Related: #139420

What Problem This Solves

Fixes an issue where users watching a Swarm in the Control UI would lose its display when all children finished, even if the parent failed before producing a final response. Deleted child sessions could no longer supply the full outcome, parent-only viewers could miss cross-agent updates, and child details depended on hover.

Why This Change Was Made

The existing collector registry now supplies a bounded summary for the authorized parent: up to four active groups and the most recent completed group. Counts include retained collectors after child-session cleanup; only a selected-parent read includes bounded member identities and statuses. Child names remain separate authorized reads. Completion counts come from frozen collector results and do not imply that the parent produced a synthesis.

The registry's existing commit boundary emits parent-scoped updates after successful changes, including archive removal. A small process-local committed snapshot prevents failed best-effort writes from suppressing a later successful notification. Atomic publication waits for owner snapshots; child completion never changes parent run status. Ordinary message and agent events retain scoped reads and omit uncomputed Swarm data.

Native details/summary replaces the hover-only display. The component loses obsolete phase wrappers, owner inference from group strings, and redundant marker-limit handling. Raw session keys follow the admitted owner; sessions.describe gains an optional agentId using the existing owner resolver, while qualified key-only callers remain valid.

User Impact

Users can see queued, running, completed and failed totals, open child details with keyboard or touch, and reload without losing retained outcome counts. Missing child details do not hide successful parent counts. Removed or denied parents clear the display, and late responses cannot restore stale ownership.

Counts expire under the existing collector retention policy. Native app widgets retain their existing active-only behavior. There is no new configuration, dependency, database schema, stored representation, retention policy, or protocol-version change. Independent database contention and provider quota failures are separate from this display repair.

Production is net +247 lines (including brace-only formatting required by CI), including the committed-parent notification owner; the component is net -98 and activity tracking -11. The remaining growth supplies the retained read projection and lifecycle delivery, rather than an additional store or compatibility path. Generated Swift adds eight lines.

Evidence

The integrated candidate passed 524 focused tests: 45 Gateway, 19 registry, 417 UI, 40 board, and three browser cases. Final reruns passed 32 summary/API tests, 19 registry tests, and three browser scenarios including literal mobile tap. Scoped lint, formatting and English verification passed. Two fixture copy lint findings and a redundant projection copy were repaired and rerun; the wire test still verifies JSON omission directly.

Original browser fixtures show the completed chart disappearing and mobile disclosure failing. Original registry tests fail at five missing-notification assertions. The candidate exercises real SQLite, parent-only versus child-only visibility, commit/retry/archive/rollback ordering, raw and qualified owners, retained totals, bounded payloads, and stale asynchronous responses. Browser fixtures cover desktop/mobile completion, cleanup, reload, and cross-agent child hydration.

Independent Codex reviews are scoped-clean through P2, including all request-option follow-ups. Generated Swift is byte-identical to the passed OpenClawProtocol build and four Codable round trips. Full UI and Gateway test types, CI-configured lint, and the shrinking assertion ratchet passed. Commit hooks preserved the verified source bytes. A patch-identical rebase picked up the separately landed main retry-test repair (#139519). Exact-head CI is green on bbda1c9eddfe7301bb253551d7d59e7bc7eb3ddb, including the native screenshot gates and the unchanged browser-job rerun.

Browser fixtures and recording-socket Gateway tests are synthetic boundary proof; live deployment evidence is reported separately. The attached before/after captures contain only inspected synthetic data. Manual remote changed-check attempts were either stopped before validation or failed the then-stale assertion baseline; none is counted as passing. The required hosted checks remain the landing gate.

CI follow-up: the initial run exposed incomplete async client fixtures, readonly-context construction, mock initialization order, and obsolete assertion-baseline entries. The second run exposed a shared test adapter dropping the third request argument, incomplete backend row fixtures, and a cloud test counting legitimate Swarm reads as placement polling. These are repaired at their test owners without production fallbacks, forced clicks, longer timeouts, or disabling Swarm. The request-tuple regression failed before the fix; 290 helper/placement tests, 72 payload-reader tests, 11 summary tests, two final cloud browser cases, full UI/Gateway test types, and lint pass. The cloud guard still rejects unpaired parent describes, and the placement owner's independent no-polling assertion remains intact.

The unchanged sidebar pointer case and dependency-download HTTP 502 both passed their Linux jobs in the second run; no speculative source or dependency change was made for them.

The final shared-adapter follow-up strengthens model-setup assertions for retained start replies, bounded cancellation, and abortable verification. The broad canonical UI run passed 14,501 cases and failed two unchanged native-fixture scenarios at their 60-second outer limit; one was observed before fixture-body entry, and the managed process owner confirmed cleanup. Both scenarios passed unchanged on the combined source in a focused replay (56.47s and 49.02s), with joined worker cleanup verified. Their slow startup remains a separate harness follow-up; the full run is not counted as passing.

The separate main suppression-inventory failure from #139483 was resolved upstream by #139576. A required conflict refresh retains that exact main helper; this PR no longer changes it. All four Swarm and request-fixture commits remained patch-identical through the refresh. Additional held-drain observer controls are being handled separately.

The latest CI run exposed one dashboard A2UI action case receiving no event after a click. The original two-theme scenario passed locally; six instrumented local cases and two instrumented Linux cases also delivered exactly one event each. Linux tracing confirmed Playwright waited through the frame's inert loading state before delivering native input. No source repair is justified for that isolated failure yet; its original CI job passed unchanged on the exact-job rerun. The Linux Testbox was stopped and independently verified completed after retrieving its four owned trace/screenshot files. The separate pre-existing fixture initializer error in opaque frames is addressed independently in #139618.