@nredd/pi-subagents

@nredd/pi-subagents is maintained in nredd/pi-harness. It was imported with history from nredd/pi-subagents commit 526fa2df76ba85df31bd14a843fae370e8efe78a, tagged v0.22.0-nredd.5, which is rebased on tintinweb/pi-subagents master commit ca93a5247e72e7053fcd10bea7125d6c4205f949. Original authorship and the MIT LICENSE are retained; subsequent local changes are maintained by nredd.

A pi extension that brings Claude Code-style autonomous sub-agents and workflow orchestration to pi. Spawn specialized agents that run in isolated sessions — each with its own tools, system prompt, model, and thinking level. Run them in the background (the default) or block on them, steer them mid-run, resume completed sessions, and define your own custom agent types. When the orchestration shouldn't be improvised, hand a deterministic JavaScript script to the SubagentWorkflow tool — agent(), parallel(), pipeline() — and scripts written for Claude Code's Workflow tool run here unchanged.

https://github.com/user-attachments/assets/8685261b-9338-4fea-8dfe-1c590d5df543

Features

• Claude Code look & feel — same tool names, calling conventions, and UI patterns (Agent, get_subagent_result, steer_subagent) — feels native
• Parallel background agents — spawn multiple agents that run concurrently with automatic queuing (configurable concurrency limit, default 10) and completion notifications that reach the parent at its next boundary, several at once when they finish together
• Live widget UI — persistent above-editor widget with animated spinners, live tool activity, token counts, and colored status icons. Configurable via /agents → Settings → Widget: all (every agent), background (default — hides foreground runs, which already render inline as the Agent tool result), or off
• FleetView — one-line summary below the editor by default; ctrl+shift+a expands the Claude Code-style navigable list of main + every running subagent (earliest-launched first). Press ↓ (or ←) at an empty prompt to jump in, ↑/↓ to move the selection, Enter to open the selected agent's live, auto-updating conversation, Esc to return. Finished agents linger briefly before dropping out, and a viewer stays open through completion so you can read the final output. Toggle via /agents → Settings → Fleet view
• Conversation viewer — select any agent in /agents to open a live-scrolling overlay of its full conversation (auto-follows new content, scroll up to pause). In pi's fullscreen mode it is a borderless, full-terminal observer with mouse support. Steer a running agent inline by pressing Enter to open a composer, typing, then Enter to send (Esc or an empty submit returns) — the message appears as a user message and redirects the agent after its current tool. Press b on a running foreground Agent call to hand that run to the background (see Sending a foreground agent to the background). Stop a still-running agent by pressing x (then x again to confirm) — steer and stop both work for background agents too. Assistant text renders as Markdown; m cycles that between off, assistant-only and everything (see Viewer markdown)
• Custom agent types — define agents in .pi/agents/<name>.md or .agents/agents/<name>.md (project) or globally, with YAML frontmatter: custom system prompts, model selection, thinking levels, tool restrictions, and Claude Code-compatible colored name badges
• Package-shipped agents — a pi package can declare its own subagents (and workflow scripts) with "pi": { "subagents": { "agents": ["./agents"] } } in its package.json, so pi install is the whole setup: no copying files into .pi/agents/ on every machine or container start. Only packages pi has actually installed are read (never node_modules), package definitions rank below every agent you wrote, and /agents shows them read-only with Eject and Disable. Narrow or switch it off with packageAgents / packageWorkflows. Full guide
• Nested subagents — opt-in, default-off delegation: a custom agent that sets allowed_subagents gets its own ownership-scoped Agent, get_subagent_result, and steer_subagent tools, depth-capped from the main session (default 2). It can control only its own children, they are stopped when it finishes, and their transcripts and token spend roll up to it. The allowlist is a privilege boundary — a child runs with its own tools, so pick it as carefully as tools: itself
• Agent mentions — subagents are first-class: type @explore also check the RPC path at the prompt and it goes to that agent instead of the main model, without a word of it entering the chat. One syntax covers the whole lifecycle — message it while it runs, resume it once it has finished, reopen its session from disk long after that, or start it if it never ran. Mentioning an agent that isn't running spawns it through an off-screen clone of the conversation, so it gets Claude Code's context-written prompt and a real Agent tool call without a word of it reaching the chat; direct mode starts it here from your text instead, with no model call at all. The orchestrator can name an agent so you address it as @auth-audit, and handles work in steer_subagent/get_subagent_result too. @ completes live agents, resumable ones, and startable types alongside pi's file completion; @main forces text back to the main model. Toggle via /agents → Settings → Agent mentions
• Scripted workflows — a SubagentWorkflow tool that runs a deterministic JavaScript script orchestrating many subagents: agent(), parallel(), pipeline(), phase(), log() and args, with a pure-literal meta block declaring the phases. pipeline() has no barrier between stages, so one item can be in a later stage while another is still in the first — unlike parallel(), which idles every fast agent until the slowest finishes. Runs in the background with a live card, inspectable via /agents → Workflows or by selecting the run in FleetView. agent() also takes gate: "npm test" to verify a child by running a command (inside its worktree, when isolated) rather than asking another model, and resume: "<label>" to continue a child instead of re-paying its context. Scripts run in a node:vm sandbox on a worker thread where Date.now(), Math.random() and eval throw. On by default, but it stands down for company: if another extension already provides a Workflow or SubagentWorkflow tool, this one warns and disables itself for the session rather than offering the model two orchestrators. Pin it either way with "workflowsEnabled" in subagents.json or /agents → Settings → Workflows. A script written for Claude Code's Workflow tool runs here unchanged: same globals, schema returns a validated object exactly as it does there, budget is present and always reports no token target (pi has no such directive) so its budget.total-guarded patterns still take the branch they were written for, and nested workflow() composes saved workflows one level deep. Full guide
• Mid-run steering — inject messages into running agents to redirect their work without restarting
• Session resume — pick up where an agent left off, preserving full conversation context. Resumes detached by default and notifies you on completion, just like a fresh spawn; pass run_in_background: false to block and get the result inline
• Graceful turn limits — agents get a "wrap up" warning before hard abort, producing clean partial results instead of cut-off output
• Case-insensitive agent types — "explore", "Explore", "EXPLORE" all work. A type that doesn't resolve to exactly one enabled agent — unknown, disabled, or ambiguous between two agents differing only by case — falls back to general-purpose with a note, or is refused outright under fallbackSubagent: none
• Fuzzy model selection — specify models by name ("haiku", "sonnet") instead of full IDs, with automatic filtering to only available/configured models
• Context inheritance — optionally fork the parent conversation into a sub-agent so it knows what's been discussed
• Persistent agent memory — three scopes (project, local, user) with automatic read-only fallback for agents without write tools
• Git worktree isolation — run agents in isolated repo copies; changes auto-committed to branches on completion
• Skill preloading — inject named skills into agent system prompts, discovered from .pi/skills/, .agents/skills/, and global locations (Pi-standard <name>/SKILL.md directory layout supported)
• Tool denylist — block specific tools via disallowed_tools frontmatter
• Styled completion notifications — background agent results render as themed, compact notification boxes (icon, stats, result preview) instead of raw XML. Expandable to show full output. Group completions render each agent individually
• Event bus — lifecycle events (subagents:created, started, completed, failed, steered, compacted) and a per-message spend ledger (subagents:usage) emitted via pi.events, enabling other extensions to react to sub-agent activity
• Cross-extension RPC — other pi extensions can spawn, stop, steer, and join subagents via the pi.events event bus (subagents:rpc:ping, subagents:rpc:spawn, subagents:rpc:stop, subagents:rpc:consume, subagents:rpc:steer). Standardized reply envelopes with protocol versioning. Emits subagents:ready on session start. Full reference
• Schedule subagents — pass schedule to the Agent tool to fire on cron / interval / one-shot. Session-scoped jobs with PID-locked persistence; results arrive as the same completion notification as a manual background agent; manage via /agents → Scheduled jobs
• Model scope enforcement — opt-in validation that subagent model choices stay within your pi enabledModels allowlist (sourced from /scoped-models, with both global and project-local pi settings honored). Caller-supplied out-of-scope → hard error to orchestrator; frontmatter-pinned out-of-scope → warning + runs anyway (frontmatter authoritative). Toggle via /agents → Settings → Scope models

Install

pi install npm:@nredd/pi-subagents

Or load directly for development:

pi -e ./src/index.ts

Requires pi 1.1.0 or newer: pi 0.86 and 0.87 moved the system prompt and conversation history into the session transcript, and the @handle conversation clone and the agent transcript files are built on that model. The peerDependencies range declares it, so npm flags an older pi at install time.

In-process factory, clock and home

src/index.ts exports createSubagentsExtension(options: SubagentsExtensionOptions = {}): ExtensionFactory. The default export is createSubagentsExtension() and keeps native time and timer reference behavior. For an in-process host, pass createSubagentsExtension({ clock }) as an inline extension. No setting or process-global clock patch is needed. The same clock object can be shared with session-kit and its now method with the subscription router:

import { createSubagentsExtension, type Clock } from "@nredd/pi-subagents/src/index.ts";

// Host-provided clock; use () => clock.now() when passing `now` separately.
declare const clock: Clock;
const extension = createSubagentsExtension({ clock });

SubagentsExtensionOptions accepts clock?: Clock and homeDir?: string. An explicit home binds user-scoped discovery, settings and child-session paths without mutating HOME; omitted homes keep call-time host defaults. Explicit agent/session directory overrides still win. The boundary audit and test seams are documented here.

The exported clock contract matches session-kit's scheduling boundary:

interface TimerHandle { cancel(): void; unref?(): void }
interface Clock {
  now(): number;
  setTimeout(callback: () => void, ms: number): TimerHandle;
  setInterval(callback: () => void, ms: number): TimerHandle;
}

Cancellation wraps native clearTimeout / clearInterval; the supplied clock owns whether its timers are referenced or unreferenced. The default clock does not add or remove unref() calls. Clock callbacks must be deferred, not invoked synchronously by setTimeout / setInterval.

The clock controls relative schedule parsing, ISO validation, one-shot/interval dispatch, quota re-parking, the Agent admission timeout, manager retention and lifecycle timestamps, model-call activity/stall ages, UI repaint tickers and the 200ms idle completion hold. Cron remains owned by Croner, including validation and next-run calculation; this seam deliberately leaves its existing behavior untouched. SubagentScheduler also accepts { clock } directly, and CompletionOutbox accepts it as its third constructor argument. Static detectSchedule(s, now?) and parseRelativeTime(s, now?) accept epoch milliseconds; omitted now uses Date.now().

This is not a whole-extension virtual clock: cron, scheduled re-admission's RPC timeout, group-join holds, workflow execution timers, child-shutdown watchdogs, detached-spool filenames, core scrollbar auto-hide and schedule-store lock retries retain native time. Queued result waits use the manager's start gate when present; legacy queued resumes retain the clock-driven fallback.

npm run check runs Biome, TypeScript, the full Vitest suite and test:bun. The four session scenario files (test/{session-scenarios,fleet-session,viewer-session,worktree-isolation-e2e}.test.ts) run under both Vitest and Bun 1.4.4 on core v1.3.0-nredd.8 release tarballs. They use exported createHarness and createInteractiveHarness/VirtualTerminal, real child sessions and git, one injected extension clock, promise-held tools/streams and raw SGR mouse input. Homes, settings and provider credentials are fixtures; no installed pi, tmux or real-clock sleeps are needed. Existing old-host strict-mode coverage explicitly removes the optional SDK export; the shared faux runtime uses the host's real withBeforeProviderRequest facade. The full older suite remains Vitest-only. Maintainer guidance on public index callback coverage and retained safety guards is here.

Other hosts

This extension is developed and tested against pi.

Third-party adapters report running it elsewhere. These are maintained independently of this project: not tested here, not covered by our CI, and compatibility may break with any release.

• DeepSeek Harness (dsh) — via an adapter that maps pi's host API onto native DSH agents. Details and reports: #258

Quick Start

The parent agent spawns sub-agents using the Agent tool:

Agent({
  subagent_type: "Explore",
  prompt: "Find all files that handle authentication",
  description: "Find auth files",
  run_in_background: true,
})

Agents run in the background by default: the call returns an ID immediately and notifies you on completion, carrying a preview of the result (use get_subagent_result for the full text). Pass run_in_background: false to block until the agent finishes and get its full output inline; while it blocks, ctrl+b hands it to the background (see below).

Sending a foreground agent to the background

Press ctrl+b while a foreground Agent call is running to hand it to the background, as in Claude Code. The same run keeps going — same session, in-flight turn, transcript and worktree — and the blocked call returns the agent ID with the same result a run_in_background: true spawn gets. Completion then arrives as the normal background notification, naming the original tool call. After 2 seconds the running call's row shows (ctrl+b to run in background).

• ctrl+b hands off every running foreground Agent call at once. To hand off just one, press b in its conversation viewer (FleetView or /agents → Running agents).
• The key is claimed only while such a call is running and the prompt editor has focus; the rest of the time — and in dialogs or the viewer's steering composer — ctrl+b is still pi's cursor-left. With FleetView disabled the extension cannot see which component has focus, so it claims the key whenever a call is running.
• One hint shows per batch of parallel calls, on the earliest one still running.
• Esc in the parent no longer reaches a handed-off agent. Stop it from /agents → Running agents with x.
• It is one-way: there is no way back to foreground.
• Other tool calls in the same assistant message still hold the model's next turn until they finish; pi runs a message's tool calls as one batch.
• Not offered for a foreground call still queued for a slot, a foreground resume, a workflow's agents or nested children. There is no automatic timeout.

Scheduling

Add a schedule field to register the agent to fire later instead of running now:

Agent({
  subagent_type: "Explore",
  prompt: "Look at recent commits and summarize what changed since last week",
  description: "Weekly commit review",
  schedule: "0 0 9 * * 1",   // 9am every Monday (6-field cron)
})

Schedule formats:

• Cron — 6-field (second minute hour day-of-month month day-of-week), e.g. "0 0 9 * * 1" for 9am every Monday, "0 */15 * * * *" for every 15 minutes.
• Interval — "5m", "1h", "30s", "2d". Fires repeatedly at that interval.
• One-shot relative — "+10m", "+2h", "+1d". Fires once at that future time.
• One-shot absolute — full ISO timestamp, e.g. "2026-12-25T09:00:00.000Z".

When a schedule fires, the spawn runs in background and its completion notification arrives the same way as a manually-spawned background agent's — your parent agent reasons about the result the same way. Model selection runs at fire time with the manual-dispatch precedence: stored model request (including a captured configured tier), current agentModels[type], agentModels["*"], definition, parent. A stored explicit model that no longer resolves or is out of scope fails the job instead of being replaced. Tier jobs retain the @tier request, including a configured tier, and re-resolve its binding at fire time. Other quota-parked jobs retain the original explicit input, not a snapshot of the configured or parent model. The job stores the resolved thinking level when you create it: an explicit Agent({ thinking }) wins over the agent definition then, while an omitted value keeps the definition's default. Editing the agent file later does not replace that stored level.

Session switches. /new, /resume, /fork and /reload do not stop running background agents. They keep showing in the fleet list of the reloaded session while they run (open, stop and steer work from there), finish on their own, and the session that spawned them gets the completion notification when you return to it (immediately if it is already open). Held notices survive the switch too: unread results are transferred to the session spool and re-enter the same boundary outbox on return, without repeating their history entries. The preview is in the notification; the full output is in the transcript file it names, since get_subagent_result cannot reach an agent from a previous session. Quitting pi still aborts everything. A step that routes through the old session's virtual model (router/*) while the switch is in flight is re-prompted up to three times with backoff; if it never recovers the agent ends error with interrupted by a session reload and its last tool calls and output as the result, not No output. Usage and compaction observers bound to the old session are dropped rather than failing the run (any other observer error is logged once and also kept from failing it), and the structured-output retry and forced final report are not attempted on a runtime that is still stale, so subagents:usage and subagents:compacted stop for an agent detached by a switch.

Schedules are session-scoped: they reset on /new and restore on /resume. List and cancel via /agents → Scheduled jobs (creation is the Agent tool's job — there is no parallel manual-create wizard). Storage at <cwd>/.pi/subagent-schedules/<sessionId>.json with PID-based file locking for cross-instance safety.

Disable the feature entirely: /agents → Settings → Scheduling → disabled removes schedule from the Agent tool spec (no LLM-context cost), hides the menu entry, and stops any active scheduler. The schema-level removal takes effect on the next pi session; the runtime kill is immediate. Re-enable from the same menu.

Restrictions:

• schedule cannot be combined with inherit_context (no parent conversation exists at fire time) or resume (schedules create fresh agents).
• run_in_background: false is refused — scheduled jobs always run in the background. Omitting it, or passing true, is fine.
• Scheduled fires bypass the maxConcurrent queue so a 5-minute interval cannot be deferred behind long-running manual agents.
• Headless pi -p doesn't wait for scheduled subagents.

Quota admission

With a subscription router installed (e.g. pi-subscription-router), a fresh Agent dispatch first asks it whether the model is out of included quota, over pi.events: router:rpc:admit { requestId, model: "provider/id" }, replied on router:rpc:admit:reply:<requestId> as { success, data: { ok, resolvedModel?, resetAt?, wait? } } (resolvedModel labels the admitted model for tiers; wait: false is the router's exhaustion policy saying fail now, absent means wait).

• Admitted, or no router answers within 100ms: the agent starts as usual
• Refused, background, reset known, wait not false, scheduling on: the dispatch is parked as a one-shot scheduled job a minute after the reset, persisted like any schedule. When it fires it asks again and re-parks at the new reset while still blocked; if the refusal can no longer park (wait: false, no reset, or a reset already past) the job is not started and ends with lastStatus: "error" and an error event. Only an explicit admission, or no router answering, starts it. If pi was closed through the reset, it asks again 10s after the session starts, once the router has warmed its usage cache
• Refused otherwise (foreground, inherit_context, scheduling off, unknown reset, or wait: false): the call fails with the reset time and no agent starts

Live resumes keep their existing selection. Tier requests are admitted on every fresh entry path, including RPC, mentions, nested and workflow children, and every scheduled fire. Only top-level Agent background calls and scheduled fires can persist a quota wait; other refused starts error. A router/* virtual model still routes and fails over inside the child.

UI

The extension renders a persistent widget above the editor showing active agents. By default it shows background runs only (widgetMode: background) — foreground agents already render inline as the Agent tool result, so the widget would otherwise double-render them. Switch to all (every agent) or off (hide the widget) via /agents → Settings → Widget:

● Agents
├─ ⠹ Agent  Refactor auth module · ↻5≤30 · 5 tool uses · 33.8k token (62%) · 12.3s
│    ⎿  editing 2 files…
├─ ⠹ Explore  Find auth files · ↻3 · 3 tool uses · 12.4k token (8%) · 4.1s
│    ⎿  searching…
├─ ⠹ Agent  Long-running task · ↻42 · 38 tool uses · 91.0k token (84% · ⇊2) · 2m17s
│    ⎿  reading…
└─ 2 queued

The token field is annotated with two optional signals inside parens:

• NN% — context-window utilization (color-coded: <70% dim, 70–85% warning, ≥85% error). Omitted when the model has no declared contextWindow, or briefly right after compaction.
• ⇊N — number of times the session has compacted, when > 0. Stays dim; the percent's color carries urgency.

The Agent tool row

pi 1.0 folds every collapsed tool row to <first line of the call> · <first line of the result> and draws the ▸/▾ marker itself, so the Agent row's call and result each put what matters on their first line and draw no marker:

Plan  design the cache · ⠹ thinking · 2m41s · 12s · 3.2k token      collapsed, running
Plan  design the cache · ✓ done · 41s · 33.8k token                 collapsed, finished

The status, elapsed time and tokens are read from the agent record on every render (and the host is asked to repaint once a second while it runs), so the row keeps moving after a background launch instead of freezing on Running in background. Expanded, the row shows the prompt, the model and thinking tags (↳ opus 4.6 · thinking: high), the transcript link, then the result.

FleetView

FleetView starts collapsed, rendering exactly one line below the editor, with no overflow rows or blank spacer. Always-collapsed is deliberate: it saves space even for one agent and avoids resizing when a fan-out crosses a threshold. ctrl+shift+a toggles it; the settled explicit choice is recorded as a versioned subagents:fleet-state custom entry, outside model context.

Auto-repeat events are ignored. A 300ms quiet period coalesces rapid toggles (net-no-change bursts write nothing); shutdown flushes the final choice before the activation loses its context. This bounds writes per burst, not all-time size of the host's append-only journal. The last valid choice on the active branch survives ticks, widget resets, /reload, and restarting/resuming that session. A fresh session defaults to collapsed. Temporary arrow navigation writes no state entry. This uses the fleet's existing focus-aware terminal-input handler, like its arrows and the foreground ctrl+b handoff; there is no subagent key-rebinding setting. Todo's ctrl+shift+t, monitors' ctrl+shift+m, and pi's ctrl+o keep their existing roles.

agents: 3 running, 2 queued, 4 done · ~$28.70 · ctrl+shift+a to expand
[1 failed, 2 parked, 1 fallback] agents: 3 running, 1 failed, 2 parked · ctrl+shift+a to expand

Counts cover top-level records, including running starts and queued/waiting agents without a session, plus workflow containers (not their children). steered counts as done; errors and aborted runs count as failed; stopped runs stay distinct. Successful/stopped records retain the existing four-second linger.

Unacknowledged errors and aborted runs do not expire from the fleet: the UI journals minimal failure identities and acknowledgements, replayed from the active session branch on reload or disk resume. Terminal callbacks and detached-result ingress record failures explicitly, including failures between render ticks. It retains at most 128 detailed alerts (id, agent type, short sanitized description, terminal state and timestamp), never complete records/results/sessions; four durable per-kind overflow counters preserve exact outstanding totals beyond that cap. Overflow contributes to both collapsed and expanded headers, but has no individual navigation rows. Specific overflow ids can still be acknowledged through result consumption. The existing session journal is the identity authority: overflow deduplication/acknowledgement scans it backwards in O(branch length) time and O(1) extra space; checkpoints restore bounded state without replaying all failures. No additional file/global registry or unbounded UI id/ack set is kept. Live record and workflow-owner observation uses weak keys, not retained records; recreated workflow display projections do not rescan acknowledged/overflow identities every tick. Each failure/ack transition writes one bounded checkpoint (up to 128 minimal details plus counters), trading journal bytes for bounded checkpoint replay; no all-time compaction is claimed. Old failures cannot leave a healthy-looking header or hide the widget.

Acknowledged means the existing resultConsumed flag was set by inline result delivery, get_subagent_result, RPC consume, or successfully closing the failed agent's FleetView conversation. A session-less/restored failure is acknowledged by selecting it to read its error or minimal-metadata toast. Restored workflow failures without a live inspector use the same toast acknowledgement. Merely delivering a completion notice, ticking, or a viewer-open failure does not acknowledge it. Failed workflow containers stay until their inspector closes successfully; paused runs stay blocked until resumed/stopped.

Enabled quota-parked schedules awaiting dispatch count as parked even with no agent record, and clear when dispatched, disabled or removed.

Failed/parked/paused-workflow/fallback counts lead in a shared, uncut flag; at narrow widths these compact to [E1 P2 B1 F1] (failed, parked, blocked/paused, fallback). Compact counts above 99 read 99+; spaces yield before alerts, so all four still fit at 20 columns ([E99+P99+B99+F99+]). Below 20 columns alerts yield in reverse priority (fallback, paused, parked), never ahead of the failure flag; if even its count cannot fit, the marker is !, never the misleading [fb]. Cost uses record-borne estimates only when showCost is on; it is the sum for the agents currently summarized, not a session ledger, and excludes workflow-container spend. Narrow lines drop trailing detail and the hint before leading alerts. An empty fleet still hides entirely; the standalone summary renderer has a one-line zero-count form.

At an empty prompt, ↓ or ← temporarily reveals the list and starts navigation. Esc, selecting main, or typing returns to the explicit disclosure choice; dialogs and conversation overlays keep their keys. ctrl+shift+a while navigating collapses and returns to the editor. The separate above-editor activity widget (widgetMode) and transcript/result/notice renderers are unchanged.

Expanded, the existing navigable list renders below the editor:

  esc to interrupt · ← for agents · ↓ to manage · ctrl+shift+a to collapse

  ● main
  ○ workflow         audit-src                    12/40 agents · 32s · ↓ 26.4k tokens
  ○ general-purpose  Sleep then report 1          ▶ running · 11s · ↓ 13.1k tokens
  ○ general-purpose  Sleep then report 2          ▶ running · 11s · ↓ 13.1k tokens
                                          ↓ 3 more: 2 running, 1 failed

Running workflows appear as a single workflow row above the agents, carrying their agent counts in place of a description. Enter on one opens the same two-pane inspector /agents → Workflows does, rather than a conversation overlay. A run's own agents are not listed separately — they belong to the run, which reports for them, so they are filtered out of the fleet list, the above-editor widget, the /agents menus and @handle resolution exactly as nested children are. They are also outside the maxConcurrent pool: the run has its own concurrency cap, and routing a fan-out through the session pool as well would let one workflow starve everything else. The agents are ordered earliest-launched first, and only agents you can actually open, or outstanding failures you can inspect, are shown (pending/queued agents with no session yet appear once they start). At an empty prompt, press ↓ (or ←) to move focus from the prompt into the list — the selected row is marked ●, the rest ○. The row block stays at five lines once the roster reaches five rows; overflowing lists show three rows plus ↑ N more: 2 running / ↓ N more: 1 queued, 1 failed indicators that count the hidden rows by state (blank at an edge), without moving the content below it. Each agent row carries its state as a glyph and a word in a fixed first column (▶ running, ◦ queued, ✓ done, ✓ limit, ✗ failed, ✗ aborted, ■ stopped), so colour is never the only cue; a spawner's name leads the description (pr-1412: Spec review ...), the configured thinking level reads think:high so it cannot be mistaken for the live thinking · 4s activity, and columns are sized from every row, not the visible window. A configured-model fallback keeps its shared leading [fallback] flag even with showModel off; column fitting reserves space for flags on hidden rows too, so scrolling cannot shift the tail. On a narrow terminal the description truncates first, then the activity, thinking and model columns drop (in that order). If a fallback row still cannot fit its stats, the shared leading flag compacts to [fb] or !, never a sliced fragment. For every row, after optional columns and any flag compaction, state words shrink to glyphs, separators and column padding collapse to spaces, and navigation padding disappear as needed. Elapsed, tokens and cost stay whole whenever they physically fit beside the marker (if present); if space is still insufficient, the state glyph and then elapsed yield before tokens or cost may clip. The selected row renders in the theme's primary text color rather than the muted/dim treatment of the others; an agent with a configured color shows its badge there too, bolded. ↑/↓ move the selection, Enter opens the selected agent's live conversation overlay (it auto-updates as the agent works), and Esc (or ↑ above main) returns to the prompt. Selecting main returns to the normal view. Inside the overlay, press Enter to steer the running agent — type a message and Enter to send it (Esc or an empty submit returns), and it redirects the agent the same way the steer_subagent tool does. On a running foreground Agent call, press b to hand that run to the background; the hint only shows where the handoff is available. A viewer stays open when its agent finishes so you can read the final output, and finished agents linger in the list for a few seconds before dropping out. Typing anything at a non-empty prompt behaves normally — the list only captures arrow keys when the prompt is empty. Disable it entirely via /agents → Settings → Fleet view. With showModel on, each row names the model and thinking level it runs at (▶ running · sonnet 4.6 · think:high · 11s · ↓ 13.1k tokens; the above-editor widget spells the level thinking: high), and a running row says what its in-flight model call is doing (see Live activity). On a virtual model such as router/default the row reads router default (routing) until the first response, then the physical model that answered. The thinking level shown is the session's selected level; a router chain entry's own thinking override is not visible to it.

Fullscreen conversation observer

In fullscreen mode (pi's default), opening an agent from FleetView, /agents, or the workflow inspector fills the terminal. Regular mode keeps the centered floating viewer and keyboard controls. Neither mode replaces the main session or pauses either agent. Escape returns to the previous view; while composing, the first Escape cancels the draft.

In pi's /settings → TUI mode → fullscreen, the observer supports:

• Mouse-wheel scrolling. Scrolling up pauses automatic following; reaching the bottom resumes it.
• A single-column scrollbar using pi's │ track, ┃ thumb, and theme colors. The thumb becomes █ while hovered or dragged. Track clicks page; holding and moving also drags.
• pi's /settings → Fullscreen scrollbar preference applies to the observer: auto appears on scrolling or hovering and hides after one second of inactivity; always reserves a column and stays visible even when the conversation fits; hidden removes the rail and its mouse target. The observer reads the main session's settings when opened, without changing them.
• Clickable footer controls for steering, stopping, Markdown mode, and closing. Stopping still takes two presses, as with x.
• Keyboard navigation and steering as before. Mouse scrolling also works while composing a steering message.
• The bottom divider stays pinned above the footer (and steering composer when open). While scrolled away, a centered [ ↓ N new messages · Ctrl+End ] indicator within the divider counts new conversation messages, not streaming updates. Before any arrive, it reads [ ↓ Jump to latest message · Ctrl+End ]. Click it or press Ctrl+End to resume following and clear the count, without losing an open steering draft.

Observer mouse events are consumed, including at scroll limits, so they do not scroll, select, or click the main view underneath. Closing restores focus without editing the main prompt. The main conversation can still advance as its agent runs.

This is an initial implementation. Mouse text selection, copy-on-select, link activation, and switching agents without leaving the observer are not implemented. pi's global shortcuts remain host-owned: its fullscreen search searches the main transcript, not the observer. The scrollbar, jump indicator, and mouse controls apply only to fullscreen mode; regular mode retains the original viewer.

Live activity

While a model call is in flight, the widget, the fleet row and the conversation viewer say what it is doing instead of thinking… for as long as it takes:

Tools running take precedence (reading…, running command…). Activity is tracked for spawns made through the Agent tool, RPC and mentions, and for foreground and background resumes. Each resume gets fresh text/model-call tracking and turn counts; physical-model labels follow the assistant messages.

Conversation viewer input

The fullscreen observer uses pi's wheel delta, draggable scrollbar and clickable footer controls (see above). Regular mode keeps the centered floating box and keyboard controls. Both stream the reply being written before it lands in the transcript, and repaint once a second while the agent runs so activity timers count.

Agent mentions

Subagents are addressable. Every agent has a typeable handle — the agent type, lowercased, numbered when instances collide (explore, explore-2) — and @handle <message> at the prompt talks to that agent, whatever state it happens to be in. Type @ to pick one:

❯ @
  @auth-audit     send message · Explore · running · audit the auth flow
  @explore-2      send message · running · find flaky tests
  @code-review    resume · code-review · check the diff
  @plan           start agent · Software architect agent for designing implementation plans.
  index.ts        src/index.ts                        ← pi's own file rows, still there
  index.d.ts      dist/index.d.ts

The handle names the agent, not one process, so a single syntax covers its whole lifecycle:

No turn is ever spent in the main conversation, and nothing about the mention enters the chat. The answer comes back as the ordinary background-completion notification, so the main model can relay it.

Starting a new agent

Claude Code does not start a mentioned agent itself. @agent-<type> becomes an attachment appending a <system-reminder> to your prompt — "the user has expressed a desire to invoke the agent X; please invoke the agent appropriately, passing in the required context to it" — and the main model makes the tool call. There is no tool forcing and no allowed-tools narrowing: the mention constrains which agent, not what it is told. So the model writes the agent's prompt, giving it the conversation context a cold spawn lacks.

The cost is a visible turn — the model's reasoning and its tool block, narrating a decision you already made by typing the handle. This extension keeps the mechanism and moves it off-screen. The conversation is copied into a throwaway in-memory session, that clone takes the turn holding only the Agent tool, and what it starts is an ordinary top-level agent:

@cyan whats your favorite color        →  (nothing in the chat)
  └─ clone of this conversation, off-screen
       └─ Agent(subagent_type: "cyan", prompt: …)
            ▸ Cyan Agent   favorite color        ← widget, fleet row, handle

It is a literal clone — the session's own entries and the same system prompt, not inherit_context's text rendering of them — taken from memory and compaction-aware, so what the copy reads is what the main model is working from. The clone gets one tool and one job; it cannot read, write or run anything, because an invisible turn with the full toolset could do invisible work. The agent it starts is attributed to the real session, so its transcript and rootSessionId land where they would have anyway, and it carries no tool-use-id — the main conversation never issued one.

Either way the started agent honours its own frontmatter — model:, thinking:, max_turns: all apply, since neither path passes them and the agent's config wins. Mentioning something as the very first thing in a session works: there is simply no history to carry, and the clone still runs on your model and system prompt. If it cannot deliver at all — a model can always answer in prose instead of calling the tool — the agent is started directly with your text and the toast says so, rather than leaving you with nothing running.

model is also the only mode that works outside the TUI: pi -p '@plan the migration' clones, spawns, and reports through the normal completion path, where a direct start would have detached the agent and printed nothing. Messaging and resuming stay TUI-only for that reason, in both modes.

Two things to weigh against direct: the clone re-sends the whole conversation, and the agent does not start until that turn finishes.

Named agents. The Agent tool takes an optional name, so the orchestrator can call one auth-audit instead of leaving you to tell @explore-2 from @explore-3. A name is additive: the type-derived handle is still assigned, so @explore keeps reaching that agent rather than starting a second one beside it. Both names share one namespace — an alias can never shadow a live handle or the reverse — and the popup shows one row per agent, under its alias, with the type moved into the description. steer_subagent and get_subagent_result accept a handle too, so you and the model address agents the same way.

Resuming much later. Because subagent sessions are persisted by default (rememberAgents), a handle keeps working after the agent's in-memory record is evicted: @explore anything else? reopens the conversation from disk. Only the definition is re-resolved, so a continuation runs under the agent type's current frontmatter, not the one the first run used. If the type has since been deleted or disabled, the resume is refused rather than falling back to another agent — re-enable it and the handle works again. Names from an evicted agent stay reserved, so a later Explore becomes explore-2 rather than shadowing something you can still reach; the 100 most recent are kept, and all of them are forgotten on /new and session switch. A resumed agent takes those names back, so @explore keeps meaning the same conversation. An agent whose session was only ever in memory leaves nothing to reopen, and the mention starts a fresh one instead; if the session file has since been deleted, the mention says so and frees the handle rather than silently sending your message to a new agent.

The grammar mirrors Claude Code's, and is deliberately narrow so nothing gets swallowed by accident:

While an agent is live its handle addresses it, so @explore never starts a second Explore alongside a running one — use the Agent tool for deliberate parallelism. @<agent-id> works too. main is reserved and can never be an agent's handle (a type slugging to it gets main-2); handles are capped at 64 characters. A handle written as typed always wins over the @agent- form, so an agent genuinely called agent-explore stays reachable. Nested subagents are not addressable — they are hidden from every top-level surface and only their owner may steer them, so a handle that would name one starts a fresh top-level agent instead of reaching through that boundary. Suggestions list live agents first, then resumable ones, then startable types — and then pi's own file rows, in the same popup: @ stays the file picker it always was, and the handles are added to it rather than replacing it. Disable the whole thing via /agents → Settings → Agent mentions.

A direct-mode start takes the non-tool spawn path shared with the scheduler and cross-extension RPC, so — like those — it writes no .output transcript. That is the trade for skipping the model call: a model-mode start goes through the real Agent tool and keeps everything. Live tool activity and the turn counter are not part of that trade — a direct start renders them like any other agent. A mention-resumed agent goes through the full resume wiring and keeps both in either mode.

Individual agent results render Claude Code-style in the conversation:

Completed results can be expanded (ctrl+o in pi) to show the full agent output inline.

By default, foreground and background agents each stream their full conversation to a per-subagent transcript — a JSON-lines file at <os-tmpdir>/pi-subagents-<uid>/<cwd>/<session>/tasks/<agent-id>.output (owner-only 0700, cleared on reboot). Set output_transcript: false on a custom agent to write no transcript path or file for it, or set outputTranscript: false in subagents.json to make transcripts opt-in for the whole project (frontmatter overrides the project default). This governs only the transcript: it is independent of persist_session (the pi session on disk), and it does not affect isolation: worktree (which commits the agent's work to a git branch) or memory: (durable files) — set those accordingly if the goal is to keep a run off disk entirely. Background agent completion notifications collapse to one line, with pi 1.0 drawing the ▸/▾ marker and toggling on click (the renderers draw none of their own):

✓ Find auth files · 3 tools · 12.4k token · ~$0.0042 · 4.1s          collapsed
3 agents finished · ✓ a, ✓ b, ✗ c                                   collapsed, grouped

Expanded, a notice adds ↻3 · 3 tools · 12.4k token · 4.1s, the result preview, and the transcript as an OSC 8 hyperlink labelled transcript ↗ (the path follows it as plain text). A group expands to one block per agent, plus a total when showCost is on. The LLM receives structured <task-notification> XML for parsing, while the user sees the themed visual.

Default Agent Types

The general-purpose agent is a parent twin — it receives the parent's entire system prompt plus a sub-agent context bridge, so it follows the same rules the parent does. Explore and Plan use standalone prompts tailored to their read-only roles.

Default agents can be ejected (/agents → select agent → Eject) to export them as .md files for customization, overridden by creating a .md file with the same name (e.g. .pi/agents/general-purpose.md), or disabled per-project with enabled: false frontmatter.

Custom Agents

Define custom agent types by creating .md files. The frontmatter name: is the subagent_type and dispatch identity, falling back to the filename when absent; display_name only changes the UI label. Claiming a default agent's name overrides it.

Agents are discovered from four locations (higher priority wins):

Project agents need project trust. .pi/agents/ and .agents/agents/ are project-supplied, so they load only when pi trusts the project (/trust, a saved decision, defaultProjectTrust: "always", or --approve). In an untrusted project (including --no-approve) both roots are skipped unread: they cannot override or disable a global or package agent, and a malformed file there neither warns nor aborts a strictAgentFiles startup. Global agents and packages configured globally stay available; packages configured only by the project's .pi/settings.json are gated by the same decision. The decision is pi's canonical one, ctx.isProjectTrusted() at session_start (so --approve/--no-approve and /trust override a saved trust.json), which this extension never reads itself. Activation runs before that answer exists, so it loads global and globally configured package agents only; the first session_start loads the project roots and re-registers the Agent tool so its description and subagent_type list match, before any provider turn. This gates reading project agent definitions (and a project agent-tool-description.md); the extension still reads operational project settings such as .pi/subagents.json before trust is known. A nested subagent applies its parent's decision, including after a spawn with a custom cwd: the decision for the parent's config root travels with that root through every later generation, and never extends to the custom working directory. The /agents menu writes new project agents to .pi/agents/, which an untrusted project does not list. Trusted projects keep the order above.

Project-level agents override global ones with the same name, so you can customize a global agent for a specific project. If both project locations define the same name, .pi/agents/ wins — .pi stays the project authority; .agents/agents/ is an additional read location for projects that keep their agent assets in the .agents workspace. The global location follows the upstream PI_CODING_AGENT_DIR env var — set it to relocate all pi-coding-agent state (agents, skills, settings) to a custom directory. An agent's name is its frontmatter name:, falling back to the filename, so two files can now claim the same one — the later load wins, and the warning below names the file that took over.

An unreadable or unparseable agent file is skipped, not fatal — a warning names the file and the error. If it was overriding a same-named agent, a second line names the file that loads instead. Set strictAgentFiles: true in subagents.json (or /agents → Settings → Strict agent files) to fail startup on a broken file instead; mid-session reloads still only warn. Strict validation runs in two phases. At activation, a malformed global agent file fails the factory as an ordinary load error; a malformed file from a package (even one configured globally) only warns, since it is a third party's file and not the user's to fix. At the first canonical session_start, once trust is known, the trusted project roots are loaded and validated, and a malformed file there throws the host's ExtensionStartupError (an untrusted project's files are never read, so they never trip it). The host class is needed for the second phase: on an older host with strictAgentFiles: true the extension refuses at activation with an upgrade message, before any agent file is read, rather than degrading to warnings. Non-strict use works on any host in the peer range.

Example: .pi/agents/auditor.md

---
color: red
description: Security Code Reviewer
tools: read, grep, find, bash
model: anthropic/claude-opus-4-6
thinking: high
max_turns: 30
---

You are a security auditor. Review code for vulnerabilities including:
- Injection flaws (SQL, command, XSS)
- Authentication and authorization issues
- Sensitive data exposure
- Insecure configurations

Report findings with file paths, line numbers, severity, and remediation advice.

Then spawn it like any built-in type:

Agent({ subagent_type: "auditor", prompt: "Review the auth module", description: "Security audit" })

Frontmatter Fields

All fields are optional — sensible defaults for everything.

Frontmatter is authoritative for max_turns, inherit_context, run_in_background, isolated, and isolation. An agent file's model and thinking are portable defaults: an explicit Agent({ model, thinking }) selection overrides either for that run. With no explicit thinking level, the definition wins; if it has none either, no level is passed and pi picks one at session creation (the per-model setting, else defaultThinkingLevel, clamped to the model), which does not follow the parent's live level. Only the @mention clone path starts from the parent's ctx.thinkingLevel. Model selection additionally follows the agentModels precedence. Other Agent tool parameters only fill fields the agent config leaves unspecified. Other spawn paths differ per field:

• isolated holds on every spawn path, including @handle mentions, workflows and cross-extension RPC.
• isolation holds on every path except workflows: as in Claude Code, only the script's agent() option decides there.
• inherit_context applies only to Agent tool and nested spawns, including default-mode @handle mentions, which spawn through the Agent tool. As in Claude Code, where only a fork inherits the conversation, every other spawn starts fresh unless its caller asks: workflow children and scheduled runs never inherit, and an RPC caller opts in with inheritContext.

Explicit non-tier tool inputs can use fuzzy names. These pick the newest model; configured pins and tier/alias members remain exact. A bare family name such as "opus" matches several ids; exact ids and aliases still win first, otherwise the candidates are narrowed to pi's enabledModels (when any candidate is enabled), then ranked by parsed version, newest first, then by match score. Versions are read from claude-<family>-<n>[-<n>] (claude-opus-4-6, claude-opus-5; a trailing -YYYYMMDD is a date, not a minor) and gpt-<n>[.<n>]-<variant> (gpt-6.1-sol); ids with no recognised scheme rank after every versioned one. Previously the shortest id won, so "opus" resolved to claude-opus-5 even with claude-opus-4-6 enabled and claude-opus-5 not.

Exact configuration, convenient tool inputs. Agent definitions, agentModels, tier members and legacy alias candidates resolve exactly, without version ranking or provider retry. An unavailable configured pin inherits the parent with a visible note by default; modelResolution.onUnavailable: "error" refuses it instead. Explicit literal/fuzzy tool inputs retain fuzzy matching and unresolved names remain errors; @tier and legacy aliases follow modelResolution. Built-in Explore uses @cheap; bind it globally or in the enterprise manifest. Unbound tiers inherit with disclosure by library default; personal configurations should set both resolution policies to error. During enterprise migration only, built-in Explore's unbound @cheap can use the existing modelAliases["claude-haiku-4-5"] exact candidates as a static cheap tier. An explicit cheap binding always wins, including when unavailable; no neighbouring tier or newer model is inferred. User-pinned legacy Explore definitions retain the exact anthropic/claude-haiku-4-5 compatibility binding. Users who pin Explore explicitly through agentModels (e.g. the owner's agentModels.Explore) are unaffected. The router's shipped examples/subscription-router.json supplies an opt-in ordered cheap profile (Haiku then Luna); no profiles are registered implicitly.

Nested subagents

Nested delegation is default-off. Set allowed_subagents only on a non-isolated custom agent that owns a real fan-out responsibility:

---
tools: read, grep, find
extensions: false
allowed_subagents: support-file-finder, support-callsite-tracer   # or `all`
---

The allowlist is a privilege boundary, not just a routing hint. A child runs with its own tools:, extensions:, and isolated: — the parent's restrictions are not inherited — so delegation grants the parent the union of what the listed agents can do. The read-only agent above can write and run commands through any listed agent that can, and all reaches every enabled agent including general-purpose. Choose the list as carefully as you would choose tools: itself; that is the main reason this is default-off.

allowed_subagents is runtime-enforced. A comma-separated list restricts nesting to those types; all (or "*" / true, matching how extensions: and skills: take booleans) allows any enabled agent; omitted, empty, none, or false means no nested tools are injected at all. Unknown, disabled, and out-of-list types are rejected rather than falling back — regardless of the project's fallback agent setting, so a configured fallback can never hand a nested caller an agent outside its allowlist — and a nested model: is validated against Model Scope exactly like a top-level spawn. Result, resume, and steering operations are ownership-scoped, so a parent can control only its own children. Nested records remain internal to that parent and do not appear in top-level tools, lifecycle events, or agent UI — so when a parent finishes, is stopped, or ends a resumed turn, its nested children are stopped with it. They do write their own .output transcript (subject to the same output_transcript gate), filed under the root session's directory alongside their ancestors', so a nested run can still be inspected after the fact. Their token usage is folded into every ancestor's totals up to the top-level agent (lifecycle events, completion notifications, /agents), so nested spend stays attributable at any depth even though the children themselves stay hidden. A nested result that ends stopped, aborted, or steered is labelled as partial, the same guarantee top-level results carry.

No background work inside a subagent. A nested Agent call always blocks and returns the child's result inline. run_in_background: true returns an error (a frontmatter run_in_background: true on the nested type is ignored and the child runs in the foreground), and the workflow / Workflow / SubagentWorkflow tools, including another extension's, return an error explaining why when a subagent calls them. A child gets a refusing SubagentWorkflow stand-in for this (one short tool definition), since the real tool is excluded and core would otherwise answer "Tool SubagentWorkflow not found" before the refusal could run; an isolated or extensions: false child gets no workflow tool at all. Both would deliver their completion to a session that ends the moment the subagent returns its answer, and a detached child is aborted with its owner anyway.

The hard cap is depth 2 by default: main session (0) → subagent (1) → nested child (2). Change it project-wide with maxSubagentDepth in subagents.json (or /agents → Settings → Nested depth); 0 or 1 turns nesting off everywhere. An agent already at the cap gets no nested tools at all — not even get_subagent_result, since it can never own a child. A child must independently set allowed_subagents to delegate again; isolated agents never receive nested tools.

Nested children occupy no slot in either per-session pool — their parent already holds one, and queueing them behind it would deadlock a parent waiting on its own child. They do take a lease from the process budget when it is on, which reserves slots for them so that cannot deadlock. The depth cap bounds how deep nesting goes, not how wide: a parent's only limit on concurrent children is that each spawn costs it a turn. Pair allowed_subagents with a max_turns on that agent if you want a hard ceiling on its fan-out.

Depth-2 review delegation. The default depth already allows main session (0) → implementation agent (1) → reviewer (2). Give the implementation agent an explicit allowlist naming only the reviewer, and leave the reviewer without allowed_subagents, so it is a leaf:

---
# ~/.pi/agent/agents/implementer.md
allowed_subagents: code-reviewer
---

The implementer then calls Agent({ subagent_type: "code-reviewer", ... }) in the foreground and gets the review inline. With the process budget on, its slot stays held while the reviewer runs, and the reserve keeps reviewers startable when every top-level slot is an implementer waiting on one.

Because a subagent session never activates this extension (that is what keeps a child from building a second agent manager, and it is why nested tools are injected directly instead), a subagent also gets none of the extension's other surfaces: no /agents command, no cross-extension RPC handlers, no subagents:ready event.

Tool & extension scoping

extensions: decides which extensions load, tools: decides which tools surface to the LLM. They compose:

# Default (both omitted): all extensions load, all 7 built-ins surface

tools: read, grep, find           # narrow to listed built-ins; extensions still load
tools: "*"                        # all 7 built-ins (alias: `all`)
tools: none                       # zero built-ins (alias: `""`)
tools: "*, ext:mcp/search"        # built-ins plus one extension tool

extensions: false                 # no extensions load
extensions: [mcp]                 # only mcp loads
extensions: [builtin:mcp, builtin:codemode]  # pi's built-in MCP + codemode
extensions: ["*", "/abs/foo.ts"]  # all defaults plus one path-loaded extension

exclude_extensions: pi-notify     # everything except pi-notify (with extensions: true)
exclude_extensions: builtin:mcp   # everything except pi's built-in MCP

# Specialist: load one extension, expose only one of its tools, keep built-ins
extensions: [mcp]
tools: "*, ext:mcp/search"

isolated: true                    # hermetic: built-ins only, no extensions/skills/context

A few rules the examples don't make obvious:

• extensions: is the sole loading authority. ext:foo in tools: narrows what surfaces; it can't load foo on its own. Mismatches fire extension-error:… warnings.
• Any ext: entry flips extension tools to an explicit allowlist — unnamed extensions still load (handlers fire) but expose no tools. So tools: "*, ext:mcp/search" exposes only search from mcp, nothing from any other extension.
• Extension names match case-insensitively ([Mcp] = [mcp]); tool names in ext:foo/bar stay case-sensitive.
• pi's built-in extensions (builtin:mcp, builtin:codemode, builtin:tool-search) load in subagents as they do in the main session, so with extensions: true every subagent starts its own set of mcp.json servers, which stay up until the subagent's session is disposed (on quit, or ~10 minutes after it finishes). Address them by the full builtin: name — bare [mcp] matches a third-party extension named mcp, not the built-in — and narrow their tools with ext:builtin:mcp/<tool>. MCP tools are reached through codemode (or tool_search for deferred servers), so an ext: allowlist must name that bridge too, as pi's --tools read,codemode,'mcp__*' does: tools: "read, ext:builtin:codemode, ext:builtin:mcp/mcp__radius__search". Opt out per agent with exclude_extensions: builtin:mcp, or everywhere with -builtin:mcp in pi's extensions setting; starting pi with --no-mcp keeps MCP out of subagents too, and --no-extensions keeps all three built-ins out. MCP tools stay undeclared until codemode or tool_search reaches them, as in the main session, and a codemode script can only call tools inside the agent's scope. Out-of-scope tools are blocked but can still appear in codemode's searchTools() and tool_search results.
• Subagents inherit the parent session's project trust: in a project you did not trust, a subagent loads none of its .pi extensions or settings and never starts the servers in its .pi/mcp.json. A subagent spawned into another directory (RPC cwd) treats that directory as untrusted.
• Extensions that register tools lazily work too. MCP-backed extensions typically can't enumerate their tools until their servers connect, so they register from session_start or before_agent_start rather than at load. Subagent scoping is re-derived as tools appear, so these surface normally — including under ext: selectors, which keep narrowing correctly no matter when a tool shows up.
• Extensions bound into a subagent see both ends of that session's lifecycle: session_start when the agent starts, session_shutdown (reason quit) when its session is disposed — on quit, and when its record is evicted (10 minutes after its result was consumed, 60 after it finished if nobody read it). Release per-session resources there; anything left armed outlives the session it belongs to. Handlers are given three seconds on quit, after which teardown proceeds regardless.
• An installed package extension matches by its package short name (@scope/pi-subagents → [pi-subagents]), in addition to its path-derived name (a package whose entry is src/index.ts also answers to [src]). A directory entry counts too: pi-web-access declares ./dist, pi loads dist/index.js, and it answers to [pi-web-access] as well as [dist]. Prefer the package name — the path-derived one is incidental.
• Plain tools: typos fail loudly: tools: reed, grep fires tools-error:… instead of silently producing an under-tooled agent.
• exclude_extensions: wins over extensions: and over ext: selectors — an excluded extension never loads and a tools: ext: entry can't pull it back. Plain names only (no paths, no *); a name matching nothing fires an extension-error:… warning.
• exclude_extensions: is not a sandbox: excluded extensions' factory code still executes once during loading. Exclusion suppresses their tools and their bound lifecycle hooks (pi.on handlers like session_start only fire for extensions bound to the session), but not other load-time side effects — a factory that subscribes directly to the shared pi.events bus stays live. Don't rely on it to contain an untrusted extension.
• Array and string forms are equivalent: [a, b] == "a, b".

How an agent's scope is advertised. The Agent tool description lists every available agent with a (Tools: …) suffix, and that suffix is what the orchestrator reads when deciding where to route work. It names the built-in scope and flags that extension tools load too, without listing them (extensions may register lazily, see above, so they can't be enumerated when the description is built):

When the effective built-in set lacks bash (including disallowed_tools: bash), the shared suffix adds ; no built-in shell: cannot run bash commands. This applies to every type in full, compact and custom-template type lists. It does not claim extensions cannot execute commands: an extension may supply ssh_exec or another command tool. Put stronger restrictions in the agent description's first sentence so compact mode preserves them, e.g. no shell/ssh: cannot run commands. Shell-less prompts should require a FIRST-turn explanation, before tool calls, when a task needs unavailable commands. This is prompt guidance, not a dispatcher classifier.

Agent has no per-call tool-override flag. Use a shell-capable type for command work; do not ask a restricted researcher to acquire bash.

The last two rows are separate because zero built-ins is not zero tools: tools: none alongside extensions: still surfaces every extension tool, so calling it none would understate what the agent can do. Without the + extension tools flag an agent that can search the web through an extension (e.g. Plan) read as web-less.

Shipping agents in a pi package

A pi package can ship subagents the same way it ships skills, so a reusable agent travels with the package that needs it instead of being copied into every project's .pi/agents/ (#109). Declare the paths in the package's package.json:

{
  "name": "my-subagents",
  "keywords": ["pi-package"],
  "pi": {
    "subagents": {
      "agents": ["./agents"],
      "workflows": ["./workflows"]
    }
  }
}

Users get the agents by installing the package with pi — no copy step:

pi install npm:my-subagents      # or git:…, or a local path; -l for project scope

Four rules define the feature:

• Declaration is required. Pi falls back to convention directories (skills/, prompts/, …) for a package with no pi key at all; this extension never does. A package carrying an agents/ folder for some other tool contributes nothing until it names it — the author's half of the agreement, where installing is yours.
• Only pi's own packages count. Discovery reads packages[] from pi's settings.json, global and project, through pi's package manager. node_modules is never scanned, and a package listed only by an untrusted project stays invisible until you trust it.
• Package agents rank last. Any .pi/agents/, .agents/agents/ or global file of the same name wins, so a package can offer a reviewer but never take the name away from you. Same tier pi gives package-provided skills, and Claude Code plugin agents.
• They are read-only in /agents, badged ▪, offering Eject (an editable copy that shadows the original) and Disable (an enabled: false stub that does the same).

packageAgents and packageWorkflows in subagents.json default to true; set either to false or to a list of package names to narrow it. Installing a package is already the trust decision — pi runs that package's extensions/ and puts its skills/ in the system prompt with no further prompt, so a declared .md agent is strictly less privileged than what it already runs.

Full guide: docs/packages.md — manifest syntax and the four accepted spellings, ! exclusions, the development loop, naming, precedence, and troubleshooting.

Tools

Agent

Launch a sub-agent.

SubagentWorkflow

Run a deterministic script that orchestrates many subagents. Returns a task id immediately; the run continues in the background and notifies on completion.

At least one of script / scriptPath / name is required; scriptPath wins over script, which wins over name. Each invocation's script is persisted to the session directory and its path returned, so iterating means editing that file and re-running rather than resending the source. A saved workflow reports its own file instead, so the same loop works on it — project .pi/workflows/ shadows a same-named global one. Those directories are ordinary folders that may hold other scripts, so only files carrying the export const meta = { name, description } declaration are listed or resolved; naming anything else reports that it is not a workflow rather than running it. The check is a regex over the source — nothing in the file is executed to make it, and even a real parse evaluates only the meta object literal, in an empty node:vm context with a 100ms bound.

export const meta = {
  name: 'auth-audit',
  description: 'Find routes missing auth checks, then verify each finding',
  phases: [{ title: 'Scan' }, { title: 'Audit' }],
}

phase('Scan')
const listing = await agent('List every route file under src/routes/. One path per line.')
const files = listing.split('\n').map(s => s.trim()).filter(Boolean)
log('auditing ' + files.length + ' files')

phase('Audit')
return await pipeline(
  files,
  file => agent(`Audit ${file} for missing auth checks.`, { label: `audit:${file}` }),
  (found, file) => agent(`Try to REFUTE this finding about ${file}: ${found}`, { label: `verify:${file}` }),
)

Concurrency is capped at max(1, min(16, cpus - 2)) — the run's own limit, independent of the session's maxConcurrent pool, which its agents do not enter. There are 1000 agents per run and 4096 items per parallel/pipeline call.

Full guide: docs/workflows.md — how the model writes the script for you, how to edit and re-run it, how to save one as a reusable named workflow, plus the complete agent() option reference, recipes and troubleshooting.

get_subagent_result

Check status and retrieve results from a background agent.

Cancelling a wait: true call (for example, with Esc) stops only the wait. The background agent keeps running, and its completion notification still arrives normally.

steer_subagent

Send a steering message to a running agent. The message interrupts after the current tool execution. A queued agent (waiting on maxConcurrent) accepts it too: the result says it was queued, and it is delivered into the agent's first turn when it starts.

Commands

/agents → Workflows (shown only when workflows are on) opens a framed two-pane inspector over a run, with two levels of depth:

 audit-src
 Dynamically discover files under src/ and audit each …                    1/3 agents · 32s

 ╭ Phases ──────────┬ Discover · 1 agent ──────────────────────────────────────────────╮
 │ ❯ ✔ Discover 1/1 │ ❯ ✔ discover:src Opus 5 (1M context) · 26.4k tok             25s │
 │   2 Audit    0/2 │                                                                  │
 │   3 Verify       │                                                                  │
 │   4 Synthesize   │                                                                  │
 ╰──────────────────┴──────────────────────────────────────────────────────────────────╯
 ↑↓ select · ⏎ open · f filter · x stop · esc close · c convo

The overview puts the phases on the left (a phase shows its number until it finishes, then ✔/✘) and the selected phase's agents on the right. ⏎ opens one: the agents move to the left pane and the right becomes that agent's Prompt, Activity and Outcome, with ⏎ now expanding the prompt and esc going back a level rather than closing. ↑↓ (or j/k) move and f cycles the state filter, naming it in the pane title. The dialog opens as a centered overlay, like the conversation viewer in regular mode; the frame sizes itself to what it holds, between six rows and twenty-two, so a three-agent run is not twenty rows of nothing and a two-hundred-agent one scrolls inside the pane. Long titles truncate with … rather than tearing it. With more than one workflow in the session it asks which, newest first.

The run itself takes five keys, and the footer offers each only while it can actually do something:

Skipping is immediate for a running agent and for one held at a pause; an agent parked behind the concurrency limit takes its skip when it reaches the front of the queue.

CLI flags

Use the = form. The bare --flag value spelling consumes the next argument, so pi --subagents-workflow-file review.js "do the thing" would take the prompt as the flag's value. Composes with headless mode: pi -p --subagents-workflow-file=review.js. With no tool call to attach to, the run renders as a session entry and its result is handed to the model as context for its next turn.

The /agents command opens an interactive menu:

Running agents (2) — 1 running, 1 done     ← only shown when agents exist
Agent types (6)                             ← unified list: defaults + custom
Create new agent                            ← manual wizard or AI-generated
Settings                                    ← max concurrency (background + foreground), max turns, grace turns, join mode

• Running agents — select one to open its live conversation viewer. While it's still running, press Enter to open the steering composer, then Enter again to send a message that redirects the agent (same mechanism as the steer_subagent tool; Esc or an empty submit returns). On a live foreground run, press b to hand the same session and in-flight work to the background without stopping or restarting it — the blocked Agent call returns the agent ID and the run notifies on completion like any background agent. Press x (then x again to confirm) to stop/abort it — including background agents, which a global Esc can't unambiguously target (Esc still stops a blocking foreground Agent call, but no longer reaches one that has been handed off — x here does). A stopped agent reports its partial output flagged as incomplete, not as a completion. m cycles how much of the transcript renders as Markdown — see Viewer markdown.
• Agent types — unified list with source indicators: • (project), ◦ (global), ▪ (package), ✕ (disabled), explained by a legend naming only the badges actually on screen. Each row shows the agent's model, and the highlighted agent's full description appears below the list — prefixed with the declaring package's name for a ▪ agent, since the badge alone doesn't say which package. The model column flags (unavailable, fallback: inherit) when a configured model can't be resolved (it would silently inherit the parent model), and shows (→ provider/id) when it resolves to a different provider or version than configured. Select an agent to manage it:
  • Default agents (no override): Eject (export as .md), Disable
  • Default agents (ejected/overridden): Edit, Disable, Reset to default, Delete
  • Custom agents: Edit, Disable, Delete
  • Disabled agents: Enable, Edit, Delete
• Eject — writes the embedded default config as a .md file to project or personal location, so you can customize it
• Disable/Enable — toggle agent availability. Disabled agents stay visible in the list (marked ✕) and can be re-enabled
• Create new agent — choose project/personal location, then manual wizard (step-by-step prompts for name, tools, model, thinking, system prompt) or AI-generated (describe what the agent should do and a sub-agent writes the .md file). Any name is allowed, including default agent names (overrides them)
• Settings — configure max concurrency (background and foreground), default max turns, grace turns, and join mode at runtime

Record retention

A finished agent's record (and the session behind get_subagent_result and steer_subagent) is kept for 10 minutes once its result was consumed (fetched with get_subagent_result, or delivered inline to a blocking caller) and 60 minutes from completion while nobody has read it. After that the record is evicted, and the manager keeps an in-memory note for the rest of the session: both tools answer Agent "<id>" (<description>) completed, evicted after N min; transcript: <path> instead of Agent not found. The path is the .output transcript when one was written, else the session file. Notes are cleared on /new and /resume, like the records. For tests, PI_SUBAGENTS_CONSUMED_TTL_MIN and PI_SUBAGENTS_UNREAD_TTL_MIN override the two windows in minutes (fractions and 0 allowed; read when the manager starts); the sweep then runs as often as the shorter window needs, but not more than once a second.

Graceful Max Turns

Instead of hard-aborting at the turn limit, agents get a graceful shutdown:

1. At max_turns — steering message: "Wrap up immediately — provide your final answer now."
2. Up to 5 grace turns to finish cleanly
3. Hard abort only after the grace period
4. Whenever a limit fired and the agent did not end on a finished answer (text, no pending tool call), one more tool-less turn asks for the full final report. The steer is only a request: an agent deep in a tool loop ignores it, and without that turn the result was whatever it narrated before the abort ("Now let me check the tests..."). That turn is tool-free even when another extension re-activates its tools from before_agent_start (e.g. pi-deferred-context-engine): the loadout is pinned empty for the turn, any call that still gets through is refused, and the turn is stopped if it calls a tool or runs past one turn

A resumed agent gets a fresh budget: max_turns is counted from the resume prompt, resolved as for a spawn -- the agent file's max_turns, else the resume call's, else the limit the agent was spawned with, else the default. The same bounded, tool-free final report applies on resume. A caller stop, including one arriving while an aborted batch drains, never launches that report.

Concurrency

There are two independent pools.

Background (maxConcurrent, default 10). Excess agents are automatically queued and start as running agents complete. The widget shows queued agents as a collapsed count. Since agents run in the background by default, nearly every spawn takes a slot; the limit was raised from 4 so that ordinary parallel fan-outs don't queue.

Foreground (maxConcurrentForeground, default 0 = unlimited). Off by default, so nothing changes unless you set it. pi dispatches a message's tool calls through Promise.all, so several Agent calls with run_in_background: false in one message have always started at once — this bounds that. Useful mainly with local models, where parallel agents thrash the prompt cache (#253). A queued foreground agent appears in /agents → Running agents as queued and can be stopped there; its Agent call says so while it waits and then returns its result normally.

The two are deliberately not one limit. A foreground agent blocks the parent anyway — the parent could have done that work itself without paying a slot — so charging it to the background pool would let a saturated pool starve the main session.

The foreground pool does not cover resume: a foreground resume reopens an existing session and never reaches the spawn path, so several blocking resumes in one message can still run at once (the process budget below does cover them). A background resume does take a background slot and queues behind them like any other background agent. A resume of an agent that is still running or queued is refused on either path, including two resumes of one agent in the same message.

Nested children and a workflow's agents are outside the pool entirely. A nested child would deadlock behind a parent waiting on it; a workflow already bounds its own fan-out at max(1, min(16, cpus - 2)), and counting its agents twice would let one run fill the session's pool and starve everything else.

Process budget

Process-wide (maxActiveSubagents, default 0 = off; nestedSubagentReserve, default 2). One ceiling on running subagents for the whole pi process — not the machine: a second pi process has its own. Unlike the two pools it counts every start, in every agent manager in the process: top-level, nested, workflow, scheduled, RPC and @handle spawns, foreground and background resumes, and agents a /new, /resume, /fork or /reload left running in the previous instance. It bounds running agents only; it does not look at CPU or memory and does not serialize commands.

A run holds its slot until it actually settles, worktree cleanup included, and a parent keeps its slot while it waits on a child: pi runs a turn's tool calls in parallel, so the parent's other tools may still be executing. Shutting a session down does not free its runs' slots either; each run frees its own when it returns. To keep held slots from deadlocking, nestedSubagentReserve slots are held back per nesting level, up to the global maxSubagentDepth L. With limit(j) = maxActiveSubagents - nestedSubagentReserve × (L - j), a run at depth d (top-level, workflow and scheduled runs are depth 1) starts only while, for every level j from d to L, fewer than limit(j) runs at depth ≤ j hold slots. With maxActiveSubagents: 12, the default reserve and the default depth 2, at most 10 top-level runs hold slots and the last 2 only ever go to depth-2 children. With maxSubagentDepth 0 or 1 top-level runs get all 12.

The guarantee needs a coherent policy: with nesting on, a reserve of at least 1 and maxActiveSubagents ≥ nestedSubagentReserve × (L - 1) + 1. A budget that breaks that (nestedSubagentReserve: 0, or 12 and 2 at depth 8) is ignored with a warning and the previous policy stays in force. A project file can still deepen its own nesting, but the reserve covers only L: while the budget is on, a nested Agent call deeper than L fails at the call instead of waiting behind its own ancestors.

Waiting starts are served first come, first served among those that fit: a child may pass an earlier top-level start that has to wait for the reserve. A waiting Agent call or resume shows as queued and can be cancelled with Esc or /agents as usual. bypassQueue starts (scheduled jobs, the /agents agent generator) skip the per-session pools but not this budget, so a scheduled job can start late. At most 128 starts wait at once while the budget is on; the next spawn or resume fails at the call.

Both settings are read from global configuration only (~/.pi/agent/subagents.json or the enterprise manifest); a project file's value is ignored with a warning, since the budget spans every session in the process. The same goes for the depth it reserves for: the global maxSubagentDepth, not a project's. The last activation to apply its settings wins, and removing the key turns the budget off. Turning it off or raising a limit applies at once; turning it on, lowering a limit or changing the global depth waits until no subagent in the process holds a slot, since running parents were admitted under the old limits. A stderr warning names a deferred policy; the previous policy remains in force, and continuous work can delay the change indefinitely. Stop starting work and let the process become idle to apply it safely.

Managers loaded from a pi-subagents build without the budget are not counted; restart pi after upgrading for a complete count. A reload between builds with incompatible budget state versions refuses the entire new activation: no tools, commands or manager are registered, and session_start fails the bind with the host's ExtensionStartupError and an explicit restart diagnostic, even if the requested policy is off. On an older host without that fatal-bind API, the factory refuses immediately with the restart and host-upgrade diagnostic. It never continues unbudgeted or creates a parallel private count; existing runs in the old build can still settle under their original state.

Sending to the background

Sending a foreground agent to the background moves its slot: it gives up the foreground slot (so a queued blocking Agent call can start) and takes a background one. If the background pool is already full it runs above capacity rather than being suspended, and queued background work waits until the count drops back under the limit.

Join Strategies

When background agents complete, they notify the main agent. As in Claude Code, a notification reaches a busy parent after its current batch of tool calls, before its next model call — running tools are not interrupted — and everything that finished during that batch arrives as one message. An idle parent gets a turn about 200 ms after a completion, together with anything else that finished in that window. If you interrupt the parent (Esc), what was waiting is kept and delivered once the interrupted turn has stopped, starting a turn — as Claude Code does. The join mode decides whether a result waits for its siblings. It applies only to background agents.

Timeout behavior (group): a 30-second timeout starts after the first agent of a group completes. If not all agents finish in time, a partial notification is sent with completed results and remaining agents continue with a shorter 15-second re-batch window for stragglers.

Configuration:

• Configure join mode in /agents → Settings → Join mode

Model tiers

Use model: "@cheap" in an agent, agentModels, Agent({ model: "@cheap" }), RPC or workflow agent() options. Explicit tool models still win. Tiers are independent of enabledModels, fuzzy matching, model versions and provider naming conventions.

Bindings live ONLY in global ~/.pi/agent/subagents.json or the subagents object of an owner-only enterprise.local.json. Project tier definitions and resolution policies are ignored with a warning. Keys match ^[a-z0-9][a-z0-9._-]*$ and are case sensitive. Each nonempty list contains either exactly one registered router/<profile>, or only exact physical provider/id members. No mixtures, nested tiers, fuzzy entries or implicit neighbouring tiers. Exact model IDs may contain revision markers such as @20240620; a bare @other-tier binding entry is still rejected. Invalid bindings are dropped with a warning; references to them are unmapped.

Routed personal setup (configure the router profiles separately):

{
  "modelTiers": { "cheap": ["router/cheap"], "default": ["router/default"], "deep": ["router/deep"], "review": ["router/review"] },
  "modelResolution": { "onUnmapped": "error", "onUnavailable": "error" },
  "agentModels": { "*": "@default", "Explore": "@cheap", "code-reviewer": "@review" }
}

A routed tier starts on its virtual model, not a physical chain member. The router owns headroom ranking, admission and mid-run failover. Exhaustion never tries router/default or another tier: it follows the existing quota parking/error rules. Missing router profiles are unavailable, not fuzzy guesses.

GPT-only setup, with no Anthropic dependency or router required:

{ "modelTiers": { "cheap": ["openai/gpt-6-luna"] }, "agentModels": { "Explore": "@cheap" } }

Enterprise manifest (machine-local, permissions 0600):

{ "version": 1, "subagents": { "modelTiers": { "cheap": ["example-provider/example-wire-model"] }, "agentModels": { "Explore": "@cheap" } } }

Ordered static list:

{ "modelTiers": { "cheap": ["provider-a/fast-model", "provider-b/fast-model"] } }

Static tiers choose the first credentialed, admitted member and retain skipped members with no-credentials or quota reasons. With a router installed, each member is admitted separately. Without one, admission fails open and disclosure explicitly says static; no quota signal.

modelResolution.onUnmapped and onUnavailable each accept error or inherit; both library defaults are inherit, with a visible note. Neither policy applies to quota exhaustion. Whitespace-only model inputs mean unset; @ alone and unknown/case-mismatched tier names are visibly unmapped.

Results disclose Model: @cheap -> provider/id via router/cheap before the first response using admission's resolvedModel, then follow the answering model. Mid-run changes say switched ... within tier. Shared leading [@cheap] flags identify healthy tiers neutrally; warning flags and grouped alert counts cover only actual switches, parent inheritance, unavailable or unmapped outcomes. Structured modelTier outcomes survive records, events, transcripts, eviction and workflow journals. Reopened evicted sessions select/admit anew and retain old outcomes separately as previousModelTier and previousModelFallback; live resumes keep the existing session's selection. Scheduled tier jobs store @cheap, not today's physical pick, and resolve it again when they fire.

Machine-local model aliases

~/.pi/agent/enterprise.local.json is an optional, owner-only (0600) machine-local manifest for provider configuration that must never enter a repository. Its subagents.modelAliases section maps a logical name to ordered canonical provider/modelId candidates. The extension loads it only from the user agent directory, never from a project, and only accepts manifest version 1. The same modelAliases key is also accepted in the global ~/.pi/agent/subagents.json (when both define it, the manifest's whole set replaces the global one); a project .pi/subagents.json value is ignored with a warning, so a cloned repo cannot redirect which provider a subagent runs on.

{
  "version": 1,
  "subagents": {
    "modelAliases": {
      "logical-coding-model": ["example-provider/example-wire-model"]
    }
  }
}

Alias candidates are resolved in order against authenticated models. Agent records retain the requested input and selection outcome alongside the selected model; schedules retain the explicit request and resolve it again at fire time. Missing, malformed, group-readable, or unsupported manifests are ignored with a safe path-and-reason warning. Aliases may not contain /; candidates must be canonical IDs, so aliases cannot recurse. Legacy aliases are exact static tiers with admission and skip disclosure, not fuzzy shortcuts. An unavailable alias follows modelResolution.onUnavailable, never exposing credential data.

Per-agent models

agentModels in the global ~/.pi/agent/subagents.json (or the subagents section of enterprise.local.json) maps an agent type, or "*" for every type, to an exact provider/modelId, @tier, or legacy alias. Agent types match case-insensitively; tier names do not.

{ "agentModels": { "Plan": "@deep", "Explore": "@cheap", "*": "@default" } }

Precedence, highest first: the Agent tool's model parameter, agentModels[type], agentModels["*"], the agent definition's model:, the parent's model. The shared selectModel module applies it on every fresh spawn: top-level and nested Agent, workflow agent(), scheduled jobs, RPC, direct @mention and reopened evicted sessions. /agents labels use the same configured selection. An unresolved configured value falls directly to the parent, not a lower-priority configured default, with disclosure: the spawn path records Note: configured model "<input>" did not resolve; running on parent model <provider>/<id>. on the agent, and every surface reads it from there: the Agent result text and expanded row, get_subagent_result, the completion notice, nested Agent results, workflow rows and agent() results (text, error and a modelNote field), the widget, fleet list and conversation viewer, completion events, and agents started by scheduler fire, cross-extension RPC or @mention. A live-session resume retains its record's selection and note, ignoring new model inputs and config. An evicted conversation's modelTier, modelFallback and every model note describe the continuation's actual outcome; previousModelTier and previousModelFallback separately retain the prior run's outcomes, never as substitutes for a missing or changed binding. Explicit unresolved literal/fuzzy model names are hard errors on every fresh dispatch, including scheduled fires; tiers and legacy aliases follow modelResolution. Configured inputs and alias members are exact; only explicit non-tier tool inputs retain fuzzy ranking and provider retry. Tier failures obey modelResolution. A workflow agent({ schema }) result stays pure JSON: the note is a sibling _modelNote on an object payload and a line in the run log. Like modelAliases it chooses providers, so a project .pi/subagents.json value is ignored with a warning.

• Intentional blank-input change: empty and whitespace-only model strings mean unset at every precedence level and on every entry path, falling through to the next default without recording a param or definition request. Previously RPC/workflow explicit blanks and runner-based configured blanks reached fuzzy resolution (and could select an available model); Agent/nested/scheduled paths skipped "" but resolved whitespace. Nonblank inputs retain their original spelling and matching.

Known limits. At terminal widths below about 60 the workflow dialog's agent-level list pane may show only a fragment of the [fallback] flag (every other row shows it first and never clips it, and withLeadingFlag degrades to [fb] then !), and a very long fallback note in the dialog's detail pane is clamped to the pane capacity. The full note stays in the result text, the record, the run log and get_subagent_result. Tracked in #158.

Model Scope

Opt-in: off by default. Enable via /agents → Settings → Scope models.

When on, each subagent spawn's effective model is validated against pi's own enabledModels list (configured via pi's /scoped-models UI). pi-subagents reads that list; it doesn't manage it. Both of pi's settings files are honored: global ~/.pi/agent/settings.json and project-local <cwd>/.pi/settings.json. Project overrides global — mirrors pi's SettingsManager deep-merge, so a tighter per-project scope (hand-edited into the project settings) is respected.

Out-of-scope handling depends on source:

Design: scopeModels is a guardrail against the orchestrator picking unexpected models at runtime, not a hard policy against user-level config. An explicit caller model is an orchestrator-level choice and is therefore refused when out of scope; an agent's default or an inherited parent model warns and proceeds. This preserves portable agent definitions while letting an orchestrator deliberately select a target-local logical alias.

Nested spawns (nested subagents) apply the same table against the parent's config root. The hard-error case is identical; the warning cases proceed silently, since a subagent session has no UI to toast to.

Pattern format: only exact provider/modelId entries are honored (e.g. anthropic/claude-haiku-4-5-20251001). Glob patterns (*sonnet*), bare model IDs, and :thinking suffixes — which pi itself supports — are silently dropped here. pi's /scoped-models picker writes exact entries, so the limitation is invisible if you configure scope through the UI. Hand-edited globs produce an empty allowed set (scope check becomes a no-op).

No-op safety: if enabledModels is missing or empty in pi's settings, scope check skips entirely — no false positives, no spurious errors.

Persistent Settings

Runtime tuning values set via /agents → Settings (max concurrency, max foreground concurrency, default max turns, grace turns, nested depth, fallback agent, default join mode, scheduling on/off, scope models on/off, disable defaults on/off, strict agent files on/off, package agents on/off, package workflows on/off, agent mentions on/off, output transcript on/off, tool description full/compact/custom, widget all/background/off, usage reporting on/off, cost display on/off, model display on/off, viewer markdown off/assistant/all) persist across pi restarts. Two files, merged on load:

• Global: ~/.pi/agent/subagents.json — your machine-wide defaults. Edit by hand; the /agents menu never writes here.
• Project: <cwd>/.pi/subagents.json — per-project overrides. Written by /agents → Settings.

Precedence: project overrides global on any field present in both. Missing fields fall back to the hardcoded defaults (max concurrency 10, max foreground concurrency 0 = unlimited, default max turns unlimited, grace turns 5, nested depth 2, join mode smart, defaults enabled).

Process budget (maxActiveSubagents, default 0 = off; nestedSubagentReserve, default 2): the process-wide ceiling described under Process budget. Global configuration only — not written by /agents, and ignored with a warning in a project file.

Nested depth (maxSubagentDepth, default 2): the hard ceiling on nested delegation, counted from the main session (main = 0, its subagents = 1). 0 or 1 disables nesting project-wide regardless of any agent's allowed_subagents. Read when a subagent session is built, so a change applies to agents started after it.

Fallback agent (fallbackSubagent, default general-purpose): the agent used when a caller-supplied subagent_type doesn't resolve to exactly one enabled agent — unknown, disabled, or ambiguous because two agents differ only by case. Name any enabled agent to route those calls there instead, or set none for strict, fail-closed dispatch: the call is refused with an error listing the available types, and nothing spawns. Strict mode matters most for background and scheduled calls, which would otherwise start executing a substituted agent before the caller learns anything. Also settable from /agents → Settings → Fallback agent. The boolean false is accepted as a spelling of none, because it would otherwise be dropped as the wrong type and silently leave the permissive default in place. Every other value is read as an agent name, so a mistaken off fails loudly at dispatch rather than meaning one thing in the settings file and another in the resolver. A fallback agent that is itself unknown or disabled is a misconfiguration and is reported rather than quietly replaced. Note the default is unchanged and stays permissive by design: with disableDefaultAgents and no general-purpose of your own, an unresolvable type still resolves to a built-in config carrying all tools — set none (or name one of your own agents) to close that.

Strict agent files (strictAgentFiles, default false): when on, an unreadable or unparseable agent file aborts extension load at startup and names the file, instead of being skipped with a warning — so a checked-in .pi/agents/ can't silently fall through to a same-named agent from another location. Startup only: the mid-session reload that runs on each Agent call keeps warning either way, since a bad edit shouldn't kill a session on an unrelated spawn. Two phases: a malformed global file fails at activation as an ordinary factory error; a malformed package file only warns, in both phases; a malformed trusted-project file fails at the first session_start, once project trust is known, with the host's ExtensionStartupError. Strict mode needs a newer pi host that exports that class. On an older supported host, activation refuses before reading any agent file; upgrade pi or set strictAgentFiles: false. Also settable from /agents → Settings → Strict agent files.

Package agents / package workflows (packageAgents and packageWorkflows, both default true): whether agents and workflow scripts declared by installed pi packages are loaded, and from which packages. true admits every declaring package, false none, and a list admits only the packages it names — matched case-insensitively against the unscoped short name (@acme/tools → tools), the full package name, or the settings source string. The two are separate so package agents can be accepted while package workflow scripts are refused; an agent is markdown handed to a model, a workflow is .js this extension executes. Toggle both from /agents → Settings; an allowlist is hand-written in subagents.json, and toggling the row replaces it (the toast names what it replaced). See Shipping agents in a pi package.

Disable defaults (disableDefaultAgents, default false): when on, the three built-in agents (general-purpose, Explore, Plan) are not registered — only your project/global custom agents are advertised and spawnable. User-defined agents are unaffected, including ones that override a default by name. The Agent tool's type list updates on the next pi session (the tool schema is registered at startup).

Agent mentions (agentMentions, default "model"): whether @handle message at the prompt addresses that subagent instead of the main model — messaging, resuming or starting it — and whether @ offers agents alongside pi's file completion. "model" and "direct" differ only in who starts an agent that isn't running: an off-screen clone of this conversation, via a <system-reminder> and a real Agent call, or this extension, immediately and with no model call. Messaging and resuming are direct in both. "off" gates all three actions plus the suggestion list, so @ means only "attach a file" again and every @… prompt reaches the main model verbatim. Toggle via /agents → Settings → Agent mentions; applied live. The booleans this setting used to take are still read — true as "model", false as "off".

Background by default (backgroundByDefault, default true): what an Agent call that doesn't say means. On — following Claude Code — the agent runs detached, the call returns its ID immediately, and a completion notification carries a preview of the result (get_subagent_result for the full text). Set false to restore the previous behaviour, where an unqualified spawn blocked the turn and returned its output inline. An explicit run_in_background on the call, or in an agent file's frontmatter, overrides this in both directions; the setting only decides what "unspecified" means. Top-level only — a nested spawn (an agent spawning its own) always defaults to foreground, because a detached child is stopped when its parent settles and has no notification path of its own. Toggle via /agents → Settings → Background by default; applied live.

Remember agents (rememberAgents, default true): whether subagents persist their pi session, which is what lets @handle reopen an agent's conversation after its in-memory record has been evicted. Two visible consequences of the default: top-level subagents write a session file, and they nest under the session that spawned them in pi's /resume. Agents spawned by another agent are excluded — they get no handle, so nothing could reopen their transcript. A custom agent's persist_session frontmatter overrides this per agent, in both directions. Toggle via /agents → Settings → Remember agents; with it off, handles expire with their record (roughly ten minutes past completion) and @explore then starts a fresh agent rather than resuming — the behaviour before this setting existed.

Output transcript (outputTranscript, default true): the project/global default for writing each subagent's .output transcript. Toggle via /agents → Settings → Output transcript, or set false in subagents.json to make transcripts opt-in project-wide — useful when run transcripts shouldn't sit on disk for backup or DLP tooling to pick up. A custom agent's output_transcript frontmatter overrides this per agent. Applied live at spawn time. Governs only the transcript, not persist_session, worktree commits, or memory files.

Worktree isolation (worktreeIsolation, default true): whether isolation: "worktree" may create a worktree at all. Toggle via /agents → Settings → Worktree isolation, or set false in subagents.json on a repo where a copy costs too much time or disk. Off, the Agent tool's isolation parameter is dropped from the schema entirely and the bullet describing it leaves the tool description with it — nothing to pass, and no context spent describing it — and worktrees are refused on every other path too (agent files, scheduled jobs, cross-extension RPC). The /agents agent-file generator stops offering the isolation: frontmatter field too, so a generated agent can't bake in a request that would be refused. A requested worktree is downgraded to a normal run rather than failing the call, since declining one is the point; there is deliberately no note on the result, which is exactly why the prose has to go when the parameter does. The refusal applies immediately; the parameter and its prose appear or disappear on the next pi session. See Turning worktrees off.

Report usage to session (reportUsage, default false): whether subagent spend is added to this session's own totals. Subagents run in their own pi sessions, so by default pi's footer, statusline and /cost count only what the main model spent — a session that delegated most of its work reads as nearly free. Turn it on and each Agent / get_subagent_result / steer_subagent result carries the spend accumulated since the last one, which pi folds into getSessionStats(); /cost attributes it to the Tools/summaries bucket. Toggle via /agents → Settings → Report usage to session; applied live.

Three things worth knowing about the numbers. Every token component is reported, cacheRead included — the cached prefix genuinely is re-read and re-billed on every call, and pi counts it the same way for the session's own messages, so withholding it would make a subagent's rows count differently from every other row in one total. (The extension's own token displays still leave it out, which is a different question: there it inflates a reading of how much work was done.) Cost is pi's own per-message figure, priced from the model's listed rates; a model pi has no rates for contributes zero rather than an estimate. And the context-window percentage is untouched: pi derives it from assistant messages alone, so a delegating session's context doesn't appear to fill up faster. Agents that finish in the background have no tool result of their own to ride on, so their spend is carried by the next one you make — the footer catches up on the following call, not the moment they finish.

Show cost (showCost, default false): whether the subagent surfaces print an estimated cost beside their token counts — the widget (running and finished lines), FleetView, the conversation viewer, foreground results, get_subagent_result, and completion notifications:

├─ ⠹ Explore  inspect code · ↻3 · 8.2k token · ~$0.0042 · 4.1s
✓ Explore  inspect code · ↻8 · 5 tool uses · ~$0.0181 · 12.3s

When several background agents finish together, their notification is topped with the batch total (3 agents · 45.1k token · ~$0.042) so the figures don't have to be added up by hand.

The ~ marks it as pi's estimate rather than a billed figure. A cost is shown only when there is one 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 — so nothing is printed at all, on every surface. For the same reason a real cost too small to render reads <$0.0001. Figures keep cents at minimum and four decimals at most (~$0.0042, ~$0.05, ~$1.24) — rounding everything to cents would print the same number for runs that differed fourfold.

Independent of reportUsage: this one is what you read, that one is what your session counts. Toggle via /agents → Settings → Show cost; applied live.

Show model (showModel, default false): whether the widget's running rows and the FleetView rows name the model driving each agent and the thinking level it is running at. Toggling it refreshes both. The widget row, then the FleetView row, which writes the level think:high so it cannot read as a duplicate of the live thinking · 4s activity beside it:

├─ ⠹ Explore  inspect code · sonnet 4.6 · thinking: high · ↻3 · 8.2k token · 4.1s
  ○ Explore  inspect code        ▶ running · sonnet 4.6 · think:high · thinking · 4s · 41s · ↓ 8.2k tokens

Off by default because the row already carries the description, turns, tool uses, tokens and elapsed time, and every character it gains is one the description loses on a narrow terminal. The other surfaces show the pair either way: the Agent tool result names the model beside its tags, and the conversation viewer's ↳ row spells out the canonical provider/model-id.

Both places report what the run actually used, read back from the child session once pi has resolved its defaults and clamped the level to what the model supports — not what the call asked for. Where Pi clamps the requested thinking level, the request is kept beside the effective value rather than dropped. An explicit model selection also overrides a frontmatter default; model resolution has no requested-versus-effective display:

  ↳ anthropic/claude-haiku-4-5 · thinking: low (asked max) · background

Toggle via /agents → Settings → Show model; applied live.

Viewer markdown (viewerMarkdown, default "assistant"): how much of the conversation viewer's transcript is rendered as Markdown rather than shown verbatim.

off        every line literal, as before this setting existed
assistant  assistant text rendered; tool results verbatim and dim   (default)
all        tool results rendered too

Scoped rather than all-or-nothing because the two kinds of content have different contracts. Assistant text is Markdown — the model writes it that way, and the viewer was the only surface showing its source. A tool result is whatever bytes the tool produced, and a Markdown pass over one rewrites things that occur constantly in real output: # section in a shell script loses its #, a --- line is swallowed as a setext heading, indented output is re-fenced, and | a | b | is redrawn as a box-drawing table. Each of those reads as the tool having misbehaved, which is why all is opt-in.

Two rewrites are suppressed outright rather than left to the mode, because they change data rather than layout: ordered-list markers keep their source numbering (3) 7) 9) stays, instead of being renumbered 3. 4. 5.) and backslash escapes are not normalised.

Turn all on for tools that genuinely emit Markdown, and off again for a diff or a log. m in the viewer cycles the three and persists the choice, so the key and this setting are the same value — the footer shows which is in force as m raw / m md / m md+. Code fences are syntax-highlighted using pi's own Markdown theme — which is also why fenced code is the one part of a result not dimmed under all; result prose still is, so the transcript keeps its hierarchy. Applied live; also settable from /agents → Settings → Viewer markdown.

Workflows (workflowsEnabled, default true): the master switch for scripted workflows. Toggle it via /agents → Settings → Workflows, or set it in subagents.json. Off, the SubagentWorkflow tool is never registered — the model is not told the feature exists and cannot call it, so it costs no tool-spec context — the /agents → Workflows entry is hidden, and --subagents-workflow-file refuses with a pointer to the setting rather than doing nothing. Read at extension load, so it applies on the next pi session; runs already in flight are left alone.

Leaving it unset is not quite the same as true. Unset means auto: on, unless another extension already provides a workflow tool, in which case this one warns and stands down for the session. Two orchestrators in one tool spec is a worse default than none — the model has to guess which to call and pays for both descriptions to find out — and the extension that was installed on purpose is the one that should survive. Setting workflowsEnabled explicitly pins the answer: true keeps ours whatever else is loaded, false is off regardless.

The match is on the exact tool names Workflow (Claude Code's) and SubagentWorkflow (ours), never a substring, so a list_workflows or github_workflow_run from some CI integration does not silently take the feature down. The check runs at session_start and nowhere earlier, because getAllTools throws during extension loading and load order means a check at registration time could not see an extension that has not loaded yet — so the tool is registered first and withdrawn from the active set through setActiveTools, which rebuilds the system prompt before any turn runs. When the other extension took the SubagentWorkflow name itself, pi's first-registration-wins rule already dropped ours, so there is nothing to withdraw and only the menu and the CLI flag come down.

Tool description (toolDescriptionMode, default "full"): which Agent tool description the LLM sees. "full" is the rich Claude Code-style prompt (~1,400 tokens with the default agents); "compact" is ~75% smaller — one-line agent type list, terse usage notes — for small/local models where tool-spec tokens are expensive. Per-option details stay in the parameter descriptions in every mode (the parameter schema is never customizable). Applies on the next pi session.

"custom" registers your own description from <cwd>/.pi/agent-tool-description.md (project) or <agentDir>/agent-tool-description.md (global; project wins). The file is read once at tool registration, so edits also apply on the next pi session. Dynamic parts stay live via placeholders — a static agent list would go stale the moment you add a custom agent:

Launch an autonomous agent. Available types:
{{typeList}}

Custom agents live in .pi/agents/ or {{agentDir}}/agents/.

Placeholders: {{typeList}} (full per-agent descriptions), {{compactTypeList}} (first sentence each), {{agentDir}}, {{isolationGuideline}} and {{scheduleGuideline}} (each expands with its own leading newline + - bullet when the matching feature is on — place them directly after your last rule line; empty when worktree isolation / scheduling is off). Unknown placeholders are left verbatim with a stderr warning; a missing or empty file falls back to "full" with a warning. Note the usual trust umbrella: a project-level file shapes the orchestrator's prompt, same as project agents and extensions do.

Starting point: copy examples/agent-tool-description.md — it reproduces the default full description exactly (a CI test keeps it in sync), so you can trim from a known-good baseline instead of writing from scratch.

Example — global defaults for a beefy machine:

mkdir -p ~/.pi/agent
cat > ~/.pi/agent/subagents.json <<'EOF'
{
  "maxConcurrent": 16,
  "graceTurns": 10
}
EOF

Every project now starts with concurrency 16 and grace 10, without ever touching the menu. Individual projects can still override via /agents → Settings.

Failure behavior: missing file is silent; malformed JSON logs a [pi-subagents] Ignoring malformed settings at … warning to stderr; invalid/out-of-range field values are dropped per-field; write failures downgrade the /agents toast to a warning with (session only; failed to persist).

Events

Agent lifecycle events are emitted via pi.events.emit() so other extensions can react:

Tier provenance on events. Payloads that disclose fallback provenance also carry modelTier and modelNote for tier selections. Outcomes are routed, static, unmapped, unavailable or exhausted; successful outcomes include the tier, bound profile/member, admitted model and latest answering model, and static outcomes list skipped members and whether a quota signal was used. subagents:record history entries and detached completion spools retain the same outcome. Persisted modelTier.model contains identity only (provider, id, name, api), never provider URLs, headers or other credential-bearing model overlays.

Model fallback on events. When a configured model did not resolve, modelFallback ({ input, used, also?, via? }) and a ready-made modelNote ride on created, started, compacted, usage and completed/failed payloads. The shared spawn path records selection before emitting started, including RPC, @mention and schedules with no explicit model. All later events read the same retained provenance.

The four agent-lifecycle events — subagents:started, :completed, :failed, :compacted — are emitted for top-level agents only; nested subagents and a workflow's children report through the parent or workflow that owns them. subagents:usage is the one event that is not scoped, on purpose: it is a ledger, not a lifecycle. It fires once per message for every agent, so summing usage.cost.total across it gives what this session's subagents spent, each message counted exactly once under the agent that spent it — unlike the ancestor records, which fold a nested child's spend into every parent. It fires regardless of reportUsage, which decides only whether the parent session's own totals include the spend.

tokens.total = input + output + cacheWrite. cacheRead is excluded — each turn's cacheRead is the cumulative cached prefix re-read on that one API call, so summing per-message would over-count it as a measure of work done. Use contextUsage.percent (surfaced as (NN%) in the widget) for current context size.

usage answers the other question — what was billed — and so does include cacheRead, because the prefix really is re-read and re-charged on every call. It is a pi Usage, the same shape pi puts on ToolResultEvent and AssistantMessage, so usage.cost.total is where a listener already expects the money and anything pi adds to Usage arrives without a change here. Neither field derives from the other; tokens is a view model, usage is the data.

Cross-Extension RPC

Other pi extensions can spawn and stop subagents programmatically via the pi.events event bus, without importing this package directly.

All RPC replies use a standardized envelope: { success: true, data?: T } on success, { success: false, error: string } on failure.

Full reference: docs/rpc.md — the complete spawn-option surface (including the fields that are silently stripped), every error string, the completion-notification race, the Symbol.for("pi-subagents:manager") registry, and what protocol version 2 does and does not promise. tintinweb/pi-tasks is the reference implementation.

Discovery

Listen for subagents:ready to know when RPC handlers are available:

pi.events.on("subagents:ready", () => {
  // RPC handlers are registered — safe to call ping/spawn/stop
});

subagents:ready fires only when pi-subagents is actually loaded and bound in the current session. A session that excludes it (via an agent's extensions:) emits no subagents:ready and does not answer the RPC channels — exactly as if pi-subagents were not installed. Treat "no subagents:ready" as "not available here" and give discovery a timeout rather than waiting indefinitely.

Ping

Check if the subagents extension is loaded and get the protocol version:

const requestId = crypto.randomUUID();
const unsub = pi.events.on(`subagents:rpc:ping:reply:${requestId}`, (reply) => {
  unsub();
  if (reply.success) console.log("Protocol version:", reply.data.version);
});
pi.events.emit("subagents:rpc:ping", { requestId });

Spawn

Spawn a subagent and receive its ID:

const requestId = crypto.randomUUID();
const unsub = pi.events.on(`subagents:rpc:spawn:reply:${requestId}`, (reply) => {
  unsub();
  if (!reply.success) {
    console.error("Spawn failed:", reply.error);
  } else {
    console.log("Agent ID:", reply.data.id);
  }
});
pi.events.emit("subagents:rpc:spawn", {
  requestId,
  type: "general-purpose",
  prompt: "Do something useful",
  options: { description: "My task", isBackground: true },
});

options is the manager's spawn-option object, not the Agent tool's parameter schema — the background flag is isBackground, and the tool's snake_case run_in_background is forwarded verbatim and ignored. Every RPC spawn returns its id immediately and runs detached either way; isBackground: true is what makes the agent occupy one of the maxConcurrent slots (and queue behind them when they are full). It does not affect subagents:created, which is never emitted for an RPC spawn at all — the first event you see for one is subagents:started. Leaving it unset starts the agent immediately regardless of that limit, though not of the process budget when it is on. maxConcurrentForeground never applies here whatever isBackground says: it bounds only spawns a caller is blocking on inline, and every RPC spawn is detached. A top-level RPC spawn renders in the widget and FleetView while it runs, with the same live tool activity and turn counter an Agent-tool spawn gets — only an explicit isBackground: false is dropped by the widget's default background mode, the way a foreground Agent call is. Nested spawns stay hidden from both.

options.model accepts either a Model object (e.g. ctx.model) or a "provider/modelId" string — strings are validated against ctx.modelRegistry at the RPC boundary and selected by the shared spawn path, which records the original input and source. Cross-extension callers can forward serializable values without losing auth context. Resolution is fuzzy, so a bare "sonnet" can land on a provider you never named: with Model Scope on, an override that resolves outside enabledModels is refused with an error envelope listing the allowed models, exactly as a caller-supplied Agent({ model }) is. null means unset — the agent inherits, same as omitting the field.

options.cwd (absolute path to an existing directory — anything else returns an error envelope; null means unset) runs the agent in a different working directory than the parent session. Its tools operate there and the prompt's environment block describes it, but .pi config still loads from the parent session's project — the target directory's .pi extensions never execute, and its agents/skills/settings are not picked up. Combined 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 stays scoped to that package), and the resulting pi-agent-* branch lands in that repo — the completion message names it. On session end, 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). Agents with memory: keep reading/writing the parent project's memory.

Stop

Stop a running agent by ID:

const requestId = crypto.randomUUID();
const unsub = pi.events.on(`subagents:rpc:stop:reply:${requestId}`, (reply) => {
  unsub();
  if (!reply.success) console.error("Stop failed:", reply.error);
});
pi.events.emit("subagents:rpc:stop", { requestId, agentId: "agent-id-here" });

Consume

Say that an agent's result has been shown to the model, so its completion notification is not delivered on top of it:

pi.events.emit("subagents:rpc:consume", { requestId: crypto.randomUUID(), agentId: "agent-id-here" });

This is the bus-side half of what get_subagent_result does when it returns a result. A caller that joins an agent on subagents:completed and reports the result itself should consume it — otherwise the notification still reaches a parent that has already seen the result, costing a model call to dismiss. Fire-and-forget is the intended use: the reply carries nothing to act on, and the channel is outside the subagents:rpc:ping version handshake, so a caller can send it unconditionally and an older extension that has no handler simply keeps notifying. Consuming a running or unknown agent is refused (success: false) and changes nothing — a running agent has no result to have been read, and its notification is still the caller's only signal that it finished.

Steer

Deliver a message to a running agent you spawned, when something it needs arrives mid-run (the bus-side half of steer_subagent):

pi.events.emit("subagents:rpc:steer", { requestId: crypto.randomUUID(), agentId: "agent-id-here", message: "A second check failed on the same head; include it before you push." });

Same ownership guard as stop; Agent is not running once the agent has settled. A successful steer emits subagents:steered.

Reply channels are scoped per requestId, so concurrent requests don't interfere.

Persistent Agent Memory

Agents can have persistent memory across sessions. Set memory in frontmatter to enable:

---
memory: project   # project | local | user
---

The user scope previously hardcoded ~/.pi/agent-memory/. If that legacy directory exists for an agent and the new location doesn't, it keeps being used — existing memories aren't orphaned.

Memory uses a MEMORY.md index file and individual memory files with frontmatter. Agents with write tools get full read-write access. Read-only agents (no write/edit tools) automatically get read-only memory — they can consume memories written by other agents but cannot modify them. This prevents unintended tool escalation.

The disallowed_tools field is respected when determining write capability — an agent with tools: write + disallowed_tools: write correctly gets read-only memory.

Worktree Isolation

Set isolation: worktree to run an agent in a temporary git worktree:

Agent({ subagent_type: "refactor", prompt: "...", isolation: "worktree" })

The agent gets a full, isolated copy of the repository. The worktree directory is removed on completion either way — what differs is whether a branch is left behind:

• No changes: worktree is cleaned up automatically, no branch
• Changes made: changes are committed to a new branch (pi-agent-<id>), and the result names the branch and the git merge command for it. The branch is the only artifact — the worktree path is gone, so nothing points into it
• Agent committed its own work: the branch is created at the agent's HEAD, preserving its commits (uncommitted leftovers are committed on top first)

The agent's system prompt names the worktree as an isolated copy and tells it to work only there, even if other instructions name the main checkout — otherwise an inherited parent prompt or a task prompt mentioning the project path walks it straight back out of the copy. This is a directive, not a sandbox: an agent with shell access can still cd out, so don't rely on isolation alone to protect the main checkout.

The automatic preservation commit uses --no-verify, so local pre-commit hooks can't block it — the commit is local-only and never pushed, and pre-push/server-side hooks still apply.

The worktree is made from the repo containing the agent's working directory, which is the session cwd unless the call passes cwd. To isolate a repo the session wasn't started in, name it:

Agent({ subagent_type: "general-purpose", prompt: "...", isolation: "worktree", cwd: "~/code/other-repo/pkg" })

The agent works at the same subdirectory inside the copy (pkg/ here), and the result names the branch and that repo's toplevel to merge in. ~/ expansion honors HOME (USERPROFILE on Windows), including on Bun; it falls back to the OS home when unset. The cwd gets the same treatment as the RPC cwd option: .pi config still loads from the session's project.

If the worktree cannot be created, the Agent call fails with an error naming the directory it tried and the cause (not a git repo, no commits yet, a git timeout, or git worktree add's stderr) instead of running unisolated — isolation: "worktree" is a strict guarantee, not a hint. The call is reported as a failed tool call, not as a subagent that ran and returned that message, so the model doesn't retry it as if the agent had merely reported a problem. Initialize git and commit at least once, or omit isolation.

A worktree is a copy, so the agent cannot see uncommitted or staged changes in the main checkout. Never use it to review a working-tree or staged diff: the agent finds an empty git diff and reports nothing wrong.

Turning worktrees off

Three levers, from narrowest to broadest:

• Per call — omit isolation, or pass isolation: "off". The explicit value exists because some models fill every optional parameter they are offered; with worktree as the only legal value they had no way to decline one (#231, #184).

• Per agent — isolation: off in an agent file. Frontmatter is authoritative, so this refuses a worktree even when the caller passes isolation: "worktree" — the only way to override a caller. Workflow agent() calls are the exception: as in Claude Code, only the script's option decides there.

• Per project — "worktreeIsolation": false in subagents.json. The Agent tool's isolation parameter disappears from the schema entirely, along with the usage-note bullet that describes it (so it costs the model no context and cannot be passed), and worktree creation is refused on every other path too: agent files, scheduled jobs, and cross-extension RPC. The /agents generator also stops offering isolation: when writing a new agent file. Use it on a repo large enough that a copy costs real time and disk. The schema and the description are both built at tool registration, so they appear or disappear in the next pi session; the refusal itself takes effect immediately.

  Schema and prose are gated together on purpose. Leaving the bullet in would teach the model to pass a field that is no longer declared — accepted silently, then dropped — and since a refused worktree carries no note on the result, the model would have every reason to go on reporting a pi-agent-* branch that was never created. A custom tool description should use the {{isolationGuideline}} placeholder rather than hardcoding the bullet, for the same reason.

Skill Preloading

Skills can be preloaded by name and injected into the agent's system prompt:

---
skills: api-conventions, error-handling
---

Discovery roots (checked in this order, first match wins):

Per root, a skill named foo resolves to the first of:

• <root>/foo.md — flat file at the top level
• <root>/foo/SKILL.md — directory skill (top-level)
• <root>/*/.../foo/SKILL.md — directory skill, found by recursive descent

Recursion skips dotfile directories and node_modules. A directory that itself contains a SKILL.md is treated as a single skill — we don't descend into it. Traversal is byte-order sorted for deterministic resolution across filesystems.

Security: project roots reject symlinks at every layer (the .pi/ or .agents/ directory, root, flat file, skill directory, SKILL.md inside a skill directory), so a cloned repository cannot point a skill at files outside it — intentional deviation from Pi. User roots belong to you and follow symlinks like Pi (linked roots, directories and files; broken links and directory cycles are skipped), so skills linked into ~/.agents/skills/ by an installer or dotfile manager load. Skill names with path-traversal characters (.., /, \, spaces, leading dot, >128 chars) are rejected.

Tool Denylist

Block specific tools from an agent even if extensions provide them:

---
tools: read, bash, grep, write
disallowed_tools: write, edit
---

This is useful for creating agents that inherit extension tools but should not have write access.

Architecture

docs/                 # Long-form guides (shipped to npm; README links out to them)
  workflows.md        # SubagentWorkflow: writing, editing, saving and re-running scripts
  rpc.md              # Cross-extension integration: pi.events, subagents:rpc:*, manager registry
  home-injection.md   # Injected home boundaries, call-time defaults and filesystem canaries
examples/
  workflows/          # Runnable examples, executed by test/workflow-examples.test.ts
  agent-tool-description.md
test/                 # Full Vitest suite; four session scenario files also run under Bun
  helpers/scenario-session.ts # Exported core host, child/provider barriers and isolated homes
src/
  index.ts            # Extension entry: tool/command registration, /agents menu, rendering
  types.ts            # Type definitions (AgentConfig, AgentRecord, etc.)

  # Agent registry
  default-agents.ts   # Embedded default agent configs (general-purpose, Explore, Plan)
  custom-agents.ts    # Load user-defined agents from .pi/agents/, .agents/agents/, global agents, and pi packages
  startup-error.ts    # Optional bridge to the host's ExtensionStartupError (strictAgentFiles)
  package-resources.ts # Agents/workflows declared by installed pi packages (pi.subagents manifest)
  agent-types.ts      # Unified agent registry (defaults + user), tool name resolution
  agent-file-toggle.ts # Locate/edit an agent's .md: enabled: toggle, eject to frontmatter
  agent-color.ts      # Claude Code/Agency Agents name color parsing and badge rendering

  # Execution
  agent-runner.ts     # Session creation, execution, graceful max_turns, steer/resume
  agent-manager.ts    # Agent lifecycle, concurrency queue, completion notifications
  nested-tools.ts     # Delegation tools handed to subagents (nested spawn/collect/steer)
  process-budget.ts   # Process-wide running-agent budget shared by every manager (maxActiveSubagents)
  child-context.ts    # AsyncLocalStorage flag marking work done for a child session
  abortable.ts        # Race a wait against Esc without cancelling the background child
  model-call.ts       # In-flight model call tracking: thinking / writing / stalled?
  completion-outbox.ts # Deliver completion notifications at the parent's next boundary, coalesced
  clock.ts            # Shared scheduler/outbox/admission/retention/UI clock and native timers
  detached-results.ts # Session-scoped spool and registry for background agents surviving reload/switch
  group-join.ts       # `group` join mode: one notification per turn's batch, with timeout
  status-note.ts      # Honest status note + salvaged partial output for non-normal outcomes
  usage.ts            # Token usage shapes, accumulators, session-stats readers

  # Invocation surface
  invocation-config.ts # Shared tool-parameter schemas (isolation, join, thinking, ...)
  model-selection.ts  # Shared precedence, scope and resolution outcome; stored on the agent record
  model-tiers.ts      # Exact static/routed tiers, admission, policy and durable disclosure
  model-resolver.ts   # Model matching: exact provider/modelId, aliases, fuzzy fallback ranked by version
  enabled-models.ts   # Read pi's enabledModels settings (project over global)
  model-scope.ts      # scopeModels allowlist policy, shared by top-level and nested tools
  mention.ts          # `@handle message` grammar: suggestion triggers and send parsing
  mention-clone.ts    # Run a mention's turn in a cloned conversation, off the main chat
  cross-extension-rpc.ts # ping/spawn/stop/consume/steer RPC handlers over pi.events

  # Scheduling
  schedule.ts         # SubagentScheduler: cron / +10m / interval / ISO dispatch
  schedule-store.ts   # PID-locked, session-scoped, atomic schedule persistence

  # Context & environment
  home-paths.ts       # Injected home normalization and Pi agent-directory defaults
  resource-loader.ts  # Fresh scoped settings paths for Pi's real resource/package loader
  memory.ts           # Persistent agent memory (resolve, read, build prompt blocks)
  skill-loader.ts     # Preload skills (Pi-standard + Agent Skills spec layouts)
  output-file.ts      # Streaming output file transcripts for agent sessions
  worktree.ts         # Git worktree isolation (create, cleanup, prune)
  prompts.ts          # Config-driven system prompt builder
  context.ts          # Parent conversation context for inherit_context
  settings.ts         # Persistent settings (~/.pi/agent/subagents.json + .pi/subagents.json)
  env.ts              # Environment detection (git, platform)

  workflow/
    meta.ts           # Extract and validate a script's pure-literal `meta` block
    worker-source.ts  # The sandbox: vm context, determinism prelude, script globals
    runtime.ts        # Worker lifecycle, RPC bridge, semaphore, caps, gate/resume
    progress.ts       # Progress event log and every derived view of it (pure)
    host.ts           # WorkflowHost adapter over AgentManager
    task.ts           # local_workflow task record and batched progress updates
    tool-description.ts # Model-facing description carrying the orchestration patterns
  ui/
    agent-widget.ts       # Persistent widget: spinners, activity, status icons, theming
    agent-row.ts          # The Agent tool row: live one-line summary, expanded prompt/tags/transcript
    notification-view.ts  # Completion notice renderer: one-line collapsed, stats/preview/link expanded
    fleet-list.ts         # FleetView: one-line summary/disclosure and navigable list below the editor
    fleet-state.ts        # Versioned journal checkpoints, bounded failure details and exact overflow counters
    conversation-viewer.ts # Live conversation overlay for viewing agent sessions
    viewer-keys.ts        # Viewer scroll keys resolved through user keybindings
    agent-mention.ts      # `@` roster (running, resumable, and startable agents) + popup rows
    schedule-menu.ts      # /agents → Scheduled jobs submenu
    select-item.ts        # Collision-safe ctx.ui.select wrapper (numbered rows)
    workflow-card.ts      # Inline workflow card (tool result and session entry)
    workflow-dialog.ts    # /agents → Workflows two-pane inspector

License

MIT — tintinweb