Morning Review — 2026-04-16

2026-04-16

Recent Commits (last 24h)

13 commits merged — steady pace of bug fixes and reliability hardening:

Commit	Issue	Description
`d08516df`	#2695	Backend fallback in list_all_by_status — handles unsynced store
`d2d10e33`	—	Capture offset increment — avoids duplicate UTF-8 content
`d3a62ec4`	—	parse_link_next panic guard — end <= start in Link header
`97c28f32`	#2690	Dead code removal — classify_final_status branch 7 unreachable
`dfeaea09`	#2691	auto_unblock dispatch key fix — internal tasks unblocked correctly
`13b3d724`	#2686	Same-agent loop detection — now actually blocks tasks
`65cdd517`	—	Stale LLM_COMMAND_TIMEOUTS reference — removed from llm.rs
`1645e055`	—	Agent prompt hardening — longer git window, search-before-file
`1c5f0e59`	—	Remove redundant llm_semaphore — raises slow-tick threshold
`27789769`	—	Revert has_leaks() scope — all patterns block posting (security)
`7984d49a`	#2685	Evening retrospective 2026-04-15
`119b7551`	#2684	Exit-0 empty-output = silent failure — model cooldown applied
`e6d63af2`	#2681	Batch session active per task — no duplicate tmux calls

Version mismatch: CLI 0.69.15, Service 0.69.18 — brew upgrade orch && brew services restart orch still pending from Apr 14
Logs: clean tick cycle (~1.4–1.6s), no persistent errors
Recurring WARN: multi-agent degradation detected (codex, kimi, glm) — expected; all three are in billing-cycle cooldowns
One HTTP retry on GitHub API sync (transient, auto-recovered)
Jobs dispatched this morning: morning-review, morning-briefing, twitter-trending-watch

Agent	Model	Success	Failed	Rate
claude	sonnet	37	14	73%
minimax	opus	30	1+4rl	86%
opencode	gpt-5-mini	25	1	96%
opencode	minimax-m2.5-free	23	1	96%
glm	opus	14	5+3rl	64%
opencode	nemotron-3-super-free	13	3+3pe	68%
opencode	gpt-5.4	1	8	11%
opencode	gemini-3.1-pro-preview	0	4	0%
claude	haiku	1	4	20%
opencode	claude-opus-4.6	0	3	0%
opencode	claude-sonnet-4.6	0	2	0%

Notable:

opencode/gpt-5-mini at 96% — continues as best github-copilot model; other copilot models still failing heavily.
github-copilot models: gpt-5.4 (11%), gemini (0%), claude-opus-4.6 (0%), claude-sonnet-4.6 (0%) — provider-side issue persists.
claude/sonnet improved — 73% (up from 67% in retro). Trend is positive.
claude/opus — not invoked in this 24h window (cooldown / circuit-breaker avoided it).
glm/opus at 64% — partially in cooldown (2h19m remaining), rate limits occurring.

Priority from Apr 15 Evening	Status
Fix version mismatch	Still pending — 0.69.15 vs 0.69.18
Investigate github-copilot failures	Ongoing — gpt-5-mini (96%) is fine; all other copilot models failing. Provider issue.
Monitor claude/opus	Sidestepped — not dispatched in 24h; #2653 closed
kimi billing cycle	Still extended — 5d23h remaining; human intervention may be needed

Throughput is healthy. Error count (61) vs dispatch (260) = ~23% error rate, consistent with github-copilot model failures.

Fix version mismatch — brew upgrade orch && brew services restart orch. Has been pending since Apr 14; now 3 versions behind (0.69.15 vs 0.69.18).
github-copilot model investigation — gpt-5-mini is 96%, but every other copilot model (gpt-5.4, gemini, claude-opus-4.6, claude-sonnet-4.6) is at 0-11%. This is a provider-level pattern. Consider routing exclusions for failing models until they stabilize.
kimi billing cycle — Still 5d23h despite expected reset. Needs human review of billing status or manual cooldown clear once confirmed.

No new GitHub issues to file — all observable problems either have existing issues or are covered by current retro priorities.
The has_leaks() revert (27789769) restores conservative security posture — good.
same-agent loop detection fix (#2686) was critical — blocked tasks weren't actually being blocked before.
Service otherwise healthy: clean ticks, active throughput, no anomalous error patterns.

Prepared by Orch automation (internal task internal:145741).