What Shipped (Last 24h)

1 commit landed on main:

The fix prevents a transient opencode models CLI failure from overwriting the free-model discovery cache with an empty list. Previously a single failed refresh poisoned the cache for the full 1h TTL, causing opencode to be falsely marked "all models cooled" for ~60 minutes at a time (observed 4x in the prior 30h, ~3-4h/day of lost routing capacity). The cache update now preserves the prior non-empty list when a refresh comes back empty, with regression tests covering both the free-model and all-models caches.

A second fix landed just outside the strict 24h window but is worth closing the loop on since it wasn't covered in a prior report: 1d4b2182 (#3408, merged 2026-07-16) added a global routing sweep so new tasks from inactive/removed repos get routed even when their project isn't in the active tick loop. Two tasks had been stuck in new for 46+ days before this landed — see Stuck Tasks below for the result.

Operational Health

Throughput

task_activity over the last 24 hours:

Zero error or rerouted events in the last 24 hours — the first fully clean window in recent reports.

Task Run Outcomes

task_runs shows 50 runs over the last 24 hours:

48 of 48 completed runs succeeded (100%). No failures, rate limits, silence detections, or parse errors in the window. The two in-progress rows are this task and the internal evening-retrospective task, both dispatched at 23:00 UTC.

Active Cooldowns

Only one active cooldown, and it's the expected persistent one from prior billing-cycle exhaustion. No new cooldowns triggered today.

Routing Accuracy

The active cooldown is scoped to minimax:opus (a model, not the minimax agent) — orch cooldown list confirms no agent-wide minimax cooldown exists. Per the settled router architecture, agent-level filtering (agent_is_routable / available_agents_for_complexity) runs before the LLM call and correctly left minimax in the candidate set, since the agent itself isn't cooled. For both internal tasks dispatched at 23:00 UTC (this review, task 155153, and the evening retrospective, task 155154), the LLM then picked minimax with the specific opus model — the one model that is cooled. The router's model-level check (is_model_in_cooldown, src/engine/router/mod.rs:947) runs after the LLM call and before dispatch, caught the cooled model in both cases, and rerouted to claude/sonnet (task_activity records "reason":"LLM selected cooled agent/model; rerouted to claude" for both). No misroute reached dispatch. Not filing — this is the documented model-level fallback path working as designed, and it's the same behavior noted in the last several reports.

Logs and Service Health

• orch.error.log is 0 bytes — no service-level errors
• Sync ticks stayed in the healthy 1.4s–3.1s range throughout the window, no watchdog alerts
• One transient GitHub connectivity gap around 22:19–22:38 UTC (HTTP send failures to api.github.com, 5xx circuit breaker opening/closing repeatedly with attempt=12 through attempt=19) — resolved on its own once connectivity returned; polling fallback and circuit breaker behaved exactly as designed, no tasks lost or stuck as a result

What Failed

Nothing. Zero failed, rate-limited, timed-out, or parse-error task runs in the last 24 hours.

Stuck Tasks

new count is now 0 — down from 2 tasks that had been stuck for 46+ days (closed by #3407/#3408's global routing sweep, which merged 2026-07-16). Confirms the fix took effect: no orphaned tasks from inactive repos remain.

The 50 blocked tasks are all downstream CI-related auto-merge blocks (per settled architecture — blocking is scoped to the single task at merge time, not the whole repo). No new blocked tasks appeared in the last 24h; this count is unchanged from recent reports.

For the orch project itself: 1,966 done, 1 in_progress (this task), 0 blocked, 0 new.

Issues

0 open issues in gabrielkoerich/orch — the queue is fully drained. #3407 and #3409 (both closed in the last ~26h) were the last two open items, and both fixes are now merged to main.

No new issues filed this run — nothing in today's operational data indicates a bug or gap not already handled by an existing mechanism. The GitHub connectivity blip resolved on its own via existing polling-fallback/circuit-breaker logic, and the one routing quirk (LLM occasionally selecting a cooled agent) is already covered by the pre-emptive routability check and doesn't need a new report.

Priorities for Tomorrow

1. Confirm the opencode cache fix holds Watch for any further "all models cooled" degraded windows for opencode with rate_limit_count=0 — the #3409 fix should eliminate them going forward. If one recurs after the fix is live, it points to a different code path than the one just patched.

2. Watch minimax:opus clear ~5d11h remaining, expected clear around 2026-07-23. Watch for immediate re-exhaustion on clear, which would suggest the billing cycle itself (not just the cooldown math) needs a longer view.

3. Keep an eye on the new-task count staying at 0 The global routing sweep (#3408) is new as of yesterday; a first quiet 24h with new at 0 is a good sign but one data point isn't a trend yet.

4. Downstream blocked-task backlog (50) is stable, not growing No orch action needed — this is CI/merge-side debt in downstream repos, already handled per-task at merge time. Continue monitoring only for growth, not investigating further unless the count moves.

Prepared by Orch automation (internal:155153) at 2026-07-17 UTC.

← All updates