feat(lineage): expose the live agent tree (siblings, children, roots) on request.lineage (#457) - #544

Fixes #457.

Why

Since #421/#444, (await agent()).lineage answers "who is my parent, what is my root, am I a subagent" for the current request, while the warm runtime's lineage registry holds the whole tree and exposes it only through snapshot() (dumps, the Workbench). A plugin that coordinates agent threads (examples/worktree-proximity) therefore hand-rolled an actor tree — kind: root|child, parentSessionId, status: active|stopped — fed from session/start, agent/start, and stop, duplicating what the registry already knew.

Design

• AgentLineage.tree?: AgentLineageTree (@agent-bundle/runtime) — { siblings, children, roots } of AgentLineagePeer ({ conversation, depth, parent?, startedAt, subagent?, resolution }). Computed at resolve time from LineageState.nodes by the same registry that placed the request, on every path that produces a placed lineage: observe() for event routes, resolveToolCall() for MCP calls correlated through a hook window or a Codex _meta whose thread the registry saw start. It rides the existing Observed<AgentLineage> axis, so unavailability semantics and the context.lineage harness seam are unchanged, and it is plain frozen data (structuredClone-safe), so the Flight worker receives it exactly as the entry process resolved it.
• Scope (the proximity model): siblings is every other live conversation under the caller's root, any depth, oldest first — the root itself included for a subagent — so a coordinator sees the whole live tree it belongs to; children the direct live children; roots the other live depth-0 conversations, on Cursor only those seen in the same workspace_roots (the rule that already scopes child binding). A Cursor child whose conversation has not spoken yet is listed under its subagent_id.
• Provenance — every peer field is a registry fact; each peer's resolution is judged exactly as it would be on a request that node made itself (nodeResolution, extracted from resolve()). Stopped nodes are never listed. No tree when the registry did not place the request: a standalone hook (resolveStandaloneLineage), or a Codex _meta naming a thread the registry never saw start, still resolves its own chain and carries no tree. Nothing derives an actor or a user (Actor principal for generated event scopes: typed unavailable reasons and the one derivable host field (parked from #233/#223) #391); request.lineage stays the only identity-adjacent surface (feat(lineage): request.lineage is the only identity-adjacent surface; scoped, self-correcting Cursor child binding #444).
• Not changed — AGENT_REQUEST_STORE_VERSION (the handle shape is unchanged; tree is an optional field of an existing value), the Workbench replay provenance (a deterministic replay has no registry and never produces a tree), the registry's mutations and journal.

examples/worktree-proximity

Deleted the actor bookkeeping the issue names: ActorSchema (kind, parentSessionId, status, the per-actor provenance.id/parentSessionId), the actorObserved and actorStopped events, and the session:<root> pseudo-id. The state definition is now worktree-proximity/intent — bindings (actor → worktree, the one fact the registry cannot know), activities, refusals — with actorBound, actorReleased, intentRecorded, edgeRefused. session/start and agent/start bind conversations under the ids request.lineage.tree lists them by; the edge itself is the registry's. tool/before takes liveness from lineage.tree.siblings (liveConversations()): an intent held by a host-identified actor the tree no longer lists under our root is stale and warns nobody; a derived worktree:<root> actor is never filtered that way; a lineage without a tree presumes nothing. New src/events/agent/stop.tsx route (so the registry marks children stopped and the route releases their binding and intent). The coordinator status tool reports agents — the tree the runtime resolved for its call (agentTree()), honestly unavailable for a client no hook window names — beside bindings, activities, refusals. The agent-topology provider stays an honest stub until #459 (providers still receive no lineage). README rewritten accordingly.

examples/host-test: dump --conversation now matches the request's own chain (conversation/parent/root/subagent.id) instead of a JSON substring, so a sibling listed in tree does not make a record match.

Tests

• rsc-runtime/tests/lineage-registry.test.ts — new lineage tree (#457) block: Claude 2.1.259 orchestration replay asserting siblings/children/roots at 15 rows (parallel + nested spawns, stops, the resume), peer resolution/startedAt/subagent against the registry's own facts and against the peer's own requests; Codex 0.147.0 (hook and _meta paths, tree absent for a cold thread); Cursor 3.18.25 (pending child under its subagent_id, bound child, nested child, stops); synthetic two-root Claude and two-workspace Cursor cases for roots; frozen + structuredClone round trip. Two toEqual assertions in lineage-codex-rollout.test.ts gained the tree.
• examples/worktree-proximity/tests/route-unit/routes.test.ts — journeys 1–8 pass on the new state shape; new: the child sees a sibling's intent while the tree lists it and stops once the tree drops it (the acceptance test), status renders the injected tree, agent/stop releases a child; tests/proximity.test.ts covers the liveness filter and the derived-actor exemption.
• packages/agent-bundle/tests/worktree-proximity-journeys.test.ts (cross-process) — children are now spawned Claude-style (Agent PreToolUse → SubagentStart → PostToolUse naming the child) and status is called inside the root's hook window: the root sees both children confirmed in the registry-fed tree, still after a server restart (the lineage journal is durable beside the intent state), and only agent-a after agent-b's SubagentStop, which also releases its binding and intent. Exact-revision equality across the restart still holds.

Docs

docs/entry-conventions.md (lineage section: shape and semantics of tree), website/docs/{en,zh}/guide/authoring/mcp.mdx (request-context lineage paragraph). Changeset: .changeset/457-lineage-tree.md (@agent-bundle/runtime patch).

Verification

pnpm typecheck, pnpm lint, pnpm test:unit (3257 passed), pnpm test:route-unit (58), worktree-proximity-journeys.test.ts (integration pool), examples/worktree-proximity check (validate, typecheck, unit, route-unit), examples/host-test typecheck + route-unit, pnpm docs:site:build.

Coordination

Rebased onto main at dacb5b0ff (#536): the proximity route-unit suite was rewritten on top of #525's mountTestState, so this PR's new tests use that harness API too. #532 (request.plugin) touches the same generated shells but not the lineage path and rebases cleanly either way. #459 builds on this PR's tree and is a separate PR; #458 (#539), #460 (#541), #532 (request.plugin), and #466 (#545, canonical.payload on event routes) landed on main while this PR was open and were merged in (6e5936018, 177869456): the proximity example now addresses notices through noticeRecipientFor(binding, worktreeRoot) — a binding whose provenance is derived gets recipient.workspace.root, every other binding recipient.conversation (its actor id is the lineage conversation, so #539's session: prefix stripping is gone with the prefix); status reports #460's notices counts beside agents/bindings/activities; and every event route, including the new agent/stop, reads canonical.payload (agentId/sessionId/toolName/toolInput) instead of the native envelope.

Review status

• 73bdb9e (PR opened): Codex review completed 2026-09-04T07:20Z with no findings (no review threads).
• Unreviewed since: 25f6f8ae9 (pins the tree on the two mcp-lineage projection assertions that failed the first CI run) and the rebase onto dacb5b0ff (head 1d22eec96, test file re-applied on mountTestState, no source changes). No PR comments are posted from this side, so no @codex review was requested; the connector does not re-review on its own after a rebase.
• 6e5936018 and 177869456 (merges of main at 6c2f8da0a and 55026f076): conflict resolution only in examples/worktree-proximity/**, packages/agent-bundle/tests/worktree-proximity-journeys.test.ts, and the two mcp.mdx pages (both sides kept: the tree paragraph followed by Expose a terminal/IO capability to routes and rendered scripts (TTY-ness, color, columns) instead of each plugin re-detecting it #511's terminal section); no packages/rsc-runtime source changes. Unreviewed; verified locally with pnpm typecheck, pnpm lint, examples/worktree-proximity check (6 unit + 16 route-unit), examples/host-test typecheck + route-unit (7), worktree-proximity-journeys.test.ts (integration pool), and pnpm docs:site:build.
• GitHub Actions created no pull_request runs for the pushes between f2894fc7c and the rebase (close/reopen and a fresh SHA did not help while other PRs' pushes did trigger); the rebased push triggered normally.