Repository navigation
0.9.0 — the captain, personas, the evidence gate, and one-click updates
The captain: an experimental voice-and-text agent that runs every project's fleet for you. It is off until you turn it on. Personas let you run any agent as someone. Remote control needs nothing more installed. The fleet admits spawns against the machine's memory, finished workers stop by themselves, and a symbol index (aisquare code) answers where code is defined and used. Work briefs and the evidence gate: a task linked to a brief's requirements is done only on a fresh pass from an asq exec report. Every first message reaches its agent, nothing the fleet types lands in your draft, agents carry far fewer tokens, restart all brings back agents that died, the agent pane keeps up with your typing, the first enabled, signed-in account is the machine default, asq offers each new release and installs it in one click with your fleet still running, every release now carries its own installer, and aisquare no longer packs your codebase or needs Node.
0.9.0 gathers everything since 0.8.2: #201 (personas, the hackathon release train), #240 (the captain), #267 (the agent info line), #272 (tmux control mode), #273 (certifi), and #274. #274 is the follow-ups, with #302 (App settings in tabs), #303 (the simple Spawn dialog) and #346 (the solo-session trim and token metrics) folded in. Its closing reviews came back clean, with no blocker or major, at 0545dc5 and again at 189dffd. Its last gate ran the full suite in an isolated home with the captain on (7439 passed, 0 failed) and off (7244 passed, 0 failed). It merged with all 24 CI checks green, windows-latest included. The release PR adds the versioned installers and the version bump, and folds in six late changes: #405 (the Spawn dialog's follow-ups), #406 (rank 1 is the machine default), #301 (the daily release check and one-click update), #408 (usage reads back off on HTTP 429), #403 (every HTTPS call verifies against the OS trust store) and #402 (the one-liner on main runs the release's own installer). It also brings in main's P0 reliability train (#295, with #409): first messages, typed writes, idle agents and account switching; and its P1 token diet (#296): leaner agents, a token ledger and a quieter manager. It also folds in #399, which removes the codebase snapshot and Repomix with it, #417 (the evidence-gate stack: command reports, work briefs and the gate on task done), and main's P2 train (#297): a symbol index, spawns admitted against the machine's memory, finished workers that stop themselves, one watcher with no model, parent managers and shared account memory.
The captain (experimental)
- One agent over every project's fleet, by voice or text. It is one Claude Code session per home. Its only tool is its own Actions server, which has 24 fixed tools, and every call is audited with your words and its receipt.
- It unblocks agents for you. It answers an agent's Claude Code permission prompt (
press yes,press no), types into an agent's pane, and spawns, stops and restarts agents. A stop, spawn or restart acts only when your words name the agent, its role or its project. Otherwise the captain asks one question first ("Stop coder-1 in alpha?"). It never types into an agent's trust dialog. - The attention queue folds every board into one ranked list of what needs you: questions, blocked or waiting agents, reviews, pull requests and stale work. See
aisquare captain attention | next | resolve | snooze. - The voice page (
aisquare captain --voice, the[voice]extra). Hold to talk, or leave it listening for the wake word; only "Captain, …" gets through. The reply is spoken back. - It ships off. Turn it on in ⚙ Settings ▸ Captain, which applies at once with no restart, or with
aisquare config set experimental.captain true, which a runningasqreads at its next start. While it is off, none of it runs;asqshows only its switch, and doctor acaptain: offrow. Guide:docs/captain.md.
Personas
- Run an agent as someone, such as a skeptic, a mentor or a minimalist. Use
aisquare launch <role> --persona NAME,fleet spawn --persona, or a default per role ([fleet.roles.<role>].persona). - A persona is a Claude Code skill.
aisquare persona list | show | new | edit | rm | validate | import | exportmanages them in three layers: the project, the user, and four bundled personas. A skill copies in and out byte for byte. - Import anything. Plain text, another tool's persona, or a web page is converted with an LLM, shown to you, and confirmed before it is saved.
- Attach to a running agent with
aisquare persona attach <name> --to <label>. A restart keeps the agent's persona. - In
asq: a Personas tab in the Project view has a live preview, an in-place editor, import, export and a two-step attach. The Spawn dialog has a Persona field, under Advanced.
The fleet in asq
- The simple Spawn dialog. It shows three fields: Role, Hand off from, and Account. Everything else folds under Advanced. There is also a new role-less option, (none) — general-purpose Claude, which is
aisquare fleet spawn agentoraisquare launch agentfrom a shell: a plain Claude Code session on the board, with no standing cycle. - App settings in tabs: Agent settings, User preferences and Remote. The Captain box sits at the top of Agent settings. Only the essentials show; the rest folds under Advanced. The first preference is Always show project directory.
- Stop an agent from its view's header, or with
xin the sidebar. - Rank 1 is the machine default. There is no separate default account any more. When nothing more specific picks one (no
--account, role binding, project default or headroom pick), a launch runs on the first enabled, signed-in account in the priority order, and its line reads[work · preferred (rank 1)].aisquare accounts default <ref>, with no--projector--role, now moves that account to rank 1, and a machine-levelaccounts default --clearis refused, since some account is always first. Project defaults and role bindings are unchanged. - A simpler Accounts table. The ★ column, the Default button and the status column are gone. ↑/↓ are the first columns of each row and the plan is the last. Why a row has no usage (not signed in, a token to renew, a usage read that failed) reads in its session and week cells, with the whole text in their tooltips.
accounts listandaccounts usagedrop the*mark (the preferred account is the first enabled, signed-in row), and the account JSON dropsis_default. - Usage reads back off when they are rate-limited. When the usage endpoint answers HTTP 429 (or 503 with
Retry-After), the account's reads wait: the endpoint'sRetry-After, or one minute doubling per 429 in a row, up to 30 minutes. The wait is recorded on disk, so noaisquareprocess, the Accounts page,--refresh, a headroom pick or a hand-over asks that account again before it ends. The wait is the signed-in account's: a token refresh or Log in again keeps it. Before, every poller kept the account rate-limited. - The last good reading stays. A read that fails keeps the most recent numbers, with their age and why the newer try read nothing:
40% · 12min the page's session cell with the reason in its tooltip, andaccount 2: rate-limited until 14:32 — …; showing the reading from 12m agounderaccounts usageandaccounts list --usage.—shows only when nothing was ever read.--jsonreadings carrykept,rate_limited,rate_limited_untilandlast_failed_reason. - A kept reading is never taken for a live one. A headroom pick uses it only inside the pick's own window with neither window reset since, and otherwise skips the account with a note giving its age. A hand-over whose live read fails with anything but a 429 or a login problem picks by a reading at most five minutes old, and says so.
- Log in again on any signed-in row of the Accounts page, so a fresh login needs no
aisquare accounts run <slot>and/loginby hand. It does not clear a rate limit, as its tooltip says.aisquare doctorlists accounts whose reads are backed off, and until when (claude-account-backoff). - The agent info line. Each agent has one line showing its label, state, account with the share of its five-hour session used, model, effort, pane and persona. The sidebar rows show the session use too.
- Two things moved. The agent view's header now shows that info line instead of the role, the task and the working directory. The command that attaches a terminal to a project's tmux session moved from the Manager tab to the project's Settings tab.
- Faster panes. A key now reaches the screen in about a millisecond; before, it took 60 ms or more. Scrolling follows your gesture while you make it. The visible pane talks to tmux through one control-mode client per server; on tmux older than 3.2, the pane polls as before.
- No more stall every two seconds. The sidebar refresh now reads in the background.
Reliability (the P0 train)
- Every first message reaches its agent, or
fleet lssays it did not. A spawn's--prompt, a restart's or a switch's line and a hand-off's were typed with one Enter about two seconds after start, and an Enter that came before Claude Code drew its input box was dropped. A Claude Code agent now gets its first message on its command line, from a file only you can read that its launch deletes; any other binary, and[fleet] first_prompt = "type", has it typed once the box is drawn and checked there.fleet spawn,restartandswitchwait up to[fleet] prompt_ack_seconds(45) for the agent's turn to take it and end with✓ first message taken, or say it was not;--jsonaddsprompt_deliveryandprompt_acked. - Nothing the fleet types lands in your draft. Every typed write (the manager's nudge,
fleet tell, a typed first message, the Remote's paste) reads the input box first and types nothing while it holds your draft, or while you typed into the pane inasqwithin[fleet] draft_quiet_seconds(10). A held nudge goes out once the box is clear; a held tell is filed as a note and says why. A note to a busy agent reaches it at the end of its turn ([fleet] deliver_notes_at_turn_end). - An idle agent reads waiting, not working. Every role reads
waitingfrom its session start until a prompt starts a turn, and aworkingrow whose pane has been quiet for[fleet] busy_quiet_seconds(30) is checked once: idle at its prompt, it reads ⏸ waiting, and afleet tellis typed into it instead of filed. - One usage limit moves an agent once, to an account that can take it. The usage-limit hand-over no longer races the manager's own
fleet switch(limited_handing_overwakes nobody; a newhandover_failedevent does), and a replacement minutes old is refused without--force. A pick reads the per-family weeks (Opus, Sonnet, Fable) and skips an account whose family week is spent for that family, and every limit an agent hits is remembered on its account until its reset.[accounts] prefer = "reset-soonest"takes, of the accounts under the line, the one whose week resets first;priority, the default, takes them in rank order, rank 1 (the machine default) first.preferorders the headroom pick only: underpick = "default"rank 1 decides.accounts usageshows each family week and its reset. - A session resumed under another account keeps its transcript, so a switched agent no longer raises a false TWO live agents banner or restarts fresh.
capture.enabled = falsestops prompt capture: nothing is stored, listed byaisquare logor spooled. A config that cannot be read records nothing (privacy fails closed).- A coder that claims a task sees its contract:
task claimandtask next --claimprint it, with why the task was last reopened. aisquare doctorgains afleet-deliveryrow.
Fewer tokens per agent (the P1 train)
- Roles launch without the tools they never use. Eight tools no role uses (Artifact, Workflow, ScheduleWakeup and others,
[fleet] deny_tools) are denied at every launch, and every role but the manager gets skill names only. A fresh coder's/contextread 9.9k tokens where it read 26.9k. Per role,deny_tools,allow_toolsandskill_listingchange it;aisquare doctorshows a<role> toolsrow for each. - A manager wakes for decisions, not news. A clean exit, a switch, a restart or its own
fleet stopwaits for its next wake; a failed or killed exit, a released claim or a hand-over that did not start still wakes it. Board updates arrive one line each, andaisquare team show <seq>prints one in full.[fleet.roles.<role>] board_context(all,relevant,none) sets what of the board an agent reads; the reviewer reads none by default. - A manager starts from its start pack (its state file, open tasks, questions, live agents and waiting events, in at most 20,000 characters) instead of the whole board, and cites briefs and rules by path (
task add --detail-file,--rules).[team] brief_blockswarns about a contract missing its objective, acceptance or boundaries. Quiet-point compaction of a manager ships off ([fleet] quiet_compact_at). - One tester per task: a tester spawned for a task verifies that task and stops; the validator stays the final gate.
fleet spawnruns the role's model ladder (--model,--effort), kept by restart and switch. aisquare metrics ledgerprints what each session spent in tokens, counted from its transcript: requests, cache reads and writes, context, wakes, compactions and tool calls.
Work briefs and the evidence gate (#417)
asq execsaves a report of any command (asq exec -- pytest): its arguments, exit status, timing and output, and with--projectthe checkout's per-file source identity before and after.asq reports list | showreads them back. Reports live under~/.aisquare/reports/, owner-only.- Work briefs turn a request into numbered requirements linked to the board's own tasks (
aisquare brief create | show | link | evidence | finding). A correction keeps every id, and marks the checks it changed as stale. - The evidence gate.
aisquare task doneon a task a brief links refuses until that task has recorded a fresh pass for each of its requirements, and names what is missing, failed, blocked or stale. A pass backed by anasq exec --projectreport must have run cleanly on the same source as now. A task no brief links is done exactly as before.
Capacity and setup (the P2 train)
aisquare code: a symbol index to ask before grepping.code find,code refsandcode similaranswer where a name is defined, who uses it and what already looks like it, also as the MCP toolcode_lookup. The index is built in the background when a project is onboarded and refreshed at session start; Python needs nothing, TypeScript and JavaScript need theaisquare-cli[code]extra. Nothing leaves the machine.- Spawns are admitted against the machine.
[fleet] max_agents_totalcaps live agents across every project, andmin_free_mb(2048) refuses a spawn when memory is short, naming the numbers. A refused spawn first stops finished workers and tries once more. - Finished workers stop by themselves. A tester, reviewer, validator or ui-tester whose task is closed and who has sat idle for 5 minutes with nothing addressed to it is stopped, waking nobody.
[fleet] auto_stop_finished = falseturns it off;aisquare fleet reap --finisheddoes it on demand. aisquare watch: one watcher for the whole fleet, with no model. Every minute it looks for agents at a dialog, a usage limit, a crash or a stall, and, with[watch] github = true, each project's pull requests, then puts one line on the manager's board and nudges it.- A parent manager over a project group. A member's questions, blocked tasks and results land on the parent's board and wake its manager, and the parent can delegate tasks into members.
- Accounts can share memory, skills, agents, commands and CLAUDE.md (
[accounts] share,aisquare accounts share), and doctor'santhropic routingrow finds a stale tracing proxy wherever it hides, with one command to remove it. asqwith many projects stays light: one store read and one tmux call per tick, andaisquare ui --project/--groupto list only some.
Updates in one click
- asq tells you when a new release is out. From 0.9.0 on, at most once a day, asq asks PyPI what
aisquare upgrade --checkasks. A newer release shows as⬆ 0.9.1 · restart to updateunder ⚙ Settings, with one toast the first time. A failed lookup (offline) is silent and tried again within the hour. - Restart & update keeps your fleet running. The row opens a dialog that counts your live agents and asks PyPI again. Restart & update closes asq, runs
aisquare upgradewithout asking again, and reopens asq on the new version in the view you had open. Agents live on the fleet's own tmux server, so the update does not touch them. Any whose pane is gone after the reopen is restarted on its session, and the toast says what it found (✓ aisquare 0.9.1 — 2 agents reattached). An agent stopped on purpose since the press, or one that exited 0, is left and named. If the install fails, its output stays on screen until Enter, and the version you had reopens. - Other installs. Homebrew, pip and pipx installs get the row, and the dialog gives the command that updates them. A source checkout or an editable install is never checked. An install pinned to this release by its own installer is checked too, and the update drops the pin.
- Turn it off with
aisquare config set updates.check false, or the Check for a new release daily switch in App settings ▸ User preferences. - It starts with 0.9.0. An 0.8.x install has no check, so it updates to 0.9.0 by hand once (
aisquare upgrade). The first release asq offers is one after 0.9.0.
Remote control
- Nothing to install. The web server, Web Push and the official ngrok Python SDK are core dependencies now. The SDK comes bundled wherever ngrok ships a wheel (macOS, Windows, and Linux on x86_64, aarch64 and armv7l); elsewhere Remote uses an ngrok on your PATH, as before. The one setup step is pasting your free ngrok token once, in ⚙ Settings ▸ Remote. An ngrok you already signed in to keeps working.
- Remote is also in ⚙ Settings, besides behind
R. While it is on, the sidebar shows📡 remote · N devices.
Less context, measured tokens
- Solo sessions carry almost nothing from aisquare. A session with no team gets no team block and a one-line MCP instruction. For a solo onboarded repo, session start drops from about 2,400 characters to about 130. Preferences that a loaded
CLAUDE.mdalready holds are no longer injected twice. aisquare metrics tokensreports exact token usage per session, project or all sessions: raw tokens, cost-weighted tokens and the list price in USD. The hooks record it, and a recording failure never fails a hook.
The codebase snapshot is gone
- aisquare no longer packs your codebase, and never runs Node. The Repomix snapshot that
aisquare project onboardmade (andinitbefore it) is removed: nothing on the AISquare platform read it, and no agent was pointed at it any more. Nothing replaces the pack; to find code, ask the new symbol index (aisquare code, above).project onboardkeeps its other job: it adds a folder on purpose and seeds one fact per marker it finds at the root (pyproject.toml,package.json, aMakefile, and others). A second run reads nothing and says so, and--refreshadds only what is new. A forgotten project onboarded again says its earlier facts are back. - Old packs are deleted once, by the first command that opens the store: every
~/.aisquare/projects/<id>/snapshot/, and nothing else. A[snapshot]section inconfig.tomlstill loads and is dropped at the next save. Doctor'srepomix,snapshotandtiktokenrows are gone, so a machine with no Node has nothing amber over them. - Node is optional, for browser tools. The one-liner no longer installs Node for aisquare. It offers the Chrome DevTools MCP instead, with its own prompt that defaults to no: Node 22.12+, then the MCP registered with Claude Code at user scope.
--with chrome-devtoolsinstalls it without asking, andclaude mcp remove -s user chrome-devtoolsundoes it.
Fixes
- HTTPS works on a python.org Python on macOS. That interpreter ships no CA bundle until its Install Certificates.command is run, so every usage read failed to verify while
curlworked. Every HTTPS call (usage, IAM, explainability, CI, persona import, the update check andaisquare upgrade, remote push) now verifies against the OS trust store (truststore), withcertifiwheretruststorecannot be imported.SSL_CERT_FILE/SSL_CERT_DIRstill name a bundle of your own, for a proxy that re-signs HTTPS, trusted as before:SSL_CERT_FILEreplaces the default CA file, so that bundle must hold the public roots as well as the proxy's CA, while a hashedSSL_CERT_DIRadds to them. A certificate that does not verify says so, with a hint.aisquare doctor --livegains atlsrow naming the trust source in use. - Restart all brings back the agents that died. This covers agents that were killed, crashed, lost their tmux server or went down in a reboot within 7 days. Agents stopped on purpose, agents that exited 0, and agents whose task is closed are left alone, and each says why. A restarted manager is told which teammates are dead and how to bring them back.
- A
fleet spawn --promptis sent, not left in the input box. A new account's first spawn gets past Claude Code's first-run screen. An agent that has had no prompt readswaiting, notworking. fleet spawn --taskwith no prompt tells the agent to start on its task.- Spawn dialog follow-ups. With Role (none) and no persona, the note says plain Claude: no role, no persona. The Pick… note is hidden while Pick… cannot be pressed.
aisquare config get/set fleet.roles.agent.*now works, so the no-role agent's permission mode, worktree and extra args can be set like any role's. App settings ▸ Agent settings shows itsagentrow. aisquare serve --stdiono longer exits mid tool call.- A tmux server the fleet starts holds no team variable, and doctor warns when one does.
- The captain's board and attention queue are safe.
project prune --captured-onlynever takes the captain's home board, and one bad row no longer empties the attention queue. - Persona folders follow your umask.
aisquare personareads the same project layer that spawn reads. A failed--forceimport keeps the old persona. - …and more. See the CHANGELOG.
Installing, and pinning this release
The one-liner is the same as before, and installs the latest release:
curl -fsSL https://raw.githubusercontent.com/AISquare-Studio/aisquare-cli/main/install.sh | sh- on Windows:
irm https://raw.githubusercontent.com/AISquare-Studio/aisquare-cli/main/install.ps1 | iex
New: pin this release with its own installers, attached below. Every release from 0.9.0 on carries an install.sh and an install.ps1 that install exactly that version:
curl -fsSL https://github.com/AISquare-Studio/aisquare-cli/releases/download/v0.9.0/install.sh | sh- on Windows:
irm https://github.com/AISquare-Studio/aisquare-cli/releases/download/v0.9.0/install.ps1 | iex
The one-liner on main runs the latest release's own installer: it downloads that release's stamped install.sh, checks it, and runs it, saying using the v0.9.0 installer, so the steps are always the ones the release shipped with. It still installs the latest release unpinned, as before. If it cannot hand off, it says why and carries on itself. --no-handoff keeps main's copy.
--version V still overrides a script's own version, and --version latest drops it. A pinned install still gets asq's update row: Restart & update or aisquare upgrade moves it to the newest release and drops the pin, which uv tool upgrade alone does not.
Upgrading:
- From 0.8.x, run
aisquare upgrade; asq's own update check starts with 0.9.0, so this one is by hand. From 0.7.0 or earlier, re-run the one-line installer. - The store migrates from schema v32 to v39 automatically on first open. v33 adds the persona columns, v34 adds how an agent's row ended, v35 adds the token-usage tables, v36 the first message's state, v37 who caused each board event, v38 a group's parent manager and a task's origin, and v39 the work-brief table. Moving back to 0.8.2 with
aisquare upgrade --versionis safe: an older build opens a v39 store as it is. - The config file moves to version 3, once. A reviewer whose
extra_argsare exactly 0.8.2's default (--disallowedTools=Edit Write NotebookEdit) loses them: that deny is the role's own now. Every other value you set stays. Moving back withaisquare upgrade --versionto 0.8.2 or earlier names each role that release would launch with tools this one denies, with theconfig setthat keeps them denied there. An agent restarted after the upgrade sends its first request uncached once. - New config keys, all with defaults, no migration:
[fleet] first_prompt(argv),prompt_ack_seconds(45),busy_quiet_seconds(30),draft_quiet_seconds(10),deliver_notes_at_turn_end(true),[accounts] prefer(priority),[updates] check(true),[fleet] deny_tools(the eight tools),[fleet.roles.<role>] skill_listing/deny_tools/allow_tools/board_context,[fleet] quiet_compact_at(0, off) and its timings,[team] brief_blocks/brief_strict,[fleet] max_agents_total(0, no budget),min_free_mb(2048),auto_stop_finished(true) andauto_stop_idle_minutes(5),[watch],[accounts] share(empty), and[code] enabled/max_file_kb/ignore. A[snapshot] ignorelist is moved to[code] ignoreat the next save.fleet spawn,restartandswitchcan take up to 45 s longer while they wait for the first message to be taken. - On a shared machine, a first message on the command line can be read by other local users in
pswhile the agent runs;aisquare config set fleet.first_prompt typekeeps it off the command line. - A machine default account moves to rank 1. A default set by an older build becomes rank 1 on the first read, the others keeping their order behind it. A machine that never set a default and has accounts signed in now launches on the first enabled, signed-in one. Two changes to check. Where headroom does not pick (
pick = "default", or no usage readable), a role with a hand-writtenteam bind <role> --env CLAUDE_CONFIG_DIR=…ran on the old machine default, and now runs on its own directory; a headroom pick still overrides the binding, as before. And an old default that is signed out or disabled no longer gets the launch: the next enabled, signed-in account does. - New core dependencies: PyYAML;
truststoreandcertifi, for HTTPS against the OS trust store; Remote's web server (starlette, uvicorn, websockets) andcryptography; and the official ngrok SDK, where ngrok ships a wheel. - New optional extras:
[llm], for persona import through the Anthropic API,[voice], for the captain's voice page, and[code], for indexing TypeScript and JavaScript.[remote]still resolves but adds nothing.
Known issues: none of these blocks 0.9.0.
- A solo session that calls aisquare's own board tools over MCP counts its own MCP client as a teammate. Until 30 minutes after its last such call, it gets the board and the full team context, as if it had a team (#45 m1, a pinned known limitation; see
docs/orchestration.md). The precise fix is planned in draft PR #397. - Fleet and team edge cases left after #274, the closing reviews' minors and nits, are tracked in #398.
The full list, with every detail, is in CHANGELOG.md → 0.9.0.
Full diff: v0.8.2...v0.9.0