What Shipped (Last 24h)

3 commits landed in the last 24 hours.

• #3385/#3386 → #3387: Yesterday's fix that skipped agent-wide cooldown for BillingCycleExhausted with a known model was applied to the main runner path. Today's follow-up (#3387) extends the same fix to the review agent path — record_review_agent_failure was applying BillingCycleExhausted as a generic agent failure when the model was known, triggering an unnecessary agent-wide cooldown on the review agent. Now returns RetryableError::ModelUnavailable in that path, keeping the penalty model-scoped.
• #3388: When a billing-limit check caused a review attempt to fast-fail before spawning a tmux session, no task_run row was created — leaving the audit trail blind. The fix inserts a task_run with the billing failure outcome before returning, so every review attempt is accounted for regardless of whether the agent ran.

The running service is orch 0.80.43.

Operational Health

Throughput

task_activity comparison vs yesterday:

Overall volume is slightly lower, but the quality signal is strong: errors dropped 57% (37 → 16) and reroutes dropped 67% (15 → 5). This tracks directly with the BillingCycleExhausted cooldown fixes — fewer spurious agent-wide cooldowns means fewer forced reroutes.

Agent / Model Outcomes (last 24h)

Claude/sonnet is carrying the heaviest load (43 successes). Codex has a 2d17h persisted cooldown, which explains low codex volume. Kimi/opus failure count (3) is on the same level as yesterday; no escalating pattern. north-mini-code-free cooldown expires in ~3h — watch whether it returns cleanly.

Active Cooldowns

The minimax:opus cooldown has ~2d11h left. The LLM router continues to select minimax as the preferred agent for both internal tasks (daily-review and evening-retrospective), triggering the routing sanity warning fallback to claude on every run. This is functioning correctly but generates noise.

Blocked Inventory

The CI-failure backlog grew by 1 (50 → 51). orch task unblock all is still the recommended manual path to drain the stale CI-failure accumulation — if tasks re-block immediately on retry they warrant investigation; if they clear, the backlog was stale.

What Failed

1. Periodic GitHub API request failures

The service log shows 3 instances of HTTP send failed, will retry if attempts remain against one downstream project's issues endpoint. Each failure adds ~12–14s to the sync tick (vs ~2s normal). Retries succeed on the next attempt. This is the same transient pattern noted yesterday; frequency is consistent, not worsening.

2. LLM router selects minimax on every internal task

Both internal:154779 (daily-review) and internal:154780 (evening-retrospective) had the LLM router select minimax (2d11h cooldown remaining) before the fallback to claude fired. This will recur every day until the minimax cooldown expires (~2026-07-09). The fallback is working; the noise is the cost.

3. opencode/north-mini-code-free: 3 timeouts + 1 parse_error + 1 failed

The north-mini-code-free model is about to exit its cooldown. The 3 timeouts in this window likely contributed to the cooldown being set. Its performance post-cooldown will be the real signal.

Routing Accuracy

Dispatch accuracy is good. No wrong complexity tier or wrong agent type visible. The LLM router's minimax selection is a false positive caught by the cooldown gate — the downstream assignment (claude/sonnet for medium, claude/opus for complex) is appropriate.

Prompt / Workflow Quality

The BillingCycleExhausted fix series (#3382 → #3383 → #3385/#3386 → #3387) reached its natural conclusion today. The full flow — from initial misclassification to model-scoped penalty to review-agent path to audit trail — is now closed. The 67% reroute reduction is a direct measurement of the fix quality.

Open Issues

gh issue list --state open returned no open issues in gabrielkoerich/orch.

No new issues filed. The minimax-routing noise and the CI-failure backlog are operational states, not code defects.

Priorities for Tomorrow

1. Watch north-mini-code-free after cooldown lifts (~3h). If it immediately re-errors, a classifier or model-quality issue may need investigation.
2. Consider running orch task unblock all to drain the 40-task CI-failure backlog. Monitor if tasks re-block — if they do, inspect the specific CI failure reason.
3. minimax cooldown expires ~2026-07-09 — expect the routing sanity warning on every internal task until then. No action needed; note it in the review.
4. kimi/opus failure rate: 3 failures / 18 successes (17%) today. If it climbs tomorrow, check for rate-limit patterns or model instability.
5. Periodic GitHub API failures: currently transient with auto-retry. If the pattern worsens to multiple failures per sync tick, investigate the network path to that project's GitHub endpoint.

Prepared by Orch automation (internal:154779) at 2026-07-06 UTC.

← All updates