[sentry-junior]

Stacked on #2499. Merge that first.

#2499 fixed the reproduced bug, where a child agent's watch-event resume went to the root worker and crashed. This PR fixes the two design gaps behind it, so the same kind of bug cannot happen again quietly.

1. The stored parent picks the worker (app/conversation-work.ts, agent-invocations/result.ts, cli/chat.ts)

Before, run picked a worker by trying a list of checks in order: message labels, the Turn Source, and agent invocation lookups. The bug came from that order. Now run reads the Conversation record once. If it has a parentConversationId, the work goes only to createChildConversationWorker. That worker always returns a result. Work that no agent invocation owns throws an error and does not fall through to a root worker. Root wakes also skip the agent invocation SQL reads they made before.

2. A resume attempt can be the final attempt (task-execution/worker.ts, agent-invocations/work.ts)

isFinalAttempt was set only from how many times a message had been delivered. A resume wake has no messages, so it was never final. When a resumed Turn kept failing, the worker gave up and dropped the mailbox, but no adapter got a chance to record the failure. For a child agent, the invocation stayed paused. The parent was never told. Later results from that child were also held back as replaced, because hasOpenChildWork still saw the stuck invocation.

Now the retry count also sets isFinalAttempt: an attempt is final when one more failure reaches CONVERSATION_WORK_MAX_RETRIES. On that attempt, the agent invocation worker fails the invocation for any error and reports it to the parent. This includes save errors without an AgentRunError wrapper, which it used to rethrow even on the final attempt. Root conversation Turns and dispatches use the same flag, so a resume that keeps failing now ends with their normal final-attempt failure path (a fallback reply or a failed dispatch), not a silent stall.

Worker crashes: this does not need a separate repair job. If a child's worker dies in the middle of a Turn, the next wake fails that Turn through projectTerminalTurn, well before the retry limit is reached.

Tests

The resume test in agent-invocation-work.test.ts goes through real production routing. It now has a fourth case: a child yields, then its resume throws on every attempt. Without the worker change, the last attempt gives up (conversation.work.retry.exhausted) and the invocation stays open. With the change, the invocation ends failed and its result is appended to the parent mailbox.

• tsc, oxlint, depcruise, the test-architecture check, and the file-length check pass.
• 25 test files / 274 tests pass: agent invocation, agent dispatch worker and recovery, task-execution component, conversation-turn, Slack runtime, and CLI.
• durable-queue.test.ts (2 tests) and agent-dispatch-work.test.ts (1 test) time out in my sandbox. They time out the same way on the fix(agent-invocations): Route subagent resumes to subagent code #2499 head without this change.

Still open (not in this PR)

• If the worker crashes several times before a child Turn even starts, the work can still give up and drop queued parent messages. That needs a heartbeat repair, and I have not seen it in production.
• The message, nested, spawn, and stop tests still build worker inputs by hand instead of going through real routing.

via David Cramer.

--

View Session [Sentry]

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
junior-docs	Ready	Preview	Nov 1, 2026 5:12pm UTC