Gabriel Koerich Orch

Daily Review — 2026-06-21

What Shipped (Last 24h)

2 commits landed today, both documentation updates.

CommitPRDescription
14058b38docs(agents): forbid naming internal projects/paths in issues; compress AGENTS.md under 40k
136d67d6#3342docs(posts): daily review 2026-06-20

Key Change: AGENTS.md Policy + Compression

The main commit (14058b38) enforces two things:

  1. New filing policy: agents must never name downstream projects or paste absolute paths in issues. Home-dir paths leak the operator's username; even non-home paths often embed internal project names. Tilde-rooted placeholders or prose descriptions required. (Cautionary tale: issue #3339 named a specific downstream project and was deleted.)
  2. Size compression: AGENTS.md trimmed from 46.1k → 37.6k to stay under the 40k character performance threshold. All operative rules preserved; only redundant restatements and verbose cautionary detail removed.

Closed Issues (Today)

No issues were closed today. All prior closed issues are from earlier in the week.


Operational Health

Volume (Last 24h)

MetricCount
Status changes278
Dispatches88
Pushes78
Branch deletes76
Routed40
Review starts39
Review decisions37
PRs created36
Errors9
Reroutes2

Healthy volume — 88 dispatches, 36 PRs, 9 errors. Error count slightly up from yesterday (8), but within normal range. Reroutes (2) are nominal.

Agent / Model Outcomes (Last 24h)

AgentModelOutcomeCount
claudesonnetsuccess26
kimiopussuccess11
codexgpt-5.4success9
codexgpt-5.5success8
opencodemimo-v2.5-freesuccess4
opencodenorth-mini-code-freesuccess4
minimaxopusrate_limit2
opencodedeepseek-v4-flash-freesuccess2
opencodenemotron-3-ultra-freesuccess2
claudesonnetfailed2
claudehaikublocked1
claudehaikufailed1
claudesonnet(no outcome)1
codexgpt-5.5blocked1
opencodenemotron-3-ultra-freefailed1
opencodenorth-mini-code-freeparse_error1

Effective pool: claude/sonnet remains the dominant workhorse (26 successes). Kimi/opus healthy (11). Codex pool strong between gpt-5.4 (9) and gpt-5.5 (8). Opencode multi-model coverage solid across four models (12 combined).

Failures: claude/sonnet had 2 failures (normal tail — transient errors, both handled by per-model retry), nemotron-3-ultra-free and north-mini-code-free each had 1 parse_error or failure — per-model cooldowns activated automatically.

Minimax: still rate-limited (2 hits). Not unexpected given the 23h3m cooldown still running.

Active Cooldowns

KeyRemainingReason
minimax23h3mpersisted (rate limits)
minimax:haiku9hpersisted

Minimax:opus cooldown has cleared since yesterday (was 3h3m). The agent-wide 23h cooldown still covers it. No other agent or model in cooldown — clean pool otherwise.

Cleanup Reconciliation Timeouts

The cleanup module continues to time out on candidate listing. Observed at 22:43, 22:46, 22:48, and 22:58 UTC today — roughly once per tick window during this period:

WARN orch::engine::cleanup: timed out listing reconciliation candidates timeout_secs=30

The throttle fix from PR #3341 is merged but not deployed — the service is still running v0.80.25 and the fix landed in a later release. These timeouts are harmless (scan is abandoned and retried next tick) but noisy. They will stop once the service is upgraded to v0.80.28.

Cron-Boundary Slow Tick

The 23:00 UTC slow tick recurred as predicted:

WARN slow tick elapsed_ms=43775

43.8 seconds — consistent with the pattern observed the prior two days. The daily-review and evening-retrospective crons fire simultaneously at 23:00:07 UTC, causing I/O contention during parallel worktree creation. No work was lost; both tasks routed and dispatched correctly.

Service Version

Running: v0.80.25 · Latest: v0.80.28 — three versions behind. The cleanup throttle from #3341 is in v0.80.28. Upgrade is overdue.


Blocked / Stuck Tasks

TaskStatusAttemptsReason
internal (external #490)new5— (no block reason, retrying)
internal (external #493)new5— (no block reason, retrying)
Many (external #432, #442, #445, #449, #494+)blocked4–6CI failure limit (3) reached during auto-merge

CI-limit blocks: Several downstream tasks hit CI failure limit (3) reached during auto-merge. This is working as designed — CI is failing repeatedly and the auto-merge guard is blocking. Requires human attention on the downstream CI failures; orch task unblock all after CI is fixed.

High-retry new tasks (external #490 and #493 at 5 attempts): router is repeatedly routing but no block reason set, suggesting routing succeeds but something upstream is retrying. Worth monitoring; if they hit the attempt ceiling, they'll be blocked and flagged.


Priorities for Tomorrow

  1. Upgrade service to v0.80.28 — three versions behind; cleanup throttle fix is waiting:
    brew update && brew upgrade orch
    brew services restart orch
    orch -V
  2. Investigate CI-failure blocks — multiple downstream tasks blocked at CI failure limit (3). Review the CI failures in the affected downstream project and unblock after fixing:
    orch task unblock all
  3. Monitor high-retry tasks (external #490, #493) — 5 attempts each with no block reason. If they stop progressing, investigate routing logs for the specific failure mode.
  4. Minimax recovery tomorrow — agent-wide cooldown clears in ~23h. Verify clean re-routing when it lifts.
  5. Watch cron-boundary slow ticks — third consecutive day at 23:00 UTC. If it persists, consider staggering the daily-review and evening-retrospective crons by 2–3 minutes to avoid concurrent worktree creation.

Prepared by Orch automation (internal:154199)

← All updates