Files
my-pi/pi-subagents/CHANGELOG.md
T

164 KiB
Raw Blame History

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]

Added

  • Native my-pi permission child lifecycle bridge. Child sessions now publish subagents:child:session-created synchronously before bindExtensions() and subagents:child:disposed after actual teardown. Every descendant keeps the interactive root session as its permission authority, and child prompts replace inherited identity tags with their stable agent key. This lets pi-permission-system apply root whole-session grants to child-fixed access facts, run auto-review against the trusted root transcript, and forward unresolved ask decisions to the root UI without treating model-generated child prompts as direct user authorization.
  • Host-owned mandatory extensions with fail-closed loading. createPiSubagentsExtension({ mandatoryExtensionPaths }) lets a bundle protect exact canonical extension entries from Agent-controlled extensions: false, isolated, and exclude_extensions. Missing or substituted entries abort before session creation. Mandatory handlers bind without automatically surfacing their tools, preserving the separation between security lifecycle and Agent tool visibility.

Changed

  • Enabled in the my-pi bundle. The root package now installs this source through file:./pi-subagents, loads the host-owned wrapper by default, and assigns the three orchestration tools to Tool Search's checked-in subagents group. Pi host peers are optional in this maintained package so the bundle continues to use the host runtime instead of installing a duplicate. The off-screen mention clone now disables extension/resource discovery so Tool Search and other active-tool owners cannot remove its single synthetic Agent tool during session_start.

[0.18.0] - 2026-08-20

⚠️ Breaking — an Agent call that doesn't say now runs in the background (backgroundByDefault). Following Claude Code, where the agent backgrounds unless the caller passes run_in_background: false. An unqualified spawn no longer blocks the turn and no longer returns the agent's output inline: it returns an ID immediately, and the completion notification carries a preview (500 chars solo, 300 grouped) with the full text behind get_subagent_result. Set backgroundByDefault: false to restore the previous behaviour globally; an explicit run_in_background on the call or in frontmatter overrides the setting in both directions. Nested spawns are unchanged — an agent spawning its own agent still defaults to foreground, because a detached child is stopped by abortOwnedChildren when its parent settles and has no notification path of its own. Two consequences worth planning for: maxConcurrent now applies to nearly every spawn where foreground used to bypass it (its default is raised 4 → 10 to compensate), and Esc interrupts the turn without stopping the agents it started — stop those from /agents → Running agents, which is also how Claude Code separates the two.

Added

  • reportUsage — subagent tokens and cost count towards this session's own totals (#193, #203 — thanks @johnstegeman and @unrelentingfox). Subagents run in their own pi sessions, so nothing they spent reached the parent's getSessionStats(): a session that delegated most of its work read as nearly free in the footer, the statusline and /cost. With the setting on, each Agent / get_subagent_result / steer_subagent result carries the spend accumulated since the last one, which pi folds into the session totals and attributes to /cost's "Tools/summaries" bucket. Off by default — it changes numbers you may already be tracking. Agents that finish in the background have no result of their own to ride on, so their spend lands on the next call rather than the moment they finish; nested agents are covered, counted once each, on the top-level result. Cost is pi's own per-message figure, and the context-window percentage is unaffected. Implemented independently in #202, whose cacheRead reasoning corrected this one: the cached prefix is genuinely re-read and re-billed per call, and pi counts it that way for the session's own messages, so the report includes it even though the extension's own token displays still exclude it (#38).
  • showCost — an estimated cost beside subagent token counts (#194 — thanks @daanzu, whose PR is the display design here). Shown in the widget, FleetView, the conversation viewer, foreground results, get_subagent_result and completion notifications, with a batch total when several background agents finish together. Off by default, and shown only when there is a cost to show: a model pi has no pricing data for reports zero, and $0.00 beside its tokens would say the run was measured and found free rather than never measured. Figures keep cents at minimum and four decimals at most (~$0.0042, ~$1.24), because rounding everything to cents printed one number for runs that differed fourfold.
  • Lifecycle events carry the run's spend as a pi Usage (#137, #138 — thanks @leset0ng). subagents:completed and subagents:failed gain a usage field alongside tokens, carrying the token components — cacheRead included — and cost.total in USD. It is pi's own Usage, the shape pi puts on ToolResultEvent and AssistantMessage, so the money is where a listener already looks and anything pi adds to Usage needs no change here. Omitted when nothing was spent, so spent nothing stays distinguishable from never ran. tokens is unchanged and still the display total, which excludes cacheRead (#38) — the two answer different questions and neither derives from the other. This is the aggregate only — the per-message {timestamp, provider, model, usage} records those asked for, and the per-model attribution they enable, remain out of scope (#137 was declined and #138 withdrawn by its author).

Changed

  • The pi peer floor moves from >=0.80.0 to >=0.81.0. reportUsage is built on pi folding toolResult.usage into getSessionStats(), which pi only started doing in 0.81.0 — every 0.80.x sums assistant messages alone and drops the field, so the setting would have been present and silently inert there. The CI floor job installs 0.81.0 and runs the suite against it, which is what caught this; the e2e that pins the behaviour runs unconditionally rather than skipping, so the range cannot quietly stop being true. Nothing else in the extension needed 0.81.
  • Top-level Agent spawns default to background, behind a new backgroundByDefault setting (see the breaking note above). Configurable at /agents → Settings → Background by default and in subagents.json; applied live. Nested spawns pass false unconditionally and ignore the setting.
  • The background concurrency limit's default is raised from 4 to 10. Foreground agents bypass the maxConcurrent pool entirely, so while foreground was the default a six-way fan-out ran six; with background as the default every top-level agent takes a slot, and a limit of 4 would have silently queued the tail of exactly the parallel fan-outs the Agent tool description tells the model to send. Still configurable — the ceiling is unchanged.
  • The agent-file writer emits run_in_background: false rather than dropping it. It wrote the field only when truthy, which was lossless while omission and false meant the same thing; under backgroundByDefault they no longer do. No current path reaches it — serializeAgentFile's only caller is Eject, which is offered for built-in defaults, and those omit the field — so this is the writer half of the format keeping pace with the loader, not a fix for observable behaviour.

Fixed

  • Subagent sessions now close the extension lifecycle they open (#242 — thanks @bsv9). Every child session binds extensions so session_start fires and they can set up per-session state, but nothing emitted the matching session_shutdown: AgentSession.dispose() only invalidates the ExtensionRunner. Whatever an extension armed in session_start — timers, fs watchers, sockets, temp dirs — therefore leaked one set per spawn, teardown a session_shutdown handler owns (killing runs started by that child, say) never ran, and the 10-minute record sweep turned the leak into a crash: disposing invalidated the runner while a leaked timer was still armed, so its next tick threw assertActive() from a bare Timeout._onTimeout — an uncaughtException that took interactive pi down with it. Both eviction and quit now emit session_shutdown (reason quit) before disposing, which is what pi itself does in AgentSessionRuntime.dispose(). Quit waits for those handlers rather than exiting past them, bounded at three seconds per session so one that hangs cannot strand you at a torn-down TUI; a session whose extensions register no handler is untouched and still disposes synchronously. Independently fixed in #228 and #170 — thanks @v-kuskov and @liu-qingyuan.
  • The Agent tool no longer tells the model that foreground agents run one at a time (#232 — thanks @willfenton). "Foreground calls run sequentially — only one executes at a time" was never true: pi's agent loop dispatches a message's tool calls through Promise.all unless the whole batch opts out via toolExecution: "sequential" or some tool in it declares executionMode: "sequential", and this extension sets neither — two foreground Agent calls in one message start within microseconds of each other. Nothing serialized them at this layer either; agent-manager deliberately exempts foreground agents from the maxConcurrent queue, since they block the parent anyway. The claim was self-inflicted rather than inherited: d0cb511 replaced a correct bullet with it, and the later upstream-alignment pass grafted it onto the end of Claude Code's real sentence instead of replacing it, which also left an invented with run_in_background: true on each qualifier in the middle — upstream says a single message with multiple tool uses runs concurrently, full stop. Both are gone, restoring upstream's wording (minus its build-validator/test-runner example, consistent with the examples already omitted for token cost). The cost of the error was steering: an orchestrator that wanted parallelism was pushed into background spawns it did not need, paying a queue slot and a notification round-trip for concurrency it already had. Investigating where the sentence came from is what surfaced the default divergence in the breaking note above, so the two ship together — the remaining foreground/background prose is now Claude Code's own, including the Don't race bullet that only earns its place once background is the default.
  • Scheduled and RPC-spawned agents show their token counts in the widget, and finished agents keep theirs in the conversation viewer. Both surfaces read spend from the live activity tracker, which only Agent-tool spawns get and which is deleted the moment an agent finishes — so those agents rendered with no token stats at all. Every surface now reads it from the agent's record instead: the record is the only total that outlives the run and the only one a nested child's spend is folded into, so the figure no longer jumped upward at completion as the read switched from one to the other. The tracker keeps what is genuinely live — tool activity, turn count, context percentage — and no longer accumulates a second copy of the totals. Spotted in #194 — thanks @daanzu.
  • Agents started outside the Agent tool say what they are doing, instead of thinking… for their whole run (#181 — thanks @xz-dev). The widget's activity line and turn counter come from an activity tracker that only the Agent tool handler created, so an agent started through cross-extension RPC (the path TaskExecute uses), through an @handle mention, or through the Symbol.for("pi-subagents:manager") registry showed a permanent thinking… while the same row's tool-use count climbed beside it and the conversation viewer showed the real work. The tracker now belongs to the one funnel all three spawn paths pass through, so none can be missed and none can supply half-wired callbacks of its own; the Agent tool reaches the manager directly, so nothing is double-tracked. Its turn ceiling is resolved exactly the way the run resolves the limit it enforces — explicit value, else the agent's own max_turns, else the project default — rather than read off the caller's options, which a mention deliberately omits so the agent file can decide: the row would otherwise render ↻3 where the tool renders ↻3≤20. Scheduled jobs still spawn through the manager and so still render without per-tool detail.

[0.17.1] - 2026-08-18

Fixed

  • A model-mode mention now actually reaches the agent it starts. The off-screen clone was built with an empty tool allowlist, which stripped its own Agent tool, so every start fell back to a direct spawn with a warning. Fixing that surfaced a second bug behind it: the clone spawned in the foreground, answering into a session disposed moments later, so the agent ran and reached nobody. Its spawn is now forced to the background.
  • @ offers files again. Handle rows are merged into pi's own suggestion list instead of replacing it — the provider asked pi only when no agent matched, and prefix matching meant an empty token matched every handle, so a bare @ listed no files at all.

[0.17.0] - 2026-08-17

⚠️ Note — an agent file's frontmatter name: now substitutes for the filename as its subagent_type. Following Claude Code, the declared name is the dispatch identity and the filename is only the fallback, so blubb.md with name: code-review is spawned, mentioned and listed as code-review. Any value is accepted except one containing :, which Claude Code reserves for plugin scoping; such a file — like any unparseable one — is skipped with a warning rather than loaded under a name nothing honours, and strictAgentFiles turns that into a startup failure.

⚠️ Breaking — subagent sessions now persist to disk by default (rememberAgents). Transcripts that used to be in-memory are written to the session dir and appear nested under their spawner in pi's /resume. Per-agent persist_session: overrides it in both directions — false keeps that agent in memory, true persists it even with the setting off. Set rememberAgents: false to restore the previous behaviour globally.

Added

  • Subagents are addressable from the prompt: @handle message goes to that agent instead of the main model. Reaching one previously meant spending a main-model turn on steer_subagent, or walking the FleetView. The handle names the agent, not one process, so a running or queued agent is messaged, a finished one is resumed in the background, and one that never ran is started — the reply arrives as the ordinary background-completion notification either way. Every agent gets a handle from its type, numbered on collision (explore, explore-2), offered by @ completion alongside pi's file completion. Only a leading @handle followed by a message is a send, so a bare @explore, a mid-sentence mention and @src/index.ts still reach the model; toggle with agentMentions.
  • Mentioning an agent that isn't running now starts it through an off-screen clone of the conversation. A throwaway copy of the session takes a turn holding only the Agent tool, so the model writes the agent's prompt from the conversation instead of forwarding your typed line — the context-aware prompt Claude Code gets from its own mention flow, without a visible turn narrating a decision the handle already made. Nothing enters the chat but a Starting @plan… toast, and what starts is an ordinary top-level agent with a transcript, widget detail and a handle. agentMentions accordingly takes "model" (default), "direct" — the previous behaviour, started from your text with no model call — or "off"; the old booleans still read as "model"/"off". Messaging a running agent and resuming a finished one are direct in both modes, a clone that cannot run falls back to a direct start rather than losing the mention, and "model" lifts the TUI-only restriction for starts, so pi -p '@plan the migration' now works.
  • @handle keeps working after the agent's record is gone. Handles used to expire with the in-memory record ~10 minutes past completion, silently flipping @explore anything else? from resume to start fresh. An evicted agent now leaves a tombstone and the mention reopens its session from disk (requires rememberAgents, below). Only the definition is re-resolved, so a continuation runs under the type's current frontmatter — and a resume whose type has since been deleted or disabled is refused rather than falling back to another agent. Tombstones cap at 100 and clear on /new and session switch.
  • The model can name an agent, and names work wherever ids do. Agent takes an optional name, so an agent can be @auth-audit instead of leaving you to tell @explore-2 from @explore-3. Naming is additive — the type-derived handle is still assigned, and both draw from one namespace, so neither can shadow the other. steer_subagent and get_subagent_result now accept a handle as well as an id, ids first, so existing calls are unchanged.
  • @main <message> forces text to the main model; @agent-<type> is a synonym. main is reserved and can never be allocated to an agent, so a leading @main is stripped and the rest passes through with its attachments — the escape hatch for text that only looks like a mention. @agent-explore fix it is Claude Code's manual spelling; the handle as written wins, so an agent actually called agent-explore stays reachable. Handles cap at 64 characters.
  • worktreeIsolation project setting — turn worktree isolation off for a whole repo (#184 — thanks @tinysnake). On a repo large enough that every worktree costs real time and disk, the only defence was asking the model not to request one. Set "worktreeIsolation": false in subagents.json (or /agents → Settings → Worktree isolation) and the Agent tool's isolation parameter is dropped from the schema entirely, along with the usage-note bullet describing it and the /agents generator's frontmatter field — the same trade schedule already makes when scheduling is disabled, and necessary, since prose left behind teaches the model to pass a parameter that is no longer declared and to report a branch that was never created. The refusal is enforced below the tool boundary as well, so it also covers agent files, scheduled jobs and the cross-extension RPC path, and a requested worktree is downgraded to a normal run rather than failing the call. Default is unchanged (true); the parameter and its prose appear or disappear on the next pi session, since both are built at tool registration, while the refusal applies immediately.

Fixed

  • The Agent tool no longer tells the model a worktree's path is returned in the result. It never was: cleanupWorktree removes the worktree directory before returning and the record keeps only { hasChanges, branch }, so a model acting on the promised path would have been sent to a deleted directory. The neighbouring claim that the worktree is "automatically cleaned up if the agent makes no changes" was misleading in the same way — the directory is removed in both cases; what is conditional is whether a branch survives. Both had been in the tool description since worktree isolation shipped, inherited from Claude Code's wording; the README's completion summary said the same thing loosely and has been corrected too.
  • isolation accepts "off", so a model that fills every optional parameter can decline a worktree (#231, #184, via #201 — thanks @Munsen). As a single-value optional literal, isolation was the one parameter whose only expressible value had an expensive side effect, and every other optional field has an inert filler — so a model that fills everything in reflexively ended up reviewing a staged diff inside a fresh copy of the repo that contained none of the work, over and over, against its own stated intent to omit the parameter. "off" is now a legal value — listed first and described as the default — and is also accepted in agent frontmatter (none, no and false too), where it becomes a genuine veto: frontmatter outranks tool-call parameters, so it refuses a worktree even when the caller asks for one, which no value could previously do. The parameter description now also warns that a worktree cannot see uncommitted or staged work, the specific trap here. One behaviour change to note: an agent file already carrying isolation: none/no/false used to parse to nothing at all, leaving a caller free to supply worktree anyway; it now parses to off and refuses one.

Changed

  • Frontmatter name: is now the agent's type rather than a display-label alias (see the note above). display_name: still sets the label, and a file setting only name: badges as that name, since the label falls back to the type. A file whose name: differs from its filename changes subagent_type; rename the field to display_name: to keep the old identity.
  • Subagent sessions are persisted by default (rememberAgents) (see the note above). They previously ran in-memory unless the agent file set persist_session: true, which left nothing for a handle to reopen once the record was evicted — the reason handles expired. Nested children are excluded, since nothing can address them and so nothing could reopen their transcript.
  • The selected FleetView row renders in the theme's primary text color (#230, #234 — thanks @kunaaal13). Agent name, description and the elapsed/token stats previously kept their muted/dim tokens while selected, so the highlighted row read as several emphasis levels rather than one selection; the marker keeps its accent. A configured agent color still badges on the selected row and is bolded there, matching Claude Code's own FleetView — the color is what identifies the row you are reading, and leaving the badge in place keeps the row's width fixed as the selection moves.

[0.16.1] - 2026-08-15

Changed

  • The pi compile baseline is declared instead of inferred. The @earendil-works/pi-* packages were peers only, so 0.80.6 came from npm's peer auto-install and the committed lockfile rather than from any decision — the emergent pin the 0.14.0 entry below had to diagnose. They are now exact devDependencies, pinned to the current release (0.84.2); the peer range stays >=0.80.0 and consumers are unaffected.
  • compat-latest-pi now runs typecheck and the full suite instead of two hand-picked e2e files, so drift can't hide in an unlisted test — against pi 0.84.2 the job reported green while typecheck and four e2e suites were failing. Still continue-on-error, but both steps pass as of the new baseline, so a failure there is now fresh breakage rather than known debt.
  • The suite runs against pi 0.84.2, so the modelRegistrymodelRuntime migration is exercised by default rather than only by a continue-on-error job. Two places had been carrying it as known-broken: the modelRuntime compat shim in src/agent-runner.ts typechecked only pre-migration (an opaque unknown read off the private facade field can't satisfy the newer ModelRuntime parameter type), and the e2e faux backend faked auth through a structural modelRegistry that post-migration pi no longer accepts, so every faux run died with "No API key found for faux". The faux registry and its post-migration counterpart — auth via getAuth/hasConfiguredAuth, the turn itself streaming through modelRuntime.streamSimple — now live together in test/helpers/faux-model-backend.ts, and both are passed to createAgentSession, each pi ignoring the option it doesn't know.
  • CI covers the floor of the declared peer range, not just the pin. >=0.80.0 used to be tested incidentally, because the inferred pin was a 0.80.x; moving the baseline to 0.84.2 would have left the range asserting support for versions nothing built against. A compat-floor-pi job now installs 0.80.1 — the earliest published release satisfying the range — and runs typecheck plus the full suite. Unlike the latest-pi canary it is blocking: that run is the only evidence behind the range, and if the floor becomes unsupportable the peer range is what should change.
  • publishConfig.access is set to public, so publishing this scoped package no longer depends on --access public or on the access already set from a prior publish.

[0.16.0] - 2026-08-14

Added

  • Custom agents can set name and color, so agent files shared with Claude Code keep their display identity (#216 — thanks @HerbertGao). name: is accepted as a display-name fallback (display_name: still wins, and the filename still determines the subagent_type and dispatch identity), and color: renders the agent name as a badge — the color is the background — in the Agent tool header, the widget, FleetView, and the conversation viewer. Claude Code's eight color names render with Claude Code's own values; six-digit hex and the Agency Agents palette aliases are accepted too, quantized to the xterm-256 palette on non-truecolor terminals the way pi's own theme does. A missing or invalid color preserves each surface's existing theme styling. Badge text is black or white by WCAG contrast against the rendered background, where Claude Code uses a single inverse color for every badge. In the Agent tool header the line paints the tool block's row tint itself and leaves it open: the badge has to close its own background, so it restores that tint, and the TUI pads a line to width after rendering it, so closing the line would leave that padding untinted. The tint is painted only when there is a badge, so an uncolored agent's header renders exactly as it did before.
  • run_in_background now works on resume, so continuing an agent no longer blocks the session (#214 — thanks @akram-ahrardi). The Agent tool's resume branch returned before its background branch, so the flag was accepted and silently ignored: every resume ran inline and held the main loop for the whole turn, however long that agent took. A resume with run_in_background: true now detaches exactly like a background spawn — it returns immediately, queues behind maxConcurrent, streams to the widget and FleetView, and notifies on completion — while a resume without the flag still runs inline and returns its result, unchanged. Detaching required the differences from a spawn to be handled rather than inherited: the run is not wired to the tool-call abort signal (a background spawn omits it too, and forwarding it would kill the agent when the parent turn is interrupted while spawns from that same turn kept going); a second resume of an agent whose run is still in flight is refused instead of replacing the live run's abort controller, which would have put it beyond the reach of /agents stop; the transcript is appended to rather than truncated, since the path is deterministic per agent and the previous run's turns — including any the session has since compacted away — are what the completion notification points at; output streaming is wired when the run actually starts, so a resume stopped while still queued leaves no subscription behind; and the subagents:created event reports the record's own type, since resume ignores subagent_type and a mirror keyed by id would otherwise re-register the agent under a type it never had. The widget's finished-agent age is cleared on resume, or an agent that had already completed once would vanish instead of showing the new run's completion line.

Fixed

  • RPC-spawned agents now appear in the live Widget and FleetView while they are running (#224 — thanks @dereknex). Cross-extension RPC calls entered through AgentManager.spawn() instead of the Agent tool handler, but only that handler started the Widget/Fleet refresh timers. The manager still tracked RPC agents and the completion callback refreshed the UI, so a long-running RPC agent was invisible until it finished and could appear only as a terminal row. The shared manager-start callback now starts and refreshes both native surfaces for every top-level spawn source, including RPC and scheduled agents; nested agents remain intentionally hidden.

Documentation

  • The cross-extension RPC spawn example named an option that does not exist. It passed run_in_background: true, the Agent tool's parameter name; the RPC handler forwards options verbatim to AgentManager.spawn, whose field is isBackground, so the flag was silently dropped — an RPC agent nobody meant to exempt never occupied a maxConcurrent slot and never queued behind one, and subagents:created reported no background flag for it. The example now passes isBackground, which the RPC test suite has always pinned, and the surrounding text says what the flag does and does not change: every RPC spawn returns its id immediately and runs detached either way. It also states the visibility rule that changed above, so the widget/FleetView behavior is documented where callers read about spawning rather than only in this file.

[0.15.2] - 2026-08-14

Documentation

  • Three README statements corrected against the code they describe, each now pinned by a test so the doc and the behavior can't drift apart again. (1) Scheduling was documented as forcing run_in_background to true; it does not — an explicit run_in_background: false alongside schedule is refused, and always has been. The 0.10.0 entry below carries the same error: git log -S shows the refusal shipped in the very commit that introduced scheduling, so both documents were written from the design sketch and never matched what was released. Refusing is also the intended behavior — silently flipping a parameter the caller explicitly set is the failure mode #37 was filed about. (2) The completion-notification example showed a transcript: .pi/output/agent-abc123.jsonl line; that path was replaced by the OS-tmpdir layout in #146 and the example was never updated, so it contradicted the paragraph six lines above it. (3) The skills frontmatter row read as though naming skills preloaded them in addition to the parent's; naming any list turns skill inheritance off and the agent gets only the named ones.

Fixed

  • /agents → Create agent produced an unloadable agent when the description contained a colon. The wizard interpolated the description and the custom model straight into YAML frontmatter, both of which come from unvalidated free-text prompts. A description like Scout: find things made the generated file unparseable — and since #212 an unparseable agent file is skipped with a warning, so the wizard reported Created <path> for an agent that then silently did not exist — and, under strictAgentFiles, aborted the next startup too, that setting gating startup only. A description containing a # was quietly truncated at the #, YAML treating the rest as a comment. Both values are now quoted, the same way the eject serializer already quoted its description. A provider/model:thinking suffix is unaffected either way — YAML only splits on a colon followed by a space — and is pinned so the quoting cannot mangle it.
  • /agents → Scheduled jobs cancelled the wrong job when two names looked alike. The menu formats each job into a row, then resolves the user's pick by matching the returned string back against the row list. Rows truncate the job name to 18 characters for column alignment, and job names come from the Agent call's description, which is LLM-authored and routinely shares a prefix — so two jobs like review the auth module A and review the auth module B, on the same schedule and agent type, produced byte-identical rows and the match always resolved to the first. Picking the second cancelled the first, and because the confirmation dialog was built from the same wrong job, its details looked correct and the user had no way to notice before confirming. Rows are now numbered and carried alongside the job they describe rather than in a parallel array, so labels are unique by construction and no data-dependent collision path exists. The same string-matching pattern in /agents → Running agents — where the consequence was only opening the wrong agent's read-only viewer — now goes through the same helper. Pi's dialog API returns the selected string with no index or value form, so uniqueness at the call site is the only available fix. The whole menu had no test coverage at all, which is how a destructive action shipped with this in it.
  • The agent widget hid queued agents without counting them. Past its 12-line cap the widget hands out a line budget — running agents first, then the one-line N queued summary, then finished agents — and increments a hidden-count for running and finished rows that don't fit. The queued summary was dropped with no counter at all, so with five running agents, one queued and one finished the widget rendered +1 more (1 finished): two things hidden, one reported, and no indication anywhere that work was waiting to start. That is precisely the state the queue matters in, since it only appears once maxConcurrent is saturated. The queued line's row is now reserved before the running agents are laid out, so it can no longer be dropped; the trade is that one running agent may yield its two rows, and that agent is counted in the overflow total. Counting the queued line instead was rejected: it summarizes N agents, so counting it as one under-reports and counting it as N inflates a total denominated in agent rows.
  • /agents Enable/Disable now agree with the loader about enabled: false (found while closing test-suite gaps). Loading parses frontmatter with a real YAML parser; the menu edited it with regex, and the two disagreed. Enable's pattern was anchored to byte 0, so it only matched when enabled: false was the first frontmatter line — the shape this extension writes, not the shape README tells users to hand-author, where description: comes first — leaving a hand-written agent disabled forever while reporting Enabled <name>. Disable's idempotence check was byte-exact, so a trailing space defeated it and inserted a second enabled: false; duplicate keys are a YAML error, and since #212 an unparseable agent file is skipped, so the agent vanished from /agents instead of showing as disabled — and would abort the next startup under strictAgentFiles, which gates startup only, every mid-session reload skipping whatever it is set to. Deciding whether a file is disabled now calls the parser rather than mirroring it — which was always a read — fixing both the spellings YAML also accepts (False, FALSE, a trailing # comment, a quoted key) and pi's fence scan, which ends the block at any line merely starting ---: below a ---- line the regex read body text as frontmatter and refused to disable a running agent. Editing stays line-wise by design, since re-serializing a parsed document would strip the comments, key order, and quoting of a file users hand-author. That leaves removal recognizing only a lowercase bare false; the other spellings are detected but not yet rewritable, and now say so instead of claiming success, with /agents → Edit as the way through.
  • The Agent tool description no longer tells the orchestrator that a zero-tool agent has every built-in. tools: none and a tools: listing only ext: selectors both parse to an empty built-in list, and the runtime honors that — the shipped e2e fixture pins that such an agent really is handed no read, bash, edit, write, grep, find, or ls. The description builder collapsed "empty list" into the same (Tools: *) it uses for an absent tools:, which genuinely does mean all built-ins, so the only capability statement the orchestrator sees before spawning contradicted what the agent would actually get, and work was routed to agents that could only fail at it. An omitted tools: still renders *. An empty one renders none only when the agent can genuinely call nothing — isolated, or extensions: false — because zero built-ins is not zero tools: tools: none alongside extensions: still surfaces every extension tool, and the shipped fixture for that case expects three of them. That configuration now renders no built-ins, extension tools only, since describing it as none would understate the agent and route work away from the only one able to do it — the same failure as the original bug, pointed the other way. The suffix has only ever described built-in scope and still does: extension tools are resolved when the agent runs (extensions may register asynchronously) and cannot be enumerated while the description is built, which is why tools: "*, ext:mcp/search" renders * and always has. The same absent-vs-empty conflation is corrected in the eject serializer, which wrote tools: all for an empty list; that path is not reachable from the menu today, since eject is offered only for built-in defaults and none ships with an empty tool list, so it is hardening rather than a fix.

[0.15.1] - 2026-08-13

Added

  • strictAgentFiles — fail startup on a broken agent file instead of skipping it. Off by default. When on, an unreadable or unparseable agents/*.md aborts extension load and names the file, so a checked-in .pi/agents/ can't silently fall through to a different agent. Startup only: mid-session reloads (one per Agent call) keep warning, since a bad edit shouldn't kill the session on an unrelated spawn. Settable in subagents.json or via /agents → Settings → Strict agent files.

Fixed

  • A persist_session subagent's session now records the session that spawned it (#205 — thanks @0xbentang). SessionManager.create(...) was called without pi's NewSessionOptions, so a persisted subagent wrote a session header with no parentSession — indistinguishable from one started by hand. pi's session picker builds its tree from exactly that field, so every persisted subagent appeared as its own top-level root: a plan → review → implement orchestration left a flat run of sessions with nothing tying them to the session that produced them, and no way to tell them apart but their first message. The spawning session's file path is now passed through, so those sessions nest under their parent in the /resume picker. Grouping still needs both sessions in the same listing — pi lists per session directory, derived from the cwd unless one is named — so a session_dir override, or an isolation: worktree agent running in the copy's cwd, lands in a different directory and keeps showing as a root; the header link is written either way and travels with the file. A spawning session with no file behind it (any in-memory session — the default for a subagent) leaves the field unset, as before.
  • A subagent that never starts now fails the tool call instead of reporting success (#179 — thanks @xz-dev). A strict isolation: "worktree" spawn in a directory that is not a git repo, has no commits, or where git worktree add fails throws before any child session exists — and the Agent tool caught that throw and returned the diagnostic as ordinary result text. Pi marks a tool result failed only when execute throws; an isError flag on a returned result is discarded. The parent model therefore saw isError: false and read the message as a subagent that had run and reported this, then retried the same doomed call. Both spawn paths now let the throw out. Nothing else changes: a subagent whose run fails, or which is aborted or stopped, still settles on its record and keeps its existing result text and status note — the only calls that now fail are the ones where no agent was ever created. The foreground spinner and widget cleanup moved into a finally so they cannot be skipped by an early return.
  • An Agent call refused before it ran no longer reports a turn limit that never happened (#199 — thanks @Tonours, fixed in #195 by @xz-dev). Pi reports every pre-execution failure — an extension tool_call hook returning { block: true }, an abort, a schema violation — as { content: [reason], details: {} } with isError set. The result renderer's guard tested only for absent details, and {} is truthy, so those results walked past every status branch into the tail and printed ✗ Aborted (max turns exceeded). The operator debugged a phantom turn limit while the real reason went only to the model, which alone received the tool result text. A result carrying no status now renders that reason verbatim, and the turn-limit wording is reached only by a genuinely aborted run instead of serving as the catch-all for anything unrecognised.
  • isolation: "worktree" agents are told to stay in the worktree (#187 — thanks @cad0p). The agent was told its working directory twice — the copy, and the main checkout named by the inherited parent prompt or by the task prompt — and followed the latter, editing and committing in the shared tree. Worktree spawns now carry a <worktree_isolation> block naming the copy as theirs and the parent as off-limits. A prompt directive, not enforcement: local trials went from 5/5 escapes to 1/7, so isolation is still not a hard sandbox.
  • One malformed agent file no longer takes pi down with it (#212 — thanks @daromaj). A YAML error in any discovered agents/*.md aborted extension activation: pi exited 1 before the TUI, naming no file. Unreadable and unparseable files are now skipped, with a warning naming the file and the error. If the skipped file was overriding a same-named agent, a second line names the file that loads in its place.

[0.15.0] - 2026-08-10

Added

  • fallbackSubagent — choose what happens when subagent_type doesn't resolve, including failing closed (#183 — thanks @0xbentang). Dispatch previously repaired every unresolvable type by substituting general-purpose, so a typo silently ran a different agent, model and tool policy — and for a background or scheduled call, execution began before the caller saw any indication. The new project setting names the substitute instead of assuming one: omitted keeps today's general-purpose behavior, any enabled agent name routes unresolvable calls there, and none (or the boolean false, which would otherwise be dropped as the wrong type and silently leave the fallback in place) disables it entirely so dispatch fails closed with an error listing the available types. Resolution now happens at one point shared by every caller-supplied spawn — the Agent tool, the scheduler at fire time, cross-extension RPC, and nested delegation — so a refused type never reaches runAgent, where the old three-tier fallback in getConfig would have substituted an agent anyway. Case-ambiguous names (two enabled agents differing only in case, reachable because agents are keyed by filename across three directories) are treated as unresolvable rather than resolved to whichever was registered first. The fallback note is emitted at resolution rather than after the run, so the background and schedule paths carry it too, and a scheduled job now stores the type that was actually requested instead of the substitute. Nested delegation keeps rejecting unconditionally whatever this is set to, so a project-level fallback cannot hand a nested caller an agent outside its allowlist. An explicitly configured fallback that is itself unknown or disabled is reported as the misconfiguration it is.
  • Opt-in nested subagent delegation — a subagent can delegate to child agents of its own (#164 — thanks @codesoda). Default-off and per-agent: a custom agent receives scoped Agent, get_subagent_result, and steer_subagent tools only when its frontmatter sets allowed_subagents (a comma-separated type list, or alltrue and "*" are accepted the way extensions:/skills: take booleans), and never when it is isolated. The allowlist is a privilege boundary rather than a routing hint: a child runs with its own tools:, extensions:, and isolated:, so delegating grants the parent the union of what the listed agents can do — all reaches general-purpose. Nesting is depth-capped from the main session — main is 0, its subagents 1, their children 2 — through maxSubagentDepth in subagents.json (default 2; 0 or 1 disables nesting project-wide) or /agents → Settings → Nested depth. A per-agent cap is deliberately not exposed yet: at the default project cap it could only restate what omitting allowed_subagents already says, and a tighten-only field is safe to add once someone raises that cap. An agent already at the cap is given no nested tools at all, rather than tools whose every call would fail. Everything a parent can do is ownership-scoped — it can fetch, resume, and steer only its own children — and unknown, disabled, or out-of-list types are refused rather than falling back to general-purpose. Nested spawns resolve agents from a registry built for their own config root (never the process-global registry the main session uses) and revalidate model: against Model Scope exactly like a top-level spawn. Nested records stay internal to their parent, absent from top-level tools, lifecycle events, and the agent UI, and are stopped when that parent finishes, is stopped, or ends a resumed turn — a hidden child must not outlive the only agent able to reach it. They still write their own .output transcript under the root session's directory (same output_transcript gate), so a nested run remains inspectable, and their token usage is folded into every ancestor's totals so nested spend stays attributable at any depth, and a nested result that ends stopped, aborted, or steered is labelled partial, the same guarantee top-level results carry. Nested children do not occupy maxConcurrent slots: their parent already holds one, and queueing them behind it would deadlock a parent waiting on its own child; the depth cap bounds how deep nesting goes, not how wide, so a max_turns on the delegating agent is what ceilings its fan-out. A subagent session no longer activates this extension at all — that is what keeps a child from building a second agent manager, and it is why the nested tools are injected directly instead; a subagent consequently also gets no /agents command, no cross-extension RPC handlers, and no subagents:ready event.

Fixed

  • A disabled agent is no longer dispatchable, and no longer runs with a mixed identity (found while implementing #183). The Agent tool resolved types with resolveType(), which reports a canonical name without checking enabled — so an agent disabled via enabled: false or /agents → Disable resolved successfully, produced no fallback note, and then split downstream: getAgentConfig returned the disabled agent's config to build the system prompt while getConfig returned general-purpose for tools, extensions, skills and prompt mode. The run therefore carried the disabled agent's instructions with a different agent's tool policy. Dispatch now requires the type to resolve to exactly one enabled agent, so a disabled type takes the fallback path — or is refused outright under fallbackSubagent: none.
  • A foreground agent result now says its output is complete, not partial — it is the task whose completion is in question (#174 — thanks @killianMei). A foreground run that ended early was annotated output may be partial, which reads as truncated, go retrieve the rest. It isn't: a foreground result already carries the agent's entire output inline, and its parent has no agent id to retrieve anything with — the id travels in the tool result's renderer details, which is never serialized to the model, and only the background path prints Agent ID: in text. A parent acting on that reading therefore passed an id it had invented and got back Agent not found: "<id>". It may have been cleaned up., which looks exactly like the record was evicted. It wasn't: foreground completion mutates the record in place, and nothing removes it outside the 10-minute cleanup timer (which keys on completedAt, so a long run doesn't age it), clearCompleted() on a session start/switch, or shutdown — none of which the report hit. Foreground results now carry their own outcome note stating that everything the agent produced is above and that it is the task whose completion is in question — hedged for a steered run, which was told to wrap up and did, so it may well have finished at the limit, and stated outright for aborted and stopped, neither of which ever delivered a final answer; the background notification and get_subagent_result keep the existing note, because their caller holds a 500-char preview and a real id, where the same sentence would be false. Every clause is a statement about state rather than an instruction to act: nothing here can measure whether a wording change improves parent behavior, so removing a false cue (which cannot induce new behavior) and adding an instruction (which can) are not equally safe — and an instruction like "ask before restarting" would also presume a user is present to ask, which is false under pi -p, in scheduled jobs, and in background-driven runs. Regression tests pin the record-retention contract, the absence of an id in foreground result text, and that a subagent session's own activation lifecycle can't evict the parent's records.

[0.14.3] - 2026-07-23

Fixed

  • Subagents can use extension tools that register asynchronously, e.g. over MCP (#141 — thanks @philipmw). MCP-backed extensions usually can't enumerate their tools until their servers connect, so they call registerTool from session_start (pi-mcp) or before_agent_start (context-mode) rather than at load — eagerly connecting during extension discovery would orphan child processes on pi's non-agent code paths (--help, config, trust probing). The agent runner, however, snapshotted extension.tools into a static tools: allowlist immediately after loader.reload() and before bindExtensions() fired session_start. Because pi's allowedToolNames gates tool registration — not merely the initial active set — a name missing from that snapshot was dropped permanently, so those tools never reached a subagent even though the main agent had them. The runner now leaves the allowlist unset so pi's live gate admits tools whenever they register, and expresses the name-stable part of the scope (this extension's own orchestration tools, built-ins the agent didn't ask for, disallowed_tools:) as a denylist that pi re-applies on every registry refresh. extensions: false / isolated: true keep the static allowlist unchanged — nothing can register asynchronously there, and a hard registry gate is the right boundary.
  • ext: selectors keep narrowing correctly when tools register late (fixes #125). Admitting late-registered tools is only half the problem: an ext: selector is an allowlist, and a not-yet-registered tool cannot be listed in one — which is why tools: "*, ext:context-mode/ctx_execute", the exact configuration reported in #125, still saw nothing. ext: narrowing is now enforced on the active tool set (what the model actually sees) rather than the registry, re-derived from the loader's live extension maps — the same maps registerTool writes into — so a tool is judged against the selectors whenever it appears. Scope is re-applied on every turn_end, and the hooks live on the session rather than the spawning call, so steered and resumed turns stay scoped too. Turn 1 is guarded at call time instead: before_agent_start fires inside prompt() and can widen the tool set after that turn's tools are already snapshotted, leaving no window to narrow in, so an out-of-scope call there is refused rather than executed. Selecting a lazy extension (ext:mcp) now surfaces its tools; leaving one out still mutes it, no matter when it registers. Vetoing a call means wrapping the beforeToolCall hook pi installs on the session (pi exposes the veto to extensions as pi.on("tool_call"), but there is no equivalent for an SDK caller constructing a child session); the wrapper chains to pi's own hook so extension tool_call handlers still fire, and a pi@latest CI job guards that the hook stays reachable, so a future pi that moves it surfaces as a test failure rather than a silently missing veto.
  • FleetView selection markers now match the rest of the UI (#155 — thanks @xz-dev, closes #122). FleetView was the only view using (U+23FA) / (U+25EF) for its selected/unselected rows; everywhere else — the conversation viewer and agent widget — uses / (U+25CF/U+25CB), the base geometric circles with the broadest terminal-font coverage ( renders visibly oversized in many fonts). Selection is already carried by the accent-vs-dim color, so the exotic glyphs added nothing. FleetView now uses selected / unselected, converging on the codebase's existing pair.
  • get_subagent_result(wait: true) is now cancellable (#159 — thanks @exoulster, fixes #158). The tool received the tool-call AbortSignal but ignored it, so pressing Esc during a wait couldn't cancel it — the call stayed blocked until the child finished, and pi could log a spurious No running process to background.. Cancelling now stops only the wait: the background agent keeps running, its result stays unconsumed, and its normal completion notification still arrives. Queued waits are abortable the same way. Because the wait no longer pre-marks the result consumed (a cancelled wait never consumes it), a successful wait still suppresses its own redundant notification by cancelling the held nudge on completion — the queued-wait poll interval was tightened so that cancellation reliably lands inside the notification hold window.
  • User-scope agent memory honors PI_CODING_AGENT_DIR (#166 — thanks @biowaffeln). The user memory scope hardcoded ~/.pi/agent-memory/<name>, ignoring PI_CODING_AGENT_DIR even though every other config consumer (custom agents, enabled models, settings, skills) already routes through getAgentDir() — so anyone who relocated their config dir got a stray ~/.pi resurrected the first time a memory: user agent ran. User-scope memory now resolves to <agentDir>/agent-memory/<name> (default ~/.pi/agent/agent-memory/). Existing memories aren't orphaned: when the legacy ~/.pi/agent-memory/<name> directory exists and the new location doesn't, it keeps being used — symlinked legacy dirs are ignored, consistent with the file's other symlink defenses — and once the new location exists it wins. Lazy fallback rather than auto-migration, so no files are moved.

[0.14.2] - 2026-07-17

Added

  • output_transcript frontmatter + outputTranscript project setting — opt out of a subagent's .output transcript (#146 — thanks @Thoughts-One). Every subagent streams its full conversation to a per-subagent JSON-lines transcript under the OS temp dir (<tmpdir>/pi-subagents-<uid>/…/<agent-id>.output, owner-only 0700, cleared on reboot); until now that write was unconditional. Set output_transcript: false on a custom agent to write no transcript file or path for it, or outputTranscript: false in subagents.json to make transcripts opt-in for the whole project (a custom agent's frontmatter overrides the project default). Useful when run transcripts shouldn't sit on disk for backup or DLP tooling to ingest. Scope is deliberately narrow — it governs only the .output transcript, not the persisted pi session (persist_session), worktree commits (isolation: worktree), or memory files — so keeping a run fully off disk means setting those too. Default is unchanged: with neither flag set, transcripts are written exactly as before. The write decision is centralized on record.outputFile, so every downstream consumer (streaming, notifications, the transcript footer) keys off a single gate.

Fixed

  • Bordered conversation-viewer rows stay exact-width at double-width truncation boundaries (#153 — thanks @xz-dev). The conversation viewer pre-pads each row to the inner width and truncates it to fit between the borders, but truncateToWidth was called without its pad flag — so when a truncation boundary fell mid-way through a double-width character (CJK, wide emoji), the result came back one column short and the right border shifted left by a column on that row. The row builder now truncates with padding on, restoring the trailing column, so every bordered row renders at exactly the box width regardless of where a wide glyph lands. A regression test sweeps a double-width character across every truncation boundary and asserts each rendered line is exactly the requested width.
  • Isolated subagents keep extension-registered custom providers on pi 0.80.8+ (fixes #151 via #152 — thanks @0xbentang). pi 0.80.8 replaced createAgentSession's modelRegistry option with modelRuntime, and the agent runner still passed the now-ignored modelRegistry — so pi built a fresh model runtime from disk for the child session. Because isolation also disables extension loading, that fresh runtime had neither the extension-registered custom provider nor its auth: an isolated: true agent pinned to such a provider failed preflight with No API key found for <provider>, while the same agent with isolated: false — or on any pi <0.80.8 — worked. The runner now forwards the parent session's ModelRuntime (read off the ModelRegistry facade on ctx.modelRegistry) as modelRuntime when the running pi exposes one, and still passes modelRegistry for the pre-0.80.8 range — so the fix spans the whole >=0.80.0 peer range without changing it: older pi omits the new field and takes the legacy path, 0.80.8+ inherits the parent runtime. Extensions and tools stay isolated exactly as before — only the model providers and their auth are inherited. A pi@latest CI job guards the reachability of that runtime accessor (a private field the fix reaches through), so a future pi that renames or hides it surfaces as a test failure instead of a silent return of this bug.

[0.14.1] - 2026-07-14

Added

  • max thinking level is now advertised in the frontmatter/tool/wizard choices (#147 — thanks @justin-ramirez-gametime). pi 0.80 added max to its ThinkingLevel, and the extension already forwarded the value unchanged, but the Agent tool description, generated-agent template, /agents creation wizard, and README all stopped at xhigh — hiding a valid capability. Those four surfaces now come from one shared list so they can't drift behind pi again. Actual availability still depends on the host pi version and the selected model; pi clamps unsupported levels down.

Fixed

  • Runs whose final assistant turn failed are now reported as error, not completed with an empty result (fixes #144 — thanks @possibilities for the diagnosis). pi resolves an exhausted-retries provider failure normally — the final assistant message carries stopReason: "error" plus an errorMessage, no rejection — so the manager mapped such runs to completed and every consumer (foreground tool result, get_subagent_result, background notifications, resume, scheduler, RPC events) saw a clean success reading No output. — or worse, an earlier turn's text presented as the fresh answer, since the history fallback walks past empty messages. The runner now inspects the final assistant message and reports a failure in two cases: the turn stopped with stopReason "error", or it hit the output-token ceiling (stopReason "length") having produced no text at all — a silent max-token death that reproduced the same empty-No output. symptom. Status still derives from how the final turn stopped, never from whether earlier turns produced text: empty-but-clean finals (tool-call-only or thinking-only endings) stay completed, a length stop that did produce text is a legitimate truncated answer and stays completed, partial-text provider errors are still failures, and the walk-back fallback keeps preserving partial output for aborted/steered runs. (An aborted final turn needs no special-casing here — the manager's hard-abort flag and stopped guard already surface it as aborted/stopped.) The response-text collector also no longer resets on user/tool-result message_start events (it tracked the last message, not the last assistant message). A failed run still surfaces any output it did produce: the tool result shows Agent failed: <error> followed by that text under a Partial output before the failure: label — and the history fallback is now bounded to the current invocation, so a failed resume no longer returns the previous turn's answer as this run's result (it returns empty). Behavior change: runs that previously ended completed with an empty result now end error carrying the provider message, and subagents:failed fires where subagents:completed did; scheduled jobs record lastStatus "error" for them.
  • A subagent can allowlist a package-installed extension by its package name, not just its source directory (fixes #143 — thanks @possibilities). Our extension-scoping (extensions: / exclude_extensions: / tools: ext:…) named an extension from its file path, and for an index.ts entry it used the parent directory — so a package installed via pi.extensions: ["./src/index.ts"] (like this one) only ever matched as src, an unstable, collision-prone name that no user would guess; extensions: [pi-subagents] silently matched nothing. An extension now also answers to its package's unscoped short name (@tintinweb/pi-subagentspi-subagents), read from the nearest package.json whose pi.extensions manifest actually declares that entry. This is an added alias — the path-derived name keeps working, so nothing that already matched src breaks — and it fires only for genuinely manifest-declared package entries, so a loose extension is never misattributed to a co-located project's name.
  • A pi-subagents activation that a child session filtered out no longer advertises or answers cross-extension RPC (fixes #142 — thanks @possibilities). pi runs every extension factory before applying an agent's extensions: filter and only delivers lifecycle events (session_start, …) to the survivors, but the pi.events bus is shared with the filtered-out activations too. Because we registered the RPC handlers and emitted subagents:ready at factory time, a child agent whose extensions: omitted pi-subagents still saw a subagents:ready broadcast and a successful subagents:rpc:ping, yet every subagents:rpc:spawn failed with No active session — its session_start never fired, so the spawn handler had no context. The RPC handler registration and the subagents:ready broadcast now happen on the first bound session_start instead of at factory time, so a session that excludes pi-subagents stays completely silent on the RPC channels — behaving like a session where it was never installed, rather than advertising a spawn service it can't provide. Emitting readiness after all factories have loaded also closes a latent race where a consumer whose factory ran after ours could miss the event. No change for sessions that do load pi-subagents: subagents:ready still fires and RPC still works, just at session_start.

[0.14.0] - 2026-07-13

Added

  • Project custom agents are also discovered from .agents/agents/<name>.md (#133 — thanks @wenerme; closes #132). Projects that keep their agent assets in the shared cross-tool .agents workspace (the same convention this extension already reads for .agents/skills/) can now define subagents there instead of duplicating files into .pi/agents/. Discovery precedence is global < .agents/agents < .pi/agents: on a name clash between the two project locations, .pi/agents/ wins.pi remains the project authority, and the /agents create/eject/disable flows keep writing there; .agents/agents/ is a read location only.

Changed

  • BREAKING: Dev/test toolchain now tracks pi 0.80.x; pi peer floor raised to >=0.80.0 (as diagnosed by @philipmw in #129). The committed lockfile had pinned the @earendil-works/pi-* peers to 0.75.5 (the latest published when it was generated), so tests and typecheck exercised an older API surface than the pi that actually runs the extension — including 22 tests that silently never ran on 0.75.5 and now do (suite skip count 26 → 4 between the two lockfiles). The lockfile now resolves the peers to 0.80.6, and the test suite imports pi-ai's relocated faux/model helpers (registerFauxProvider, getModel/compat-only since pi-ai 0.80) through a single re-export module, test/helpers/pi-ai.ts, so the next relocation touches one file. Because the test surface no longer resolves on ≤0.75.x, peerDependencies move from >=0.74.0 to >=0.80.0 to match what is actually tested. Migration: installs under pi <0.80 may emit unmet-peer warnings (or fail under strict peer resolution such as npm --strict-peer-deps / pnpm defaults) — upgrade pi to ≥0.80; runtime behavior itself is unchanged on any pi version, since pi substitutes its own bundled modules for extension imports at load time. Live-mode e2e (PI_E2E_LIVE=1) now fails fast with a clear error when the pinned PI_PROVIDER/PI_MODEL isn't in pi-ai's builtin catalog, instead of silently letting the session resolve a different model.

Fixed

  • Output-file streaming survives session compaction (fixes #145 — thanks @possibilities for the diagnosis). Compaction replaces session.messages with a shorter, summarized array, which stranded the streamer's write index past the new end — the flush loop never matched again and the agent's output file silently froze for the rest of the run, exactly on the long background runs that auto-compact. The streamer now flushes any not-yet-written tail when compaction starts and re-anchors its index to the rebuilt array after a successful compaction — deferred one microtask, because on the overflow-retry path pi trims the trailing error assistant message after emitting compaction_end, and a synchronous anchor would skip the first post-compaction message. Aborted and failed compactions leave the session untouched and change nothing. Verified against a real pi AgentSession driving a real session.compact() in the new e2e regression.
  • FleetView no longer steals arrow/Enter/Esc keys from other interactive components (fixes #123 — thanks @TommyC81 for the report). pi delivers terminal input to extension listeners before the focused component, and selector/input dialogs (ctx.ui.select & co.) swap the prompt editor out while getEditorText() still reads the detached — empty — editor. So while subagents were running, FleetView's empty-prompt gate passed and it consumed the navigation keys that belonged to whatever dialog was actually focused: other extensions' pickers (e.g. rpiv-ask-user-question), pi's own menus, and even /agents → Settings itself. FleetView now checks that pi's prompt editor is the focused component before touching any key (pi's editor is an Editor subclass; every dialog is not), and an in-progress list navigation is dropped the moment something else takes focus. Unknowable focus errs toward the editor, so list activation keeps working; non-Editor custom editor components simply flow keys through untouched.
  • Widget and viewer lines keep their dim styling after nested color annotations (#136 — thanks @xz-dev). pi themes close a foreground color with a bare \x1b[39m, so the threshold-colored context-fill percent inside formatSessionTokens (warning at ≥70%, error at ≥85%) also terminated the surrounding dim style — the closing ) and any following text on the widget's running-agent line and the conversation-viewer header bled to the terminal's default color. A small fgPreservingNestedStyles helper re-opens the outer color after each nested reset (derived from the theme itself, so it stays theme-agnostic and is a no-op under NO_COLOR).
  • get_subagent_result with wait: true now waits for queued agents (#127 — thanks @benrhodeland). A background agent past the concurrency ceiling sits in the queue with no run promise yet, so the wait path (gated on status === "running" && record.promise) skipped it and returned No output. immediately — the orchestrator read a queued agent as "finished with nothing". The wait now also covers queued: it polls until the queue starts (or stops) the agent, then awaits the run like any running agent. Most visible with parallel background spawns in one message, where spawns beyond maxConcurrent (default 4) queue.
  • Subagent activations no longer clobber the Symbol.for("pi-subagents:manager") registry (#128 — thanks @benrhodeland). Child sessions re-activate the extension in the same process (session.bindExtensions in the agent runner), and every activation overwrote the global registry slot — pointing cross-package consumers (RPC extensions, headless hosts) at a short-lived child manager whose shutdown could then delete the root session's entry entirely. The first activation now claims the slot, child activations leave it alone, and only the owning activation releases it (identity-checked) on shutdown.

[0.13.0] - 2026-06-30

Added

  • Steer a running agent from the conversation viewer. The live conversation overlay (FleetView's Enter, or /agents → Running agents) now lets you redirect an agent without leaving the view: press Enter to open an inline composer, type a message, Enter to send — Esc or an empty submit just returns. The message is delivered through the same path as the steer_subagent tool (AgentManager.steer()session.steer, or queued onto pendingSteers if the session isn't ready yet), so it appears as a user message and redirects the agent after its current tool execution; feedback is the message showing up in the live transcript you're already watching. The affordance is offered only while the agent is still running/queued (mirrors the x/stop affordance), and the viewer stays modal — every existing shortcut (x/x stop, arrows/j/k scroll, q close) is untouched, and Enter was previously inert here so nothing is overridden. The idle footer was reorganized to actions-left / navigation-right so the full scroll-key hint (↑↓ scroll · PgUp/PgDn or Shift+↑↓ · Esc close) stays fully visible down to 80-column terminals; the N lines · % readout returns on the left whenever there's spare width.
  • Forgiving model resolution for agent model: pins. resolveModel now tolerates cosmetic id variations and falls back across providers, so a qualified or date-pinned config resolves widely instead of silently dropping to the parent model: . and - are treated as equivalent in version numbers (claude-haiku-4.5claude-haiku-4-5); a trailing -YYYYMMDD date stamp is optional (anthropic/claude-haiku-4-5-20251001 matches an undated registry id); and a provider/modelId that isn't available under the named provider retries the bare id against every provider (the named provider is still preferred when present). An exact match still wins over a tolerant one, so dated snapshots aren't conflated — the precedence is exact → fuzzy-under-named-provider → same model under any provider → unavailable.
  • /agents → Agent types shows each agent's full description and what its model resolves to. The list now renders with SettingsList (like the Settings menu) instead of a flat selector: the highlighted agent's full description shows on its own line below the list, so long descriptions no longer wrap and push the agent names out of alignment. The model column shows the configured model, flags it (unavailable, fallback: inherit) when it can't be resolved against the registry (it would silently inherit the parent model at runtime), and surfaces the resolved target (→ provider/id) when resolution lands on a different provider or version than configured.
  • widgetMode setting — control what the above-editor widget shows: all / background / off (#117 — thanks @Alan-TheGentleman; fixes #118). Foreground agents already render inline as the Agent tool result, so also listing them in the persistent widget double-rendered the same run (most visible in tmux/zellij). widgetMode (via /agents → Settings → Widget, or subagents.json) selects the widget's contents: all shows every agent (the previous behavior), background shows background/queued/scheduled/RPC runs but hides foreground, and off hides the widget entirely (agents still appear inline and in FleetView). Applied live — toggling refreshes immediately. Filtering keys off a new tri-state AgentRecord.isBackground captured at spawn (true = background, false = foreground, undefined = undeclared, e.g. a cross-extension RPC spawn), independent of the UI-only invocation snapshot — so scheduler- and RPC-spawned background agents stay visible instead of vanishing; only runs known to be foreground are dropped. The running-status line was also refactored from one multiline Text into two component rows, so rapid partial updates replace cleanly instead of leaving stale rows behind in terminal multiplexers — identical on-screen output, no more ghost lines.

Changed

  • The built-in Explore agent's model pin is now anthropic/claude-haiku-4-5 (was …-4-5-20251001). The stale date stamp is dropped to match the anthropic/claude-sonnet-4-6 / anthropic/claude-opus-4-6 convention used by the create-agent wizard; combined with the forgiving resolution above, Explore now picks up its fast/cheap model across registry variations (dated, undated, dotted, or under a non-anthropic provider) instead of silently falling back to the parent model.
  • The above-editor widget now hides foreground agents by default (widgetMode defaults to background). Foreground runs still render inline as the Agent tool result (and in FleetView); set /agents → Settings → Widget to all to restore the previous show-everything view, or off to hide the widget. Existing subagents.json files load unchanged (absent → background).

[0.12.0] - 2026-06-24

Added

  • FleetView — a Claude Code-style subagent navigator below the editor. A persistent, navigable list of main + every active subagent renders beneath the editor whenever agents are running — auto-shown, no keypress needed — mirroring Claude Code's bottom fleet bar: / selection markers, agent type + description, right-aligned elapsed · ↓ tokens, and a ↓ N more overflow once past five rows. Press (or ) at an empty prompt to move focus into the list, / to select, Enter to open the selected agent's live, auto-updating conversation overlay, and Esc (or above main) to return to the prompt. Implemented as a belowEditor widget with all key handling routed through onTerminalInput (which fires before the editor); it only captures arrow keys at an empty prompt — and acts on key-press only (kitty-protocol release events are ignored, otherwise each tap moved twice) — so typing, history, and cursor movement are untouched. Rows are ordered earliest-launched first; only openable agents (those with a session) are shown, so pending/queued agents appear once they start and Enter never dead-ends; finished agents linger ~4s before dropping out (their elapsed freezes at completion); and a viewer stays open through its agent's completion so the final output remains readable. Selection follows the viewed agent by id, so closing a viewer returns you to the same agent even if the list reordered while it was open. Every rendered line is width-clamped — the narrow-terminal crash/flicker class previously fixed in v0.2.7 and #7. Toggle via /agents → Settings → Fleet view (default on; pure-UI, so no LLM-context cost).

[0.11.0] - 2026-06-23

Added

  • persist_session / session_dir agent frontmatter — persist a subagent as a real pi session (#111 — thanks @codesoda). persist_session: true runs the subagent through SessionManager.create(...) instead of SessionManager.inMemory(...), so its full transcript is written to pi's normal session location (~/.pi/agent/sessions) — inspectable and resumable after the fact, like a top-level session — rather than living in memory only. Useful for long-running, multi-round orchestrations (plan → review → implement → verify) where each subagent's conversation is worth keeping. session_dir optionally overrides where the persisted session is written (absolute, ~, or agent-cwd-relative path); omitted, persistence follows pi's own precedence — PI_CODING_AGENT_SESSION_DIR, then the settings manager's getSessionDir(), then pi's default location. Both default off/unset, so existing agents are unchanged: the in-memory path is byte-identical to before, and the sidechain .output transcript is still written either way.

[0.10.4] - 2026-06-23

Fixed

  • Background agent records lost before result is read (#108 — thanks @philipmw). On session switch or /new//resume, clearCompleted() removed completed agent records regardless of whether the LLM had retrieved the result, causing get_subagent_result to return "Agent not found" for agents that had finished but hadn't been checked yet. clearCompleted() now accepts a skipUnconsumed flag; session event handlers pass true, so records with resultConsumed=false are preserved across session transitions. The 10-minute cleanup timer handles eventual eviction. Note: a full session shutdown (session_shutdown) calls dispose() which clears all records unconditionally — that path is not affected by this fix.

Added

  • Foreground agent lifecycle completion and conversation logging (#105 — thanks @benrhodeland). Two gaps closed: (1) onComplete now fires for foreground agents, emitting subagents:completed / subagents:failed lifecycle events and writing a subagents:record entry to the parent JSONL — previously only background agents emitted these, leaving cross-extension observers with an orphaned subagents:started event and no matching completion. resultConsumed is pre-set so the callback skips notifications (the result is returned inline); no change to the tool's return value. (2) Foreground agent conversations are now streamed to .output files (same .pi/output/agent-<id>.jsonl path as background agents) — inline subagent transcripts were previously permanently lost after spawnAndWait returned.

[0.10.3] - 2026-06-12

Added

  • SpawnOptions.cwd — spawn a subagent in a different working directory (#96 — thanks @madeleineostoja). For RPC/programmatic callers (not exposed on the Agent tool — the LLM-visible surface is unchanged). The agent's tools operate in the target directory and the prompt's environment block describes it, but .pi config keeps loading from the parent session's project (new RunOptions.configCwd split): the target's .pi extensions never execute, and its agents/skills/settings/memory are not picked up — spawning into an untrusted directory sends a worker there with the parent's toolbox, rather than "opening pi there." Composes with isolation: "worktree": the worktree is created from the target directory's repo, the agent works at the equivalent subdirectory inside the copy (a monorepo-package cwd keeps its scoping instead of silently widening to the repo root — new WorktreeInfo.workPath), and the resulting pi-agent-* branch lands in that repo, with the completion message naming it so the orchestrator merges in the right place. Validation is strict, typed, and early — non-strings, relative paths, nonexistent paths, and files all throw curated errors at spawn() (before queueing) and are re-checked at queue drain, surfacing as RPC error envelopes (null is treated as unset). On dispose, worktree registrations are pruned in every repo that received one; only a hard crash can leave a stale entry (then: git worktree prune in the target repo).

[0.10.2] - 2026-06-10

Added

  • exclude_extensions: agent frontmatter — extension denylist for subagents (#94 — thanks @ramhaidar). Applied after the extensions: include set; exclude wins, including over tools: ext: selectors (an excluded extension never loads, so its ext: reference becomes the usual orphan warning). The key use case: extensions: true + exclude_extensions: pi-notify — all extensions except a noisy one, without hand-maintaining an allowlist. Plain canonical names only (case-insensitive); paths, *, and unmatched names fire extension-error:… warnings (warn-not-abort, as with extensions: mismatches); extensions: false + an exclude warns that the exclude has no effect. Not a sandbox: excluded extensions' factory code still executes once during loading — exclusion suppresses handler binding and tool registration, not load-time side effects. The negation syntax extensions: ["*", "!name"] was deliberately rejected: an unquoted !name is a YAML tag and silently mis-parses.
  • toolDescriptionMode setting — opt-in compact Agent tool description (#91 — thanks @tiberiuichim). The full Claude Code-style description costs ~1,400 tokens with the default agents and grows with each custom agent (the type list embeds full agent descriptions) — significant for small/local models. toolDescriptionMode: "compact" (via /agents → Settings → Tool description or subagents.json) swaps in a ~75% smaller description: one-line type list (first sentence of each agent description), terse usage notes, per-option details left to the parameter descriptions. Default "full" is byte-identical to before — the rich description's guardrails are deliberately load-bearing and stay the default. A third mode, "custom", registers a user-authored description from <cwd>/.pi/agent-tool-description.md (project) or <agentDir>/agent-tool-description.md (global; project wins), with {{placeholder}} substitution keeping the dynamic parts live — {{typeList}}, {{compactTypeList}}, {{agentDir}}, {{scheduleGuideline}} — so a hand-written description can't drift out of sync with the registered agents (the advertised-vs-spawnable staleness #92 just fixed). Unknown placeholders are left verbatim with a stderr warning; a missing/empty file falls back to "full". Only the prose is customizable — the parameter schema stays code-owned. A ready-made starting point ships at examples/agent-tool-description.md, reproducing the full description exactly (CI-enforced byte-identical, so the example can't go stale). Like schedulingEnabled, the mode is read at tool registration — changing it applies on the next pi session. The issue's original ask (move the description to a skill) isn't possible in pi: tools must register their description in the tool schema for the model to call them; skills are lazily-loaded instructions, not tool registrations.

Fixed

  • Conversation viewer honors custom tui.select.* keybindings (#99 — thanks @owenniles). The viewer hardcoded its scroll keys and discarded the KeybindingsManager pi injects into ctx.ui.custom(), so user bindings (e.g. emacs-style ctrl+p/ctrl+n on tui.select.up/down) worked in pi core selectors but not here. Scrolling now resolves through tui.select.up/down/pageUp/pageDown; the viewer-specific k/j and shift+arrow aliases still work alongside, and behavior without custom bindings is unchanged (the tui.select.* defaults are the previously hardcoded keys).

[0.10.1] - 2026-06-10

Added

  • disableDefaultAgents setting (#92 — thanks @TommyC81). When on, the three built-in default agents (general-purpose, Explore, Plan) are skipped at registration — only user-defined .pi/agents/*.md agents are advertised and spawnable. User agents are unaffected, including ones overriding a default by name; with no user agents defined, spawning falls back to the hardcoded generic config. Off by default; toggle via /agents → Settings → Disable defaults or disableDefaultAgents in subagents.json. Like schedulingEnabled, the Agent tool's type list reflects the change on the next pi session (tool schema is registered at startup).

Fixed

  • Agents with enabled: false are no longer advertised in the Agent tool description (#92). buildTypeListText listed every registered agent, including disabled ones that isValidType then refused to spawn — the LLM was offered types it could never use. The type list now filters through getAvailableTypes(), matching the subagent_type parameter description.
  • Agent tool type list no longer built from pre-settings state. The description text was captured into a variable before persisted settings were applied; it's now built at tool-registration time, after subagents:settings_loaded.
  • Committed work from isolation: "worktree" subagents is now preserved (#68 — thanks @rylwin). If an isolated subagent creates its own commit, cleanup previously saw a clean git status, treated it as "no changes", and removed the detached worktree — silently discarding the commits. The worktree now records its base SHA at creation, and cleanup creates the expected pi-agent-* branch whenever HEAD moved past it, even with a clean tree.
  • Automatic commits in isolated worktrees skip local Git hooks (#68). The preservation commit at worktree cleanup now uses --no-verify, so a failing local pre-commit hook can't abort it (which previously surfaced as hasChanges: false — the agent's work lost).

[0.10.0] - 2026-06-01

⚠️ Breaking: extensions: and tools: in agent frontmatter semantics changed. The extensions: [...] array now selects which extensions load, not which tool names surface. Agents that previously used the array form will behave differently — see migration below. The tools: field also grew new ext: and * selector forms; existing tools: values without these selectors are unchanged.

  • extensions: [...] is now an extension allowlist applied at load time, not a tool-name substring filter. Each entry is an extension name, a path (absolute, ~/-prefixed, or relative-to-cwd), or "*". Migration: extensions: ["mcp"] previously loaded every extension and then surfaced only tools whose names contained mcp. To keep all extensions, use extensions: true or extensions: "*". To narrow, name the extensions or point at their files. "*" composes: extensions: "*, /abs/path/extra-ext.ts" is all defaults plus one path-loaded.
  • tools: now accepts ext: selectors and *. Gotcha: a tools: value containing only ext: entries yields zero built-in tools — add * (e.g. tools: "*, ext:foo") to keep the built-ins. And any ext: entry flips extension tools to an explicit allowlist (non-listed extensions stay loaded but expose no tools). A tools: with no ext: entries is unchanged.
  • extensions: is the sole loading authority. ext:foo only narrows tool exposure within the already-loaded set; it cannot pull an extension in. extensions: false + tools: "ext:foo" loads nothing and warns that ext:foo is orphaned. To expose one extension's tool from an otherwise-narrow agent, name the extension explicitly: extensions: [foo] + tools: "ext:foo/bar".

⚠️ Heads-up — widget glyphs changed (visual only): turn count now renders as ↻N (was ⟳N) and compaction count as ⇊N (was ↻N). Fix for #84 overflowed its cell in common monospace fonts. No API, behavior, or output-format changes — only the glyphs. If you grep agent stats lines or pipe widget output through scripts, update your patterns: (turns), (compactions).

Added

  • tools: accepts ext: extension-tool selectors and a * built-in wildcard. Entries in the tools: CSV are now partitioned: plain names are the built-in allowlist (unchanged); * expands to all built-ins (symmetric with extensions: "*"); ext:foo / ext:foo/bar select extension tools. Any ext: entry flips extension tools to an explicit allowlist — only tools named by an ext: selector reach the LLM, and extensions not named stay loaded (their session_start etc. handlers still fire) but expose no tools. ext:foo exposes all of foo's tools; ext:foo/bar narrows foo to just bar (multiple ext:foo/x entries union; a bare ext:foo alongside ext:foo/bar lets narrowing win). ext: is narrowing-only — it does not load extensions. extensions: remains the sole loading authority; an ext:foo against an extension that extensions: excluded (including extensions: false) is orphaned and warns via onToolActivity (extension-error:ext:foo …). With no ext: entry present, extension-tool behaviour is unchanged. ext: is name-only (matched by canonical name, so it composes with path-loaded extensions); paths still go in extensions:. isolated: true ignores ext: selectors.
  • Stop a running agent from the conversation viewer. In /agents → Running agents, select an agent and press x (then x again to confirm) to abort it. The two-press guard prevents an accidental kill; the footer shows x stopx again to STOP. This works for background agents — which a global Esc can't unambiguously target — while Esc still stops a blocking foreground Agent call. Wires the existing AgentManager.abort(id) to the viewer (onStop callback); the affordance only appears while the agent is running/queued. Addresses the common "how do I stop a background subagent?" question (#88).

Changed

  • BREAKING: extensions: [...] in agent frontmatter is now a loader-level extension allowlist, not a tool-name filter. Previously a string[] value filtered exposed tool names by substring (t.startsWith(e) || t.includes(e)) while every discovered extension still loaded and ran its handlers. Now each entry selects an extension: a bare name keeps the matching default-discovered extension, a path (absolute, ~/-prefixed, or relative-to-cwd) loads that extension fresh via additionalExtensionPaths, and "*" keeps all default-discovered extensions. Entries compose — ["*", "/abs/foo.ts"] is all defaults plus foo, ["mcp", "/abs/foo.ts"] is just those two. Excluded extensions no longer bind handlers or register tools (their factory still runs once during reload()). Directory extensions (foo/index.ts) match by the parent directory name. Extension names match case-insensitively (extensions: [Mcp] resolves the same as [mcp]); tool names within ext:foo/bar selectors remain case-sensitive (they're matched against pi-mono's registered identifiers). Unmatched names and failed paths warn via onToolActivity but do not abort the subagent (see the heads-up above for migration).
  • Non-normal subagent outcomes are now stated explicitly in the text delivered to the parent, so the orchestrator can't mistake a stopped/incomplete agent for a completed one. The foreground Agent result, get_subagent_result, and the <task-notification> summary all append a clear note for stopped (user abort) → (STOPPED BY THE USER before completion — output is partial; the task was NOT finished), aborted (turn limit) → (aborted — hit the turn limit before completion; output may be incomplete), and steered(wrapped up at the turn limit — output may be partial). stopped (human intervention) is kept distinct from aborted (turn-budget cutoff); a clean completed adds no note. Extracted as getStatusNote in src/status-note.ts.
  • BUILTIN_TOOL_NAMES is derived from pi's tool factories (createCodingTools + createReadOnlyTools) rather than a hardcoded list, so the built-in set tracks pi-mono automatically. Internal; no behavior change (the resolved set is the same seven names).

Fixed

  • Turn-count glyph in the agent widget no longer overflows its monospace cell (#84 — thanks @linozen). formatTurns used (U+27F3 CLOCKWISE GAPPED CIRCLE ARROW) from the Miscellaneous Mathematical Symbols-A block, where common monospace fonts (Iosevka Nerd Font Mono, Menlo, SF Mono, JetBrains Mono) draw the glyph visually wider than one cell despite its Neutral East Asian Width — making the next character (the digit) overlap the glyph. Replaced with (U+21BB CLOCKWISE OPEN CIRCLE ARROW) from the standard Arrows block, which renders cleanly at one cell in those fonts. To avoid colliding with the existing compaction indicator (which previously also used ), the compaction glyph moves to (U+21CA DOWNWARDS PAIRED ARROWS) — same Arrows block, also single-cell, visually distinct. Widget vocabulary now reads: ↻5≤30 for turns, ⇊2 for compactions. Pi UI consumers / scripts grepping for the glyph in stats lines must update.
  • tools: none now actually yields zero built-in tools. getToolNamesForType treated an explicit empty builtinToolNames ([], produced by tools: none) as "unspecified" and fell back to all 7 built-ins. It now distinguishes an omitted field (undefined → all built-ins, for default agents) from an explicit empty list ([] → zero), consistent with getConfig. Same fix makes tools: values containing only ext: selectors yield zero built-ins as documented.
  • tools: typos no longer silently break tool-calling (#75). Two parts: (a) all was previously parsed as a literal tool name, producing a one-element allowlist of the non-existent tool "all" — the model then returned an empty response or emitted raw XML tool calls, all with status: completed and no error. parseToolsField now treats all (case-insensitive) as an alias for the * wildcard, both standalone and inside a CSV. (b) Plain entries in tools: are expected to be built-in names (extension tools route through ext:), so an unknown name there is unambiguously a typo. runAgent now emits a tools-error:tool "X" requested by agent "Y" is not a known built-in event via onToolActivity for each unrecognized plain entry — same surfacing channel as the existing extension-error: warnings.
  • Subagents with extensions: true now actually expose extension-registered tools (MCP, etc.) (#47). runAgent previously passed only the built-in tool names as the tools: allowlist to createAgentSession, so pi-mono's allowedToolNames gate rejected every extension-registered tool at registration — extensions: true agents silently got only the 7 built-ins. runAgent now enumerates extension tool names from the resource loader after reload() and builds the full master allowlist (built-ins + permitted extension tools), so pi-mono's gate admits them from the first instant of the session. disallowedTools and the internal Agent/get_subagent_result/steer_subagent exclusions are applied uniformly to built-in and extension tools at construction — no post-construction setActiveToolsByName narrowing.
  • Append-mode subagents no longer defeat the LLM's KV cache (#73 — reported by @jeffutter). The assembled child prompt placed the per-spawn-varying <active_agent> tag and # Environment block before the ~8k-token inherited parent prompt, and wrapped the parent prompt in <inherited_system_prompt> tags. Because KV caches key on a byte-identical prefix, every subagent spawn reprocessed all ~8k shared tokens from scratch (~40s on slower hardware). The parent prompt is now emitted verbatim at the start of the prompt (wrapper dropped), so it forms an identical, cacheable prefix with the parent session and across every spawn; the static <sub_agent_context> bridge follows, then the varying <active_agent> tag and env block. replace mode is unchanged (it inherits no parent prefix). The <active_agent> tag stays present and is parsed position-independently, so downstream permission resolution is unaffected. Mirrors the fix in gotgenes/pi-packages#180.

[0.9.1] - 2026-05-30

Added

  • Agent, get_subagent_result, and steer_subagent now surface in pi's default system prompt (#87 — thanks @that-yolanda). Adds promptSnippet to all three (a line in the prompt's Available tools: section) and promptGuidelines to Agent (bullets in Guidelines:). The tools were always callable via the tool-call API; this only adds system-prompt reinforcement for prompt-following models. No schema or tool-call changes.

[0.9.0] - 2026-05-30

Heads-up — orchestrator behavior may shift. This release substantially rewrites the Agent tool description and the three default-agent descriptions (general-purpose, Explore, Plan) to mirror Claude Code's upstream wording. No API, schema, or tool-call shape changes — purely a prompt-engineering shift, but a load-bearing one:

  • Agent selection may drift. The new agent descriptions carry richer positive ("Use it to …") and negative ("Do NOT use it for …") guidance plus search-breadth hints for Explore ("quick" / "medium" / "very thorough"). For ambiguous tasks where the orchestrator previously picked one default agent, it may now pick another — typically more correctly, but the choice may differ from prior releases.
  • Subagent briefings will skew longer and more contextual. The restored upstream guardrails and the new ## Writing the prompt section actively coach "smart colleague who just walked into the room"-style prompts. Expect more context, more constraint, more upfront framing in the prompt: field the orchestrator passes to subagents.
  • Parallel/background patterns more strongly enforced. The merged bullet on parallel execution now explicitly says run_in_background: true is required on each tool call for actual concurrency, and that the orchestrator MUST send a single message with multiple tool uses when the user says "in parallel." Workflows relying on sequential-foreground default behavior are unaffected.
  • If you have tests or workflows that depend on the prior agent-selection or briefing behavior, pin to a v0.7.x release.

Added

  • scopeModels setting — opt-in subagent model-scope enforcement (off by default). New setting toggleable via /agents → Settings → Scope models. When enabled, the effective model of each subagent spawn is validated against enabledModels from pi's settings (which pi manages via its own /scoped-models UI; pi-subagents only reads it). Both pi settings files are honored: global <agentDir>/settings.json plus project-local <cwd>/.pi/settings.json, with project overriding global — mirrors pi's SettingsManager deep-merge and our own subagents.json precedence. Out-of-scope handling depends on source: caller-supplied via Agent({ model: "..." }) → hard error to the orchestrator with the allowed list; frontmatter-pinned or parent-inherited → warning toast + the agent runs anyway (preserves "frontmatter is authoritative" guarantee from v0.5.1; scopeModels is a guardrail against runtime LLM choices, not user-level config). Limitation: only exact provider/modelId entries in enabledModels are honored — globs (*sonnet*), bare model IDs, and :thinking suffixes that pi itself supports are silently dropped here. Matches pi's /scoped-models picker output, so the limitation is invisible to UI users.

Changed

  • Agent tool prompt restructured to mirror Claude Code's upstream Agent tool description format. Section headings now match upstream (## When not to use, ## Usage notes, ## Writing the prompt); the auto-generated agent list renders as a flat list (no Default agents: / Custom agents: sub-headers) with a per-agent (Tools: …) suffix derived from each agent's builtinToolNames (or * when the agent has the full built-in set). Restored upstream's load-bearing guardrails that were missing or compressed in the old prompt: "result is not visible to the user → summarize", "trust but verify", "fresh agent / self-contained prompt" on resume, "tell the agent whether to write code or do research", "use proactively when the description says so", "MUST send a single message for parallel", and the worktree auto-cleanup behavior detail. The three redundant "Use Explore / Plan / general-purpose for …" shorthand bullets were dropped — the agent descriptions themselves now carry the canonical (and richer) selection guidance. Upstream's two <example> blocks at the end of "Writing the prompt" are also intentionally omitted: the per-orchestrator-turn token cost is recurring, the abstract guidance + the now-rich agent descriptions cover the same pedagogical ground, and the examples embed Anthropic-specific <thinking> framing that doesn't generalize across pi-ai's provider surface (OpenAI, Bedrock, Gemini, Mistral, …). All pi-specific bullets (resume, steer_subagent, model, thinking, inherit_context, isolation: "worktree", ${scheduleGuideline}) preserved.
  • Default agent descriptions (general-purpose, Explore, Plan) replaced with upstream Claude Code's verbatim wording. Previously one-line labels (e.g. "Fast codebase exploration agent (read-only)"); now multi-sentence descriptions that include positive ("Use it to …") and negative ("Do NOT use it for …") guidance plus, for Explore, search-breadth hints ("quick" / "medium" / "very thorough"). The LLM-facing selection signal is now substantially stronger.
  • /agents → Eject now emits YAML-safe description: frontmatter. The new Explore description contains a : colon-space pattern (the search-breadth hint) and embedded quote characters — emitting it raw would have produced malformed frontmatter that the yaml parser would mis-parse. ejectAgent now wraps the description with JSON.stringify (a valid YAML 1.2 double-quoted scalar), so any description string round-trips cleanly through eject → re-load. Latent bug: previously unreachable because old descriptions were YAML-plain-safe.
  • /agents → Settings UI rewritten to inline-editable SettingsList. Replaces the previous modal ctx.ui.select chain. All settings visible at once; / to navigate, Space to cycle preset values on numerics (Max concurrency, Default max turns, Grace turns), Enter to type a custom value, Esc to exit. Functionally equivalent — same fields, same valid ranges, same persistence behavior — but the interaction model is different. Users scripting against the old screen flow may notice.
  • .gitignore additions. Added .pi/subagents.json (project-local subagents settings — written by /agents → Settings, shouldn't be committed) plus pi-runtime working files (progress.md, AGENTS.md, CLAUDE.md). Migration: if you previously committed .pi/subagents.json to your repo, run git rm --cached .pi/subagents.json to untrack — gitignore only blocks new additions.

[0.8.0] - 2026-05-26

⚠️ Breaking: peer dependencies moved from @mariozechner/pi-* to @earendil-works/pi-*. The upstream Pi runtime relocated npm scopes on 2026-05-07; the @mariozechner/pi-* packages are deprecated. This release pins @earendil-works/pi-{ai,coding-agent,tui} at >=0.74.0. Hosts on the old scope must update their pi installation first (pi update --self handles the rename automatically) before installing this version.

Note on Node: this release is tested against @earendil-works/pi-coding-agent@latest (currently 0.75.x), which requires Node >=22.19.0 because its bundled undici calls Node 22+ APIs. CI runs on Node 22. The peer range (>=0.74.0) technically also matches the upstream legacy-node20 line (0.74.x, Node 20 compatible) and this extension contains no Node 22+ API calls of its own, but the legacy line is not exercised in CI — consumers pinning it do so at their own risk.

Changed

  • Peer deps migrated from @mariozechner/pi-* to @earendil-works/pi-* (#76 — thanks @SEHANTA for the report). On 2026-05-07 the upstream Pi runtime moved npm scopes — @mariozechner/pi-coding-agent@0.73.1 was the final publish (now deprecated on npm), and @earendil-works/pi-coding-agent@0.74.0 shipped 30 minutes later from the same monorepo (same author, same code). peerDependencies now target @earendil-works/pi-{ai,coding-agent,tui} at >=0.74.0, and all src/** and test/** imports are renamed to the new scope — pure rename, no API changes. Consumers pinning the new scope no longer hit the peer-dep conflict warnings reported in #76.
  • ThinkingLevel now imported from @earendil-works/pi-ai instead of …/pi-agent-core. src/types.ts previously reached past the public API into pi-agent-core (an internal package), which only resolved because npm flat-hoisted it as a transitive of pi-coding-agent — under pnpm or strict-resolver setups the import failed (TS2307: Cannot find module '@mariozechner/pi-agent-core'). pi-ai re-exports ThinkingLevel from its public surface (export * from "./types.ts"), so the import goes through the documented entry point and no extra peer dep is needed.

Fixed

  • .pi/subagent-schedules/ is no longer created in every working directory. ScheduleStore's constructor previously ran mkdirSync unconditionally, so any session with scheduling enabled left an empty .pi/subagent-schedules/ dir behind even when nothing was ever scheduled. Directory creation is now lazy — deferred to a new private ensureDir() invoked at the top of withLock, so the dir (and its <sessionId>.json) appear only when a job is actually persisted. Additionally, update/remove now short-circuit on an unknown id (in-memory jobs.has(id) check) before taking the lock, so no-op mutations never touch disk. Read-only use (list/get/hasName) and constructing the store never create the dir. Pre-existing leftover dirs are not cleaned up — remove them manually.

[0.7.3] - 2026-05-14

Added

  • <active_agent name="…"/> tag prepended to every child system prompt (#73 — thanks @chris-lasher). buildAgentPrompt now emits <active_agent name="${config.name}"/> as the first line of the assembled prompt in both replace and append modes, before the env block. Downstream extensions (e.g. permission/policy systems) can parse it from inside the child session to resolve per-agent policy. The tag uses the agent's config.name verbatim — no escaping or normalization — and does not couple this extension to any specific downstream consumer; ignoring it is harmless.

Changed

  • Subagent sessions now get a stable, type-derived name with an id suffix for parallel spawns (#51 — thanks @forcepushdev). runAgent calls session.setSessionName(agentConfig?.name ?? type), and when the manager assigns an agentId (always, in production), the name is suffixed with an 8-char slice — e.g. Explore#a1b2c3d4 — so concurrent spawns of the same agent type are distinguishable in the overlay instead of all collapsing onto the same bare name. Direct runAgent callers without an agentId (e.g. tests) get the bare name.

Fixed

  • Cross-extension spawn RPC now accepts a string options.model (#59, fixes #60). Cross-extension callers (e.g. @tintinweb/pi-tasks@>=0.4.3's TaskExecute) naturally forward model as a serializable "provider/modelId" string. Previously the spawn handler passed strings straight through to runAgent(), which expects a Model object — the spawned agent then crashed with No API key found for undefined. The handler now resolves strings via the same resolveModel(ctx.modelRegistry) path the scheduler uses; Model objects pass through unchanged. Unresolved strings surface the human-readable Model not found: "…" error instead of the auth-lookup crash. Thanks @any-victor.

[0.7.2] - 2026-05-12

Heads-up — behavior changes in skill preloading:

  • .txt and extensionless flat skill files are no longer loaded. Only <name>.md flat files and <name>/SKILL.md directory skills resolve now. Rename any <name>.txt or extensionless skill files to <name>.md.

Added

  • Pi-standard <name>/SKILL.md directory layout is now discovered alongside flat <name>.md files. Top-level and nested matches both resolve via BFS — for skill foo, the loader checks <root>/foo/SKILL.md, then recursively descends looking for */.../foo/SKILL.md. Recursion skips dotfile directories and node_modules; a directory that itself contains SKILL.md is treated as a single skill (Pi's "skills don't nest" rule).
  • Five discovery roots, checked in precedence order:
    • <cwd>/.pi/skills/ (project, Pi)
    • <cwd>/.agents/skills/ (project, Agent Skills spec)
    • $PI_CODING_AGENT_DIR/skills/ — default ~/.pi/agent/skills/ (user, Pi)
    • ~/.agents/skills/ (user, Agent Skills spec)
    • ~/.pi/skills/ (legacy global, kept for backward compatibility)
  • Symlink rejection broadened to the new layouts: symlinked skill roots, nested skill directories, and SKILL.md files inside otherwise-real directories are all rejected (intentional deviation from Pi, which follows symlinks).
  • Deterministic traversal order — entries are sorted byte-order so collisions resolve identically across filesystems. Pi's iteration order is readdirSync-dependent.
  • Resolved spawn args are now shown in the dedicated conversation viewer (#62). Open /subagent → Running Agents → select an agent: a second header row displays the effective invocation — model override (when different from parent), thinking: <level>, isolated, worktree, inherit context, background, and max turns: N. Tags appear when the resolved value is notable (e.g. isolated: true), not just when the caller explicitly set it; max turns is the one exception and shows only when explicitly configured. Lets you verify the parent agent honored your spawn instructions without scrolling back through the chat. Snapshot stored on the new AgentRecord.invocation field. The same tag set is also surfaced on the Agent tool-call result render (which previously showed a narrower subset).
  • Shift+↑ / Shift+↓ scroll a full page in the conversation viewer — same behavior as PgUp / PgDn. Note: some terminal emulators intercept Shift+arrows for text selection or tab switching, in which case PgUp/PgDn remain available.

Changed

  • .txt and extensionless flat skill files are no longer loaded. Pi only supports .md; we now match. Migration: rename any <name>.txt / <name> skill files to <name>.md.
  • Conversation viewer no longer fills the full screen. The overlay is now capped at 70% of terminal height (90% width unchanged), and the viewer's internal viewport mirrors that cap so the footer/scroll indicator can't be clipped.

[0.7.1] - 2026-05-07

Heads-up — behavior change:

  • isolation: "worktree" now fails loud (returns an error) instead of silently falling back to the main tree. Affects users running pi in a non-git directory or a fresh repo with no commits.

Changed

  • isolation: "worktree" now fails loud instead of silently falling back. Previously when createWorktree returned undefined (not a git repo, no commits yet, or git worktree add failed), the agent ran in the main cwd with a [WARNING: ...] block prepended to its prompt — visible only to the LLM, never surfaced to the caller. Now the failure throws a structured error that propagates back to the Agent tool response; no agent record is created. Failed scheduled fires are recorded as lastStatus: "error" with the reason in the subagents:scheduled error event. Queued background spawns whose worktree creation fails when they dequeue are marked terminal-error and don't block the rest of the queue.

Fixed

  • Headless pi --print runs no longer hang or crash after background subagents complete. Cleanup timers no longer keep the process alive, and stale completion notifications are treated as best-effort shutdown side effects.

[0.7.0] - 2026-05-04

Heads-up — behavior changes:

  • subagents:completed/failed event tokens.total now excludes cacheRead (previously double-counted across turns) — see Fixed [#38].
  • Cron ? is now a wildcard (same as *), not "current time value" — affects Quartz-style expressions only.

Changed

  • @mariozechner/pi-{ai,coding-agent,tui} moved to peerDependencies (>=0.70.5). Avoids duplicate framework instances when the host loads this extension.
  • @sinclair/typebox pinned from latest to ^0.34.49 so installs are reproducible.
  • croner bumped 8 → 10. Heads-up: in cron strings, ? now means wildcard (same as *) instead of "current time value" — affects Quartz-style expressions only.

Added

  • Master switch for scheduling — new schedulingEnabled setting (default true) under /agents → Settings → Scheduling. When set to false: the schedule parameter and its guideline are stripped from the Agent tool spec at registration (zero LLM-context cost), the scheduler does not bind to the session, the /agents → Scheduled jobs menu entry is hidden, and any in-flight scheduler is stopped immediately. The schema-level removal applies on next pi session; the runtime kill (menu, fire path) takes effect immediately. Persisted at <cwd>/.pi/subagents.json.
  • Schedule subagent spawns — the Agent tool now accepts an optional schedule parameter. When set, the spawn registers a job that fires later instead of running immediately. Three formats: 6-field cron ("0 0 9 * * 1" — 9am every Monday), interval ("5m", "1h"), or one-shot ("+10m" or ISO timestamp). Returns the job ID. Schedules are session-scoped — they reset on /new, restore on /resume (mirrors the persistence model of pi-chonky-tasks). Storage at <cwd>/.pi/subagent-schedules/<sessionId>.json, with PID-based file locking + atomic temp+rename for concurrent-instance safety. Result delivery is identical to today's background-spawn completions: when the scheduled agent finishes, the existing subagent-notification followUp path emits the result to the conversation — no new delivery code, no new message types. Concurrency: scheduled fires bypass maxConcurrent so a 5-minute interval can't be deferred behind 4 long-running manual agents. Management: /agents → "Scheduled jobs" lists active jobs and lets you cancel any one of them. Creation is via the Agent tool only — no parallel manual-create wizard in this iteration. Events: subagents:scheduled ({ type: "added" | "removed" | "updated" | "fired" | "error", … }) and subagents:scheduler_ready for cross-extension consumers. Restrictions: schedule is incompatible with inherit_context (no parent at fire time) and resume (schedules create fresh agents); forces run_in_background: true. Scheduler engine mirrors pi-cron-schedule (croner for cron, setInterval/setTimeout for interval/once); past one-shot timestamps and invalid cron expressions are caught at create time.
  • Context-window utilization indicator in the subagent overlay — token count is now followed by a colored (NN%) showing how full the subagent's context is right now (estimateContextTokens(messages) / model.contextWindow * 100, sourced from upstream contextUsage.percent). Threshold colors: <70% dim, 7085% warning, ≥85% error. Gracefully omitted when the model has no contextWindow declared, or right after compaction before the next assistant turn (tokens is null in that window). The same annotation slot also surfaces a compaction count ↻N when the agent has compacted at least once — e.g. 12.3k token (84% · ↻3) (percent + compactions joined with ·), 12.3k token (↻1) (compactions only, immediately post-compaction while percent is still null). The compaction glyph stays dim regardless; the percent's threshold color carries the urgency signal. Two live overlays get the annotations (running stats line; inspect-overlay header); post-completion notifications and result/event payloads only get the count (the indicator is no longer actionable once the agent is done).
  • Token usage and context% exposed to the parent agent at every interaction surface — get_subagent_result adds Context: NN% to its stats line; steer_subagent returns a Current state: 12.3k token · 5 tool uses · context 72% full line so the steering agent knows whether it has room before sending more context; task-notification XML adds <context_percent>NN</context_percent> (omitted when null). All plain-text, no ANSI codes — designed for LLM consumption, not human display.
  • New subagents:compacted lifecycle event fires when a subagent's session successfully compacts. Payload: { id, type, description, reason: "manual" | "threshold" | "overflow", tokensBefore, compactionCount }tokensBefore is upstream's pre-compaction context size estimate; compactionCount is the running total for this agent (also persisted on AgentRecord.compactionCount and surfaced in get_subagent_result / steer_subagent / task-notification when > 0). Aborted compactions don't fire. Routed through a new manager-level onCompact constructor callback, matching the existing onStart / onComplete pattern.

Fixed

  • Subagent token count was inflated 515× and reset mid-run (#38). Two distinct bugs in the same field. (1) Upstream getSessionStats().tokens.total sums per-turn cacheRead across every assistant message — but each turn's cacheRead is the cumulative cached prefix re-read on that one API call, so summing N turns counts the prefix N times (quadratic inflation, very visible on long sessions). (2) Even with that fixed, anything derived from session.state.messages resets at compaction because upstream replaces the array via this.agent.state.messages = sessionContext.messages. Fix replaces all six display readers with a lifetime accumulator (AgentRecord.lifetimeUsage and AgentActivity.lifetimeUsage{ input, output, cacheWrite }) fed by a new onAssistantUsage callback dispatched from message_end events in both runAgent and resumeAgent. The accumulator is independent of state.messages mutation, so it survives compaction; total = input + output + cacheWrite by construction (cacheRead deliberately excluded — same prefix-double-counting reason). The subagents:completed/failed event payload's tokens field is now also lifetime-accumulated for input, output, and total together (was: total lifetime, input/output session-derived → inconsistent after compaction).
  • ESC during a foreground Agent call now actually stops the subagent (#44 — thanks @Zeng-Zer). Pi's interrupt path is esc → agent.abort() on the parent → AbortSignal delivered to every tool's execute(toolCallId, params, signal, …), but the Agent tool dropped that signal on the floor: subagents ran on their own independent AbortController inside AgentManager, so the parent abort was invisible and the subagent kept running until natural completion or max_turns. Fix threads signal through Agent.executemanager.spawnAndWait()SpawnOptions.signal, and AgentManager.startAgent() now attaches an { once: true } "abort" listener that calls this.abort(id) (which sets status: "stopped" and aborts the child controller). The listener is detached in both .then and .catch to avoid leaking on natural settle. Scope: foreground only — background agents intentionally outlive the parent tool call, so their spawn deliberately does not forward signal. Resume path (AgentManager.resume()) has the same blind spot and is tracked as a follow-up.

0.6.3 - 2026-04-28

Fixed

  • run_in_background: true (and inherit_context, isolated) silently ignored on default agents (#37 — thanks @kylesnowschwartz for the diagnosis). The three built-in defaults (general-purpose, Explore, Plan) baked runInBackground: false, inheritContext: false, and isolated: false into their configs. resolveAgentInvocationConfig uses agentConfig?.field ?? params.field ?? false, and ?? only falls through on null/undefined — so an explicit false from the agent config silently won over the caller's true. Calling Agent({ subagent_type: "general-purpose", run_in_background: true }) returned the result inline instead of backgrounding, blocking the parent UI for the agent's full runtime. Fix drops the three lines from each default (and from the unreachable defensive fallback in agent-runner.ts) — the type already declared each as field?: boolean with JSDoc "undefined = caller decides", so the runtime now matches the documented contract. Behavior: custom agents that explicitly set these fields in frontmatter still lock as before (the v0.5.1 "frontmatter is authoritative" guarantee is preserved); the fix only stops defaults from spuriously claiming an opinion on callsite-strategy fields they don't actually have. The unreachable fallback now spreads DEFAULT_AGENTS.get("general-purpose") instead of duplicating the config inline, so future drift is impossible.

0.6.2 - 2026-04-28

Fixed

  • Agent tool fails on Windows with ENOENT creating output directory (#27 — thanks @sixnathan for the diagnosis). The cwd-encoding regex in output-file.ts only handled POSIX / separators, so on Windows cwd = "C:\\Users\\foo\\project" survived unchanged and path.join(tmpRoot, encoded, …) produced an invalid nested-absolute path. Now extracts a small encodeCwd() helper that handles both / and \\ separators, strips the Windows drive-letter prefix, and preserves UNC server/share segments. The chmodSync(root, 0o700) call is also wrapped in a try/catch that swallows errors only on Windows (where chmod is a no-op and can throw on some filesystems); on Unix the error still propagates so umask-defeating 0o700 enforcement is preserved.

0.6.1 - 2026-04-25

Added

  • Persistent /agents → Settings (#24) — the four runtime tuning values (maxConcurrent, defaultMaxTurns, graceTurns, defaultJoinMode) now survive pi restarts via a two-file dual-scope model mirroring pi's own SettingsManager. Global ~/.pi/agent/subagents.json provides machine-wide defaults (edit by hand; the menu never writes here); project <cwd>/.pi/subagents.json holds per-project overrides (written by /agents → Settings). Load merges both with project winning on conflicts. Invalid fields are silently dropped per field; malformed JSON emits a warning to stderr and falls back to defaults so startup always proceeds; write failures downgrade the settings toast to a warning with (session only; failed to persist) so changes aren't silently reverted on next restart.
  • New lifecycle eventssubagents:settings_loaded (emitted once at extension init with the merged settings) and subagents:settings_changed (emitted on each /agents → Settings mutation with the new snapshot and a persisted: boolean flag so listeners can react to write failures).

Fixed

  • AGENTS.md / CLAUDE.md / APPEND_SYSTEM.md no longer leak into sub-agent prompts (#26 — thanks @mikeyobrien for the diagnosis). Upstream buildSystemPrompt() re-appends contextFiles and appendSystemPrompt after our systemPromptOverride runs, which silently defeated prompt_mode: replace and isolated: true — parent project context (e.g. autoresearch-mode blocks) was bleeding into fresh Explore / custom sub-agents regardless of frontmatter. Fix uses upstream's noContextFiles: true flag (skips the load entirely, introduced in pi 0.68) plus appendSystemPromptOverride: () => [] (no flag equivalent for append sources). Behavior change: subagents no longer implicitly inherit parent AGENTS.md/CLAUDE.md/APPEND_SYSTEM.md. To get parent project context into a subagent, use prompt_mode: append (parent's already-built system prompt flows in via systemPromptOverride), or inherit_context: true (parent conversation), or inline the content into the agent's own frontmatter.
  • Custom agent discovery respects PI_CODING_AGENT_DIR (#35, closes #23 — thanks @Amolith for the diagnosis). Two remaining hardcoded ~/.pi/agent/agents/ paths in custom-agents.ts and index.ts bypassed the env var, so users who relocated their agent directory (e.g. via PI_CODING_AGENT_DIR) still had global agents loaded from the default location and help text referencing the wrong path. Both now use upstream getAgentDir(), consistent with agent-runner.ts and settings.ts; tilde expansion is handled by upstream.

0.6.0 - 2026-04-24

⚠️ Breaking: drops support for pi < 0.68. The upstream pi-coding-agent package shipped breaking API changes in v0.68 (and further ones in v0.70). This release migrates to ^0.70.2 and is not backward-compatible with hosts on pi 0.620.67. Users on those versions must upgrade their pi installation (npm install -g @mariozechner/pi-coding-agent@latest) before updating this extension.

Changed

  • Bumped peer @mariozechner/pi-coding-agent to ^0.70.2 (#28) — crosses the v0.68 breaking-change line upstream. Specifically: tools are now passed as string[] (was Tool[]); cwd/agentDir are mandatory on SettingsManager.create() and DefaultResourceLoader; session_switch event renamed to session_before_switch; ToolDefinition.params widens to unknown under contextual typing, requiring defineTool(...).
  • Tool registrations wrapped with defineTool(...) — preserves TParams inference so execute handlers get properly-typed params instead of unknown. Applies to the Agent, get_subagent_result, and steer_subagent tools.

Removed

  • Cwd-bound tool factory registry — the internal TOOL_FACTORIES closure table and create{Bash,Edit,Read,Write,Grep,Find,Ls}Tool imports are gone. Exported helpers renamed: getToolsForType(type, cwd)getToolNamesForType(type), getMemoryTools(cwd, set)getMemoryToolNames(set), getReadOnlyMemoryTools(cwd, set)getReadOnlyMemoryToolNames(set) — all returning string[] instead of Tool[]. The host binds cwd when resolving tool names, so the extension no longer instantiates tools directly.

Fixed

  • Subagent SettingsManager read wrong project settings in worktree mode (#30) — SettingsManager.create() was called without arguments, defaulting cwd to process.cwd(). When the subagent's effective cwd differed (worktree isolation or explicit cwd override), its settings manager read .pi/settings.json from the parent's cwd rather than its own, diverging from the loader and session manager. Now passes effectiveCwd and agentDir explicitly, keeping all three managers consistent.

0.5.2 - 2026-03-26

Fixed

  • Extension session_start handlers now fire in subagent sessions (#20) — bindExtensions() was never called on subagent sessions, so extensions that initialize state in session_start (e.g. loading credentials, setting up connections) silently failed at runtime. Tools appeared registered but were non-functional. Now calls session.bindExtensions() after tool filtering and before prompting, matching the lifecycle used by pi's interactive, print, and RPC modes. Also triggers extendResourcesFromExtensions("startup") so extension-provided skills and prompts are discovered.

0.5.1 - 2026-03-24

Changed

  • Agent config is authoritative — frontmatter values for model, thinking, max_turns, inherit_context, run_in_background, isolated, and isolation now take precedence over Agent tool-call parameters. Tool-call params only fill fields the agent config leaves unspecified.
  • join_mode is now a global setting only — removed the per-call join_mode parameter from the Agent tool. Join behavior is configured via /agents → Settings → Join mode.
  • max_turns: 0 means unlimited — agent files can now explicitly set max_turns: 0 to lock unlimited turns. Previously 0 was silently clamped to 1.

Fixed

  • Final subagent text preserved from non-streaming providers — agents using providers that return the final message without streaming text_delta events no longer return empty results. Falls back to extracting text from the completed session history.
  • effectiveMaxTurns passed to spawn calls — previously params.max_turns was passed raw to both foreground and background spawn, bypassing the agent config entirely.

0.5.0 - 2026-03-22

Added

  • RPC stop handler — new subagents:rpc:stop event bus RPC allows other extensions to stop running subagents by agent ID. Returns structured error ("Agent not found") on failure.
  • abort in SpawnCapable interface — cross-extension RPC consumers can now stop agents, not just spawn them.
  • Live turn counter — all agents now show a live turn count in the widget, inline result, and completion notification. With a turn limit: ⟳5≤30 (5 of 30 turns). Without: ⟳5. Updates in real time as turns progress via onTurnEnd callback.
  • Biome linting — added Biome for correctness linting (unused imports, suspicious patterns). Style rules disabled. Run npm run lint to check, npm run lint:fix to auto-fix.
  • CI workflow — GitHub Actions runs lint, typecheck, and tests on push to master and PRs.
  • Auto-trigger parent turn on background completion — background agent completion notifications now use triggerTurn: true, automatically prompting the parent agent to process results instead of waiting for user input.

Changed

  • Standardized RPC envelope — cross-extension RPC handlers (ping, spawn, stop) now use a handleRpc wrapper that emits structured envelopes ({ success: true, data } / { success: false, error }), matching pi-mono's RpcResponse convention.
  • Protocol versioning via ping — ping reply now includes { version: PROTOCOL_VERSION } (currently v2). Callers can detect version mismatches and warn users to update.
  • Default max turns is now unlimited — subagents no longer have a 50-turn default cap. The default is unlimited (no turn limit), matching Claude Code's main loop behavior. Users can still set explicit limits per-agent via max_turns frontmatter or the Agent tool parameter, or globally via /agents → Settings (0 = unlimited).
  • Stale dist in published package — added prepublishOnly hook to build fresh dist/ on every npm publish.

Fixed

  • Tool name displaygetAgentConversation now reads ToolCall.name (the correct property) instead of toolName, resolving [Tool: unknown] in conversation viewer and verbose output.
  • Env test CI failuredetectEnv test assumed a branch name exists, but CI checks out detached HEAD. Split into separate tests for repo detection and branch detection with a controlled temp repo.

0.4.9 - 2026-03-18

Fixed

  • Conversation viewer crash in narrow terminals (#7) — buildContentLines() in the live conversation viewer could return lines wider than the terminal when wrapTextWithAnsi() misjudged visible width on ANSI-heavy input (e.g. tool output with embedded escape codes, long URLs, wide tables). All content lines are now clamped with truncateToWidth() before returning. Same class of bug as the widget fix in v0.2.7, different component.

Added

  • Conversation viewer width-safety tests — 17 tests covering render() and buildContentLines() across varied content (plain text, ANSI codes, unicode, tables, long URLs, narrow terminals). Includes mock-based regression tests that simulate upstream wrapTextWithAnsi returning overwidth lines, ensuring the safety net catches them.

0.4.8 - 2026-03-18

Added

  • Cross-extension RPC — other pi extensions can spawn subagents via pi.events event bus (subagents:rpc:ping, subagents:rpc:spawn). Emits subagents:ready on load.
  • Session persistence for agent records — completed agent records are persisted via pi.appendEntry("subagents:record", ...) for cross-extension history reconstruction.

Fixed

  • Background agent notification race conditionpi.sendMessage() is fire-and-forget, so completion notifications sent eagerly from onComplete could not be retracted when get_subagent_result was called in the same turn. Notifications are now held behind a 200ms cancellable timer; get_subagent_result cancels the pending timer before it fires, eliminating duplicate notifications. Group notifications also re-check resultConsumed at send time so consumed agents are filtered out.

0.4.7 - 2026-03-17

Added

  • Custom notification renderer — background agent completion notifications now render as styled, themed boxes instead of raw XML. Uses pi.registerMessageRenderer() with the "subagent-notification" custom message type. The LLM continues to receive <task-notification> XML via content; only the user-facing display changes.
  • Group notification rendering — group completions render each agent as its own styled block (icon, description, stats, result preview) instead of showing only the first agent.
  • Output file streaming for background agents — background agents now get the same output file transcript as foreground agents, with onSessionCreated wiring and proper cleanup on completion/error.
  • NotificationDetails type in types.ts — structured details for the notification renderer, with optional others array for group notifications.
  • buildNotificationDetails() helper — extracts renderer-facing details from an AgentRecord.

Changed

  • Notification deliverysendIndividualNudge and group notification now use pi.sendMessage() (custom message) instead of pi.sendUserMessage() (plain text), enabling renderer-controlled display.
  • Steered status rendering — steered agents show "completed (steered)" in the notification box instead of plain "completed".

Fixed

  • Output file cleanup on completionagent-manager.ts now calls record.outputCleanup() in both the success and error paths of agent completion, ensuring the streaming subscription is flushed and released.

0.4.6 - 2026-03-16

Fixed

  • Graceful shutdown aborts agents instead of blockingsession_shutdown now calls abortAll() instead of waitForAll(), so the process exits immediately instead of hanging until all background agents complete. Agent results are undeliverable after shutdown anyway.

Added

  • abortAll() method on AgentManager — stops all queued and running agents at once, returning the count of affected agents.

0.4.5 - 2026-03-16

Changed

  • Widget render-once pattern — the widget callback is now registered once via setWidget() and subsequent updates use requestRender() instead of re-registering the entire widget on every update() call. Eliminates layout thrashing from repeated widget teardown/setup cycles.
  • Status bar dedupsetStatus() is now only called when the status text actually changes, avoiding redundant TUI updates.
  • UICtx change detectionsetUICtx() detects context changes and forces widget re-registration, correctly handling session switches.

Refactored

  • Extracted renderWidget() private method — moves all widget content rendering out of the update() closure into a standalone method that reads live state on each call.
  • update() is now a lightweight coordinator: counts agents, manages registration lifecycle, and triggers re-renders.

0.4.4 - 2026-03-16

Fixed

  • Race condition in get_subagent_result with wait: trueresultConsumed is now set before await record.promise, preventing a redundant follow-up notification. Previously the onComplete callback (attached at spawn time via .then()) always fired before the await resumed, seeing resultConsumed as false.
  • Stale agent records across sessions — new clearCompleted() method removes all completed/stopped/errored agent records on session_start and session_switch events, so tasks from a prior session don't persist into a new one.
  • steer_subagent race on freshly launched agents — steering an agent before its session initialized silently dropped the message. Now steers are queued on the record and flushed once onSessionCreated fires.

Changed

  • Extracted removeRecord() private helper in AgentManager — deduplicates dispose+delete logic between cleanup() and clearCompleted().

Added

  • 8 new tests covering resultConsumed race condition and clearCompleted behavior (185 total).

0.4.3 - 2026-03-13

Added

  • Persistent agent memory — new memory frontmatter field with three scopes: "user" (global ~/.pi/), "project" (per-project .pi/), "local" (gitignored .pi/). Agents with write/edit tools get full read-write memory; read-only agents get a read-only fallback that injects existing MEMORY.md content without granting write access or creating directories.
  • Git worktree isolation — new isolation: "worktree" frontmatter field and Agent tool parameter. Creates a temporary git worktree so agents work on an isolated copy of the repo. On completion, changes are auto-committed to a pi-agent-<id> branch; clean worktrees are removed. Includes crash recovery via pruneWorktrees().
  • Skill preloadingskills frontmatter now accepts a comma-separated list of skill names (e.g. skills: planning, review). Reads from .pi/skills/ (project) then ~/.pi/skills/ (global), tries .md/.txt/bare extensions. Content injected into the system prompt as # Preloaded Skill: {name}.
  • Tool denylist — new disallowed_tools frontmatter field (e.g. disallowed_tools: bash, write). Blocks specified tools even if builtinToolNames or extensions would provide them. Enforced for both extension-enabled and extension-disabled agents.
  • Prompt extras system — new PromptExtras interface in prompts.ts; buildAgentPrompt() accepts optional memory and skill blocks appended in both replace and append modes.
  • getMemoryTools(), getReadOnlyMemoryTools() in agent-types.ts.
  • buildMemoryBlock(), buildReadOnlyMemoryBlock(), isSymlink(), safeReadFile() in memory.ts.
  • preloadSkills() in skill-loader.ts.
  • createWorktree(), cleanupWorktree(), pruneWorktrees() in worktree.ts.
  • MemoryScope, IsolationMode types; memory, isolation, disallowedTools fields on AgentConfig; worktree, worktreeResult fields on AgentRecord.
  • 177 total tests across 8 test files (41 new tests).

Fixed

  • Read-only agents no longer escalated to read-write — enabling memory on a read-only agent (e.g. Explore) previously auto-added write/edit tools. Now the runner detects write capability and branches: read-write agents get full memory tools, read-only agents get read-only memory prompt with only the read tool added.
  • Denylist-aware memory detection — write capability check now accounts for disallowedTools. An agent with tools: write + disallowed_tools: write correctly gets read-only memory instead of broken read-write instructions.
  • Worktree requires commits — repos with no commits (empty HEAD) are now rejected early with a warning instead of failing silently at git worktree add.
  • Worktree failure warning — when worktree creation fails, a warning is prepended to the agent's prompt instead of silently falling through to the main cwd.
  • No force-branch overwrite — worktree cleanup appends a timestamp suffix on branch name conflict instead of using git branch -f.

Security

  • Whitelist name validation — agent/skill names must match ^[a-zA-Z0-9][a-zA-Z0-9._-]*$, max 128 chars. Rejects path traversal, leading dots, spaces, and special characters.
  • Symlink protectionsafeReadFile() and isSymlink() reject symlinks in memory directories, MEMORY.md files, and skill files, preventing arbitrary file reads.
  • Symlink-safe directory creationensureMemoryDir() throws on symlinked directories.

Changed

  • agent-runner.ts: tool/extension/skill resolution moved before memory detection; ctx.cwdeffectiveCwd throughout.
  • custom-agents.ts: extracted parseCsvField() helper; added csvListOptional() and parseMemory().
  • skill-loader.ts: uses safeReadFile() from memory.ts instead of raw readFileSync.
  • Agent tool schema updated with isolation parameter and help text for memory, isolation, disallowed_tools, and skill list.

0.4.2 - 2026-03-12

Added

  • Event bus — agent lifecycle events emitted via pi.events.emit(), enabling other extensions to react to sub-agent activity:
    • subagents:created — background agent registered (includes id, type, description, isBackground)
    • subagents:started — agent transitions to running (includes queued→running)
    • subagents:completed — agent finished successfully (includes durationMs, tokens, toolUses, result)
    • subagents:failed — agent errored, stopped, or aborted (same payload as completed)
    • subagents:steered — steering message sent to a running agent
  • OnAgentStart callback and onStart constructor parameter on AgentManager.
  • Cross-package manager now also exposes spawn() and getRecord() via the Symbol.for("pi-subagents:manager") global.

0.4.1 - 2026-03-11

Fixed

  • Graceful shutdown in headless mode — the CLI now waits for all running and queued background agents to complete before exiting (waitForAll on session_shutdown). Previously, background agents could be silently killed mid-execution when the session ended. Only affects headless/non-interactive mode; interactive sessions already kept the process alive.

Added

  • hasRunning() / waitForAll() methods on AgentManager.
  • Cross-package manager access — agent manager exposed via Symbol.for("pi-subagents:manager") on globalThis for other extensions to check status or await completion.

0.4.0 - 2026-03-11

Added

  • XML-delimited prompt sections — append-mode agents now wrap inherited content in <inherited_system_prompt>, <sub_agent_context>, and <agent_instructions> XML tags, giving the model explicit structure to distinguish inherited rules from sub-agent-specific instructions. Replace mode is unchanged.
  • Token count in agent results — foreground agent results, background completion notifications, and get_subagent_result now include the token count alongside tool uses and duration (e.g. Agent completed in 4.2s (12 tool uses, 33.8k token)).
  • Widget overflow cap — the running agents widget now caps at 12 lines. When exceeded, running agents are prioritized over finished ones and an overflow summary line shows hidden counts (e.g. +3 more (1 running, 2 finished)).

Changed - changing behavior

  • General-purpose agent inherits parent prompt — the default general-purpose agent now uses promptMode: "append" with an empty system prompt, making it a "parent twin" that inherits the full parent system prompt (including CLAUDE.md rules, project conventions, and safety guardrails). Previously it used a standalone prompt that duplicated a subset of the parent's rules. Explore and Plan are unchanged (standalone prompts). To customize: eject via /agents → select general-purpose → Eject, then edit the resulting .md file. Set prompt_mode: replace to go back to a standalone prompt, or keep prompt_mode: append and add extra instructions in the body.
  • Append-mode agents receive parent system promptbuildAgentPrompt now accepts the parent's system prompt and threads it into append-mode agents (env header + parent prompt + sub-agent context bridge + optional custom instructions). Replace-mode agents are unchanged.
  • Prompt pipeline simplified — removed systemPromptOverride/systemPromptAppend from SpawnOptions and RunOptions. These were a separate code path where index.ts pre-resolved the prompt mode and passed raw strings into the runner, bypassing buildAgentPrompt. Now all prompt assembly flows through buildAgentPrompt using the agent's promptMode config — one code path, no special cases.

Removed

  • Deprecated backwards-compat aliases: registerCustomAgents, getCustomAgentConfig, getCustomAgentNames (use registerAgents, getAgentConfig, getUserAgentNames).
  • resolveCustomPrompt() helper in index.ts — no longer needed now that prompt routing is config-driven.

0.3.1 - 2026-03-09

Added

  • Live conversation viewer — selecting a running (or completed) agent in /agents → "Running agents" now opens a scrollable overlay showing the agent's full conversation in real time. Auto-scrolls to follow new content; scroll up to pause, End to resume. Press Esc to close.

0.3.0 - 2026-03-08

Added

  • Case-insensitive agent type lookup"explore", "EXPLORE", and "Explore" all resolve to the same agent. LLMs frequently lowercase type names; this prevents validation failures.
  • Unknown type fallback — unrecognized agent types fall back to general-purpose with a note, instead of hard-rejecting. Matches Claude Code behavior.
  • Dynamic tool list for general-purposebuiltinToolNames is now optional in AgentConfig. When omitted, the agent gets all tools from TOOL_FACTORIES at lookup time, so new tools added upstream are automatically available.
  • Agent source indicators in /agents menu (project), (global), (disabled) with legend. Defaults are unmarked.
  • Disabled agents visible in UI — disabled agents now show in the "Agent types" list (marked ) with an Enable action, instead of being invisible.
  • Enable action — re-enable a disabled agent from the /agents menu. Stub files are auto-cleaned.
  • Disable action for all agent types — custom and ejected default agents can now be disabled from the UI, not just built-in defaults.
  • resolveType() export — case-insensitive type name resolution for external use.
  • getAllTypes() export — returns all agent names including disabled (for UI listing).
  • source field on AgentConfig — tracks where an agent was loaded from ("default", "project", "global").

Fixed

  • Model resolver checks auth for exact matchesresolveModel("anthropic/claude-haiku-4-5-20251001") now fails gracefully when no Anthropic API key is configured, instead of returning a model that errors at the API call. Explore silently falls back to the parent model on non-Anthropic setups.

Changed

  • Unified agent registry — built-in and custom agents now use the same AgentConfig type and a single registry. No more separate code paths for built-in vs custom agents.
  • Default agents are overridable — creating a .md file with the same name as a default agent (e.g. .pi/agents/Explore.md) overrides it.
  • /agents menu — "Agent types" list shows defaults and custom agents together with source indicators. Default agents get Eject/Disable actions; overridden defaults get Reset to default.
  • Eject action — export a default agent's embedded config as a .md file to project or personal location for customization.
  • Model labels — provider-agnostic: strips provider/ prefix and -YYYYMMDD date suffix (e.g. anthropic/claude-haiku-4-5-20251001claude-haiku-4-5). Works for any provider.
  • New frontmatter fieldsdisplay_name (UI display name) and enabled (default: true; set to false to disable).
  • Menu navigation — Esc in agent detail returns to agent list (not main menu).

Removed

  • statusline-setup and claude-code-guide agents — removed as built-in types (never spawned programmatically). Users can recreate them as custom agents if needed.
  • BuiltinSubagentType union type, SUBAGENT_TYPES array, DISPLAY_NAMES map, SubagentTypeConfig interface — replaced by unified AgentConfig.
  • buildSystemPrompt() switch statement — replaced by config-driven buildAgentPrompt().
  • HAIKU_MODEL_IDS fallback array — Explore's haiku default is now just the model field in its config.
  • BUILTIN_MODEL_LABELS — model labels now derived from config.
  • ALL_TOOLS hardcoded constant — general-purpose now derives tools dynamically.

Added

  • src/default-agents.ts — embedded default configs for general-purpose, Explore, and Plan.

0.2.7 - 2026-03-08

Fixed

  • Widget crash in narrow terminals — agent widget lines were not truncated to terminal width, causing doRender to throw when the tmux pane was narrower than the rendered content. All widget lines are now truncated using truncateToWidth() with the actual terminal column count.

0.2.6 - 2026-03-07

Added

  • Background task join strategies — smart grouping of background agent completion notifications
    • smart (default): 2+ background agents spawned in the same turn are auto-grouped into a single consolidated notification instead of individual nudges
    • async: each agent notifies individually on completion (previous behavior)
    • group: force grouping even for solo agents
    • 30s timeout after first completion delivers partial results; 15s straggler re-batch window for remaining agents
  • join_mode parameter on the Agent tool — override join strategy per agent ("async" or "group")
  • Join mode setting in /agents → Settings — configure the default join mode at runtime
  • New src/group-join.tsGroupJoinManager class for batched completion notifications

Changed

  • AgentRecord now includes optional groupId, joinMode, and resultConsumed fields
  • Background agent completion routing refactored: individual nudge logic extracted to sendIndividualNudge(), group delivery via GroupJoinManager

Fixed

  • Debounce window race — agents that complete during the 100ms batch debounce window are now deferred and retroactively fed into the group once it's registered, preventing split notifications (one individual + one partial group) and zombie groups
  • Solo agent swallowed notification — if only one agent was spawned (no group formed) but it completed during the debounce window, its deferred notification is now sent when the batch finalizes
  • Duplicate notifications after polling — calling get_subagent_result on a completed agent now marks its result as consumed, suppressing the subsequent completion notification (both individual and group)

0.2.5 - 2026-03-06

Added

  • Interactive /agents menu — single command replaces /agent and /agents with a full management wizard
    • Browse and manage running agents
    • Custom agents submenu — edit or delete existing agents
    • Create new custom agents via manual wizard or AI-generated (with comprehensive frontmatter documentation for the generator)
    • Settings: configure max concurrency, default max turns, and grace turns at runtime
    • Built-in agent types shown with model info (e.g. Explore · haiku)
    • Aligned formatting for agent lists
  • Configurable turn limitsdefaultMaxTurns and graceTurns are now runtime-adjustable via /agents → Settings
  • Sub-menus return to main menu instead of exiting

Removed

  • /agent <type> <prompt> command (use Agent tool directly, or create custom agents via /agents)

0.2.4 - 2026-03-06

Added

  • Global custom agents — agents in ~/.pi/agent/agents/*.md are now discovered automatically and available across all projects
  • Two-tier discovery hierarchy: project-level (.pi/agents/) overrides global (~/.pi/agent/agents/)

0.2.3 - 2026-03-05

Added

  • Screenshot in README

0.2.2 - 2026-03-05

Changed

  • Renamed package to @tintinweb/pi-subagents
  • Fuzzy model resolver now only matches models with auth configured (prevents selecting unconfigured providers)
  • Custom agents hot-reload on each Agent tool call (no restart needed for new .pi/agents/*.md files)
  • Updated pi dependencies to 0.56.1

Refactored

  • Extracted createActivityTracker() — eliminates duplicated tool activity wiring between foreground and background paths
  • Extracted safeFormatTokens() — replaces 4 repeated try-catch blocks
  • Extracted buildDetails() — consolidates AgentDetails construction
  • Extracted getStatusLabel() / getStatusNote() — consolidates 3 duplicated status formatting chains
  • Shared extractText() — consolidated duplicate from context.ts and agent-runner.ts
  • Added ERROR_STATUSES constant in widget for consistent status checks
  • getDisplayName() now delegates to getConfig() instead of separate lookups
  • Removed unused Tool type export from agent-types

0.2.1 - 2026-03-05

Added

  • Persistent above-editor widget — tree view of all running/queued/finished agents with animated spinners and live stats
  • Concurrency queue — configurable max concurrent background agents (default: 4), auto-drain
  • Queued agents collapsed to single summary line in widget
  • Turn-based widget linger — completed agents clear after 1 turn, errors/aborted linger for 2 extra turns
  • Colored status icons — themed rendering via setWidget callback form ( green, yellow, red, dim)
  • Live response streamingonTextDelta shows truncated agent response text instead of static "thinking..."

Changed

  • Tool names match Claude Code: Agent, get_subagent_result, steer_subagent
  • Labels use "Agent" / "Agents" (not "Subagent")
  • Widget heading: when active, when only lingering finished agents
  • Extracted all UI code to src/ui/agent-widget.ts

0.2.0 - 2026-03-05

Added

  • Claude Code-style UI renderingrenderCall/renderResult/onUpdate for live streaming progress
    • Live activity descriptions: "searching, reading 3 files…"
    • Token count display: "33.8k token"
    • Per-agent tool use counter
    • Expandable completed results (ctrl+o)
    • Distinct states: running, background, completed, error, aborted
  • Async environment detection — replaced execSync with pi.exec() for non-blocking git/platform detection
  • Status bar integration — running background agent count shown in pi's status bar
  • Fuzzy model selection"haiku", "sonnet" resolve to best matching available model

Changed

  • Tool label changed from "Spawn Agent" to "Agent" (matches Claude Code style)
  • onToolUse callback replaced with richer onToolActivity (includes tool name + start/end)
  • onSessionCreated callback for accessing session stats (token counts)
  • env.ts now requires ExtensionAPI parameter (async pi.exec() instead of execSync)

0.1.0 - 2026-03-05

Initial release.

Added

  • Autonomous sub-agents — spawn specialized agents via tool call, each running in an isolated pi session
  • Built-in agent types — general-purpose, Explore (defaults to haiku), Plan, statusline-setup, claude-code-guide
  • Custom user-defined agents — define agents in .pi/agents/<name>.md with YAML frontmatter + system prompt body
  • Frontmatter configuration — tools, extensions, skills, model, thinking, max_turns, prompt_mode, inherit_context, run_in_background, isolated
  • Graceful max_turns — steer message at limit, 5 grace turns, then hard abort
  • Background executionrun_in_background with completion notifications
  • get_subagent_result tool — check status, wait for completion, verbose conversation output
  • steer_subagent tool — inject steering messages into running agents mid-execution
  • Agent resume — continue a previous agent's session with a new prompt
  • Context inheritance — fork the parent conversation into the sub-agent
  • Model override — per-agent model selection
  • Thinking level — per-agent extended thinking control
  • /agent and /agents commands