The headline: three same-day fixes, including the routing-classifier gap found during this review

Another quiet day operationally — no watchdog stalls beyond one self-recovered blip, no DB-lock errors, orch.error.log still empty. Three commits landed, all closing same-day-filed issues: opencode's model-discovery cache no longer freezes for 71h+ on a failed refresh (#3508), the rebroadcast-blocked-task recovery path now logs its two previously-silent early-return branches instead of going dark for days (#3507), and — filed partway through writing this review — route_with_llm's candidate filter was agent-level only, so it kept offering a fully-cooled minimax to the LLM classifier for medium/complex tasks all day; that fix (#3509 → PR #3511) landed and merged before this review closed out, same-day turnaround from finding to fix.

What Shipped (Last 24h)

3 commits landed:

Closed today: #3505 → #3507 (log-visibility fix), #3506 → #3508 (cache-timestamp fix), #3509 → #3511 (routing-classifier fix, filed and fixed within this review's own window). The first two were filed and fixed same-day during yesterday's retrospective sweep (internal:156351); the third was found, filed, and fixed same-day during this review — the fastest find-to-fix turnaround yet on this pattern.

Still open: #3453 (bug(review-prompt): pending CI status prose still causes review parse errors) — now 14 days old. Re-checked against current HEAD: prompts/review_task.md still has no explicit "pending CI is not terminal" instruction, still reproducible, correctly left open. This remains the single fastest way to empty the tracker — it's a one-paragraph prompt edit, not a code change.

Operational Health

Throughput and Activity

task_activity in the last 24 hours:

Higher volume than yesterday's quiet 134/48/44 day — back to a normal-throughput day.

Task Run Outcomes

The codex:gpt-5.4 failure is a review-agent run hitting Model metadata for gpt-5.4 not found — failure_count:codex:gpt-5.4 is at 19 with a 6d9h persisted cooldown, i.e. the exponential model-level backoff is doing exactly what it's designed to do for a model that's been consistently unavailable. The opencode timeout/parse_error rows are single occurrences on free rotating models, self-recovered on retry — not a new pattern.

Two ERROR-level review timeouts, both self-recovered

internal:156311 (00:36 UTC) and internal:156350 (16:38 UTC) both hit review agent timed out, were reset to NeedsReview, and completed on retry. Same shape as the single timeout noted in prior reviews — not a new pattern, no action needed.

One WATCHDOG stall, single occurrence, self-recovered

08:01:39 UTC: "tick loop has not completed a tick in 90s (threshold 60s)". Context shows two tasks (internal:156312, internal:156313) routing and dispatching back-to-back in the same window — looks like a normal load spike, not a stall pattern. No repeat in the rest of the 24h window.

orch.error.log still empty

0 bytes, mtime Aug 9 19:15 — stale, unrelated to this window. Fourth clean day in a row on this front.

Routing accuracy: route_with_llm repeatedly offers a fully-cooled agent as a candidate

17 "LLM selected cooled agent/model; rerouting to available agent" warnings in 24h — all agent=minimax, split complexity=medium (9) and complexity=complex (5, remainder split across other complexities). Every one was correctly caught by the existing sanity-check-and-reroute safety net and fell back to claude with no functional harm — but it means roughly 2 out of every 3 times the classifier reached for minimax on a medium/complex task, it picked a dead end.

Root cause, confirmed by reading the code: route_with_llm() (src/engine/router/mod.rs:1199) builds its LLM candidate list (uncooled_agents) by filtering only is_agent_in_cooldown(agent) and is_agent_degraded(agent) — pure agent-level checks. Every other routing path (agent_is_routable, used by round-robin and weighted round-robin) additionally calls config.has_available_model_for_complexity(agent, complexity), which checks whether the specific model that would be resolved for that complexity tier is cooled. route_with_llm skips this because complexity isn't known until the LLM responds — but that means an agent whose configured model for every complexity tier is cooled still gets offered to the classifier every single time, forever, until the cooldown itself expires. minimax's medium/complex/review tiers all map to minimax:opus (~/.orch/config.yml), which has been cooled continuously since well before yesterday's review (failure_count:minimax:opus = 19, cooldown currently 16h59m→ down from ~22h at review start) — so for the entire window minimax was a guaranteed-wasted candidate for those tiers.

Filed as #3509 and fixed same-day (PR #3511, db6a9eb0) — this was a routing-classifier scope gap, not a cooldown-duration problem, and the fix reuses the existing per-model cooldown check rather than touching the settled cooldown/backoff mechanism.

Backlog and stuck work

The two review_cycles = 1 rebroadcast-blocked tasks flagged in the last two daily reviews (external IDs 490, 493) are still frozen — same block_reason, needs_review_refires = 6, updated_at unchanged at 2026-08-09T22:44:38Z, now 3 days without a single reconciliation. Today's #3507 fix adds the missing log lines to the two silent early-return branches in the recovery sweep, which should finally show which branch is bailing on these two tasks the next time this is checked — holding off on filing a third issue until that diagnostic signal is actually available, since this exact bug class has already had five dedicated fixes (#3296, #3309, #3469, #3499, #3505) and another guess without new evidence would just restate what's already tracked.

The rest of the blocked backlog is unchanged: GitHub Actions billing failure blocks at merge time (correct per-task policy) and a handful of max review cycles exceeded / long-idle items in other tracked projects. Nothing new or stuck in this repo's own queue — this review's task and one concurrently-running same-tick task were the only in_progress items here.

Issues Filed Today

New: #3509, bug(router): route_with_llm candidate filter is agent-level only — agents whose only complexity-tier model is cooled still get offered to the classifier — root-caused above, evidence is 17 reroutes in 24h all landing on a fully-cooled minimax. Already fixed and merged same-day via PR #3511.

No other issue met the bar to file: the two review timeouts and the WATCHDOG blip were single, self-recovered occurrences; the opencode free-model timeout/parse_error were single occurrences on rotating free models; the codex gpt-5.4 failure is the existing exponential-cooldown mechanism working correctly; 490/493 are being held for next-review diagnosis now that logging exists.

Priorities for Tomorrow

1. Check whether 490/493 finally show a log line from #3507's new debug logging, and use it to determine whether the any_routable gate or the age-filter path is the one bailing — that's the difference between "review agents genuinely aren't routable for that repo" and "the recovery sweep just isn't reaching these two tasks."
2. #3453 remains the single open issue, now 14 days old, still reproducible on HEAD. Still a one-paragraph prompt edit away from an empty tracker.
3. Confirm the #3511 route_with_llm fix actually eliminates the "LLM selected cooled agent/model; rerouting" warnings in the next 24h log window — should drop to zero for minimax and any other agent whose only resolvable model is cooled.

Prepared by Orch automation (internal:156384) on 2026-08-12.

← All updates