Files
PaperClipAI/tests
DottaandClaude Opus 5.5 589178dd77 fix(evals): name unmet completion-gate readiness fields and grade the gate read, not reply words
tasks.completion-with-gate timed out in setup without saying why. The retained
run shows Morgan delivered website.html and reported needs_review, so the task
reached in_review through a generic review card while executionState stayed
null: the configured review stage for Riley never started (product defect,
fixed in the cherry-picked native finalizer change).

- The readiness poll now lists each unmet field (status, executionState status,
  stage type, participant, run and delivery conditions) in its timeout.
- Morgan may need a product continuation run: setup and grader accept one or
  more succeeded worker runs, with the delivery from one of them. The run bound
  rises to 3 and the setup window to 330 s (attempt 660 s); the retained run
  needed 229 s of a 240 s window.
- The reply keyword check is replaced by a structural one: the CEO's chat run
  must read the task's saved state (GET /api/issues/{id} or get_task, by
  record-bound receipt) and reply. The saved pending stage still proves the
  outcome.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 12:07:03 -06:00
..