Repository navigation
v0.80.70
·
2728 commits
to main
since this release
0.80.70 turns the agent into a team and makes sessions background-capable: Office adds persistent named coworkers with huddles and errands, and switching away from a session no longer kills its in-flight turn — including across projects. Plan mode is reworked around a persisted plan file with pre-approved read-only commands and an auto-accept path, and the TUI gains /context, /diff, /memory, /permissions, /think, and output styles. Under the hood: pi auth credential commands, in-place self-updates, opt-in analytics, and a round of compaction/RPC/provider fixes.
Highlights
- Your Office — a society of named coworkers with faces, huddles (group conversations), and errands (scheduled jobs with memory) living in the engine, persisted under
~/.a-coder/cli/office/with a canonical session per project so memory survives restarts. Drive it via/office(hire,huddle,tell,say,stop,errands) or the new desktop Office panel. - Background sessions — switching away from a session with an in-flight turn no longer aborts it; the runtime detaches, keeps streaming, and re-attaches when you return. The engine now hosts sessions from multiple projects without restarting (see Sessions).
- Plan mode rework — a persisted plan file the model can write, approval-free read-only bash while planning, an "auto-accept rest of plan" option on permission prompts, an interactive exit-plan dialog, and
allowedPromptspre-approvals on exit. - Running-tasks viewer & background bash notifications — one unified viewer (Down from an empty editor, Shift+Down, Ctrl+Shift+B,
/subagents) over background bash processes and sub-agents with live output andkto kill; background bash terminations now wake the agent with a<task-notification>report, rendered as compact status pills. - New TUI controls —
/contextwindow bars,/diff [n]with rewind preview,/memorypicker,/permissionsrule manager,/think [level]plus "ultrathink"-style keyword escalation, switchable output styles (/output-style), astatusLinesetting, and intent tags on bash tool cards. pi auth& in-place updates —auth print-api-key,auth print-bearer-token --min-expiry <duration>, andauth check [--json]resolve credentials without starting a session; managed installs stage the release tarball, verify its SHA256, and atomically activate it.
Details
Engine
AGENTS.override.mdper-directory context overrides: when a directory containsAGENTS.override.md, it is loaded instead of that directory'sAGENTS.md/CLAUDE.md; other directories layer normally (upstream pi earendil-works#7681 parity).SessionManager.inMemory(cwd, options, entries)restores sessions held outside the filesystem (e.g. a database): entries are adopted with their tree, labels and compaction structure intact, a stored session header carries its identity and version through migration, and headerless entries become the body of a fresh session (upstream pi earendil-works#8980 parity).ui_prompt_start/ui_prompt_endextension events (upstream pi earendil-works#8355 parity): hosts can distinguish active agent work from time spent waiting on blockingctx.uiprompts (select/confirm/input/editor/custom), with nested prompts tracked by depth; RPC/desktop clients receive them as extension events for waiting-state rendering.ctx.modelRegistry.stream()andstreamSimple()for extension model calls through configured providers with resolved authentication (upstream pi earendil-works#8964/earendil-works#9272 parity); setup failures surface as error events on the returned stream.- RPC
clear_queueretrieves and removes queued steering and follow-up messages for interactive Esc behavior (upstream pi earendil-works#8432 parity): clear beforeabort, then restore the returned text in the client editor. - The MCP client negotiates the
rootscapability and reports the session's working directory toroots/list(root defaults to cwd; resource-loader passes the session cwd). Roots-aware servers like chrome-devtools-mcp scope their file-writing tools to the project instead of the OS temp directory, sotake_screenshotwith afilePathinside the session's project works; paths outside stay denied. - Grep tool falls back to grep(1) when ripgrep cannot be provisioned (offline mode, Termux, failed download):
grep -RInEwith-F/-i/--include/-Cmapping, samepath:line: textoutput shape, with a notice that.gitignorerules are not applied; exposed asrunGrepFallback/buildGrepFallbackArguments.
Sessions
- Background sessions (runtime registry; roadmap in
desktop-app/SESSION-ARCHITECTURE.md): a mid-turn runtime moved off the active slot keeps streaming to its own session file and is reaped once it settles; switching back re-attaches the same live runtime./clearstill aborts (it rewrites the session file in place). A background turn blocked on a tool-approval prompt or extension dialog flags the runtimeneedsInput(viamarkSessionNeedsInput), cleared when answered. - New RPC:
get_sessions_status(per-sessionrunning/pendingMessageCount/active/needsInput),abort_session(abort a background turn by session path), and a pushedsessions_updateevent whenever the registry changes. - Multi-project (Phase 3): the engine hosts sessions from any project — each runtime's cwd-bound services follow the session's recorded cwd — so the desktop no longer restarts the engine when switching projects;
newSession/new_sessionaccept acwd, and switching to a project resumes its most recent session or starts a new one. Extension-initiated replacements (ctx.newSession()/ctx.switchSession()/ctx.fork()) passdiscardPreviousto preserve the stale-ctx contract (earendil-works#2860). - Keep-alive and targeting (Phase 2): detached runtimes stay alive after their turn settles (idle keep-alive, 30-minute default reap,
setDetachedIdleTimeoutMsoverride);prompt/steer/follow_up/abortaccept an optionalsessionPathtargeting a background session;switch_sessionresponses carryreattached+ asnapshot(running/needsInput/pendingMessageCount); blockingextension_ui_requestevents (confirm/select/input/editor/question + permission prompts) are tagged with thesessionFileof the session that opened them. background_processes_updateRPC event: headless/RPC mode streams background bash-process snapshots to clients, alongside the existingsubagents_updateevent.
Office
turn_endactivity events now carry the coworker turn's stop reason and awillRetryflag, so clients can chime on natural turn ends only — aborts, in-session errors and retryable failures stay quiet.- While a coworker's turn runs, the engine pushes
office_activityRPC events (turn lifecycle, tool start/end with tool name, completed assistant speech — never streaming deltas) so clients can render live office scenes; failed turns surface as a system line in the huddle log with the reason instead of reading as a silent pass. - Coworker side runtimes are created via the new
AgentSessionRuntime.createSideRuntime, outside the active/detached registries so they never count against background-session caps.
Plan mode
- Entering plan mode surfaces a persisted session plan file (
<agentDir>/plans/<sessionId>.md) that the model can write while planning — the only write auto-approved in plan mode. - Obvious read-only bash commands (pwd/ls/cat/head/tail/tree/wc/stat/du/df/which/rg/grep/find/git status|diff|log|show, guarded against redirection,
&&/||chaining, and destructive flags like-delete/-exec) run without an approval prompt while planning. plan_modeexit acceptsplan(content persisted to the plan file) andallowedPrompts(pre-approved commands, e.g.npm test), which become arg-scoped session allow rules taking precedence over the permission mode without triggering the auto-mode classifier.- "Auto-accept rest of plan": when the agent requests permission for a mutating tool in plan mode, the prompt offers a third option — "Yes, auto-accept rest of plan" — which approves the call and subsequent plan-mode mutations until plan mode is disabled (the flag resets whenever plan mode turns off).
- Exit-plan approval dialog: previews the plan file and offers "Yes, proceed", "Yes, switch to ask mode" (permission mode switches to ask so every edit still prompts), or "No, keep planning" with typed feedback returned to the model while plan mode stays on. Headless/RPC sessions without the handler exit directly, as before.
TUI
statusLinesetting: setstatusLinein settings.json to a shell command and its first stdout line renders as an extra row above the footer. Runs on a 300ms trailing debounce on session events with a 2s timeout, receiving JSON context on stdin (model,cwd,permissionMode,contextPercent,tokens)./contextvisualizes the context window split (System prompt / Tool definitions / Conversation history / Free space) as proportional bars with token counts and a low-space hint./diff [n]shows the git working-tree status, shortstat and a budgeted colored patch, plus the file-history section: what rewinding the last n user turns would change./memorypicker lists global MEMORY.md, project (workspace) MEMORY.md and the current session's memory file with "(new)" markers; Enter opens the selected file in your external editor (new files get a minimal# Memorytemplate), and the agent reads the file fresh on its next turn./permissions(TUI): interactive manager for the "auto" permission-mode policy rules (allow / soft-deny / hard-deny) backed by the existingPermissionPolicyConfigengine; lists rules grouped by category, removes with confirm, and adds via "Add …" rows (e.g.bash,edit,mcp__server:*,$defaults) — no hand-editing settings.json.- Output styles: switchable system-prompt overlays (
default/Explanatory/Learningplus custom.mdfiles under<agentDir>/output-styles/and<cwd>/.a-coder-cli/output-styles/) that reshape how the agent answers without changing its tools;/output-style [name]opens a selector with no arg, the choice persists tosettings.outputStyleand applies on the next prompt;/output-style <name>now offers fuzzy argument completion (active style filtered out). - Thinking controls: new
/think [level]slash command (with argument completion) next to the existing keybinding cycling, and thinking-keyword escalation at the session layer — "ultrathink" raises to the model's highest supported level, "megathink"/"think harder" to high, "think hard" to medium (raise-only, never lowers, reasoning models only); works in interactive, print, and SDK prompts. /rewindsurvives a session resume: file-history snapshots are mirrored into the session transcript asfile_history_snapshotentries (invisible in LLM context; hidden in the default tree view, labeled checkpoints in "all" view), restored on load and re-hydrated when navigating session-tree branches.- Reduced-motion setting (
terminal.reducedMotion, envA_CODER_CLI_REDUCED_MOTION=1, toggle in/settings): spinner frames render a static frame; the message shimmer sweep and in-flight tool blink dots go solid. - Bash tool cards gain intent tags:
[Git] $ git status,[Test·npm test] $ npm test,[Build] $ tsc -p .,[Search·"pattern"] $ rg …,[List] $ ls src— the full command stays visible for approval transparency. - Session picker 1-9 quick keys: with an empty search field, pressing 1-9 jumps straight to that row and resumes it.
- Background-task notifications render as compact status pills instead of raw multi-line notes: a completed/failed/killed note collapses to one colored line (glyph + command/subagent summary + duration); the wake turn renders pills only, with the boilerplate instruction line suppressed from display. The model still receives the full note text.
- Historical thinking folds away: only the latest round's thinking trace renders; the verbose transcript (Ctrl+Shift+T) temporarily unfolds everything, and the global hide-thinking setting still takes precedence.
ask_user_questiongains an implicit "Other" row: an extra choice after the model's options that opens a free-text input — type a message and press Enter to send it as the answer (single-select advances, multi-select adds it to the staged selection), Esc returns to the options; the tool description tells the model the UI provides Other automatically.- Nested slash-command suggestions:
/officesuggests the verbs the command consumes (hire/huddle/tell/say/stop/errands) with usage descriptions, and/exportsuggests the output formats; both filter as you type. - Tool cards show a "waiting for approval" phase while their permission prompt is open: the session emits
tool_permission_updateevents around the user prompt, and the pending card renders a warning badge until the call executes. - Running-tasks viewer (Claude Code-style): press Down from an empty editor (also Shift+Down / Ctrl+Shift+B, and
/subagents) to open a unified picker over running background bash processes and in-process sub-agents; Enter drills into a live view (bash output tail or sub-agent transcript), Esc/Left goes back,kkills. Sub-agent records now carry theirgoal(spawn prompt) andmodel, shown in rows and detail view. Replaces the two separate sub-agent / background-process viewers (sub-agents are flat — they cannot spawn children — so the list is flat, not a tree). - The running-tasks viewer's agent timeline is fully scrollable: ↑/↓ pages back through the complete event history (3 events per tick) with a hidden-events count and a "(↓ returns to live)" affordance, instead of a fixed 12-event tail.
- Background-process output notifications are throttled to ~10Hz (leading + trailing edge) in the store: a chatty background command no longer drives per-chunk renders or per-chunk
background_processes_updateRPC events to the desktop; completion events still notify immediately.
Agents, skills & memory
- Skill
allowed-toolsandpathsfrontmatter:allowed-tools: Bash(git commit *), Read(CSV string or list) grants those entries as arg-scoped session allow rules when the skill is invoked — by the model via the skill tool or by the user via/skill:name; entries match the permission-policy rule syntax.paths:(gitignore-style patterns; glob values starting with*must be YAML-quoted) makes a skill conditional: it stays out of the model's skill listing until a read/write/edit/find call this session touches a matching path (relative to cwd; matches stick for the rest of the session), then joins the listing on the next prompt; conditional skills stay invocable by name throughout. - Sub-agent
permissionModefrontmatter is now enforced at spawn (planenforces read-only,autoapplies policy rules and the classifier under the normal auto path,defaultinherits the session mode); the before-tool-call pipeline is shared between parent and sub-agents so the override covers hooks, permission resolution, file-history tracking, and extension handlers. - Persistent-memory auto-injection (opt-out via
settings.memory.enabled): global and workspaceMEMORY.mdcontents fold into the system prompt as a<persistent_memory>section on prompt (re)build, capped per scope (default 2,000 chars,settings.memory.maxCharsPerScope); the section notes the snapshot can lag mid-session writes and to re-read via the memory tool; unreadable/missing files are skipped silently. - Teammate mail is no longer spawn-only: a running named teammate drains its inbox after each turn chain and, when mail arrived mid-run, wakes for another turn to act on it (turn budget still respected; mail arriving after the budget is spent stays unread for inspection). Mailbox writes/drains and team-file member updates are serialized through a per-file keyed in-process mutex, so a
send_messagelanding while a teammate drains cannot interleave read-modify-write and drop messages.
Tasks & background processes
- Background bash processes notify the agent when they terminate (done/error — by exit or timeout): the same
<task-notification>wake used for detached sub-agents fires a turn reporting the outcome, including exit code, duration, line/byte totals, the output tail and the full-output log path. Multiple terminations within one wake window batch into a single turn; a persistently failing wake prompt backs off 2s instead of spinning. A user kill from the running-tasks viewer stays quiet and only rides along on the next submission, as does a sub-agent killed via the viewer or kill_subagent tool (the killing party already knows). task_createbatch creation: the tool now takes atasksarray (each withsubject,description, optionalactiveForm/metadata) and creates the complete task list in one call instead of one round-trip per task; ids are assigned sequentially and the result text lists every created id.- Persistent task graph:
task_createand relatedtask_*tools (entry truncated in the source changelog).
Reliability
- Full-turn auto-retry honors provider
Retry-After: error assistant messages can carry aretryAfterMshint (extracted from Retry-After/-Ms headers or SDK fields); the retry waits that long instead of the backoff schedule, capped by the newretry.maxDelayMssetting (default 60000ms, same cap as the backoff tiers). - Auto-retry hardening: a turn whose failed attempt already streamed visible output (non-empty text or thinking, or any tool call) is no longer silently replayed by the auto-retry — the error surfaces instead, avoiding duplicated visible output; consecutive overloaded/529 failures stop auto-retrying after 3 attempts (reset by any successful turn or a non-overload retry).
Fixed
- Fixed direct RPC
steerandfollow_upcommands bypassing extensioninputhandlers (#8718): both now run through the same input-handler/queue path as interactive submissions, withsource: "rpc". - Fixed RPC
abortreporting success without cancelling an in-progress manual compaction (#8920): abort now also cancels compaction and branch summarization, and the session exposesisIdle/waitForIdle()so abort settles before responding. - Fixed large tool results crossing the auto-compaction threshold being sent to the provider before compaction: threshold compaction now runs between tool execution and the next assistant response in the same run, and interactive progress resumes when the run continues (#6879, upstream earendil-works#8782).
prepareNextTurnnow runs only when a next turn actually happens, aftershouldStopAfterTurn. - Fixed extension messages sent with
triggerTurn: falsewhile the agent is running being inserted between a tool call and its result, which made providers that validate message order reject the replayed history; they are now appended once the turn's tool results are in (#8537). - Fixed compaction and branch summaries forcing
toolChoice: "none", which OpenAI-compatible gateways reject when no tools are sent (#8649, #8638). - Fixed
read,write,edit,grep,find,ls, andbashtools ignoringctx.cwdwhen registered by extensions with a different session cwd (#8627). - Fixed proxied plain-HTTP provider requests hanging after a tool call by tunneling them with CONNECT (
proxyTunnel, upstream earendil-works#8134).
Analytics
- Opt-in product analytics via PostHog: when
enableAnalyticsis on in settings.json (asked during the experimental first-time setup; toggle also in the desktop Settings panel), the CLI captures anonymouscli_session_start/cli_session_endevents (version, mode, OS/arch, model/provider, aggregate token/cost/tool-call counters; never prompt content, paths, or personal data) with thetrackingIdUUID as distinct ID, pluscli_model_usage— one event per (provider, serving model) pair actually used, with per-model turns/tokens/cost so mid-session model switches are counted accurately (requested_modelrecords the picked model when the provider serves a different concrete one, e.g. OpenRouter auto-routing) — and acli_installevent alongside the existing install ping. Events fire in interactive, print, RPC (desktop engine), and ACP modes; RPC end-of-session is captured on engine shutdown. Gate overrides:A_CODER_CLI_ANALYTICS=1/0; offline mode always disables. Ingest key is baked intocore/analytics.tswithA_CODER_CLI_POSTHOG_KEY/A_CODER_CLI_POSTHOG_HOSToverrides for self-hosting/testing (an explicitly empty key disables the client).
Dependencies
- Upgraded Composio SDK to
@composio/core0.18.0 +@composio/experimental0.2.3 (matching pair), silencing the SDK's runtime "Upgrade available" banner; the experimental provider base changed shape in 0.18, so both move together.