Summary

Exceptional throughput day: ~20 issues closed, 34 commits landed. The dominant theme was review system correctness — response parsing false positives/negatives, hardcoded timeouts, stale state corruption, and systematic default-branch detection failures across multiple subsystems. By end of day the review pipeline is substantially more robust, though two issues required human escalation (blocked).

Accomplished Today

Review Response Parsing (5 fixes)

• False approval on negation phrases (#1059) — "not approved" was triggering the approval path because the keyword match was a substring search. Fixed with phrase-level negation checks.
• Substring match too broad in synthesize_response (#1055, #1058) — "incomplete" / "not done" were being marked as done due to broad substring matching; "approved" was similarly over-matching.
• Missing approval keywords (9d9df5a) — "all checks passed", "no issues found", etc. were not recognized as approval signals. Tasks were being re-routed unnecessarily.
• Issue/comment plain-text acceptance (#1053) — runner now accepts issue-creation and comment-posting acknowledgments as done status, preventing false needs_review states.

Hardcoded Timeouts Eliminated (4 fixes)

All four were the same root cause — timeout was hardcoded rather than reading workflow.timeout_seconds:

• Session runner hardcoded 30min (#1073)
• Review agent hardcoded 600s (#1040 / #1042)
• gh pr checks --watch had no timeout at all (#1060 / #1063)

Default Branch Detection Fixed Systematically (#1044, #1045, #1046, #1048, #1057)

origin/main was hardcoded or config-read in five separate locations. All now use detect_default_branch() against the remote. This was causing silent failures on non-main default branches across rebase, build_git_log, build_git_diff, and auto-merge.

Review Infrastructure Hardening

• Stale output.json corruption (#1065) — review attempt directory was reused on failure, causing crash detection to read stale data from a previous attempt.
• Review agent excludes all previous agents (210afdb) — was only excluding agents that failed, not all prior review agents. Could re-dispatch to the same agent on non-failure cycles.
• Pre-existing failure baseline (#1069 / #1078) — review prompt was blocking PRs for test failures that existed before the PR. Now checks a baseline before requesting changes.
• CI trust qualification (4f0e942, f434909, 195f667) — review prompt now only trusts CI when branch is up to date with main; simplified to CI-first without contradictory fallback logic.
• Review failure comments posted on PR (a62fd7b) — failures were not being surfaced to the PR, making debugging opaque.
• PR kept open when task is blocked (0d8af1c) — blocked tasks were closing their PR.

Runner / Infrastructure

• DispatchGuard key leak on panic (#1072) — currently in review.
• Service version written at startup (#1037 / #1043) — enables CLI/service drift detection.
• Kill orch tmux sessions on graceful shutdown (0b2ad9d) — orphaned sessions were surviving service restarts, causing stuck task re-dispatch.
• Pricing model coverage extended (#1061) — github-copilot, free-tier, deepseek, gpt-5 codex models were missing from cost reports.
• DispatchGuard counter separation (#1056) — push failures were depleting pr_create_failures, causing premature task blocking.
• Test isolation (#1070 → #1079) — build_runner_script tests were sharing ~/.orch/state/ with the running service, causing flakes and state pollution. Fixed with isolated tmp dirs.
• Migration rename (ea75e32) — duplicate migration 009 renamed to 010 to prevent silent schema skip.

Agent Prompt Quality

• Checklist item 2 qualified for non-code tasks (#1064) — mandatory commit requirement was confusing read-only / issue-creation tasks into infinite retry loops.
• NDJSON parser uses production parser in integration tests (#1038 / #1049) — tests were validating a copy of the parser, not the production path.

What Failed / Needed Escalation

Blocked: #1039 — Stale remote refs before review

Closed as blocked. The review agent was starting without a git fetch, so build_git_diff and baseline checks were operating on stale remote refs. No fix landed — escalated to human. This could cause false diffs or missing baseline data on repos with active main branch.

Status: Needs a follow-up fix. Should be straightforward: add git fetch --prune at the start of the review agent invocation, before any diff or baseline check runs.

Blocked: #1070 — Superseded by #1079

Closed as blocked but the root cause was resolved by the follow-up fix in #1079 (test isolation via tmp dirs). The blocked status is a naming artifact — effectively done.

Routing Accuracy

Today's closed issues used a mix of agents:

Complexity routing: most issues tagged simple or medium. No complex tasks today. Routing accuracy appeared good — no obvious misroutes. The kimi dispatch (#1041) for a simple bare-repo branch name fix is appropriate.

The #1039 blocked case is worth noting: agent:claude + complexity:medium — the task was feasible but was closed as blocked rather than fixed. Could indicate the task prompt was under-specified or the agent gave up rather than solving it.

Patterns & Health

Positive:

• Volume sustained: pipeline processed ~20 issues in a single day with no human intervention.
• The review system is now substantially more robust — false approvals, false rejections, and stale state corruption were all addressed in one day.
• Hardcoded timeout elimination was systematic — all four were found and fixed.

Concerning:

• Prompt drift: review_task.md received 4+ commits today (CI trust, baseline check, timeout, simplification). This level of churn suggests the review prompt is complex and fragile. Each fix adds a constraint; the prompt may be accumulating contradictions.
• Two blocked closes: #1039 and #1070. While #1070 was a naming artifact, #1039 is a real unresolved issue. Blocking and closing rather than filing a follow-up is a data-loss risk.

Open at End of Day

Both are in-flight. #1076 is the broader test isolation fix (beyond build_runner_script); #1072 is the DispatchGuard leak.

Tomorrow's Priorities

1. #1039 follow-up — stale remote refs before review agent. File a new issue or re-open #1039. Simple fix: git fetch --prune at review agent start. High impact: affects diff correctness.
2. #1076 completion — broader test isolation pass across all test files touching ~/.orch/state/.
3. #1072 review — DispatchGuard key leak; monitor review agent outcome.
4. Review prompt stability — after 4+ changes today, do a read-through of review_task.md and review_system.md for contradictions or over-constraints that could cause the next round of unexpected behavior.
5. CLI version drift — brew upgrade orch to sync CLI with service version (#1037 landed service-side version reporting; CLI should now reflect the correct version post-upgrade).

← All updates