Repository navigation
v0.80.87
·
2728 commits
to main
since this release
0.80.87 turns the agent into a team: Your Office brings persistent named coworkers with huddles and errands into the engine, and A-Coder Cloud v0 adds a self-hosted, always-on agent fleet. The transcript is redesigned around uniform tool result cards and a consolidated status rail, the /model picker now refreshes live provider catalogs, and plan mode, memory, and thinking controls gain new commands and behaviors. A long tail of engine fixes lands around compaction timing, RPC abort/steer behavior, and provider compatibility.
Highlights
- Your Office: named coworkers with faces, huddles, and errands live inside the engine and persist under
~/.a-coder/cli/office/, each keeping its own per-project session so memory survives restarts. Drive it from/office(hire,huddle,tell,say,stop,errands) or the newoffice_*RPC commands behind the desktop Office panel. - A-Coder Cloud (Sovereign Edition) v0: a self-hosted, always-on agent fleet —
a-coder cloud spawn <repo> <prompt...>runs an agent in a cloned workspace with git-synced checkpoints and structured reports, supervised by acloud servedaemon that re-queues orphaned tasks. - Transcript overhaul: settled tool calls render as uniform three-layer cards (outcome headline with per-tool stats like
+12 −4orexit 0 · 1.2s, a 3-line preview, errors always visible — MCP/extension tools included); background agents and processes consolidate into one status rail below the editor (⚡ 2 running · backend (5 tools, 8s) · npm run dev (12s)), and errors/warnings render as a grouped notice block instead of nakedError:lines. - Live
/modelpicker: the cached list paints instantly, then every fetchable provider (Ollama Cloud, OpenAdapter, LM Studio, llama.cpp, local Ollama) is re-fetched past its TTL cache and the list repaints — re-applying your search so a late refresh never clobbers typing. - New introspection commands:
/context(context-window split as proportional bars),/diff [n](working-tree patch plus what rewinding n turns would change),/memorypicker, and thinking controls (/thinkingselector with Ctrl+S save,/think [level], "ultrathink"-style keyword escalation). - Plan mode enrichment: a persisted plan file the model can write, read-only bash commands run without approval,
allowedPromptspre-approvals on exit, and an interactive exit-plan dialog (proceed / switch to ask mode / keep planning with feedback).
Details
Office
- Huddles run as a bounded round-robin drive (max 3 rounds, max 10 messages): who speaks each round is a deterministic @mention parse, every coworker is fed only the room delta since its last turn, replying exactly "(pass)" is silence, and a round where everyone passed settles the room. User sends carry hold directives ("stop @atlas" holds a coworker until re-addressed); turns that outlive their window are harvested late instead of lost.
- Supervised coworkers route tool approvals and questions to the user as office prompts (
office_respondanswers them); auto coworkers run with the auto permission mode. - Errands fire on a 30s ticker with every/daily/once schedules and, with continuity, run inside the coworker's own session so scheduled agents actually learn between runs.
- Coworker side runtimes are created via the new
AgentSessionRuntime.createSideRuntime, outside the active/detached registries so they never count against background-session caps. 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.
Cloud
- New
@earendil-works/pi-cloudpackage with a task store, a git sync layer (ac-cloud/<task-id>branches, WIP checkpoint commits, diff summaries, optional origin push), and a daemon (a-coder cloud serve, unix-socket IPC, auto-started on demand, orphaned tasks re-queued after daemon restart). - The task runner drives an a-coder RPC agent in a cloned workspace — checkpointing, wall-clock budget, extension-UI prompts auto-declined for headless runs. The agent writes
report.mdinto the repo itself; the control plane adds a structured report (commits, diffstat, changed files, token/tool usage, warnings) plus session continuity (session JSONL copied into the task artifacts). - CLI surface:
a-coder cloud spawn <repo> <prompt...> [--base --model --timeout --push],status [--json],review <id>,stop <id>,serve.
Agents & background tasks
- Background bash processes now notify the agent when they terminate (done/error — by exit or timeout): the
<task-notification>wake fires a turn reporting 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, and user kills from the running-tasks viewer stay quiet (riding along on the next submission, as with sub-agents killed via the viewer orkill_subagent). - 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 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.
- Sub-agent
permissionModefrontmatter is enforced at spawn:planenforces read-only,autoapplies policy rules and the classifier,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. task_createaccepts atasksarray (each withsubject,description, optionalactiveForm/metadata) and creates the complete task list in one call; ids are assigned sequentially and the result text lists every created id.- The running-tasks viewer's agent timeline is fully scrollable: ↑/↓ page 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.
TUI & commands
/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 MEMORY.md and the 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 it fresh on its next turn./thinkingopens a searchable thinking-level selector marking the current level and the startup default; Ctrl+S (newapp.thinking.savekeybinding) saves the selection as the startup default, while a level argument sets it for the session only.AgentSession.setThinkingLevel/cycleThinkingLevelgained{ persist }options — session changes no longer rewrite the saved default implicitly./think [level]slash command with argument completion, plus 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, reasoning models only; works in interactive, print, and SDK prompts)./rewindnow survives a session resume: file-history snapshots are mirrored into the 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 across session-tree branches./output-style <name>offers fuzzy argument completion over registered style names (the active style is filtered out).statusLinesetting: set a shell command in settings.json and its first stdout line renders as an extra row above the footer; runs on a 300ms trailing debounce with a 2s timeout, receiving JSON on stdin (model,cwd,permissionMode,contextPercent,tokens) — Claude-Code-style custom footers without writing an extension.- Session picker 1-9 quick keys: with an empty search field, pressing 1-9 jumps straight to that row and resumes it.
- 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. - Historical thinking folds away: only the latest round's thinking trace renders in the live transcript; Ctrl+Shift+T temporarily unfolds everything, and the global hide-thinking setting still takes precedence.
- Background-task notifications render as compact status pills (glyph + command/subagent summary + duration) instead of raw multi-line notes; the wake turn renders pills only, and the model still receives the full note text.
- 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. - Nested slash-command suggestions:
/officesuggests its verbs (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 prompt and the pending card renders a warning badge until the call executes. ask_user_questiongains an implicit "Other" row after the model's options, opening a free-text input — Enter sends it as the answer (single-select advances, multi-select adds it to the staged selection), Esc returns to the options; the tool description now tells the model the UI provides Other automatically.- Compaction, branch-summarization, and retry spinners render inside the editor top border alongside the working indicator;
@earendil-works/pi-tuieditor borders gained centered scroll indicators via overridablerenderTopBorder/renderBottomBorderhooks, and custom editors opt in with{ embedWorkingStatus: true }. - Status rail and notice block specifics: the rail replaces the three scattered surfaces (inline AGENTS transcript card, agents bar above the editor widgets, processes bar below the footer) with a dim right-aligned
↓ taskshint for the running-tasks viewer; the notice block pins at the end of the transcript with theme-colored icons, hanging-indent wrapping, identical consecutive notices collapsed with a ×N count, and only the 6 most recent shown. /scoped-models,/model <term>and the post-login flow already fetched fresh via the registry cache reset; model cycling keeps its TTL cache on purpose so rapid cycling stays instant.
Skills, memory & permissions
- Skill frontmatter:
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 session), after which it joins the listing on the next prompt. Conditional skills stay invocable by name throughout. - 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. - Plan-mode enrichment: 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;plan_modeexit acceptsplan(content persisted to the plan file) andallowedPrompts, which become arg-scoped session allow rules taking precedence over the permission mode without triggering the auto-mode classifier. - Exit-plan approval dialog: when the model leaves plan mode, an interactive 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.
Engine & RPC
pi authcommand surface:auth print-api-key,auth print-bearer-token --min-expiry <duration>, andauth check [--json] [--credentials] [--no-refresh]resolve credentials through the model registry and auth storage without starting a session.--no-refreshonly refreshes tokens that are already expired;--min-expiryrequires validity for the requested window and refreshes proactively.ctx.modelRegistry.stream()andstreamSimple()let extensions make model calls through configured providers with resolved authentication; 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: clear beforeabort, then restore the returned text in the client editor. - In-place managed-install updates stage the release tarball, verify its SHA256, and atomically activate it.
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.AGENTS.override.mdper-directory context overrides: when a directory contains one, it is loaded instead of that directory'sAGENTS.md/CLAUDE.md; other directories layer normally.ui_prompt_start/ui_prompt_endextension events let hosts 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 for waiting-state rendering.- Background-process output notifications are throttled to ~10Hz (leading + trailing edge) in the store, so a chatty background command no longer drives per-chunk renders or
background_processes_updateRPC events to the desktop; completion events still notify immediately. - Background sessions (runtime registry, Phase 1 of the session architecture roadmap).
Providers & models
- Built-in
read,bash,edit, andwritetools prefer strict JSON-schema sampling by default (constrainedSampling: { type: "json_schema", strict: "prefer" }): providers that support strict tools enforce the schema, others fall back silently; extensions can re-register tool definitions withconstrainedSampling: false. - The MCP client now negotiates the
rootscapability and reports the session's working directory toroots/list(root defaults to cwd; the resource loader passes the session cwd). Roots-aware servers like chrome-devtools-mcp scope their file-writing tools to the project instead of falling back to the OS temp directory; paths outside stay denied. - Grep falls back to grep(1) when ripgrep cannot be provisioned (offline mode, Termux, failed download): the search runs via
grep -RInE(with-F/-i/--include/-Cmapping) and returns the samepath:line: textoutput shape, with a notice that.gitignorerules are not applied. Also exposed asrunGrepFallback/buildGrepFallbackArgumentsfor tests and remote backends. - 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). - 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 — the error surfaces instead; consecutive overloaded/529 failures stop auto-retrying after 3 attempts (reset by any successful turn or a non-overload retry).
- Composio SDK upgraded to
@composio/core0.18.0 +@composio/experimental0.2.3 (matching pair), silencing the SDK's runtime "Upgrade available" banner.
Analytics
- Opt-in product analytics via PostHog: with
enableAnalyticsin settings.json (asked during the experimental first-time setup; also toggleable 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), per-modelcli_model_usageevents (one per provider/serving-model pair, so mid-session model switches count accurately;requested_modelrecords the picked model when the provider serves a different concrete one, e.g. OpenRouter auto-routing), and acli_installevent. Fires in interactive, print, RPC (desktop engine), and ACP modes. Gate overrides:A_CODER_CLI_ANALYTICS=1/0; offline mode always disables;A_CODER_CLI_POSTHOG_KEY/A_CODER_CLI_POSTHOG_HOSTfor self-hosting/testing (an explicitly empty key disables the client).
Fixed
- Fixed
/reloadnot picking up MCP server changes: the inline extension factory list (includingcreateMcpExtensionFactory) was assembled once at ResourceLoader construction, so MCP server changes in settings.json only took effect after restarting the CLI.reload()now rebuilds the inline factories from fresh settings; previous connections close cleanly via thesession_shutdown(reasonreload) event that fires before the rebuild. - 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).