Surface live subagent rosters and interaction attribution

[bradhallett]

Human comments

This is the workflow I actually miss: I run orchestrator fan-outs through one acp-omp thread (scouts, reviewers, implementers), and today that session is a black box — I can't see which children exist or which child is asking me to approve something. This PR is the visibility layer for that.

The judgment calls I'd most like maintainer eyes on are called out in the deviations section, but to surface them: the capability ships dark on purpose (no shipped provider emits rosters today — I verified omp 18.1.7 emits nothing over ACP — and I'd rather reviewers trust a capability bit than see empty rosters), snapshot-replace instead of per-agent deltas, and the daemon-wire agentLabel stays dormant until the bridge protocol grows it. If any of those read wrong to you, the contract layer is where to push back — everything downstream follows those choices.

It's stacked on #3040 because the ACP translation sits on the plan-capability plumbing; happy to rebase independently or trim the ACP half entirely if you'd rather land the core slice first. Agent-built with the same local gate discipline as #3040 (clean-main worktree classification for every residual failure); I reviewed the schema and migration work myself.

What was wrong

BB had no visibility into a provider session's subagent mesh: a thread's child agents were invisible to clients, and pending interactions raised inside a subagent could not be attributed to the agent that asked. OMP threads running orchestrator workloads (parallel scouts, reviewers, task fan-out) looked like opaque single-agent sessions, and the desktop/CLI surfaces had no roster to inspect. The bridge protocol, daemon contract, and persistence layer had no vocabulary for any of this. Fixes #3020.

What changed

Contract layer (domain, provider-bridge-protocol, host-daemon-contract):

• New provider event thread/subagents/updated carrying a full roster snapshot (ThreadSubagent { id, label, state, summary, transcriptRef }), thread-scoped, snapshot-replace semantics (latest event wins, mirrors thread/goal/updated).
• Delta grammar v3 extended in place (additive, capability-gated — same precedent as thread.goal.clear): delta kind thread.subagents, new bridge handshake capability subagents (default false). Snapshot regenerated.
• HOST_DAEMON_PROTOCOL_VERSION 178 → 179: HostDaemonBridgeLaunch capabilities gain required supportsSubagents.
• pendingInteractionBaseSchema gains agentLabel: string | null (default null) — a single source inherited by every provider and plugin interaction variant; legacy payloads parse unchanged.
• Domain providerCapabilitiesSchema gains supportsSubagents (default false) so clients can see which providers report rosters.

Persistence: migration 0113_subagent_roster_state (pending_interactions.agent_label; partial index widened to include thread/subagents/updated), drizzle meta regenerated, migrate.ts validator and test rewind helpers per repo convention.

Server + capability chain: agent-runtime adapter derives supportsSubagents live from the handshake (getter, mirroring approvalEnforcedBy); plugin-sdk declaration normalizes absent → false; server surfaces it through ProviderInfo.capabilities and the daemon launch payload; timeline read path gains subagents (mirrors goal) with latest-page extraction in thread-view.

ACP bridge (dormant translation): wire schema for sessionUpdate: subagents_update reusing the domain ThreadSubagent schema; translation to the thread.subagents delta (thread-scoped, safe outside turns — usage_update precedent); emission gated per-provider by the acpSubagents bridge option. Honesty rule: no shipped agent enables it — binary-verified that omp 18.1.7 emits no roster over ACP (subagent state exists only on its internal persistence format; the ACP session-update validator is a closed switch), so acp-omp ships with the capability off. The fake ACP agent gains FAKE_ACP_SUBAGENTS=1 roster emission for tests.

CLI + SDK + docs parity: bb thread subagents <threadId> [--json] with honest empty states; SDK method, guide template chapter, bb-cli skill entry, configuration.md. Plugin SDK 0.4.40 → 0.4.41 (bump-plugin-sdk.mjs --patch, mandated by the npm version guard since the shipped artifacts change); bundled d.ts regenerated.

Desktop app: ThreadSubagentsCard roster panel (mirrors ThreadTodoCard) — collapsible, running-count header, rows sorted running > idle > parked > aborted with state badges, summary lines; hidden when no roster; events-appended invalidation refetches so snapshots arrive live. Pending-interaction banners render Agent: {label} when attributed.

Deviations from the issue's proposal: none material — the daemon wire does not carry an agent label yet, so plugin-origin interactions stay null until the bridge protocol grows the field (dormant plumbing, same honesty rule as the roster capability).

How you verified

Full CI parity locally: turbo run typecheck test lint --concurrency=4 + turbo run build (19/19) + app bundle budget + provider-literal ratchet + plugin-SDK npm version guard — all green. Every residual test failure reproduced identically on a clean worktree of main (tasks/theme-preview jsdom localStorage, integration-harness esbuild, desktop electron binary, host-workspace, server install-machine-script) — local-environment baseline, none in this PR's blast radius; plugin-build failures dropped from 4 to 2 vs main after the bundled-d.ts regen. oxfmt --check: zero files from this diff in the failing set (192-file head drift is un-gated pre-existing main state). Migration round-trip incl. fabricated-history replays; conformance subagents/roster-state; grammar snapshot; bridge tests cover emission → assembler thread scope, snapshot replace, and silent drop without the capability; UI tests cover render, sort, badges, hidden states, and attribution.

Depends on #3040 (stacked: only commit 853f7b5fe is new here; merge that first — or review this as a single combined diff).

Fixes #3020

AGENT GENERATED