[TheGreatAxios]

When a helper agent (writer, designer or coder) finishes or fails with an unanswered assignment, the hub sends the project's coordinator a note saying so. Interchange doesn't send this notice itself yet, so the hub has a small loop that checks ended runs in order (ended_at), one at a time across the hub.

• The bug: If one notice couldn't be delivered (mailing the coordinator returned a 5xx error) or its stored mail was malformed, the loop retried that same oldest run on every 5-second pass forever. Every notice queued behind it—across every project on the hub—was stuck, so coordinators never heard that their helpers had finished or failed.
• The fix: A notice whose mail fails is now marked settled (notified: false), logged, and skipped on future passes so the loop moves on to the next one. Malformed mail headers/MIME are caught inside the same error handling so they can't crash the pass either.
• Test: An e2e test (test/e2e/specialist-end-notices.test.ts) on real Postgres with an older notice whose coordinator mail fails (500) plus a newer one with malformed multipart mail. Fails on main, passes with the fix.
• Why this is a stopgap: Once the hub accepts a mail, Interchange already queues and retries delivery. Longer term, run-ended notices belong in the harness itself (Interchange onFailure / pushing the notice back) rather than a hub polling loop.

Supersedes #43.
Closes CL-11918

[linear-code]

CL-11918

[corbits-triage]

Triage: Unsure

Asking a maintainer about:

• added code matches a code pattern: test/e2e/specialist-end-notices.test.ts:188 A hand-rolled sleep; wait on an event, or on an Interchange wait, retry or timeout where one exists.

Changed since the last verdict (Fail at a5dfe01): 1 new commit.

[corbits-triage]

Triage requests changes:

• Invented fallbacks: added code matches a code pattern

Nothing triage asked to change still fails, so this request no longer stands.