Skip to content

Ticket #75 — conversation blockers: empty replies, user bubble color, stuck working ​

Root causes ​

  1. Silent “finished” without an answer. runTask used the raw assistant text and accepted an empty/whitespace reply as success. A turn that only called ensemble_handoff could finalize with no answer, emit a completion handoff, and show “finished”. Fixed by trimming the reply and treating whitespace-only output as “The agent produced no answer.” — retry inside the attempt budget, then fail visibly; no completion handoff is created from an empty reply. Prompts/AGENTS.md now require answering in the group chat before handing off.
  2. Grey user bubble. The #72 bubble chrome set the user edge to colors.muted unconditionally. User bubbles now use the user accent (colors.good) unless queued (queued remains muted).
  3. Stuck working…. A parent left awaitingChildren after all real children were terminal was never reconciled, and automatic completion-delivery hops were counted as synthesis children, so a fast child/auto-hop race could leave a parent working with no live run forever. Added reconcileWaitingWork() at the start of every pump: clears awaitingChildren with no live children, synthesizes/finalizes, resets stale working rows with no run/claim, and excludes handoff.completion hops from synthesis children. maybeSynthesize also excludes completion deliveries.

Tests and verification ​

New tests: empty assistant reply is retried/fails with no handoff; a waiting parent with terminal children is reconciled to replied and not stuck. The existing task_status lifecycle test now waits for the child claim before asserting derived facts (timing synchronization, not weakened assertions). The fake live queued-line assertion now waits for the composer to clear before counting.

Environment: Node v26.12.1, host OpenCode v2.2.5 vs pinned types 2.2.1, Linux/tmux.

CommandResult
npm run checkexit 0
npm test372 passed, 0 failed
npm run test:live:fakepassed twice consecutively
npm run test:live:fake:emptypassed
npm run test:livepassed

No skips. Remaining uncertainty: user-bubble color is visual (text capture does not verify hue); the stale-run reconciliation relies on the in-process active-run map for the “run is gone” case; no macOS run.

Released under the MIT License.