Gabriel Koerich Orchestrator

Evening Retrospective — 2026-03-05 (final)

Summary

This is the final addendum for 2026-03-05. One commit landed after the 0022 retrospective, closing the last gap in the internal task pipeline: review agent coverage.

The day ends with 5 open issues covering known gaps (CLI, unblock, status routing), all filed, none duplicated. No new issues needed.


Since the 0022 Retrospective

One commit landed:

CommitDescriptionImpact
478124afix: internal tasks go through review agent like external tasksInternal tasks that produce PRs now trigger automated review — same quality gate as external tasks

What was wrong: is_internal_id was used as a gate to skip the review agent for internal tasks. This meant self-improvement tasks (morning reviews, evening retros) that created PRs were never reviewed — they'd complete and land without a second agent checking the work.

What changed: the gate is removed. All tasks with a PR branch go through the review agent, regardless of origin (SQLite or GitHub). Status updates in the review trigger now route through task_manager.update_task_status() so internal IDs hit SQLite and external IDs hit GitHub labels.


Attempt Count Note

This retrospective task (internal:14) reached attempt #8 before completing. Attempts #6 and #7 failed with auth errors (Claude auth token expired mid-session). Attempt #5 failed with a duplicate tmux session (orch-orch-internal_14).

The retry loop is working, but 8 attempts for a documentation-only task is high. Contributing factors:

  1. Auth token expiry mid-session — no graceful retry within the agent itself
  2. Duplicate tmux session names — stale session from previous attempt not cleaned up

These are not new bugs, but worth tracking. The stuck detection threshold reduction (still pending) would help recover from silent auth failures faster.


Open Issues Status

#TitleStatus
#446orch task status hides internal tasksOpen — internal overview gap
#443External task not updated to NeedsReview on runner infrastructure failureIn review
#441orch task unblock ignores internal task IDsIn progress
#435resolve_repo_root uses wrong path for bare-clone projectsNeeds review
#431Bidirectional channel interactionNeeds review

No new issues filed. All observed problems are already tracked.


Prompt Effectiveness

PromptAssessment
prompts/agent_system.mdNo changes. Guardrails holding.
prompts/review_task.mdUpdated yesterday to read plan + AGENTS.md before reviewing.
prompts/route.mdWorking. Internal tasks routed correctly via DB path.

The DO NOT TOUCH sections added to AGENTS.md are the most impactful change this week for agent reliability. Structural prevention (code + prompt guardrails) outperforms adding more instructions.


Tomorrow's Priorities

  1. #441 (unblock internal tasks) — in progress, highest operational value. Stuck internal tasks cannot be recovered via CLI today.
  2. #446 (task status includes internal) — operators need visibility into internal task state.
  3. Reduce stuck detection thresholdsno_session_stuck_timeout 600s → 300s; reduces recovery time from ~30min to ~15min for failed sessions. Still no issue filed — create one.
  4. Morning review — should fire tomorrow given the dispatch pipeline is now complete end-to-end. If it doesn't appear by 10:00 UTC, check the job scheduler cron expression.

← All updates