# Extension and integration APIs

Public seams for other Pi extensions and host integrations: the in-process RPC, the structured delegation API, launch preflight, capability ceilings, the background-work provider contract, and the Herdr integration.

## Trusted workflow resources

Loaded trusted TypeScript extensions can import `registerWorkflowResource` from `pi-subagents/workflow-resources`. This subpath does not load the main extension and exposes no resolver or permit constructor. Its exported types are `RegisterWorkflowResourceInput`, `WorkflowResourceDefinition`, and `WorkflowResourceRegistration`:

```typescript
registerWorkflowResource({
  sessionId: string,
  definition: {
    name: string,
    version: number,
    resolve(args: Readonly<Record<string, unknown>>):
      | { script: string; hostCommands?: readonly { key: string; command: string }[] }
      | { error: string },
  },
}): { dispose(): void }
```

Names are case-sensitive, at most 128 characters, and match `[A-Za-z0-9][A-Za-z0-9._-]*`; use an extension prefix. Versions are positive safe integers. Registration throws for invalid input, protected builtins (`review`, `run-ci`), or duplicate names within the same session. Different sessions may register the same name. Dispose before replacement; there is no silent overwrite.

Register in `session_start` using **`ctx.sessionManager.getSessionId()`**, not the session file path or a tool argument. Dispose in `session_shutdown`. New/resumed/forked sessions and reloads need registration from the replacement runtime's `session_start`; do not retain old `pi`/`ctx` references. The extension owns cleanup, not an automatic registration lifecycle manager. Disposal is idempotent and cannot remove a newer replacement. Missing cleanup can cause a duplicate-registration failure on reload.

`resolve` must do synchronous, bounded validation and string construction, without I/O, SDK calls, timers or process work. Core deep-copies plain JSON args: at most 16 KiB encoded, nesting depth 8, 16 fields per object, 64 items per array, finite numbers, and nonempty strings of at most 16 KiB. The extension must additionally reject unsupported fields and validate resource-specific semantics. Throws, promises/thenables and malformed expansions fail before authority is issued; errors are bounded to 4096 characters.

Host grants bind **exact key/trimmed-command pairs**, not independent sets of keys and commands. At most 32 grants are accepted, with unique safe workflow keys and nonempty commands bounded to 16 KiB without NUL. Omitted grants give no host authority. Core snapshots the expansion and grants at resolution. Disposing stops future lookup, but already-admitted workflows retain captured grants, even after replacement. Use existing stop/deadline controls for cancellation; this does not promise survival of host shutdown or durable named scheduling. Existing child admission and capability ceilings still apply.

### Mixed child and finite host example

This extension owns two fixed commands; `scripts/finite-check.mjs` must be an existing trusted finite helper in the workflow cwd. It runs a reviewer first, then a check. No command or flags come from free-form public args.

```typescript
import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
import { registerWorkflowResource } from "pi-subagents/workflow-resources";

export default function (pi: ExtensionAPI) {
  let registration: { dispose(): void } | undefined;
  pi.on("session_start", (_event, ctx) => {
    registration?.dispose();
    registration = registerWorkflowResource({
      sessionId: ctx.sessionManager.getSessionId(),
      definition: {
        name: "acme.review-check",
        version: 1,
        resolve(args) {
          if (Object.keys(args).some(k => k !== "task" && k !== "check"))
            return { error: "Only task and check are supported." };
          if (typeof args.task !== "string" || !args.task.trim() || args.task.length > 4000)
            return { error: "task must contain 1–4000 characters." };
          if (args.check !== "quick" && args.check !== "full")
            return { error: "check must be quick or full." };
          const command = args.check === "quick"
            ? "node ./scripts/finite-check.mjs --mode quick"
            : "node ./scripts/finite-check.mjs --mode full";
          const host = { kind: "command", command, timeoutMs: 120000 };
          return {
            hostCommands: [{ key: "check", command }],
            script: `
              const review = await runs.run("review", {
                agent: "reviewer", task: ${JSON.stringify(args.task)}
              });
              if (!review.ok) throw new Error("Review child failed");
              const check = await runs.host("check", ${JSON.stringify(host)});
              return { review: review.output, check };
            `,
          };
        },
      },
    });
  });
  pi.on("session_shutdown", () => {
    registration?.dispose();
    registration = undefined;
  });
}
```

Invoke through the public `subagent` tool (use `async: false` for foreground):

```json
{
  "workflow": "acme.review-check",
  "args": { "task": "Review the current change; return findings only.", "check": "quick" },
  "async": true
}
```

The parent evaluates child findings and ordinary command logs/status/terminal receipts; child success is not approval or proof of a clean review. The timeout above bounds the host command, not the whole workflow.

**Trust boundary:** this API composes already-loaded trusted code; it is neither authentication nor a sandbox. Session IDs scope lookup, not authorization between malicious extensions. Core owns opaque permits and provenance; caller-supplied issuer/trust/permit metadata cannot grant authority. Raw public scripts and script paths do not gain host authority, and registration is not an arbitrary-command entry point for public args.

The existing shell runner uses workflow cwd and inherited environment. Exact matching does not pin PATH resolution, executable bytes, repository helpers, or credentials. Those remain operator/extension trust responsibilities. `JSON.stringify` embeds data in JavaScript source; **it is not shell escaping**. Keep commands fixed as above, or validate strictly bounded numeric/hex tokens before binding known positions; never concatenate arbitrary task text or flags into shell commands. No new runner, cwd confinement, CI/merge policy, or SDK lifecycle framework is provided.

## In-process event-bus RPC

Other Pi extensions can use the in-process event-bus RPC instead of scraping slash output or calling internal modules. Listen for `subagents:rpc:v1:ready`, send requests on `subagents:rpc:v1:request`, and read replies from `subagents:rpc:v1:reply:<requestId>`.

```typescript
const requestId = crypto.randomUUID();
pi.events.on(`subagents:rpc:v1:reply:${requestId}`, (reply) => {
  // { version: 1, requestId, success: true, data } or
  // { version: 1, requestId, success: false, error: { code, message } }
});
pi.events.emit("subagents:rpc:v1:request", {
  version: 1,
  requestId,
  method: "spawn",
  params: {
    workflowScript: `return runs.run("main", { agent: "reviewer", task: "Review the current diff" })`,
    context: "fresh"
  }
});
```

The RPC methods are `ping`, `status`, `manage`, `spawn`, `steer`, `interrupt`, `stop`, and `resume`. `status`, `manage`, `steer`, `interrupt`, and `resume` reuse normal package-owned actions.

Method notes:

- `manage` exposes a narrow schedule-only allowlist: `schedule.list`, `schedule.show`, `schedule.history`, `schedule.pause`, `schedule.resume`, `schedule.run`, and `schedule.delete`. All actions except `schedule.list` require `id`. Mission, agent, config, worktree, and arbitrary management actions are rejected before executor dispatch. `ping.capabilities.managementActions` advertises the exact allowlist.
- `spawn` accepts structured single-child execution (`agent`, `task?`), inline `workflowScript`, or `workflowScriptPath` and is async-only: omit `async` or set `async: true`, omit `clarify`, and do not pass management `action` values. Relative script paths resolve against the request `cwd`. It goes through the same executor as the `subagent` tool, so agent discovery, validation, session attribution, configured spawn caps, child-safety depth, artifacts, and async status all behave the same.
- `steer` requires an async run `id` (plus optional child `index`) and a non-empty `message`; its reply preserves the normal acknowledged-delivery result. Optional `mode` values are `steer` (default), `follow_up`, and `auto`, and receipts include `deliveryStatus: "delivered" | "queued"`. RPC steering disables the direct tool's pause-and-revive recovery in every mode so an extension keeps authority over the exact child it spawned; `ping.capabilities.nonRecoveringSteer` advertises this guarantee.
- `resume` requires a run target and non-empty `message`. It delegates to the existing revival path, which validates current-session ownership, persisted session/recovery metadata, stopped/live state, capability ceilings, and the exclusive session lease before returning the new async run details. Callers may request a `file-only` output path for the revived result without overriding its model, tools, or budgets. `ping.capabilities.resume` advertises this seam.
- `stop` targets current-session top-level async runs through the stop control channel and records a `stopped` lifecycle instead of reporting a timeout.
- `status` keeps targeted and rich requests on the executor-backed path. A request with no `id`, `runId`, `dir`, `index`, `view`, or `lines` may use the restored in-memory projections and a short summary; when the live state is missing, stale, session-mismatched, or not restored, it falls back to normal executor status. Status `view`, `lines`, and `index` are forwarded for targeted transcript/fleet requests. Successful replies retain `text`, `details`, `fleet`, and `asyncSnapshot`; the short summary intentionally omits canonical filesystem details, wait subscriptions, and budget annotations.

Capability advertisements on `ping`:

- `events.asyncComplete` — exact process-local completion correlation after RPC `spawn`.
- `managementActions` — exact schedule management actions accepted by RPC `manage`.
- `launchResolvedExtensions` — the optional launch-resolved extension projection in status details.
- `runtimeAcknowledgedExtensions` — the optional child-runtime acknowledgement projection and event name.
- `processTerminalProof` — the process-terminal proof status (see [observability.md](observability.md#process-terminal-proof)).
- `nonRecoveringSteer` — RPC steering never pauses-and-revives.
- `resume` — the revival seam described above.
- `statusProjection: { version: 1, untargeted: "in-memory-when-ready", targeted: "executor" }` — untargeted status may use restored bounded projections; targeted or rich status remains executor-backed.
- `fleetStatus: { version: 1 }` — successful `status` replies additionally include `data.fleet`.

Structured delegation progress updates carry `runId` as soon as foreground execution allocates it, so a caller can retain the package-owned revival target even if its own tool turn is interrupted before the terminal response. Foreground `details.results[]` rows also include a numeric `index` that is unique within the run and stable across partial progress snapshots and the final result; use `(runId, index)` instead of row position to correlate single, counted parallel, and chain children.

### Fleet status DTO

When `ping.capabilities.fleetStatus` is `{ version: 1 }`, successful `status` replies include `data.fleet`: `{ version: 1, entries, totalActive, omitted }`.

Entries are bounded, current-session public display records with an opaque reconciliation `key`, resolved `agent`, optional `role`, `model`, `effort`, caller-facing `goal`, safe `startedAt`, and `{ input, output, total }` tokens. `totalActive` and `omitted` preserve overflow information beyond the bounded entry window.

The DTO intentionally never exposes run, async, or tool IDs. Clients must ignore unknown fields and fall back to status text when the capability is absent.

`data.asyncSnapshot` is a separate bounded projection included on successful status replies when available. Its `runs[].id` contains the async run id; unlike the fleet DTO, it is not an opaque display key. Fleet keys remain opaque and must not be interpreted as run or async identifiers.

### Scope

`pi.events` is in-process only. It does not reach separate Pi processes or child subagents; use the file lifecycle artifacts or `pi-intercom` for cross-process coordination.

## Runtime agent registration from independent extensions

An independently installed Pi extension can register an agent with the installed `pi-subagents` owner through the process-local `pi-subagents:runtime-agent-register:v1` event. Emit after extension setup, such as during `session_start`. Event delivery is synchronous, so the owner writes the result onto the request before `emit()` returns.

```typescript
const request: {
  version: 1;
  name: string;
  definition: {
    description: string;
    systemPrompt: string;
    tools?: readonly string[];
  };
  result?:
    | { ok: true; registration: { dispose(): void } }
    | { ok: false; error: Error };
} = {
  version: 1,
  name: "runtime-probe-agent",
  definition: {
    description: "Agent registered by an independent extension",
    systemPrompt: "Return the words runtime probe.",
    tools: [],
  },
};

pi.events.emit("pi-subagents:runtime-agent-register:v1", request);
if (!request.result) throw new Error("pi-subagents is not installed or not ready");
if (!request.result.ok) throw request.result.error;
const registration = request.result.registration;
// Call registration.dispose() during your extension cleanup.
```

If `pi-subagents` is a resolvable dependency of the consumer package, `pi-subagents/agents` exports `RUNTIME_AGENT_REGISTER_EVENT`, the request/result types, and `registerAgentViaEvents()` for the same contract. A separately installed Pi package is not automatically a Node dependency of another package. In that case, use the event contract directly instead of a runtime import. A type-only development dependency is optional.

The installed owner applies the existing runtime-agent validation, collision checks, limits, runtime source metadata, and cleanup. If more than one owner listens, the first handler that writes `request.result` wins. Unsupported versions, malformed requests, and registration failures return `{ ok: false, error }`. No result means no compatible owner handled the event.

This contract is process-local. It does not register agents in child sessions or other Pi processes, and it does not change package discovery or package resolution.

## External jobs in FleetView

Use `pi-subagents/external-runs` to publish display-only current-session jobs owned by another extension:

```ts
import {
  registerExternalRun,
  updateExternalRun,
  unregisterExternalRun,
} from "pi-subagents/external-runs";

registerExternalRun({
  id: "dependency-review",
  sessionId: ctx.sessionManager.getSessionId(),
  source: "interactive-shell",
  label: "Dependency review",
  state: "running",
  startedAt: Date.now(),
  currentAction: "Inspecting package metadata",
});

updateExternalRun(ctx.sessionManager.getSessionId(), "dependency-review", {
  state: "completed",
  updatedAt: Date.now(),
  endedAt: Date.now(),
  preview: "No dependency blockers found.",
  reportPath: "/tmp/dependency-review.md",
});

unregisterExternalRun(ctx.sessionManager.getSessionId(), "dependency-review");
```

The API validates and caches bounded display fields when the caller registers or updates a job. FleetView reads that cache only. It does not poll caller code. `snapshotExternalRuns(sessionId)` and `listExternalRuns(sessionId)` return bounded current-session snapshots. Snapshots filter the session-qualified cache key before inspecting record fields; API-written records avoid repeated normalization through module-private provenance, while records replaced or mutated through the process-local registry are validated on demand. By default, malformed records for the requested session throw with the validation error. Display-only Fleet callers can pass `{ ignoreMalformed: true, onMalformedRecord }` to remove bad records and keep rendering with a programmatic diagnostic.

External jobs are observational. The caller owns execution, persistence, cancellation, and result delivery. FleetView does not expose stop, steer, resume, cancel, or Herdr controls for them. Supplied report and transcript paths are shown as bounded text only; FleetView does not read arbitrary external paths.

## Launch contract preflight

Use `pi-subagents/preflight` when an extension needs to inspect the resolved child launch contract before deciding whether to run anything:

```ts
import { resolveSubagentLaunchContract } from "pi-subagents/preflight";

const result = await resolveSubagentLaunchContract({
  agent: "reviewer",
  task: "Review the current diff.",
  context: "fresh",
  cwd: ctx.cwd,
  sessionRoot: "/tmp/my-extension-preflight-session-root",
  availableModels: ctx.modelRegistry.getAvailable(),
});

if (!result.ok) {
  // missing_agent, ambiguous_agent, missing_skill, denied_required_tool,
  // invalid_artifact_dir, invalid_cwd, or unsupported_mode
  throw new Error(result.message);
}

console.log(result.contract.digest, result.contract.tools.effectiveAllowlist);
```

Preflight covers ordinary single-agent launch resolution:

- Selected agent identity and shadowed candidates.
- A parsed-definition digest, including system prompt and launch-affecting model, tool, skill, extension, output, and memory fields.
- Fresh/fork context, effective model and thinking, skill and tool resolution, direct MCP selections, runtime/configured extensions.
- Artifact/session paths, async lifecycle/status/result/event/process-terminal paths, package/lifecycle versions, capability-ceiling audit data, and stable digests.

`launchContractDigest` is the canonical digest of the caller task, effective system prompt, model candidates, effective tools/extensions/MCP (including inherited capability ceilings), output binding, and structured-output schema that ordinary foreground and async execution report in results/status/events and metadata.

Boundaries:

- Runtime acceptance prose and output-task annotations are intentionally excluded because side-effect-free preflight does not resolve those host/runtime augmentations; the launch and task digests make that boundary explicit.
- Raw prompts are not exposed in public contract output.
- It is side-effect-free for launch state: it does not create child sessions, temp prompt files, structured-output runtimes, tool-diagnostic files, or run artifacts.
- Some host-owned facts, such as exact fork snapshots, nested async roots, and live model registries, can only be proven by the Pi host; those appear as `host_required` diagnostics instead of silently pretending to be exact.
- Preflight reads the extension config, so `defaultSubagentContext: "fresh"` or `"fork"` affects omitted context in the same way as execution. Explicit `context` still wins.

## Structured delegation API

Other Pi extensions can ask `pi-subagents` to run one configured foreground leaf agent through the structured delegation API. It uses the established `prompt-template:subagent:*` event family and the same executor as the `subagent` tool; it does not add another launcher.

```ts
import {
  SUBAGENT_DELEGATION_REQUEST_EVENT,
  SUBAGENT_DELEGATION_RESPONSE_EVENT,
  type SubagentDelegationRequest,
  type SubagentDelegationResponse,
} from "pi-subagents/delegation";

const request: SubagentDelegationRequest = {
  requestId: crypto.randomUUID(),
  ownerRunId: workflowRunId,
  nodeId: "review-accuracy",
  agent: "reviewer",
  task: "Review the supplied evidence.",
  context: "fresh",
  cwd: ctx.cwd,
  thinking: "high",
  result: {
    kind: "structured",
    schema: {
      type: "object",
      properties: { verdict: { type: "string" } },
      required: ["verdict"],
      additionalProperties: false,
    },
  },
};

const unsubscribe = pi.events.on(SUBAGENT_DELEGATION_RESPONSE_EVENT, (payload) => {
  const response = payload as SubagentDelegationResponse;
  if (response.requestId !== request.requestId) return;
  if (response.ownerRunId !== request.ownerRunId || response.nodeId !== request.nodeId) return;
  unsubscribe();
  // Inspect response.status, response.result, response.usage, model, and thinking.
});
pi.events.emit(SUBAGENT_DELEGATION_REQUEST_EVENT, request);
```

Identity:

- `ownerRunId` plus `nodeId` is the active logical identity; `requestId` identifies one attempt.
- A second active attempt for the same logical node receives `duplicate_node` without disturbing the original.
- Started, update, response, and cancellation payloads carry the full tuple. Cancellation affects only an exact tuple, including cancel-before-start races.
- Each attempt emits at most one terminal response.

Results:

- Result mode is explicit. Text remains literal even when it looks like JSON. Structured mode returns the separately captured, schema-validated JSON value.
- Terminal usage reports input, output, cache-read, cache-write, cost, turns, tool calls, and duration alongside the effective model and thinking level when known.

Live update events are bounded progress snapshots, not patches, so consumers should replace the prior snapshot rather than merge it as a delta. Structured delegation coalesces heartbeats whose delegation-visible progress and recent output are unchanged; a duration-only heartbeat therefore does not produce another update. The terminal response remains authoritative for the complete result, error, and final usage details.

Bounds:

- Schemas are capped at 64 KiB; tasks and returned text/structured values are capped at 1 MiB, with smaller bounds on identity/configuration strings and a maximum `timeoutMs` of 2,147,483,647.
- Structured delegation accepts `toolBudget: { hard: 0, block: "*" }` to block the first tool call and run a zero-tool leaf; ordinary model-facing/configured budgets keep their existing minimum of one.
- The foreground bridge retains up to 8,192 exact pending-cancellation and settled-attempt identities per extension context. If either history fills, it fails closed with `unavailable_context` for later starts rather than evicting identity facts; lifecycle reset clears the bounded history.

Constraints:

- Delegation requires an active extension context. Emit requests from a supported event callback or queued application step, not by recursively invoking the `subagent` tool inside another tool's `tool_call` hook.
- The caller selects a configured agent, but agent discovery and effective tools remain package-owned. A request cannot grant arbitrary tools, and tool restrictions are not an operating-system sandbox.
- The detached RPC remains async-only; this API is foreground-only.

Unversioned prompt-template payloads with `requestId`, `agent`, `task`, `context`, `model`, and `cwd` are rejected as legacy direct delegation. New integrations must use the structured owned-leaf request above. `pi-subagents/delegation` is the canonical contract for extension integrations.

## Capability ceilings

Parent extensions can enforce an out-of-band, session-scoped capability ceiling without adding a model-visible field to `subagent`:

```ts
import { registerSubagentCapabilityCeiling } from "pi-subagents/capability-ceiling";

const restriction = registerSubagentCapabilityCeiling({
  sessionId: ctx.sessionManager.getSessionId(),
  source: "plan-mode",
  ceiling: {
    allowedAgents: ["plan-scout", "plan-researcher", "plan-reviewer"],
    allowedTools: ["read", "grep", "find", "ls"],
    denyExtensions: true,
  },
});
// restriction.update(...) replaces this provider's policy atomically.
// restriction.dispose() removes only this provider's registration.
```

Semantics:

- Active registrations intersect their `allowedTools` and `allowedAgents` sets and OR `denyExtensions`.
- An explicit empty list means no caller-facing tools or launchable agents for that field; an omitted list does not restrict names.
- `allowedAgents` entries are canonical agent names and are case-sensitive.
- Launching a non-allowlisted agent fails before spawn, and `{ action: "list" }` keeps restricted agents visible in a separate non-executable section instead of silently hiding them.
- The resolved snapshot is propagated monotonically to nested and async children and is retained for recovery.
- `structured_output` may remain as a package-owned internal protocol tool when an output schema requires it; it is not a caller capability.
- A denied lazy-skill `read` requirement fails before spawn rather than widening the ceiling.

`denyExtensions` suppresses ambient, configured, and MCP provider extensions while retaining the package runtime needed for child protocol enforcement. This is a same-process policy boundary, not a sandbox against malicious code already running in the parent process.

Schedules created while a ceiling is active are rejected until durable schedule persistence is available; unrestricted schedules remain subject to any policy active when they fire. Public status exposes bounded audit counts and sources, never full extension paths.

## Background-work provider API

Other Pi extensions can make their current-session jobs visible to `bg_wait` through the process-local provider contract:

```ts
import { registerBackgroundWorkProvider } from "pi-subagents/background-work";

const dispose = registerBackgroundWorkProvider({
  name: "my-background-extension",
  wakeChannels: ["my-extension:job-finished"],
  listActiveWork: (context) => jobs
    .filter((job) => job.status === "running" && (!context || job.ownerSessionId === context.sessionId))
    .map((job) => ({ id: job.id, sessionId: job.ownerSessionId })),
  reconcile: ({ sessionId, nowMs }) => reconcileJobs(sessionId, nowMs),
});
```

Semantics:

- Each item needs a stable provider-local ID and the exact Pi session ID that owns it. `bg_wait` captures those identities rather than a count, so one job finishing while another starts still satisfies first-completion waits without losing the replacement.
- `listActiveWork` receives an optional `{ sessionId, nowMs }` context during snapshots. Providers can use `sessionId` to avoid scanning unrelated work; existing zero-argument `() => items` providers continue to work, and returned items are still validated and filtered to the exact requested session.
- It filters snapshots to the active session, fails closed if a provider disappears while its work is tracked, and surfaces malformed snapshots or provider errors with provider context.
- Wake channels only shorten polling; validated snapshots remain authoritative.
- Providers share a registry through `Symbol.for("pi-subagents.background-work.v1")`, allowing independently loaded extension modules to meet in one Pi process.
- Registration is reload-safe: a new provider with the same name replaces the old callback, and the old disposer cannot remove the replacement. Call the disposer during extension shutdown when possible.

Children do not gain provider tools or extensions automatically. Add `bg_wait` to the child agent's `tools` allowlist and load each provider through `extensions` or `subagentOnlyExtensions`. The parent's effective `waitTool` setting reaches every child through its typed runtime config; `PI_SUBAGENT_WAIT_TOOL_ENABLED` keeps precedence in the parent.

Foreground children never load the parent's ambient extensions: they share the parent's process, and loading them would start a second copy of every ambient extension, including this one, inside it. Agents that need MCP tools (`mcpDirectTools`, or MCP tools from an ambient adapter such as pi-mcp-adapter) or models from a provider extension must run as background children (`async: true`), which load the ambient extensions inside the detached runner process unless the agent sets `extensions` or the capability ceiling denies extensions.

## External job provider bridge

Extensions that own long-running advisor jobs can register a process-local provider for `runner.type: external-job` agents:

```ts
import { registerExternalJobProvider } from "pi-subagents/external-job-provider";

const dispose = registerExternalJobProvider({
  name: "surf-oracle",
  start: ({ prompt, promptDigest, cwd, runId, stepIndex, agent, options }) => startSurfJob({ prompt, promptDigest, cwd, runId, stepIndex, agent, options }),
  followUp: ({ prompt, parentProviderJobId, requestId, requestDigest, options }) => followUpSurfJob({ prompt, parentProviderJobId, requestId, requestDigest, options }),
  status: (providerJobId) => getSurfJobStatus(providerJobId),
  result: (providerJobId) => getSurfJobResult(providerJobId),
  reattach: (providerJobId) => reattachSurfJob(providerJobId),
});
```

The provider returns handles with `providerJobId`, `state`, optional `handleUrl`/`conversationUrl`, optional `failureCode`/`failureMessage`, and optional `blockingJobId` for capacity conflicts. `result` can also return `output` and/or `artifactPath`.

`followUp(input)` is optional. When it is present, a completed external-job run can be continued with `subagent({ action: "resume", id: "<run>", message: "..." })`. Pi sends the completed parent provider job id plus a stable `requestId` and `requestDigest`. The provider must continue that parent conversation or fail closed. It must not open a fresh thread when the parent conversation is missing.

The async runner process does not import provider internals. It writes operation requests into its async run directory. The parent Pi process services those requests against the registered provider and writes operation responses. If the provider is not registered, the bridge fails closed with an actionable error. If a run is recovered after provider job metadata exists, the runner calls `reattach` and `result`; it does not call `start` or `follow-up` again.

## Herdr integration

When Pi runs inside [Herdr](https://herdr.dev), pi-subagents automatically reports active async-run counts through Herdr pane metadata.

- The bridge is enabled only when Herdr supplies `HERDR_ENV=1` and `HERDR_PANE_ID`; outside Herdr it registers no listeners or timers.
- It restores current-session active runs after `/reload` or `/resume`, refreshes metadata while work is active, and clears it on completion or shutdown.
- The bridge uses Herdr's existing `herdr:blocked` sibling event when an async child needs attention, and emits `herdr:busy` while async work remains. Herdr versions that support the sibling event keep the pane's semantic state `working`; older versions ignore it safely and still display the metadata label while the Pi integration remains the lifecycle authority.
- The owning Pi session is the only publisher for its own pane metadata. When an active workflow has an explicit bounded `label`, the newest active label appears in the summary and compact `title-suffix`; overlapping completion restores the previous active label. Raw task and goal prompts never enter Herdr metadata. Without a label, one active run uses its agent name and two or more use the active-run count. Attention adds `⚠`, and the suffix is cleared when active work reaches zero.

To show the reported label in the expanded Agent sidebar, include `state_text` or `$summary` in its row layout:

```toml
[ui.sidebar.agents]
rows = [
  ["state_icon", "workspace", "tab"],
  ["agent", "state_text"],
]
```

### Inspector panes

Herdr 0.7.5+ can open an on-demand inspector for an existing async run:

```ts
subagent({ action: "inspector.open", id: "<run-id>", index: 0, focus: true })
subagent({ action: "inspector.status", id: "<run-id>", index: 0 })
subagent({ action: "inspector.close", id: "<run-id>", index: 0 })
```

The inspector is a raw dashboard pane, not the child session and not a literal attach. It reads lifecycle/status/output/mission artifacts and sends `steer` or `stop` through pi-subagents' existing control inbox. Closing it never stops the run.

Herdr remains optional. Ordinary launches stay headless, and missing/older Herdr versions affect only Herdr-specific inspector and project-pane actions. FleetView opens the selected active async child with `H`. Use `focus` only with `inspector.open`; Herdr 0.7.5 cannot focus an arbitrary existing raw pane id.

### Project panes

For substantial work in another codebase, Herdr 0.7.5+ can open a project-owned Pi pane rooted in that repository:

```ts
subagent({ action: "project.open", cwd: "/path/to/repo", message: "Own the auth refresh mission for this project." })
subagent({ action: "project.status", cwd: "/path/to/repo" })
subagent({ action: "project.close", cwd: "/path/to/repo" })
```

A project pane runs its own Pi session in the target directory, so subagents launched from that pane use that project's config, agents, skills, files, git state, and missions. The parent session keeps coordination authority, but it does not own or control the subagents inside the peer pane. Existing headless runs are not moved into the pane. Pane bindings live under `<projectRoot>/.pi/subagents/project-panes/herdr.json` and are only a local pointer to the Herdr pane.

Other Pi extensions should use the versioned public TypeScript surface instead of invoking the model-facing tool or importing inspector internals:

```ts
import {
  PROJECT_PANES_API_VERSION,
  openProjectPane,
  getProjectPaneStatus,
  focusProjectPane,
  closeProjectPane,
} from "pi-subagents/project-panes";

const opened = await openProjectPane({ cwd: "/path/to/repo", focus: false });
const status = await getProjectPaneStatus({ cwd: "/path/to/repo" });
const focused = await focusProjectPane({ cwd: "/path/to/repo" });
const closed = await closeProjectPane({ cwd: "/path/to/repo", requireIdle: true });
```

The API returns discriminated structured results with canonical project root, binding path, pane identity, bounded Herdr runtime fields, stable error codes, and `PROJECT_PANES_API_VERSION: 1`.

- Close fails closed unless the saved pane id is still verified for that project and Herdr explicitly reports `agent_status: "idle"`. `requireIdle` is retained for callers that already pass it, but it cannot weaken that rule.
- Focus uses the saved pane id, asks Herdr for its `tab_id` or `workspace_id`, and then calls the matching Herdr focus command.
- The API reports `trust: "human-verification-required"`. It never bypasses or claims to attest Pi's project-trust prompt.

## Host session lifetime and completion wakes

A host that embeds this extension owns whether completion wakes can be delivered at all.

Ordinary async and foreground completion wakes use `registerSubagentNotify` and `sendCompletion`. They listen for completion events and deliver through `pi.sendMessage(..., { triggerTurn })`. Session shutdown stops the result watcher and disposes this completion notifier. `createWaitSubscriptionManager` is separate: it is the explicit non-blocking `bg_wait` subscription path for work without native notification, not the ordinary completion wake path.

Detached children do not stop when the session does. They are the host process's children, not the session's, so the run keeps going, completes, and notifies nobody. What is lost is the notification, not the work.

This matters because "is the parent busy?" is the wrong idle signal. A parent that launches a detached run and hands control back — which is what the async launch output tells it to do — is not prompting, streaming, compacting, or running a shell command. A host that reaps sessions on those signals alone will dispose exactly the session that was waiting to be woken.

When pi-subagents runs inside a compatible pi-web host, it discovers the versioned `Symbol.for("@agegr/pi-web/session-liveness/v1")` registry and registers one provider for the current session. The provider reports live `queued`/`running` async jobs, active nested descendants (including foreground routes retained after their direct parent settles), foreground controls that still have a scheduling owner or active child, and completion notifications waiting for their batch-delivery timer. Retained terminal history, future schedules, and wait subscriptions do not make a session live by themselves. The registration is replaced on session changes and released during runtime shutdown or reload; other hosts remain unaffected.

If your host reclaims idle sessions, keep a session alive while it still has live detached work:

- Read run state from the status files under the async run directory rather than from event traffic. A long, quiet workflow sends almost nothing to the parent, so recent-activity heuristics conclude the wrong thing.
- Treat `queued` and `running` as live, matching `isActiveAsyncState`. An interrupted run that is `paused` is finalized. A workflow that paused because a child used `contact_supervisor` still has a live child; keep that parent session until reconcile writes `complete` or `failed`.
- Do not treat `lastUpdate` as a heartbeat. The runner advances it in memory every second but only rewrites `status.json` when the activity classification changes, so a live run inside one long quiet tool call leaves a stale file behind. Judging liveness by file age will reap exactly the run you meant to protect.
- Prefer the recorded runner `pid`, which stays true through a silent tool call and goes false when the runner dies. Keep file age only as a fallback for runs that record no pid, and give it a wide window.
- Match `sessionId` in `status.json` against both forms. It is resolved as `getSessionFile() ?? getSessionId()`, so it is normally the parent's session *file path*, but a session that is not persisted records a bare session id instead.

The symptom when this is missed is quiet and easy to misattribute: subagents appear never to report back, which looks like a fault in this extension rather than in the host that disposed the listener.

## Runtime files

The main runtime files in this repository:

| File | Purpose |
|------|---------|
| `src/extension/index.ts` | Extension registration, tool registration, message/render wiring. |
| `src/integrations/pi-web-session-liveness.ts` | Optional pi-web idle-eviction liveness bridge. |
| `src/agents/agents.ts` | Agent and chain discovery, frontmatter parsing. |
| `src/runs/foreground/subagent-executor.ts` | Main execution routing for single, parallel, chain, management, status, interrupt, and doctor actions. |
| `src/runs/foreground/execution.ts` | Core foreground `runSync` handling: drives one in-process child session per attempt. |
| `src/runs/shared/child-session.ts` | In-process child session factory (`createAgentSession` behind an injectable seam) and the shared model runtime; used by both launch paths. |
| `src/runs/shared/child-launch.ts` | Builds the tool plan, typed child runtime config, and session launch for a child in either host process. |
| `src/runs/shared/child-tool-plan.ts` | Tool, MCP, and extension resolution for a child launch. |
| `src/runs/shared/child-runtime-config.ts` | `ChildRuntimeConfig`: everything the child-side hooks need. |
| `src/runs/shared/child-hooks.ts` | The inline hook extensions every child gets (prompt runtime, fast mode, fanout). |
| `src/runs/background/subagent-runner.ts` | Detached async runner; hosts background child sessions in its own process. |
| `src/runs/background/run-child-session.ts` | Drives one background child session and mirrors its events into the run artifacts. |
| `src/runs/background/runner-aliases.ts` | Aliases the host peer packages to the installed pi package for the runner (`JITI_ALIAS`). |
| `src/runs/background/async-execution.ts` | Background launch support. |
| `src/runs/background/async-status.ts` | Status discovery and formatting for async runs. |
| `src/workflows/scripted-workflow.ts` / `src/runs/foreground/subagent-executor.ts` | Scripted workflow orchestration and child launch routing. |
| `src/shared/settings.ts` | Chain behavior, instructions, and config helpers. |
| `src/runs/shared/worktree.ts` | Git worktree isolation. |
| `src/intercom/intercom-bridge.ts` | Runtime intercom bridge instructions and diagnostics. |
| `src/extension/schemas.ts` / `src/shared/types.ts` | Tool schemas, shared types, and event constants. |
| `test/unit/` / `test/integration/` | Unit and loader-based integration tests. |