face · Duck · Pumpkin · Winter · Dobie, with clear borders, and Home and End jump to the first and the last tab. The dobermans are now called Dobies.
  • — pricing: the daily price watch now also reads DeepSeek, MiniMax and Kimi pricing, OpenCode Zen (for the rows whose figure is Zen's), Cursor's models page (its own models' prices, and the names Caprock cannot map to a priced model), Codex's model catalog (a model its picker lists with no price is drift), the Anthropic, OpenAI and Gemini model lists (context windows, new models, retirement dates) and OpenRouter's model list (new models, as a hint only). Findings that need no change to the table are reported "for information" and never open the "Pricing drift" issue. A new test fails any pull request whose launch menus offer a model the pricing table cannot price.
  • — pricing: the context fill for six Gemini models now uses their real 1,048,576-token window, and the older OpenAI models Codex can run (GPT-5.5, 5.4, 5.2, 5.1, 5, o3, GPT-4.1) show a context fill at all (pricing table 2026-11-17.2).
  • — cursor: Cursor turns now carry the model Cursor billed for them, not the one its model picker happens to show now: a chat billed to one model files every turn under it, and a chat billed to several says the model is not known rather than guess. Chats stored before are corrected once on the next start, with their total tokens and cost unchanged.
  • — cursor: the Context meter of a Cursor chat reads Cursor's own measure of how full the window is (a 1M window in max mode), instead of a figure worked out from tokens that were never one request's size.
  • — cursor: Cursor chats now record what each tool call returned, so they find the pull requests they opened and the branches they pushed (and with them the PR and CI cells), and show failed and rejected tool calls in the timeline and loop alerts, as Claude Code sessions do.
  • — cost: Cursor's spend is read per request, not per usage record. Cursor keeps its cost per chat and model, and Caprock stored it on records that are not turns, so a model of 2,153 requests and $623 read as "7 turns" and "$89 per turn". Those records no longer count as turns anywhere — the model mix, All time, the agents and the Week card — and a Cursor model shows its requests instead: 2,153 requests · $0.29/request. Hermes's usage records stop counting as turns too.
  • — session: a figure an agent does not record reads the same for every agent: "—" with Cursor doesn't record cache reads or writes, Gemini doesn't record cache writes, Hermes doesn't record each request's tokens under the tile it belongs to. Cursor's Context tile no longer measures a sum of requests against a window (or names its model as missing from the pricing table), and a Cursor chat with no cost says Cursor recorded none rather than "price unknown".
v0.180.0 2026-11-17

ci

  • — a daily pricing-watch run compares the Anthropic, OpenAI and Gemini pricing pages with pricing/pricing.json and opens one "Pricing drift" issue when a page lists a model the table lacks, a model the table prices as a different one by prefix (Claude Fable 5.2 as Fable 5), or a different price. It never edits the table; prices are still copied by hand with their source and date.
  • — pricing: new models priced: Claude Fable 5.2 and Mythos 5.2, Gemini 4 Carbon, Argon and Hydrogen, Gemini 3 Flash Preview, MiniMax M3.1 Flash, and the older OpenAI models Codex can run (GPT-5.5 back to GPT-5, GPT-4.1, o3).
  • — pricing: Kimi is priced: Claude Code sessions on Kimi K3 and K2.7 Code now show a cost, including the ones already recorded.
  • — pricing: DeepSeek V4 Pro is priced at DeepSeek's current rate; turns already recorded keep the cost they were given.
  • — pricing: announced price changes are entered ahead of their date, and today's price is shown until the day comes: Gemini 4 Carbon and Argon on December 1, Gemini 2.5 Pro billed as 3.8 Flash from December 1, Gemini 3.8, 3.7 and 3.6 Flash on January 1.
  • — pricing: Claude Fable 5.2 is no longer priced as Fable 5. A model is now matched only by its own name (or a dated or [1m] form of it), so a newer model is shown as unpriced rather than charged at an older model's rate. Turns already priced the wrong way are corrected on the next start.
  • — pricing: long prompts are priced at their higher rate: Claude Haiku 5.5 over 100,000 tokens, GPT-6 and GPT-5.6 over 272K, Gemini 3.1 Pro and 2.5 Pro over 200k, MiniMax M3 over 512K. Earlier turns on these models are corrected on the next start.
  • — pricing: OpenCode, Hermes and Cursor turns keep the cost their own agent recorded, or none, when Caprock reprices or fills in stored turns.
v0.179.1 2026-11-17

deck

  • — the right panel's CI and PRs now also find a pull request by the branch the session pushed, so CI no longer sticks on an older pull request when the new one was opened by a script or without the agent running gh pr itself.
v0.179.0 2026-11-17

ui

  • — two new sets of faces for Claude's portrait: eight for the New Year (Santa, a snowman, a reindeer, a Christmas tree, an elf, a bauble, a Santa duck and a gingerbread man) and eight dobermans in the breed's black, red, blue and fawn coats with rust markings (one with floppy ears, one with its tongue out, a cool one in sunglasses, a sleepy one). Each has its own tab in the face menu and its own pool in Settings → Appearance (A New Year face per session, A doberman per session). Nobody's current face changes.
  • — ui: subagents now draw their faces from every set — faces, ducks, pumpkins, New Year and dobermans — whenever Claude's face is set to a random pool.
v0.178.2 2026-11-16

cost

  • — the Cost screen reads at a glance. The total is whole dollars ($7,562) and the token tiles three figures (31.2M, 13.1B), exact on hover. Per project folds agents' throwaway folders into one row per stem, chats into chats and /, src and the like into other folders; Model mix and Per project show eight rows and fold the rest. Cursor's overhead table puts each figure and its sample count in their own columns.
  • — inspector: the right panel's Release section no longer vanishes after a release: with nothing new since the last tag it says Nothing new since v0.178.1 under its heading.
v0.178.1 2026-11-16

jarvis

  • — Jarvis is back to the 0.177.0 screen. The decision cards, Handled for you, the chat, hold to talk, the model behind it and the automatic continue after a limit reset are withdrawn.
v0.178.0 2026-11-16

jarvis

  • — Jarvis is a chat now. What you write goes to the right agent, starts one where none is free, or answers a question card, and the thread keeps every message and reply. With a brain chosen, your own Claude Code (or Codex, on the sign-in you pick) reads plain speech and answers routine multiple-choice questions for you, never ones about pushing, deploying, deleting, paying, credentials or sending messages; without one, @project routing works as before. When an agent Jarvis gave work to finishes, Jarvis says so with its last words. What a brain run is sent and where: docs/jarvis-brain.md.
  • — jarvis: a question an agent asks in a session Caprock started can be answered from Jarvis in one tap: the digest carries the dialog's options with the recommended one marked, and POST /v1/jarvis/answer presses the option (or types an answer on its Other row) with the keys Claude Code itself reads, only after checking the dialog is still the one on the screen (ADR-096).
  • — jarvis: Handled for you: a log of what Caprock did on its own — a session moved at its usage limit (with Undo) and moved back, a session resumed after its limit reset, ended when done or idle, the morning summary sent — at GET /v1/jarvis/handled and as a live jarvis frame; the digest counts today's.
  • — limits: a session Caprock started that a usage limit stopped is sent "continue" once the limit has reset, through its message queue as Jarvis's Continue sends it. On by default; limit_auto_resume in settings switches it off.
  • — jarvis: Needs you is now decision cards: an agent's question shows its options as buttons, the one it recommends first with a star, and one tap answers it in a session Caprock started; Something else… types a free answer. Permissions and plans to approve are one tap too (Allow, Deny, Approve and start). Each card has Later (an hour) and Dismiss (until the agent asks something new), kept per browser.
  • — jarvis: Handled for you lists what Caprock did on its own today — limit moves with Undo, finished and idle sessions ended, the morning summary, routine questions answered — and the summary adds I handled N things myself today.
  • — jarvis: the command line is a chat. @project is optional: without it the message goes to the project picked on top, else the one you worked in last, and the field shows where (→ caprock) before you send. Replies, Done: when an agent finishes, and replies still coming show in the thread.
  • — jarvis: hold to talk — hold ⌥Space (Ctrl+Shift+Space on Windows and Linux) or the mic, speak, release to send; Esc cancels. In browsers, the iOS app and the macOS app (which now asks for speech recognition once).
  • — jarvis: Jarvis can understand plain words on your own Claude or Codex subscription. A one-time card says what is sent where before anything is; Use my Claude subscription is one click, Choose… picks the agent and sign-in, and Settings → Jarvis has the same choice.
v0.177.0 2026-11-16

cursor

  • — Cursor support. Caprock reads Cursor's Agent and Composer chats from the editor's own database, read-only, with the cost Cursor recorded per chat and model; a chat Cursor recorded no cost for shows its tokens and no price. New agent → Cursor starts cursor-agent with the model you type, and an ended CLI chat continues with --resume. cursor-agent chats you start yourself appear once you turn on Cursor's hooks on the Cost screen, merged into ~/.cursor/hooks.json beside other tools' entries and removable. The Cost screen's new Cursor panel shows Cursor's figures by model and estimates how many tokens Cursor sends per request beyond what you typed, beside Claude Code's figure for the same model; a Cursor chat's page shows its own. Pricing table 2026-11-16.1 adds Cursor's Composer 2.5 models. See docs/cursor.md.
v0.176.0 2026-11-15

daemon

  • — GET /v1/jarvis, one digest of every live agent for the Jarvis screen: what needs you (permission prompts, questions, task approvals, sessions stuck in a loop, most urgent first), what each agent is doing now, how far its plan is, and what today cost. Built from what Caprock already records, with no model call and no tokens spent; a paired phone may read it.
  • — ui: Jarvis, one screen over every agent on the machine (⇧⌘J, Ctrl+Alt+Shift+J on Windows and Linux; the first row of the app's sidebar; #/jarvis in a browser). One line with the projects and the day's figures, what needs you with its action right there (the permission card, a reply through the queue, Approve or Deny, Stop), one row per agent with what it is doing, its plan's progress and its cost today, and a command line: @project text goes to that project's agent through its queue, or starts a new session there when none is running. A project's × mutes it on this screen. Sessions Caprock did not start offer Open only.
  • — mcp: Claude Desktop can add Caprock in one click. Each release carries caprock.mcpb, an MCP bundle that starts the caprock already installed on the computer, and Caprock is listed in the MCP Registry as io.github.dspv/caprock for clients and catalogs that install from there. The bundle holds no copy of Caprock and sends nothing; without Caprock installed it says where to get it.
  • — ios: pairing the iPhone app by its scanner works. The code was read, but the app's web view ignored the address change that carries the link, so the page reloaded back to the welcome screen and nothing happened. The same fix covers pasting a link and adding a second computer in the app.
v0.175.2 2026-11-12

ui

  • — the Updated to X · What's new line no longer sits over the open version panel, which says the same; it steps aside while the panel is open, as the update card already did.
v0.175.1 2026-11-11

ui

  • — a text field given its own width keeps it. The shared field style stretched every field to the full row, so the daily spend cap's dollar field ran across the whole panel and squeezed "Stop the day at" onto two lines.
  • — ui: a long link in a terminal opens whole on ⌘-click (Ctrl+click elsewhere), not cut where it wraps. Claude Code breaks a URL wider than the screen onto the next line itself, and each line was read as a separate link; the rows of one URL are now read together, whichever row you click.
v0.175.0 2026-11-07

machines

  • — a prompt on another of your computers can be approved or denied from the app's notification, as one on this computer can (macOS and Linux); the answer goes to that computer through the link, and a prompt answered there first takes the notification away.
  • — machines: the Cost screen has This Mac / All computers: with another computer linked, All computers shows each computer's spend, priced with its own table, and the total of those that answered. A computer that did not answer is named as not counted, never as $0.
  • — machines: a linked computer's access is now bound to that computer's own key, with nothing to do: a copy of it taken off that computer works from nowhere else.
  • — docs: docs/machines.md says what the hosted relay measured, how and when to run your own relay, and how to measure a link between two computers.
  • — storage: The status strip at the bottom of the window shows how far compression has got (Compressing · 40% · 224.6MB saved), and once it has caught up and packing the file would free a meaningful amount, offers Pack database · ~85 s. A click opens the details: how long it takes and what it frees, both estimated from this database, that sessions keep running and new events wait, and Pack now. Every event that arrives during the pack is recorded when it ends. The pack shows a progress bar measured from what SQLite has written. GET /v1/storage/pack reports the job and the pack, POST starts one, and GET /v1/storage's compression gains the job's progress.
  • — storage: Compress tool output now works on a computer whose agents never stop. It waited for 30 minutes with nothing recorded, which agents running around the clock never leave, so with the setting on all night it compressed nothing. Its steps — and the base64 compaction's, and incremental vacuum's, which give free space back to the disk — now run whenever writes pause for 5 seconds, give way to any write that arrives, and rest nine times as long as each step took, so they use at most a tenth of one core.
v0.174.1 2026-11-07

machines

  • — GET /v1/machines says what each linked computer's channel has cost the relay (traffic: messages and bytes each way since the daemon started or paired it), and bench/peer.mjs --idle N measures a window left open — so the hosted relay's free-plan headroom is read, not guessed.
  • — ui: the Agent panel's header shows the session's portrait again: the face you picked for it (a duck, a crown, a pumpkin…) or its own, large, where 0.174.0 drew only the agent's mark. A click on it still picks the face, and the mark is a small badge on its corner, so Claude and Codex still tell apart. The header is symmetric: the portrait is as tall as the figures beside it, so both columns start and end on the same lines; the subagents' faces sit centred beside the portrait in at most two rows (past eight, seven and a +N); and when the state and the permission mode wrap under a long name they stand at the left, not over the figures. The 1 working · peak 4 line now floats under the faces on hover instead of holding empty room under them.
  • — machines: after the other computer's owner pressed Allow, the sidebar, Now and New agent kept saying "waiting for Allow" until Settings → Computers was opened, and the next look came a minute later. A computer that lets this one in — a socket opens, a request is answered — is now asked again by itself, one that waits for Allow is looked at again after 5 s, 10 s, 20 s…, and the same clears "removed" when that computer knows this one again. In the two-daemon smoke the laptop learned of the Allow about 3 s after it.
  • — machines: a prompt said on another computer while its sessions were being read could miss the one list for up to 10 s: the read from just before it was kept as fresh. Such a read is now answered but not kept.
  • — machines: after another computer's Caprock restarted, this one kept asking it for frames past its new count, got a reset on every reconnect, and lost the prompts said there while the link was down; it now resumes from where that computer stands. A computer that lets the watch in and hangs up at once is waited out (2 s doubling to 60 s) instead of re-dialled every 2 s.
v0.174.0 2026-11-06

plugin

  • — Caprock is a Claude Code plugin. claude plugin marketplace add dspv/caprock, then claude plugin install caprock@caprock, adds the read-only caprock mcp server for every project. Its launcher finds caprock even when the desktop app put none on your PATH, and when Caprock is not installed it says so, with the install link, instead of a bare "failed". It downloads and sends nothing.
  • — ui: the file tab reads more kinds of file. Markdown and HTML get Rendered / Source / Split, Split showing the source and the page side by side, their scrolling kept together for Markdown; the choice is remembered. An HTML page is drawn in a sandboxed frame that runs no script and loads nothing from the network, its own stylesheets and pictures read from the project. CSV and TSV files show as a table, pictures (SVG too) as themselves, and code and config files are syntax-highlighted by a part of the app loaded only when such a file is opened; a file over 256 KB stays plain.
  • — ui: the Agent panel's header puts the pictures at the left and the figures at the right. The agent's mark is a large tile (still the button that picks the session's face), and beside it the subagents' faces in a grid of four columns: the working in colour, the recently finished grey, a +N tile past eight. The small mark on the name row is gone in this look, so the agent is drawn once. Burn, tokens and cost stand right-aligned, under the pictures in a narrow panel. Hovering the grid shows 1 working · peak 4, which opens the last hour's chart, and the three dots that switch looks. This Matrix look is the default and replaces Crew; a viewer who had picked the crew gets the matrix. Faces as chips and Strip stay.
  • — ui: the Agent panel (⌘I) no longer says $0.00 for a session on a model Caprock has no price for, such as Kimi or DeepSeek after a limit moved the session there. Its cost figures read — with the model named on hover; a session with some turns priced shows the priced sum with a + ($1.84+) and says on hover how many turns are not in it.
  • — daemon: "Return after the reset" survives a restart. A session that Caprock moved to another provider or sign-in at a usage limit still goes back once the limit resets, even when Caprock restarted in between; a return more than an hour overdue at the restart is dropped.
v0.173.0 2026-11-06

sidebar

  • — on a large database with Compress tool output off, the sidebar's bottom block says so in one quiet line — for example Database 2.5GB · compressing frees about 1.0GB, both figures from your own database, the saving an estimate from a sample of your tool calls. It shows only when the estimate is at least 300 MB. A click opens the switch in Settings → Storage, lit, and turns nothing on; × hides it for 30 days. Folded, it is a database icon with a dot. It goes as soon as compression is on.
  • — storage: Compress tool output, under Status → Storage, stores tool calls and their results older than a day zstd-compressed, a few hundred rows at a time while Caprock is idle — never at start and never while anything is being recorded. Prompts and replies stay as they are, so search is unchanged. Off until you turn it on; the panel says how much smaller your own database would be (an estimate from 2,000 of your tool calls), and once it runs, how much it is. On a copy of a 2.5 GB database it saved 818 MiB of tool output, and the file went to 1.5 GB after the one VACUUM that follows, which runs only when the disk has room for it. Turned off, the same job restores the rows. An older Caprock refuses to open the database after this upgrade (with a one-line message, changing nothing). Proposed and measured by [@SergeAx](https://github.com/SergeAx) in #672.
  • — ios: scanning the computer's pairing code with the iPhone's own Camera app now opens Caprock for iPhone and pairs it, instead of opening Safari and asking you to scan again inside the app. Only the pairing link opens the app; every other m.caprock.dev address still opens in Safari, and without the app the link opens the phone's page as before. The link is read on the phone; nothing new is sent anywhere. Needs the next App Store build of the app.
  • — phone: m.caprock.dev in Safari on an iPhone shows Apple's banner, Caprock · Open in the App Store, so someone who scanned the pairing code with the camera learns that Caprock for iPhone exists. Only the phone's page carries it; other browsers, the app itself and a home-screen copy never show it.
  • — ui: Settings → Use on your phone offers Caprock for iPhone in a clear row, iPhone? Caprock for iPhone adds notifications, voice and photos. Get it on the App Store shows a QR the iPhone's camera opens the App Store with; before, a small grey link opened the listing on the computer, where it installs nothing on the phone.
  • — cockpit: the Context meter says what a compaction is doing. While one runs — started from Caprock, typed in the terminal or Claude Code's own — stripes move over the bar and the line reads Compacting · 0:16 · usually ~1 min, the time since it began and this machine's median (no "usually" until five compactions were timed; Claude Code reports no progress, so there is no percent). The moment it ends the meter shows the size it left, 2% full · was 332k, the old fill shrinking away once. The empty track is tinted where the bar turns amber (60%) and red (85%), and the quarter marks read the same over the fill and the track.
  • — ui: Settings on a computer is large and easy to read, the way the phone's is: section titles 19px, every option's name 15.5px with its explanation under it, every on/off a switch, fields and buttons at least 34–36px tall, and a step up for all the small print in every section. Nothing moved and no setting changed; the phone layout is as it was.
  • — ui: the Agent panel's subagents take far less room. By default the subagent chips under the figures are now the subagents' faces — in colour and pulsing while one works, grey once it finished — with 1 working · peak 4 beside them, which opens the last hour's chart under the row; there is no framed block. A Strip look draws one line instead: the working faces big, the finished small and grey, and the hour as a small sparkline that shows the peak on hover and opens the full chart on a click. The Crew look stays, with less padding and a shorter chart. The Chart of faces and Numbers looks are gone; a viewer who had picked one gets the new default. With no subagents in the session, nothing is drawn.
  • — ios: dictation in the iPhone app waits up to two seconds for the final words after you stop it, instead of dropping the recogniser at once (the speech plugin 8.4.6, on Capacitor 8.6.0).
  • — cockpit: the Context meter drops to the size a compaction left as soon as /compact finishes, instead of showing the size before it until the model answers again.
  • — ui: attaching on a phone offers two clear choices. The Attach button (the phone's New chat and Save, a chat, and a session's chat on the Full dashboard) opens a menu with Photos, the photo library or the camera, and Files, which opens the Files browser straight away; before, one picker left both to iOS's own sheet. With a mouse it opens the file dialog as before. A HEIC photo is now always sent as JPEG, so one picked from Files is no longer refused. The iOS app names why it opens the photo library, from its next App Store build.
v0.172.0 2026-11-06

ui

  • — Settings → Use on your phone has one line under Connect a phone, iPhone: get Caprock on the App Store, linking to Caprock for iPhone, now on the App Store; it opens in the default browser and sends nothing. How it works names the app as a way to keep Caprock on an iPhone, and the iOS README, docs/app.md and .ai/ no longer call the app ready for review.
  • — telegram: a Morning summary — one Telegram message each morning from your bot: how many of the new mail conversations Caprock Mail's sorting marked important, with the sender and subject of the newest five; the sessions that finished overnight per project, with those that stopped at a usage limit or exited with an error; and the sessions and tasks waiting for you. Buttons open mail, Home and History on your phone when it can reach this computer. Off until turned on in Settings → Telegram, at 08:00 unless you pick another time; a computer asleep then sends it on waking, until noon, and never twice in a day. Settings says what it sends, and Send a test summary now shows it at once. Free (ADR-092).
  • — chats: on a phone, New chat asks who works — Claude on your subscription, another Claude sign-in, DeepSeek or Kimi on your key (only while it is set), Codex — and where: a chat of its own, a recent folder or any other folder under home. The chat then works in that folder, opens as a chat and is listed in Chats, with a line saying what it sees and that only its model's provider gets what it reads. On a desktop New chat still starts at once, and In a folder… opens the same steps. Folders can be labelled Work or Personal from the Where step; the tags show on the Chats list, which filters by them (ADR-091).
  • — ui: the app's sidebar is calmer and says what is open and in which phase. New agent sits in its own band at the top, with Chat, Project and a ⋯ menu (New shell, Search, History, Share your week) in one row under it; the keys are the ones they always had. A new Open section lists exactly your open tabs in tab order, each with its project, its ⌘ key and a coloured phase: working, needs you, idle, done or failed. Projects are folded to one line with a short summary ("1 needs you", "2 running", "3 idle") and unfold to their sessions as before. Recently closed folds to one line. The Today card became one thin line above Dashboard — spend, running, what waits and the fuller plan window — with a × to hide it; Settings or the ⋯ menu brings it back. The Projects | Tabs switch left the heading, since Open lists the tabs; Settings → Appearance keeps the choice of what sits under Open.
v0.171.0 2026-11-03

agents

  • — when a Claude Code session Caprock started hits its plan's usage limit, it can be continued elsewhere: on DeepSeek or Kimi as the same conversation, on another Claude Code sign-in as the same conversation (its transcript copied there, never moved), or on Codex with a summary of the work. POST /v1/sessions/{id}/continue-on; sessions carry limit_hit with the reset time; Settings can move a session automatically and bring it back after the reset (ADR-090).
  • — terminal: at the limit, one quiet line under the session's terminal says which window and when it reopens ("5-hour limit · back at 18:40"), with a small chip for each place it can go on — another Claude sign-in, DeepSeek or Kimi when their keys are set, Codex — and Wait. A click moves it; the tab follows the new session, and a note says where it went, with Undo.
  • — settings: When a limit is reached: Ask me, Continue automatically on an ordered list of targets, or Wait for the reset, and Back on Claude after the reset.
  • — subscriptions: Claude Code on Kimi, on your own Kimi platform key, kept in the keychain like the DeepSeek key, with Moonshot's published settings for Claude Code; its key field sits under DeepSeek's.
  • — plan: Claude's plans carry Anthropic's new names: Max 5× is now Max ($100) and Max 20× is now Ultra ($200), as renamed at the end of October 2026 (claude.com/pricing, read 2026-11-03). The plan picker offers Pro, Max and Ultra, and a plan Caprock's picker stored under an old name reads back under the new one, so the header chip that said Max 20× · $200.00/mo now says Ultra · $200.00/mo. A plan name you typed yourself is left as you wrote it. A verified Ultra account has no five-hour limit, so the Plan limits card shows its week alone, and a five-hour reading left from before the move is no longer shown as a stale window forever.
  • — session: a session on a model with no published price (Claude Code on Kimi, kimi-k3) no longer reads "Cost $0.00". The Cost tile, the session header and the Now card show "—" and "price unknown", or the priced sum with "+ N unpriced turns" when only some turns could be priced, and the cumulative cost panel counts priced turns only. The session API carries stats.unpriced_turns.
v0.170.1 2026-11-02

app

  • — the macOS .dmg opens on its branded window again (in fact for the first time from a release): the background picture, the app and the Applications folder side by side, the 660×400 window. Tauri's bundler lays that window out with a Finder script it skips on CI, so every .dmg release.yml built since 0.97.0 opened as a plain white folder. The image is now made with dmgbuild, which writes the window's settings directly, and the release and CI checks fail without them.
v0.170.0 2026-11-02

ui

  • — the all-time strip on Now is a row of tiles like Cost's Totals — a label over each figure, one even row — instead of a running line that wrapped sessions a day onto a second line; on a phone the Quick chat and New session buttons go under it.
  • — ui: the Agent panel's empty subagent frame says no subagents · 1h on one line instead of wrapping no subagents in the last hour.
  • — app: the macOS installer window is graphite with Caprock's characters round the arrow — Your agents are waiting, Drag Caprock onto Applications — instead of the plain paper ground.

← October 2026 All months

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.

Caprock is a desktop app: it installs on your computer, not your phone.

Or email yourself the link:

macOS, Windows and Linux. Your phone pairs with it by QR code. Install options →

Download on the App StoreDownload on the App Store

On an iPhone, pair it with the app.

Download for Mac

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

Download for Windows

Windows 10 22H2 or later, x64. Windows may ask once whether to run it: click More info, then Run anyway.

Paste it into Claude Code, Codex or any coding agent.

See what it says ▾
Install the Caprock desktop app for me. It is open source (Apache-2.0), a local
app for my coding agents: https://github.com/dspv/caprock

What the install does and how to remove it:
https://github.com/dspv/caprock/blob/master/docs/install-agent.md

Run the install script for my system: https://caprock.dev/mac.sh, /win.ps1 or
/linux.sh. Read it before you run it if you like. Leave my Claude Code settings
alone; the app asks me itself. Tell me when it's open.

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

Or run it yourself:

irm https://caprock.dev/win.ps1 | iex

Paste in PowerShell: it installs Caprock for you and opens it.

Or run it yourself:

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

Paste in a terminal: it installs Caprock for you, no sudo. Or download the AppImage