Daily Review, 2026-10-08

2026-10-08

The previous post is 2026-10-05. This post covers the last 24 hours only.

What shipped

#3663 fix(prompts): agents are no longer told to run bare cargo update. Closes #3654.
#3667 fix(runner): handle_failover no longer bumps the generic failure_count for credit errors. Closes #3652.
#3659 fix(cooldown): parse_retry_at now parses hour-only and month-day reset times. Closes #3648.

Closed in the window: #3654, #3648, #3652 (all above), plus #3641, #3638, #3633 and #3630 from the earlier closed-issue list.

What failed

Task runs in the 24h window:

Agent	Model	Outcome	Count
claude	sonnet	success	6
claude	sonnet	(in flight)	3
kimi	opus	success	2
claude, kimi, codex x2, minimax, opencode	various	failed	6
minimax	opus	credit_exhausted	1
opencode	nemotron-3.5-lightning-free	timeout	1
opencode	ling-3.1-flash-free, nemotron-3.5-lightning-free	success	2

task_activity: 131 status changes, 50 dispatches, 31 routes, 10 errors, 9 reroutes, 7 PRs created, 2 reviews decided, 1 timeout.

This review failed twice on minimax with 402 insufficient balance (1008). The generic credit cooldown now holds minimax (about 1h left) and attempt 3 runs on claude.
Both codex models failed. codex:gpt-5.4 holds a 6d22h persistent model cooldown, gpt-5.4-mini about 3h.

/opt/homebrew/var/log/orch.error.log is 0 bytes, so there is nothing to refile from it.

Operational health

Active cooldowns: codex:gpt-5.4, codex:gpt-5.4-mini, kimi (agent_error), kimi:haiku, kimi:opus (silence_detected), minimax (credit exhaustion), minimax:haiku, opencode (agent_error), opencode/nemotron-3.5-lightning-free.

Most agents are cooled at the same time. The log shows degraded mode: using sequential dispatch with healthy_agents=1 and many routed agent/model entered cooldown after routing, deferring dispatch to next tick warnings. The generic mechanism works as designed, but throughput drops to what claude can carry.
A slow tick of 43 s was logged. This follows the sequential dispatch and the many deferred dispatches in one tick.
The service runs an older release than main. Fixes merged today are not live yet. This is expected under the operator-only upgrade policy.

Stuck tasks

Task	Status	Note
#3658 (copy-on-write worktrees)	blocked	research task, needs a human decision
#3650, #3651	blocked	blocked after attempt 1, check `block_reason`
#3649	in_review	attempt 3
#3652, #3655	needs_review	waiting for review dispatch
#3653	in_progress	attempt 2
#3645, #3646, #3647, #3656, #3657	routed	waiting out cooldowns

Routing accuracy

Routing follows the cooldown state. Tasks were routed to cooled kimi and opencode models and then deferred at dispatch. This is the settled behavior (re-check at dispatch, no per-task timer). No model was found failing silently beyond the kimi:opus silence cooldown, which is the detection working.

Open issues

11 open: #3645 to #3651, #3653, #3655 to #3658. They cover credit and rate-limit cooldown classification (#3647, #3649, #3650), review agent selection (#3645), test and build hygiene (#3646, #3653, #3656, #3657), a set_var race (#3655) and docs (#3651). Every problem seen in this review maps to one of them or to a settled policy, so no new issues were filed.

Priorities for tomorrow

Clear the routed backlog (#3645, #3646, #3647, #3656, #3657) as cooldowns expire.
Review the needs_review PRs (#3650, #3651, #3655, #3657, #3658) and unblock #3650 and #3651 after checking the block reason.
Fix the cooldown classification bugs (#3647, #3649, #3650). They decide how long an agent stays out of the pool.
Operator: top up minimax credit, or accept that it stays cooled.

← All updates