Gabriel Koerich Orch

Daily Review — 2026-07-12

What Shipped (Last 24h)

2 commits landed on main in the last 24 hours:

CommitPRSummary
d727c9f3#3400fix(parser): recognize "failure" status as canonical error
b6849791#3398Daily review post (2026-07-11)

1 issue closed in the last 24h: #3399bug(parser): agent status "failure" not recognized (the fix from #3400). The parser now treats failure as a canonical error instead of misclassifying it as a parse error.

Service is running v0.80.48 (up from v0.80.46 at the previous review) — the #3400 parser fix is deployed.


Operational Health

Throughput (full day, vs 2026-07-11)

Event2026-07-122026-07-11Change
status_change343307+12%
push111101+10%
dispatch10391+13%
branch_delete7862+26%
review_start5651+10%
review_decision5449+10%
pr_create5348+10%
routed4034+18%
error76+1

Steady throughput growth day-over-day. Error count rose by 1 (7 vs 6) — still very low against 103 dispatches.

Agent / Model Outcomes (last 24h)

AgentModelOutcomeCount
claudesonnetsuccess54
codexgpt-5.4success22
kimiopussuccess16
opencodemimo-v2.5-freesuccess6
opencodedeepseek-v4-flash-freesuccess5
opencodehy3-freesuccess3
opencodenorth-mini-code-freesuccess3
claudehaikusuccess1
claudesonnetpush_failed1
claudesonnetfailed1
codexgpt-5.4failed1
kimiopusfailed1
opencodedeepseek-v4-flash-freefailed1
opencodenemotron-3-ultra-freefailed1
opencodenorth-mini-code-freeparse_error1
minimaxsonnetrate_limit1
opencodehy3-free1
claudesonnet1

120 runs total; ~110 successes (~92% success rate). Claude/sonnet leads with 54 successes. Codex/gpt-5.4 solid at 22. Kimi/opus clean at 16 (continuing its strong streak).

Active Cooldowns

KeyRemainingNotes
minimax:opus~3d11hpersisted — LLM router keeps selecting, immediate fallback to claude
minimax:sonnet~1d10hpersisted — Token Plan usage limit
opencode:opencode/nemotron-3-ultra-free~16h53mpersisted — single failure today

Fallback routing handles all three cleanly. No agent-level cooldowns.


What Failed

1. opencode/nemotron-3-ultra-free: failed (1)

Single failure, no successes today → ~16h53m cooldown. Model remains unreliable.

2. opencode/north-mini-code-free: parse_error (1)

1 parse_error amid 3 successes. Short cooldown. Looks like a one-off parse failure.

3. opencode/deepseek-v4-flash-free: failed (1)

1 failure amid 5 successes. No cooldown triggered (threshold not met).

4. claude/sonnet: 1 failed + 1 push_failed

Agent completed successfully; push step failed once (distinct from task failure). One outright failure amid 54 successes (98% success rate). Both isolated.

5. codex/gpt-5.4: 1 failed

1 failure amid 22 successes (95.5%). No cooldown. Transient.

6. kimi/opus: 1 failed

1 failure amid 16 successes. Isolated; no cooldown.

7. minimax/sonnet: rate_limit (1)

"Token Plan usage limit reached" — consistent with the minimax:sonnet cooldown. Handled by generic rate-limit cooldown path.

8. opencode "Streaming response failed" in review path — classified as failed

An opencode review run failed with network error: "Streaming response failed" and was classified as failed rather than NetworkError. This is the same pattern as #3378 (opencode/nemotron), but it surfaced in the review runner path. Worth verifying the review error classifier maps streaming-failure network errors to NetworkError so the model isn't silently penalized.

9. Security: GitHub token embedded in stored git error messages

A push_failed error captured the full git stderr, which echoes the insteadOf-rewritten remote URL including x-access-token:gho_…@github.com. That token is persisted verbatim in task_runs.error_message and task_activity.details in orch.db. security::has_leaks() would block it from being posted to GitHub, but the token still sits in plaintext in the local DB. Filed as issue below.


Routing Accuracy

No wrong complexity tier observed. The minimax LLM-router preference for complex internal tasks still selects minimax (now minimax:sonnet/minimax:opus, both cooled) and falls back to claude immediately — fully explained by the persistent minimax cooldowns, not a routing bug. No silent failures (silent-agent detections correctly reverted tasks to new). No dead models re-selected.


Log Health

Service running v0.80.48. Brew error log is 0 bytes (clean). Sync ticks completing in 1.5–2.5s. Cooldown KV sync sub-millisecond. One slow-tick warning (38s) when two internal tasks (daily-review + evening-retrospective + weekly-review) were dispatched simultaneously in the same tick — expected under burst multi-dispatch, not a health concern.


Open Issues

gh issue list --state open returned no open issues in gabrielkoerich/orch at review time (prior issues all closed).

1 issue filed during this review: #3401 — git push/remote error strings persist the insteadOf-injected x-access-token:TOKEN URL into task_runs.error_message / task_activity.details in plaintext. Root cause: git echoes the rewritten remote URL on push failure and orch stores it verbatim. Fix: scrub the token from git error strings before persisting.


Priorities for Tomorrow

  1. nemotron-3-ultra-free cooldown clears in ~17h — model is unreliable (0 successes, 1 failure today); monitor whether it re-enters rotation and fails again.
  2. minimax:sonnet cooldown clears in ~1d10h (~2026-07-14); minimax:opus runs ~3d11h (~2026-07-16). Routing sanity warnings on complex internal tasks continue until then — no action needed.
  3. Token-in-DB leak — apply redaction scrubbing to stored git error strings.
  4. opencode "Streaming response failed" in review path — confirm the review error classifier maps streaming network errors to NetworkError (per #3378 precedent).
  5. internal:154863 blocked at PR #3392 — stale daily-review PR blocked on CI failure limit. Close or merge manually; CI may never pass.
  6. Blocked inventory — run orch task unblock all to drain the CI-failure blocklist; inspect any that immediately re-block.
  7. push_failed category — watch for recurrence on claude/sonnet; if it repeats, investigate git push timeout vs branch conflict.

Prepared by Orch automation (internal:154993) at 2026-07-12 UTC.

← All updates