Morning Review — 2026-04-15

2026-04-15

Recent Commits (last 24h)

12 commits merged — steady day with emphasis on bug fixes and push semantics:

Commit	Issue	Description
`353287ac`	#2657	set_fields duplicate ALLOWED_FIELDS check — dead code removed
`86a990de`	#2652	Fallback message for empty task_runs.error — improves diagnostics
`f87bc031`	#2647	Hoist batch_session_active() to tick() — eliminates duplicate tmux subprocess per cycle
`50dc043e`	#2646	PushResult::NoCommits variant — semantic fix for no-commits case
`fedaf9f5`	#2645	Security: has_leaks only on high-confidence patterns — reduces false positives
`190c086a`	—	set_model/set_agent_cooldown now async — guarantees KV persistence
`f96a1cc9`	#2641	RouterConfig::from_config() per review cycle — not called repeatedly
`976ef4f1`	#2610	files_modified falls back to empty — retry agents retain file context
`ddf5636e`	#2623	Ollama local model routing — new feature
`b26b7ab2`	#2636	GitHub API fire-and-forget — reduces tick latency
`d33f4463`	#2635	Corrupted worktree index recovery
`f4d48adb`	#2633	LLM routing budget cap — prevents tick loop stalls

Version mismatch: CLI 0.69.0, Service 0.69.2 — brew upgrade orch && brew services restart orch needed
Logs: clean tick cycle (~1.5s), no persistent errors, one HTTP retry noted (rate limit)
Jobs executed: morning-review, morning-briefing, twitter-trending-watch

Agent	Model	Success	Failed	Rate
claude	sonnet	39	27	59%
claude	opus	3	8	27%
minimax	opus	28	4 + 4 rl	70%
opencode	gpt-5-mini	17	0	100%
opencode	minimax-m2.5-free	19	1 + 1 empty	90%
opencode	gemini-3.1-pro-preview	2	8	20%
opencode	gpt-5.4	1	7	13%
opencode	nemotron-3-super-free	10	5	67%
glm	opus	17	6 + 2 rl	68%

Notable:

claude/opus at 27% — unchanged from yesterday (3/8 → 3/8). Issue #2653 is open.
opencode/gpt-5-mini 100% — continues best free model performance.
kimi cooldown extended unexpectedly — was expected to expire ~03:00 UTC but still shows 6d23h (billing cycle from Apr 14, not model-level). Still in extended outage.
opencode:github-copilot failures — gpt-5.4 (7 failed), gemini-3.1-pro-preview (8 failed), claude models struggling.

Key	Remaining	Reason
codex	6d	Billing cycle exhausted
kimi	6d23h	Billing cycle (still extended unexpectedly)
kimi:haiku	3h4m	Model cooldown
opencode:github-copilot/gpt-5.4	2h59m	Persistence

Priority from Apr 14	Status
Fix CLI version mismatch	Still pending — CLI 0.69.0 vs Service 0.69.2
Investigate claude/opus 27% rate	Unchanged — #2653 open
Monitor kimi recovery (Apr 15)	Extended unexpectedly — still 6d23h cooldown
Investigate "no PR/code changes"	Not queried in this review
Audit 47 blocked tasks	Not yet conducted

Fix version mismatch — brew upgrade orch && brew services restart orch. Was pending from Apr 14.
Investigate kimi extended cooldown — kimi was expected to recover ~03:00 UTC but shows 6d23h remaining. Either billing cycle did not reset or cooldown was re-applied. Check KV store or logs for re-trigger.
Continue monitoring claude/opus — issue #2653 is open. Track whether error patterns are now visible after #2652 fix.
Audit blocked tasks — 2-3 blocked from current list. Categorize for review.

kimi's unexpected billing cycle extension is unusual — was supposed to expire overnight. Needs investigation.
No new GitHub issues created this morning beyond existing #2653.
Service is otherwise healthy with clean tick cycles and steady throughput.

Prepared by Orch automation (internal task internal:145498).