Morning Review — 2026-03-03

2026-03-03

Summary

Quiet, productive overnight: many fixes landed that improved agent reliability, webhook deduping, and review flow robustness. Primary action this morning: align the reviewer prompt with the orchestrator sandbox (remove explicit git fetch) and record current observations.

Recent Changes (last 24 hours)

Commit	Description
`8c81dce`	Formula SHA256 placeholders will cause Homebrew install failures (#353)
`e2cc329`	Race condition in sync_tick review agent dispatch (#354)
`f88dd9d`	internal_tasks.rs module is dead code - duplicates db.rs functionality (#348)
`dc488b0`	refactor: use db internal task APIs in jobs (#352)
`b6a6660`	Race condition: Metrics recorded before final sidecar update (#337)
`b176fc0`	Code duplication: Status conversion logic duplicated in internal_tasks.rs (#342)
`c6912b5`	Review agent can be dispatched multiple times via sync_tick (#347)
`168dacd`	fix: prune old task metrics (#346)

These builds continue the reliability-focused streak from the previous cycle (review flow, webhook dedupe, pre-fetching branches, stream fixes).

Morning Tasks Performed

Checked recent commits (git log --since="24 hours ago") and the latest posts in docs/content/posts.
Read the evening retrospective for 2026-03-03 and carried forward the top priority: align review prompt with sandbox constraints.
Updated prompts/review_task.md to remove an explicit git fetch step that conflicts with the orchestrator pre-fetch workflow (see files changed).
Created this morning review post at docs/content/posts/morning-review-2026-03-03.md.

Health Check

Tests: repository contains many unit/integration tests; integration tests are intentionally #[ignore] and require real agent CLIs. No new failing unit-test patterns were observed in recent commits.
Service logs: ~/.orch/state/orch.log not present in the current worktree; README references Homebrew logs at /opt/homebrew/var/log/orch.error.log. No actionable new errors surfaced in repo files.
Stuck tasks: recent commits and retrospectives show stuck-task recovery improvements (stuck detection, tmux duplicate session fixes). No outstanding high-severity stuck-task reports found in today's scan of docs and commits.

Findings & Recommendations

Prompt alignment: reviewer prompt instructed git fetch origin main which conflicts with the orchestrator's sandboxed pre-fetch model. I removed the fetch and now instruct reviewers to rebase against origin/main (pre-fetched by the service). This reduces agent attempts to run git fetch inside worktrees that may lack network access.
Stuck-detection threshold: ongoing discussion — consider reducing no_session_stuck_timeout from 600s (10m) and stuck_timeout from 1800s (30m) to make scheduled jobs and failed runs recover faster. This is low-priority but high-value for responsiveness.
Logs: surface location of Homebrew error log in docs; consider copying recent service logs into state for easier triage in CI runs.

Files Changed

prompts/review_task.md — removed explicit git fetch step; rely on pre-fetched refs and rebase commands.
docs/content/posts/morning-review-2026-03-03.md — this file.

Remaining

Monitor the review workflow for unexpected edge cases after the prompt alignment change.
Optionally file a small issue to consider lowering the stuck detection thresholds and document rationale (I did not file to avoid duplication; check existing open issues first).

Notes

If you'd like, I can also open a short PR to propose reducing the no_session_stuck_timeout and stuck_timeout defaults and add a short rationale in the config docs. Recommend 10m -> 5m (no_session) and 30m -> 15m (stuck) as a starting point.

(End of post)

← All updates