Skip to content

v0.80.135

Choose a tag to compare

@github-actions github-actions released this 07 Oct 00:16
· 2728 commits to main since this release

0.80.135 is a large TUI-and-automation release: the terminal UI now runs fullscreen by default, cron brings scheduled and event-triggered agent runs with a run-history inbox, and script workflows orchestrate multi-agent runs from plain JavaScript. Also new: a mobile serve bridge, an experimental computer-use tool, ChatGPT sign-in, vLLM/SGLang/Unsloth providers, six new themes, and Mermaid rendering in the chat.

Highlights

  • Fullscreen TUI by default — the UI now runs in the terminal's alternate screen with a fixed editor/status dock and a scrolling, searchable transcript. --tui-mode <regular|fullscreen> overrides per invocation, /settings switches live (editor/scroll state carried across), and new tuiMode, fullscreenExitOutput, fullscreenScrollbar, fullscreenCopyOnSelect, and fullscreenWheelScrollLines settings plus tui.altScreen.* keybindings and app.message.copy (ctrl+x) are configurable.
  • Cron — scheduled tasks for the main agent — jobs pair a name, prompt, and schedule (every:<n>m|h|d min 5 minutes, daily:HH:MM, once:<date>) or fire on workspace events (on:turn-end, on:commit). Managed via /cron, new RPC commands, and a desktop Scheduled panel, with per-job run history and a new Activity inbox tab.
  • Script workflows — plain JavaScript files that orchestrate background sub-agents (agent(), pipeline(), parallel()) with structured outputs, replayable/resumable runs, and concurrency caps; saved workflows register as /<name> slash commands.
  • Mobile bridge — a-coder-cli serve exposes the RPC engine protocol to mobile clients over an authenticated WebSocket, with QR-code pairing, mDNS discovery, and idle-exit.
  • Experimental computer tool (computerUse setting, default off; A_CODER_CLI_COMPUTER_USE=1 still overrides) — desktop control on macOS/Windows/Linux via the external cua-driver binary over MCP-on-stdio: screenshots with element overlays, click/type/scroll/key actions, permission-aware gating, hard-blocked destructive inputs, and a free doctor diagnostics action.
  • Sharper transcript — settled tool results render as uniform three-layer cards (outcome headline, 3-line preview, errors always visible), background agents/processes consolidate into one status rail below the editor, and mermaid code fences render as themed Unicode diagrams.

Details

TUI

  • Settled tool call results are now uniform three-layer cards: an outcome headline (✓/✗ glyph plus tool-specific stats — 40 lines for read, 6 matches in 2 files for grep, +12 −4 for edit, N lines for write, N entries/N results for ls/find, exit 0 · 1.2s for bash replacing the trailing Took line), a 3-line content preview with one unified ... N more <unit>, ctrl+shift+o to expand footer, and errors that always break through. Compact-classified resource reads (AGENTS.md, SKILL.md, docs) stay headline-only. MCP and renderer-less extension tools collapse from a raw output dump to the same headline + preview block, and their args JSON only shows when expanded.
  • Reorganized layout: background sub-agent and background-process status render as one consolidated status rail flush below the editor (⚡ 2 running · backend (5 tools, 8s) · npm run dev (12s)) with a dim right-aligned ↓ tasks hint — replacing the three scattered surfaces (inline ⚙ AGENTS transcript card, ⚡ agents bar, ▸ processes bar). Errors and warnings render as a grouped notice block pinned at the end of the transcript (✗/⚠ icons, theme colors, hanging-indent wrapping, identical consecutive notices collapsed with a ×N count, only the 6 most recent shown) instead of naked Error: lines spliced into the conversation.
  • The running-tasks viewer (Down on an empty editor, /subagents) now scrolls: a windowed list sized to the terminal (4–18 rows with … N more above/below markers) whose viewport follows the cursor, so a workflow's swarm of sub-agents or a pile of background processes is navigable at any size. The bash-process live view gained ↑/↓ scrolling over the retained output tail.
  • The /model picker fetches the live catalog from every provider on open: the cached list paints instantly (never blocks on the network), then refreshDynamicModels(force) re-fetches Ollama Cloud, OpenAdapter, LM Studio, llama.cpp, and local Ollama past their TTL caches, repainting with the user's live search query re-applied. /scoped-models, /model <term>, and the post-login flow already fetched fresh; model cycling keeps its TTL cache so rapid cycling stays instant.
  • Six new built-in color themes beyond dark/light: Nord, Gruvbox (dark + light), Catppuccin (Mocha + Latte), Tokyo Night, and Solarized (dark + light) — hand-tuned 56-color palettes covering the full theme surface (thinking ramp, syntax, diff, markdown, export tokens), selectable via /theme or settings.theme.
  • Mermaid: mermaid fences render as themed Unicode diagrams in the chat (markdown.mermaid setting: off / final / streaming, default streaming; oversized or unsupported diagrams fall back to the original fence, and partial-render warnings are summarized on one line once the render settles). Adds the registerMarkdownTransformer extension API — extensions can rewrite user/assistant/thinking markdown before terminal rendering with width awareness (MarkdownTransformContext).
  • Clipboard modernization: copy goes through pi-tui's native host-clipboard module when present (command fallbacks and OSC 52), plus readClipboardText / readClipboardFilePaths; pasting now handles Finder file paths, images, and text.

Agents, tasks & workflows

  • Bounded task graphs: the graph is now a work frontier instead of ever-growing history. Completed tasks get a completedAt stamp; past the 15 most recent they auto-archive to archive.jsonl (read-only history — ids stay resolvable via task_get, task_list gains include_archived: true, archived tasks refuse task_update mutations except explicit deletion, and delete cascades reach archived rows). Tool-result snapshots carry archivedCount, so the /tasks panel and desktop Task panel show the open frontier first with a "+N archived" tail. Stale per-session task-list dirs are removed once per process at startup after 30 days untouched — taskRetentionDays setting (0 disables, mirrored in the desktop settings under the same name). Also fixed along the way: deleting an archived task left dangling references in live siblings.
  • Self-extensible sub-agents are documented and shipped with examples: docs/subagents.md covers the two built-ins (general-purpose, read-only Explore), the Markdown definition format (frontmatter name/description/tools/disallowedTools/model/maxTurns/permissionMode/isolation with worktree filesystem isolation), project-vs-user precedence, per-spawn overrides, and the spawn_subagent/wait_subagent/get_subagent_status/kill_subagent tool surface. examples/agents/ ships pr-reviewer, test-loop, and an Explore override (with a test asserting they parse), and the agent can scaffold new definitions on request.
  • Script workflow mechanics: a workflow is a plain JavaScript file — export const meta = { name, description } plus a top-level-await body — orchestrating in-process background subagents through sandboxed runtime primitives: agent(prompt, {schema, label, model, agentType}) (structured outputs validated against JSON Schemas with retries; proven schema self-contradictions fail statically; stops/failures resolve to null), pipeline(list, fn) (one agent per item, null slots preserved), parallel(tasks), phase(title), log(msg), and the args global. The script holds the loop, branching, filtering, and intermediate results — the model's context only sees the final answer. Determinism guards (Date.now(), Math.random(), no-argument new Date() throw) make runs replayable: run state persists a start-ordered agent log under the session directory, and run_workflow { resume: "<runId>" } re-executes the script — completed agents with unchanged prompts replay from cache, everything from the first changed prompt onward reruns (refusing while its agents are live). Concurrency is capped by workflowMaxConcurrentAgents (1–256, default 16) via a run-wide semaphore; runs are bounded by a 4096-item cap per pipeline/parallel call, a 1000-agent cap, and a 50-minute budget, with a large-run advisory past 25 agents. Saved workflows load from project .a-coder-cli/workflows/*.js (wins) over the personal dir over package-provided workflows/*.js, launch through normal permission evaluation with arg-scoped rules pre-approving by name (run_workflow(<name>)) or for all (run_workflow(*)), and /workflows lists runs.

Computer use

  • One tool, action discriminator: capture (som = screenshot with numbered element overlays + AX element list, vision = plain screenshot, ax = tree only), click/double_click/right_click/middle_click, drag, scroll, type, key, set_value, wait, list_apps, list_windows, focus_app. Element-index clicking with snapshot element_token staleness detection; background-first input delivery (no focus steal) with delivery_mode: "foreground" gated on the driver's live schema (refused, never silently downgraded); a sticky-target guard returns input_target_mismatch instead of typing into the wrong window; every result carries a verdict (done / verify fresh state / escalate). Destructive system shortcuts (logout/lock/empty-trash, alias- and hyphen-proof) and dangerous type payloads (pipe-to-shell, sudo rm -rf, fork bombs) are hard-blocked before approval. Permission-aware: read actions (capture/list/wait) are free; input actions weigh like mutating tools in every mode (blocked in read-only, prompt in ask, classifier-checked in auto). Fail-closed recovery: a timed-out driver call's outcome is unknown — never replayed, transport rebuilt before the next call. Driver resolution: A_CODER_CUA_DRIVER_CMD override, then PATH; child env strips provider API keys and disables cua-driver telemetry by default. The TUI settings screen gates enabling behind a disclaimer submenu (labeled screenshots, no focus steal, permission-mode gating, hard blocks) with an explicit "I understand — enable" choice.
  • doctor action (free, no approval): runs the driver's health_report when advertised (schema v1 contract: overall status, per-check pass/fail/skip with remediation hints — binary version, platform, session, bundle identity, TCC accessibility + screen recording, AX and capture capability), falling back to a minimal check_permissions report on older builds. Backend failures now carry a hint pointing at the doctor action and A_CODER_CUA_DRIVER_CMD.
  • Screenshot dedup (context economics): a capture of the same target (app, window) whose bytes are identical to the last delivered frame omits the image block in favor of an explicit "screen unchanged" note (element indices and tokens remain current). Append-only, bounded by a streak cap of 2 consecutive omissions so full pixels are re-delivered before state can drift; ax captures never update dedup state.
  • Actionable install guidance: failed computer calls (including action: "doctor") return a backend_unavailable payload with an install object — platform/arch-matched release asset (tag cua-driver-rs-v0.33.3), direct URL, the agent bin dir, and copy-pasteable install commands — instead of a vague "install it from the releases" message. Driver auto-resolution also checks ~/.a-coder/cli/agent/bin/ (cua-driver.exe on Windows) after PATH, where is used instead of which on Windows, and a new docs/computer-use.md documents enabling, backend resolution, install commands, doctor diagnostics, and the security posture.

Providers & models

  • Sign in with ChatGPT (/login openai): the openai provider gains an OAuth method (ChatGPT subscription) via the new pi-ai flow — fixed-port browser callback with manual paste fallback, port-in-use failures instead of state-mismatch races, and a device ID (settings deviceId) sent as the agent host ID.
  • vLLM and SGLang local servers: served models discovered from /v1/models with real context windows (max_model_len when reported), listed and selectable like other keyless local providers (/login vllm / /login sglang set a base URL; defaults http://localhost:8000/v1 and http://localhost:30000/v1), refreshable on a 2-min cadence, and configurable via the desktop Settings "Local models" section (VLLM_BASE_URL / SGLANG_BASE_URL env or localProviders.vllmBaseUrl / sglangBaseUrl settings).
  • New Unsloth provider (keyless local OpenAI-compatible server, default http://localhost:8888/v1, UNSLOTH_BASE_URL override): display name and ModelRegistry dynamic refresh with the same best-effort/retry policy as the other local providers; /login unsloth shows a base-URL dialog and desktop Settings → Models → Local models gains an "Unsloth address" field (localProviders.unslothBaseUrl).
  • Re-added the OpenAdapter provider (removed in v0.80.101): default model mapping, display name, and the dynamic /v1/models refresh.
  • Fresh installations come with the chrome-devtools MCP server pre-configured: the first run seeds a mcpServers entry (chrome-devtools-mcp@latest via npx, cmd /c npx on Windows) with --isolated (throwaway per-session Chrome profile) and --no-usage-statistics / --no-performance-crux (the server's telemetry off), so browser control works straightaway. The seed persists immediately — subsequent partial settings saves keep it — and is then an ordinary user-owned entry: deleting it sticks, and any pre-existing settings file (even {}) disables seeding entirely.

RPC & serve

  • Mobile bridge: a-coder-cli serve spawns the engine (--mode rpc, same contract as the desktop app) lazily on the first authenticated client and forwards NDJSON frames verbatim in both directions, injecting only "type":"bridge" control frames (connected, engine_exited, engine_failed, shutting_down, server_version). Auth accepts a Bearer header or ?token=. Pairing: a per-machine 256-bit token (timing-safe compare, persisted under ~/.a-coder-cli/agent/serve/token, --rotate to regenerate) printed to the terminal as a QR code plus 4×8 manual chunks; discovery via mDNS (_a-coder._tcp, --no-announce to disable); failed handshakes rate-limited per IP (10 per 15 min); slow clients dropped above an 8 MiB buffer; crash-restart policy max 2 restarts; default idle-exit after 30 clientless minutes (--idle-exit 0 disables). Flags: --cwd, --port (8787 with auto-probe), --bind, --provider/--model, --continue, --token, --max-clients, --idle-exit, --no-auth (loopback only), --no-announce, --rotate.
  • The serve bridge sends a per-client server_version control frame immediately after the WebSocket handshake completes (clients treat it as auth-complete and start their initial sync; matches the mobile app's pairing protocol).
  • RPC engines announce themselves: the first stdout frame in --mode rpc is an engine_info event carrying the engine version and a protocol contract number (RPC_CONTRACT in rpc-types.ts, currently 1), so embedded clients can detect engine/client skew early. Clients predating the event ignore it and treat the engine as contract 1; the serve bridge's server_version handshake frame and /info endpoint carry a matching contract field.
  • Serve bridge: client lines the engine cannot accept (not started, exited, or crash-restarting) are refused back to the sender with a new engine_unavailable bridge control frame instead of being dropped silently, so an approval answered during an engine crash can no longer vanish while the agent blocks on it forever. A stdin error handler also keeps a write racing the engine's death (EPIPE) from taking down the whole serve process.
  • New queue_remove RPC command ({ kind: "steering" | "followUp", index }) with AgentSession.removeQueuedMessage and per-index steering/follow-up removal in the agent queue — clients (the desktop queued-messages panel) can dequeue individual items instead of only clearing the whole queue via clear_queue.
  • set_session_name now accepts an optional sessionPath so clients can rename background sessions without switching to them; without it, the active session is renamed as before.
  • get_state now includes the active session's workspace (cwd), and switch_session responses carry the authoritative session identity (sessionFile, sessionId, sessionName, cwd) so clients can adopt tab labels and workspace indicators immediately without depending on session_start event ordering.

Cron & activity

  • Cron core: jobs persist in ~/.a-coder/cli/agent/cron/jobs.json scoped to the project they were created in and fire on a process ticker — the prompt is delivered to the project's active session (queued behind a running turn) or, when the project has no open session, to a continuity background session that repeat runs reuse. Fires are bounded by a 15-minute abort and run unattended with automatic tool permissions. Surface: /cron (list/add/run/pause/resume/remove), RPC cron_list/cron_create/cron_update/cron_delete/cron_run_now with live cron_update snapshots, and a Scheduled panel in the desktop right sidebar (run-now, pause/resume, edit, delete).
  • Run history + activity inbox: every cron run is persisted (~/.a-coder/cli/agent/cron/runs.json, last 20 per job: trigger, delivery, status, error, session file) and pushed to RPC clients as cron_run started/finished events; cron_runs fetches a job's history. The desktop Scheduled panel grows a per-job run history (status dots, expand, and Continue — one click reopens the run's session to pick the conversation back up), and a new Activity right-sidebar tab aggregates the inbox: pending approvals and run outcomes sort to "Needs attention" with an unread badge; anything with a session opens it on click.
  • Event triggers (Loops-style): jobs can fire on workspace events beside schedules — on:turn-end (after the agent finishes a turn in the project, coalesced by a quiet period, default 30 min) and on:commit (once per new commit, deduped by a persisted head watermark). Surface: /cron add <name> on:turn-end|on:commit <prompt>, an "On event" schedule kind in the desktop editor, and the event schedule kind over RPC.

Engine

  • Binary builds hardened against the upstream Bun non-AVX2 regression (oven-sh/bun#30613): x64 engine targets are compiled with bun 1.3.8 — the last version that works on CPUs/VMs without SSE4.2/POPCNT (bun >= 1.3.9's WebKit -march=nehalem change makes compiled engines crash with 0xC0000005 at startup) — while arm64 targets (including the windows-arm64 cross-compile, which exists only since bun 1.3.10) build with bun 1.3.13. CI supplies both (BUN_BASELINE_BIN), and build-binaries.sh enforces the per-ISA policy with instructions, keeping local release builds from shipping crashing engines.
  • Upstream tier-1/2 port: new before_provider_headers extension hook (upstream earendil-works#6350) — extensions mutate per-request provider headers in place after static assembly and before the HTTP call, with throwing handlers isolated. Session metadata reaches bash as A_CODER_SESSION_ID / A_CODER_SESSION_FILE / A_CODER_PROVIDER / A_CODER_MODEL / A_CODER_REASONING_LEVEL, resolved per command start (upstream earendil-works#6967), with a new environment-variables doc and a bash exposeSessionEnvironment opt-out. Skills stay available with bash-only toolsets (read OR bash is a loader; upstream earendil-works#8552). Tool-result images auto-resize per the existing autoResize setting, including images injected by extensions (upstream earendil-works#7330). AI_AGENT env marker for child-tooling attribution (upstream earendil-works#7493); shellPath expands ~ (upstream earendil-works#6470); sub-agents inherit the session's model/per-model overrides and agent config (upstream earendil-works#7897); edit accepts a single {oldText, newText} object in every caller shape it arrives in (upstream earendil-works#8011); image MIME detector export (upstream earendil-works#8600); truncation-recovery failures now say whether context-overflow compaction or an output-cap cut failed (upstream c7c763f); length-truncated summaries are rejected rather than checkpointed (upstream earendil-works#7048 semantics, already present, now tested).
  • Auto-update handoff marker: the TUI's self-update writes update-handoff.json before relaunching, and the next startup reports the outcome — "updated to vX" when the relaunch landed on the new version, or "an update to vX did not complete" with a retry hint when it didn't. The auto-update also no longer tears down the TUI mid-compaction (previously only mid-response was guarded).

Desktop

  • Onboarding tour: first launch plays a ~2-minute guided walkthrough video (chat and sessions, the model picker, approvals, workflows/ultracode, Your Office, scheduled tasks, the activity inbox, voice mode, and CLI engine parity) before the folder picker; it autoplays muted with burned-in captions and can be skipped. Settings > General gains a "Welcome tour" control that replays it anytime; if the bundled video is missing the modal falls straight back to the folder picker.

Docs

  • Documentation sweep (2026-10-05): all docs pages verified against the current feature surface. Fixed shipped git-conflict markers in settings.md, stale .pi/ project paths, pi CLI invocations and product naming, dead pi.dev endpoint references, and wrong shortcut claims (Shift+Tab is the permission-mode selector; thinking cycles on Shift+Ctrl+M; tool-output toggle is Ctrl+Shift+O); removed obsolete env rows (PI_CACHE_RETENTION is CLI-managed; session env is A_CODER_SESSION_*). settings.md gained the TUI/fullscreen, computer-use, MCP servers, memory, status line/output style, workflows, updates, and local-model-server setting tables; providers.md documents the keyless local model servers (Ollama/Ollama Cloud/LM Studio/llama.cpp/Unsloth/vLLM/SGLang), the new OAuth providers, and the new API-key providers; rpc.md documents all previously-undocumented methods (queue management, permission/plan mode, auth/models, session management, Composio apps, office errands) and the new events; extensions.md gained registerMarkdownTransformer, the before_provider_headers hook, and ui_prompt_start/end; the desktop README now matches the shipped app (MCP editor, local model servers, mobile access, computer use, shortcuts, settings panels); quickstart/index got the current defaults, install script, and browser-control-out-of-the-box note.

Fixed

  • Fixed a crash of the running TUI when a long-running bash command finished: the partial-update path returned the live progress component, which the finished-render path then reused — type-confused — as the bash result component, so the output-preview cache closure rendered against a missing .state and the first synchronous repaint (the fullscreen renderer's mouse-dispatch hit-testing) threw an uncaught TypeError and exited pi. The result renderer now reuses a lastComponent only when it actually is a BashResultRenderComponent.