fix(server): report delegated tasks stopped by a server restart

[saphid]

What Changed

When the server restarts, a parent now hears about delegated tasks the restart stopped as soon as the server is back up, not at the next restart.

On startup, runtime recovery cancels the runs of delegated children that were still working, and any parent completion wake that was running. Those cancellations use runtime-reconcile commands, which the live terminal-run reactor deliberately ignores. The orchestrator did recover unreported child results and interrupted completion wakes, but it did so while being constructed, before runtime recovery ran. Anything recovery cancelled was left until the following server start.

This change moves both recovery passes, child results and completion wakes, into one recoverDelegatedTaskReports step. It runs once, in the startup recover phase, right after runtime recovery and before the effect worker starts. Results go through the existing completion delivery.

Restart continuation ("continue threads after update") is respected:

• A child whose restart continuation is still pending in the outbox is not reported. If the continuation starts a run, the continued run reports as usual.
• If the continuation is skipped (for example, the setting is off or newer work exists), or it fails for good, the executor publishes the cancelled result then. continueRestartedRun now returns whether it dispatched a run.
• Because nothing runs during construction any more, a child with a pending continuation is never reported cancelled first, including after a crash between recovery and the continuation.
• Cancelled completion wakes are finalized without deferral. Their claimed tasks go back to pending and get a follow-up delivery. A restart continuation of a wake does not carry its delivery ownership, so deferring would strand those tasks.

Why

In our database, 11 delegated children were cancelled across two parents at 23:42 UTC. Their parents only received the results at 03:47, when the server next started. That's four hours in which each parent assumed its children were still working.

This is one of a few focused fixes to make sure a parent learns when a child stops. #13938 fixed results that were never delivered, and #13343 covers children blocked on a question or approval.

Tests

• DelegatedCompletionDelivery.test.ts, "recovers a child that runtime recovery cancelled, deferring pending continuations": nothing is reported before the pass runs; a child with a pending continuation is skipped; otherwise its result is published once, and a second pass publishes nothing. If the deferral is removed, the test fails.
• DelegatedCompletionDelivery.test.ts, "re-offers a parent wake that startup recovery cancelled with a sibling": a running wake and a running sibling are both cancelled by recovery. The pass publishes the sibling's result and reserves one follow-up delivery carrying both tasks. If the wake pass is dropped, the test fails.
• RestartContinuation.test.ts, "reconciles a delegated child when its restart continuation will not run": nothing happens when the continuation starts or will retry; the result is published when it is skipped or fails for good.
• The earlier head passed 138 tests in serverRuntimeStartup, ProviderRuntimeRecoveryService (and its regression file), RestartContinuation, EffectWorker, DelegatedCompletionDelivery, AgentAwarenessRelay, ThreadManagementService, OrchestratorMcpToolkit.integration, SteeringCompletion.integration, ThreadDeletion, and runtimeLayer. The latest recovery-error fix passes vp test run apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.test.ts apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts (19 tests) and pnpm run typecheck in apps/server.

Limits:

• No test rebuilds the orchestrator over saved state. That construction no longer recovers anything is a structural change, verified by reading the code.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes (no UI change)
• I included a video for animation/interaction changes (no UI change)

This change was made by Claude Opus 5.5 with Claude Code and reviewed by GPT-6 Astra with Codex.

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This production fix changes startup recovery ordering and delegated-task orchestration across multiple shared runtime components. It also gates child-result publication and parent wake processing on restart-continuation state, creating a broader runtime impact than a small self-contained bug fix.

^{You can add or adjust custom eligibility rules. Learn more.}