CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

0.9.0 - 2026-08-24

Added

• taskScope: "session-global" — per-session tasks kept outside the workspace. Session scope writes .pi/tasks/tasks-<sessionId>.json into the workspace, so every repository pi runs in picks up an untracked directory holding a file keyed by a session ID: data of no use to anyone else cloning it, and one more .gitignore rule per project. The new scope keeps the same per-session isolation but stores the file at <agent-dir>/tasks/sessions/<project-key>/tasks-<sessionId>.json (~/.pi/agent/tasks/… by default, following pi's configured agent path like every other piece of user-level state), where <project-key> encodes the workspace path the way pi encodes it for its own session logs (--Users-me-work-repo--), so a project's tasks sit under the same name as its transcripts and same-ID sessions in different workspaces stay apart.

  session remains the default and is unchanged — upgrading moves nothing. Opting in is non-destructive too: it only changes where new session files are created, and a session that already has a file in <workspace>/.pi/tasks/ keeps using it, so switching back strands nothing. (#61 — thanks @kunaaal13 — for #53 — thanks @rfgamaral — and #57)

• Every glyph the widget draws is configurable. A glyphs map in tasks-config.json at either scope replaces the status marks (✔ / ◼ / ◻), the active-task spinner frames, and the widget's furniture — header bullet, overflow and clipping markers, blocked-by marker, token arrows, stats separator, active-row ellipsis, and the glyph on the collapsed N completed line. Enough to run the widget in pure ASCII on a terminal without a Nerd Font, or in a set of your own. Every key is optional and falls back on its own, so a partial or malformed map still leaves every task marked, and glyphs merge per key across the global and project files rather than as a block. Glyphs are the first config value to reach the terminal as free text, so control characters and bidi overrides are rejected per glyph — a .pi/tasks-config.json arriving with a cloned repository cannot retitle your window or break a widget line in two. Config file only, not in /tasks → Settings. See CUSTOMIZING.md. (#60 — thanks @ningw42)

Changed

• TaskCreate is told how to create a batch in one turn. Its description now says that several tasks are created by issuing one TaskCreate call per task in a single response — pi runs independent tool calls in parallel — rather than leaving the model to serialize them across turns. No schema change: subject and description stay required, and there is no batch parameter to get wrong. (#58 — thanks @elecnix)

Fixed

• TaskOutput returns a finished subagent's result, and dismisses its completion notification. It reported only Task #N [status] — subagent <id>, so a model that had just blocked on the join still had to call TaskGet to read what the agent produced — and pi-subagents, never told the result had been read, delivered its completion notification anyway, after the parent had already answered, costing another turn. The result (or the error) now comes back under the status line, and handing it over consumes it through the new subagents:rpc:consume RPC. The suppression half needs a pi-subagents that answers that channel; one that does not keeps notifying rather than failing, and an agent that is still running is never consumed — its notification is the only thing that will announce it. (#62 — thanks @felipe3dfx)
• A failed subagent run no longer leaves the previous run's result behind. The failure listener recorded lastError and reverted the task to pending but kept metadata.result from an earlier successful run, so a retry that failed carried both — and every reader had to guess which was current. A task back to pending now has no result, which is what TaskOutput, TaskGet and a cascaded agent's injected prerequisite output all already assumed. An intentional stop still keeps its partial result: that task is completed, not retried.
• The test suite no longer reads or writes the real ~/.pi/. Shared named lists and the global tasks-config.json both resolve from the home directory, so the suite left files in the home directory of whoever ran it and picked up their global settings. HOME now points at a scratch directory for the duration of the run, and PI_CODING_AGENT_DIR is cleared — getAgentDir() consults it first, so redirecting the home directory alone would not have contained a contributor who has it set.
• vitest.config.ts is no longer published to npm. It refers to test/, which .npmignore excludes, so the tarball carried a config file pointing at files that were not there.

0.8.0 - 2026-08-16

Added

• Custom widget sort orders. sortOrder now also accepts a sort spec — an ordered list of comparison keys (id / status / updatedAt), each with an optional direction and, on status, a custom rank — written into tasks-config.json at either scope. A new active preset (in-progress → pending → completed, then by id) covers the common case from /tasks → Settings without editing a file. Specs are data, not code: anything unrecognised falls back to id order, which also tightens a typo'd sortOrder name that previously left tasks in storage order. See CUSTOMIZING.md. (#35 — thanks @Minh-Ng — and #54 — thanks @Rizhiy)
• Completed tasks can be collapsed to a single line. collapseCompleted (default false, toggleable from /tasks → Settings) replaces finished tasks with one ✔ N completed line at the bottom of the widget. It composes with the visible limit rather than overriding it: collapsing decides what is in the list, then maxVisible / showAll / hiddenAt apply to what is left. The header still counts every task. (#54 — thanks @Rizhiy)
• A customization guide — CUSTOMIZING.md, linked from the README: config scopes and precedence, the display settings and the order they apply in, sort presets and specs, recipes, and troubleshooting.

Fixed

• An unrecognised hiddenAt value no longer blanks the task widget. Anything other than top or bottom in a hand-edited tasks-config.json threw on every render; the widget's render guard caught it, so the widget silently showed nothing at all with tasks still in the list. The value now falls back to bottom.
• A new batch of tasks no longer lands under the completed list from the last one. Both auto-clear modes count in turns and only tick at turn_start, so the countdown stops when the agent does — and a run that ends right after its final completion, which is the usual shape, freezes one mid-count. The finished list then stayed on screen, and the next batch was appended to it (#4 under three completed rows) rather than starting fresh, which reset the countdown again; across a session of short batches the list never cleared. A task created after the agent has stopped now retires a list that has nothing left to do, before it is added to it. The finished list still stays visible while you read it and through follow-up questions, a list with unfinished work in it is untouched, and task IDs remain monotonic. Tasks the agent adds while working through a list are unaffected — that is the same batch taking shape, not a new one, which is why the run boundary rather than the list's state decides. Also fixes the same problem on a resumed session, where the countdown never started at all. (Reported in #51 — thanks @ElJeskos — and #56 — thanks @kunaaal13)
• An auto-cleared session task file is deleted rather than left behind empty. Clearing writes the emptied list back to .pi/tasks/tasks-<sessionId>.json; the manual and session-start clear paths already removed the file afterwards, the automatic one did not.

0.7.3 - 2026-08-16

Fixed

• The active-task spinner animates at a steady rate. Its frame advanced inside the widget's redraw, which also runs on every task mutation and tool execution — so the animation raced ahead during bursts of activity and stalled when nothing was happening. The frame now advances only on the widget's own 150 ms timer. (Diagnosed in #41 — thanks @xz-dev)
• Task paths resolve from the session's workspace, not the directory pi was started in. The two match in the ordinary terminal flow, but not in a host that serves sessions from elsewhere, or when a session is opened by an explicit path from another project — those wrote every session's tasks into whichever directory pi happened to start in. Task files, .pi/tasks-config.json and relative PI_TASKS paths now all follow the workspace pi reports for the session, the same one its own file tools use. This also makes the taskScope setting take effect on the next session start as its description promises; it previously needed a full restart. If you open sessions by path from outside their project, an existing .pi/tasks/ in your launch directory will no longer be read — the files are untouched, just no longer the ones in use. (#45 — thanks @yeyaowei)
• A malformed task file no longer corrupts task IDs or silently empties the list. normalizeTask hardens each record, but nothing checked the envelope around them. A file that parsed but had no nextId produced the task ID "NaN", and after the next reload IDs restarted at "0" — colliding with tasks already in the list — while a file with no tasks array threw mid-load and left the store wiped. The counter is now derived from the records when the stored one is missing or stale, entries that aren't task records are skipped, and a file that isn't usable leaves the current state untouched.
• Subagents are reattached to their tasks after a reload. The agent→task map lives only in the extension instance, so a reload started with an empty one while the subagents kept running. Every completion or failure event for an agent spawned before the reload was dropped and its task stayed in_progress permanently — and with auto-cascade on, the rest of the chain stalled behind it. The map is now rebuilt on session start from the agentId already recorded in task metadata. Only in_progress tasks are relinked, so a late event cannot resurrect a task that was reverted to pending. Fixes #47.
• A finishing subagent can no longer close an unrelated task in the next session. The agent→task map was never cleared on /new, /resume or /fork, and task IDs restart at 1 in every session — so a mapping held over from the previous session pointed at a different task, which the agent's completion then marked completed and overwrote the metadata of. The map is cleared on every session switch and rebuilt from the incoming session's own tasks.
• pi --no-session no longer leaves a task file behind. pi mints a session ID even when it is not persisting the session, so a session-scoped store wrote .pi/tasks/tasks-<id>.json for a session that could never be resumed, and left it on disk. Tasks now stay in memory whenever pi reports no session file. Fixes #43.
• TaskOutput and TaskStop now act on the task an agent ID refers to. Both tools accept an agent ID in place of a task ID and resolve it through the agent→task map, but then kept querying the store with the original argument. TaskStop therefore stopped the subagent and reported success while leaving the task in_progress and its widget spinner running, and TaskOutput reported the status captured before the wait — visible on file-backed stores, where each read deserializes a fresh object.
• Releasing a file lock no longer deletes a lock that has changed hands. Stale-lock detection asks the local process table whether the holder is alive, so a session in another PID namespace — a container, or a list shared over NFS — can read a live holder's PID as dead and reclaim the lock. The original holder then deleted the successor's lock on the way out, leaving two sessions writing the file at once. Lock files now carry a <pid>:<uuid> token and are only removed by the session that wrote them.
• A lock file left behind without a PID no longer wedges a shared task list. acquireLock() writes the lock file and then its PID, so a crash in between leaves an empty lock naming no process. Stale-lock detection only recognised locks naming a dead process, so every later mutation burned the full 5-second retry budget and threw — permanently, until the file was deleted by hand. An unreadable PID now counts as stale too, after a couple of polls so a live acquirer in that same window is not evicted.
• TaskOutput rejects an empty task_id. It accepts agent IDs and matches them by prefix, and every ID starts with the empty string — so an empty argument silently reported on whichever background agent the internal map yielded first. It now fails with task_id is required, matching TaskStop.

0.7.2 - 2026-07-22

Added

• Global task defaults can be set in <agent-dir>/tasks-config.json (~/.pi/agent/tasks-config.json by default). Project settings in <cwd>/.pi/tasks-config.json override global values key by key, and the settings menu persists only project-level differences. (#36)
• The system-reminder now catches stale in_progress tasks. Previously the reminder could only fire from the tool_result hook, so an agent that finished a turn with text only (no tool call) was never nudged and could leave tasks stuck in_progress. That case is now detected on turn_end and a reminder is scheduled for the next LLM call. (#37)

Changed

• The auto-cascade setting is easier to find in /tasks → Settings because its label now uses the same "auto-cascade" terminology as the documentation.
• System-reminder reshaped after Claude Code's todo reminders — it is now either an empty-list nudge or a JSON echo of the current task list. The echo is capped at 10 tasks (completed dropped first) to bound its size on large or long-lived lists, and says so when truncated rather than presenting itself as the full list. Also restores the cheap store.list() guard on the hot tool_result path and stops mutating the shared cadence config. (#37)

Fixed

• Legacy task files no longer crash the host on load. Task files written before the dependency feature have no blockedBy/blocks/metadata fields; TaskStore.load() deserialized them as-is, so unguarded reads (e.g. task.blockedBy.length in the widget) threw a TypeError. Because the widget renders on a pi-tui timer, that throw was uncaught and killed the whole pi process on the first render after upgrade. Every record is now normalized at the load() boundary (normalizeTask, with type guards for hand-edited files), and the widget render is wrapped in a try/catch returning a safe fallback so a render error can never again escape to the TUI timer. Fixes #33. (#39)
• Persisted tasks are rehydrated correctly on reload, and forks get an independent copy. session_switch is never emitted by the SDK, so the /new and /resume reset logic was dead code and forks silently lost (or leaked writes into) the parent's tasks. All session lifecycle handling is consolidated into session_start, keyed on event.reason: new/resume/fork reset session-scoped state (store pointer, shown flag, cadence, auto-clear) on the cached closure, while startup/reload start clean; on /fork the parent's tasks are snapshotted and seeded into the new session file so writes no longer leak back. Adds TaskStore.snapshot()/seed(). (#34)
• Task-store writes survive removal of the parent directory — acquireLock() and save() now mkdir -p the target directory before writing, so a task write after the .pi/tasks/ directory was deleted mid-session recreates it instead of throwing. (#27)
• The task directory is created lazily, on first write — the TaskStore constructor no longer eagerly creates .pi/tasks/. A session that never persists a task now leaves no .pi/tasks/ directory behind. (#38)

0.7.1 - 2026-06-24

Changed

• TaskExecute now offers a recovery path when subagents are unavailable. When @tintinweb/pi-subagents isn't loaded (or its protocol version mismatches), the tool previously dead-ended with "ensure the extension is loaded". It now tells the agent it can run the work as plain Agent-tool spawns, with the explicit caveat that pi-tasks won't track those — status stays pending, cascade won't fire, and TaskOutput stays empty. Scoped to the unavailable branch so it doesn't conflict with the success-path guideline. Message only; no behaviour change. (#26)

0.7.0 - 2026-05-30

Changed

• ⚠ Behaviour change — system-reminder delivery. The periodic <system-reminder> nudge was previously appended onto the content of whatever unrelated tool (read, bash, grep, …) happened to run when it was due. That misattributed host policy text as tool output and persisted a now-stale reminder into session history, so it reappeared on every later turn even after task tools were used. It is now injected via the context hook as a transient <system-reminder>-tagged user message on the one request where it's due — not persisted, and not attached to any tool result. tool_result is now used solely for cadence tracking and never mutates tool output. The cadence (when a reminder fires) is unchanged; only the delivery mechanism and persistence differ. Cadence logic was extracted into a pure, unit-tested src/reminder-cadence.ts. (#19)

Added

• Configurable widget display settings — four new options in /tasks → Settings (persisted to .pi/tasks-config.json), all defaulting to the previous behaviour: (#22)
  • sortOrder (default id) — id (creation order), status (completed → in-progress → pending), recent / oldest (by last-updated time). Sort logic lives in TaskStore.list(sortOrder).
  • maxVisible (default 10) — caps how many task lines the widget shows (5–100).
  • showAll (default false) — when true, every task is shown regardless of maxVisible.
  • hiddenAt (default bottom) — controls whether the … and N more overflow collapses from the bottom or top of the list; top pairs with sortOrder: status to keep active work visible.

0.6.1 - 2026-05-30

Fixed

• CI test failures on webidl.util.markAsUncloneable is not a function — bumped the GitHub Actions node-version from 20 to 22. Node 20's bundled undici lacks webidl.util.markAsUncloneable, which pi's event/RPC layer (exercised by subagent-integration.test.ts) requires. Matches the same fix in @tintinweb/pi-subagents.

0.6.0 - 2026-05-30

Changed

• Migrated pi peer dependencies to the @earendil-works scope — @mariozechner/pi-coding-agent and @mariozechner/pi-tui (>=0.70.5) → @earendil-works/pi-coding-agent and @earendil-works/pi-tui (>=0.74.0). Package was renamed/rescoped across package.json and src imports; lockfile regenerated. Matches the same migration in @tintinweb/pi-subagents. (#21)

0.5.0 - 2026-04-28

Changed

• Bumped @mariozechner/pi-coding-agent and @mariozechner/pi-tui ^0.62.0 → ^0.70.5. Picks up the TypeBox 1.x validator path (pi 0.69), session-replacement context invalidation (pi 0.69), the cwd-bound built-in tool removals (pi 0.68), and the working-indicator / autocomplete-provider APIs along the way. No public surface used by this extension was renamed or removed.
• Migrated from @sinclair/typebox to typebox ^1.1.34 per the pi 0.69 extension guidance. One-line import change in src/index.ts; all Type.* calls (Object, String, Optional, Boolean, Number, Array, Record, Any, Unsafe) work unchanged.
• Toolchain bumps: TypeScript ^5 → ^6, @types/node ^20 → ^25, @biomejs/biome ^2.3.5 → ^2.4.13, vitest ^4.0.18 → ^4.1.5.

0.4.3 - 2026-04-28

Added

• Cascade dependency-result injection — when autoCascade is enabled, a cascaded subagent's prompt now includes a ## Prerequisite task results section listing each completed blocker's stored metadata.result (capped at 4 KB per dep, with a truncation marker pointing at TaskGet). Cascaded agents previously had no context from their prerequisites. (#7)

Performance

• Spinner render rate — reduced widget animation interval from 80 ms (12.5 fps) to 150 ms (~6.7 fps). pi-tui's requestRender() triggers a full component-tree re-render with no scoped invalidation, so the spinner alone could drive sustained ~70-100% single-core CPU on long sessions. ~47% fewer renders, well above the perceptual threshold where the twinkling-star animation reads as alive. (#11)

Fixed

• TaskUpdate.status schema — replaced the anyOf of enum+const shape with a single flat enum: ["pending", "in_progress", "completed", "deleted"]. Some LLMs (notably Gemini and earlier Claude variants) parsed the previous shape into double-quoted values like "\"completed\"", causing TaskUpdate calls to silently fail validation. The accepted value set is unchanged. (#13)
• TaskExecute model parameter now actually forwards — the model option was declared on the tool and captured into the cascade config, but silently dropped at both spawnSubagent call sites (initial spawn and cascade). Now propagated end-to-end. (#7)

0.4.2 - 2026-03-24

Added

• Task numbers in widget — each task line now shows its #id (e.g., ◻ #3 Do something), making blocker references like blocked by #3 easy to cross-reference at a glance. Numbers are dimmed so they stay readable without competing with the task subject.

0.4.1 - 2026-03-22

Added

• Auto-clear completed tasks — new autoClearCompleted setting with three modes:
  • never: completed tasks stay visible until manually cleared
  • on_list_complete (default): completed tasks are cleared after all tasks are done and a few turns pass — user sees the "all done" state before cleanup
  • on_task_complete: each completed task is cleared individually after a few turns
  • Both auto-clear modes use a turn-based delay (matching REMINDER_INTERVAL) for consistent, non-jarring UX — tasks linger briefly so the user sees the completion before they disappear
• AutoClearManager — extracted, testable class (src/auto-clear.ts) handling turn-based clearing logic with per-task and batch countdown tracking
• 20 new unit tests — full coverage of all three auto-clear modes, turn delays, dependency cleanup, batch reset, dynamic mode switching, session reset, and store swap

Changed

• Settings — /tasks → Settings now shows "Auto-clear completed tasks" toggle with never / on_list_complete / on_task_complete values. Also configurable via .pi/tasks-config.json.

Fixed

• /new and /resume now correctly switch session state — storeUpgraded and persistedTasksShown flags were never reset on session_switch, causing the store to stay pointed at the old session file and the widget to not refresh. All session-scoped state (turn counters, reminder flags, auto-clear tracking) is now reset on both /new and /resume. Memory-mode tasks are explicitly cleared on /new.

0.4.0 - 2026-03-22

Added

• RPC-based subagent spawning — TaskExecute now communicates with @tintinweb/pi-subagents via a standardized RPC envelope (rpcCall helper) with protocol version negotiation and timeout handling.
• RPC-based subagent stopping — stopSubagent sends stop requests via subagents:rpc:stop event bus RPC.
• TaskOutput supports subagent tasks — can wait for subagent completion with blocking/timeout, using subagents:completed and subagents:failed events.
• TaskStop supports subagent tasks — stops running subagents via RPC and marks the task as completed.
• Debug logging — set PI_TASKS_DEBUG=1 to trace RPC communication (request/reply/timeout) and spawn errors to stderr.
• TaskExecute prompt guidelines — agents are instructed not to use the Agent tool for tasks already launched via TaskExecute.
• Biome linter — added Biome for correctness linting.

Changed

• TaskOutput/TaskStop accept agent IDs — both tools now resolve agent IDs (including partial prefixes) to task IDs via agentTaskMap, fixing the mismatch where TaskExecute returns agent IDs but TaskOutput/TaskStop only accepted task IDs.
• TaskGet shows metadata — non-empty metadata is now displayed in TaskGet output as JSON.
• TaskGet filters completed blockers — consistent with TaskList, TaskGet now only shows open (non-completed) blockers instead of all dependency edges.
• TaskExecute success message — now includes guidance to use TaskOutput for progress and not spawn duplicate agents.
• Softened TaskExecute description — removed "Requires @tintinweb/pi-subagents extension" from the tool description to prevent agents from refusing to use it when the extension is loaded.
• Stopped subagents handled gracefully — subagents:failed listener now distinguishes intentional stops (status "stopped" → mark completed, preserve partial result) from actual errors (revert to pending).

0.3.3 - 2026-03-17

Added

• Session-scoped task storage — new taskScope config with three modes: memory (in-memory only), session (per-session file, default), project (shared across sessions). Session mode uses tasks-<sessionId>.json, surviving session resume while keeping sessions isolated.
• Session resume support — session_switch event handler reloads persisted tasks on resume without auto-clearing completed tasks (user may want to review).
• Session file cleanup — empty session task files are automatically deleted when all tasks are cleared, preventing stale file accumulation.
• "Clear all" in /tasks menu — wipe all tasks regardless of status, not just completed ones.

Changed

• Unified storage setting — replaced persistTasks (boolean) with a single taskScope: "memory" | "session" | "project" setting. The persistTasks field is no longer recognized.
• Auto-clear completed on new session start — when all persisted tasks are completed, they are silently cleared instead of showing stale completed work. On resume, completed tasks are preserved.
• Widget only shows on start if there's unfinished work — sessions with only completed tasks start with a clean slate.
• Settings moved to last position in /tasks menu for better UX (actions first, config last).

Fixed

• Robust session store upgrade — store upgrade from in-memory to file-backed triggers on turn_start, before_agent_start, session_switch, and tool_execution_start — whichever fires first.

0.3.2 - 2026-03-17

Fixed

• Completed tasks no longer vanish from the list — completed tasks are now persisted to disk so they survive reloads and show as strikethrough instead of disappearing. Use "Clear completed" in /tasks to explicitly remove them.

0.3.1 - 2026-03-16

Added

• Local-by-default task persistence — tasks now auto-persist to <cwd>/.pi/tasks/tasks.json on every mutation and reload on restart. No config needed. Set PI_TASKS=off to opt out (CI/automation).
• Settings persistence — persistTasks and autoCascade settings survive restarts via <cwd>/.pi/tasks-config.json.
• "Persist tasks" toggle in Settings — /tasks → Settings now shows two toggles: auto-execute and persist. Both are saved immediately to tasks-config.json.
• Completed tasks excluded from disk — only pending and in_progress tasks are written to disk. Completed tasks are in-memory only and pruned on restart. (Reverted in 0.3.2 — completed tasks are now persisted.)
• Absolute path support — TaskStore now accepts an absolute file path in addition to a short list ID.

Changed

• ⚠ BREAKING: PI_TASKS_FILE / PI_TASKS_LIST → PI_TASKS — two env vars consolidated into one. Values: off (in-memory), sprint-1 (named list → ~/.pi/tasks/sprint-1.json), /abs/path (absolute), ./rel/path (relative to cwd). PI_TASKS_LIST=name users: rename to PI_TASKS=name.
• Settings menu — extracted to src/ui/settings-menu.ts and rebuilt using ui.custom() + SettingsList for native TUI rendering: keyboard navigation, live toggle, per-row descriptions, theme-consistent styling.
• autoCascade setting — now loaded from tasks-config.json on startup so the toggle survives restarts.
• Hardened TaskUpdate description — added "Before starting work on a task: mark it in_progress BEFORE beginning" as an explicit use case. Previously this rule only appeared in TaskCreate; now it lives in the tool actually used to set that status.
• Removed before_agent_start system prompt injection — task state is no longer injected into the system prompt on every agent loop. Analysis showed this creates wallpaper noise that trains the model to ignore the task block. Claude Code itself does not do this: the workflow contract lives in tool descriptions (read at decision time) and the periodic <system-reminder> nudge (fired when task tools haven't been used recently). Removed the corresponding 3 tests.
• Widget render-once refactor — TaskWidget now registers the widget callback a single time and uses tui.requestRender() for subsequent updates instead of calling setWidget() on every tick. Rendering logic extracted to renderWidget(). Eliminates redundant callback re-registration and keeps a cached tui reference for lightweight invalidation.

0.3.0 - 2026-03-14

Changed

• Eventbus RPC for subagent communication — replaced the Symbol.for global registry bridge with a proper eventbus RPC protocol. pi-tasks now communicates with @tintinweb/pi-subagents via scoped request/reply channels (subagents:rpc:spawn, subagents:rpc:ping), eliminating shared mutable global state and enabling reliable cross-extension coordination regardless of load order.
• Presence detection — two-path handshake: (1) ping RPC on init with scoped reply channel, (2) subagents:ready broadcast listener. Works whether pi-subagents loads before or after pi-tasks.
• Agent-task mapping — in-memory agentTaskMap (agentId → taskId) replaces linear store.list().find() scans for O(1) completion event lookup.
• Spawn error handling — spawnSubagent() returns a Promise with 30s timeout. Failed spawns revert tasks to pending with error in metadata instead of silently failing.
• Removed SubagentBridge type — the types.ts interface for the global registry bridge is no longer needed.
• Widget icon colors — completed tasks show green ✔, in-progress tasks show accent-colored ◼ (matching Claude Code's UI).

0.2.0 - 2026-03-12

Added

• TaskExecute tool — execute tasks as background subagents via @tintinweb/pi-subagents. Tasks with agentType metadata are spawned as independent agents; validates status, dependencies, and agent type before launching.
• agentType parameter on TaskCreate — opt-in field (e.g., "general-purpose", "Explore") that marks tasks for subagent execution.
• Auto-cascade — when enabled via /tasks → Settings, completed agent tasks automatically trigger execution of their unblocked dependents, flowing through the task DAG like a build system. Off by default.
• Subagent completion listener — listens to subagents:completed and subagents:failed events to automatically update task status. Failed tasks revert to pending with error stored in metadata.
• READY tags in system prompt — pending tasks with agentType and all dependencies completed are marked [READY — use TaskExecute to start] in the system prompt.
• Agent ID in widget — in-progress tasks backed by subagents show the agent ID (e.g., ✳ Writing tests (agent abc12)…).
• Settings menu — /tasks → Settings → toggle "Auto-execute tasks with agents".
• SubagentBridge type — typed interface for the cross-extension Symbol.for bridge.

Changed

• @tintinweb/pi-subagents global registry now exposes spawn() and getRecord() in addition to waitForAll() and hasRunning().
• @tintinweb/pi-subagents emits lifecycle events on pi.events: subagents:created, subagents:started, subagents:completed, subagents:failed, subagents:steered.
• AgentManager accepts an optional onStart callback, fired when an agent transitions to running (including from queue).

0.1.0 - 2026-03-12

Initial release — Claude Code-style task tracking and coordination for pi.

Added

• 6 LLM-callable tools — TaskCreate, TaskList, TaskGet, TaskUpdate, TaskOutput, TaskStop — matching Claude Code's exact tool specs, descriptions, and schemas.
• System-reminder injection — periodic <system-reminder> nudges appended to non-task tool results when tasks exist but task tools haven't been used for 4+ turns. Matches Claude Code's host-level reminder mechanism.
• Prompt guidelines — promptGuidelines on TaskCreate injects persistent guidance into the system prompt, nudging the LLM to use task tools for complex work.
• Task state in system prompt — before_agent_start event appends current task state to the system prompt on every agent loop, ensuring task awareness survives context compaction.
• Persistent widget — live task list above editor with ✔ (completed, strikethrough + dim), ◼ (in-progress), ◻ (pending), animated star spinner (✳✽) for active tasks with elapsed time and token counts (e.g., ✳ Running tests… (2m 49s · ↑ 4.1k ↓ 1.2k)).
• Multiple parallel active tasks — widget supports multiple simultaneous spinners.
• /tasks command — interactive menu: view tasks with actions (start, complete, delete), create tasks, clear completed.
• Bidirectional dependency management — addBlocks/addBlockedBy maintain both sides automatically. Edges cleaned up on task deletion.
• Dependency warnings — cycles, self-dependencies, and dangling references produce warnings in TaskUpdate responses. Edges are still stored, matching Claude Code's permissive behavior.
• File-backed shared storage — set PI_TASK_LIST_ID env var for multi-session coordination at ~/.pi/tasks/<id>.json. File locking with stale-lock detection prevents race conditions.
• In-memory session-scoped mode — default when no env var is set, zero disk I/O.
• Background process tracker — output buffering (stdout + stderr), waiter notification, graceful stop with timeout escalation (SIGTERM → 5s → SIGKILL).
• 78 unit tests — task store CRUD, dependencies, warnings, file persistence; widget rendering, icons, spinners, token/duration formatting; process tracker lifecycle.