Make one store own worker lifecycle and wait results

[TheGreatAxios]

Summary

• One stored WorkerLifecycle owns worker state, report/error, retention, and physical handles. fail() is failed, not shutdown; closeOne after fail returns immediately.
• wait_agents JSON is a projection of that lifecycle plus a per-install FleetMailbox (membership, pin, collected, interrupt override), not a second terminal store.
• Operator cancel and close/send_input interrupt report wait status interrupted rather than failed. Headless exec finally cancels live sub-agents the same way TUI runtime-shutdown does.

Verification

• bun run typecheck, bun run build, and bun run test pass
• bun run check in this worktree: lint, typecheck, build, and 5744 tests pass

Fixes CL-7269

[linear-code]

CL-7269

[TheGreatAxios]

skywalker · approve

One stored WorkerLifecycle owns worker state; wait_agents projects from it plus a per-install FleetMailbox; headless exec finally cancels live workers.

No blocking findings.

[TheGreatAxios]

critic · comment

Wait JSON is a projection of stored lifecycle plus mailbox overlay; leftover resolve/reject methods are gone.

• docs/ARCHITECTURE.md:222 — operator cancel docs omit the in-flight window: cancelled+inFlight still projects running until settleRun
• src/exec/runner.ts:133 — cancelAll is not try/caught; a throwing subscriber would skip close/dispose

[TheGreatAxios]

greybeard · comment

Session store owns lifecycle; FleetMailbox is overlay membership, pin, collected, and interrupt — not a second terminal store. fail() is failed, not shutdown.

No architectural blockers.

[TheGreatAxios]

Review · Request changes

Worker wait results now project one stored lifecycle through a per-install mailbox, and headless exec shutdown cancels live workers.

Findings

• src/subagent/agent-fleet.ts:203 — take() unpins before returning a fresh snapshot, so synchronous retention pruning can delete the session and drop the report from its first delivery. With the default cap, completing 21 non-retained workers and collecting them in reverse order returns the oldest result as {"status":"done","collected":true} without report-0; maxCompleted: 0 drops every first report. Could this return the pre-unpin snapshot, or otherwise retain the payload through delivery?
• src/subagent/agent-fleet.ts:127 — each FleetMailbox permanently subscribes to the shared session store and discards the unsubscribe handle. Nested orchestrator runs create these mailboxes at src/subagent/run.ts:602 over the shared store passed at src/subagent/agent-fleet.ts:651, so completed orchestrators remain retained and every later session notification invokes all stale mailboxes. Could mailbox/toolset disposal unsubscribe this listener?
• 306c252e — the commit body contains a 203-character line, exceeding the repository's 72-character body-line limit.

Notes

• bun run check exits 0 across lint, typecheck, build, and tests.