Daily Review — 2026-06-13

2026-06-13

What Shipped (Last 24h)

Five commits landed, all focused on runner/review pipeline correctness:

Commit	PR	Description
`69a75714`	#3315	bug(runner): apply persistent cooldown for review `parse_error` and `InvalidResponse`
`4440b673`	#3312	enhancement(sync): auto-recover rebroadcast-blocked tasks when review agents return
`08496c5c`	#3311	bug(runner): `find_auth_line_scan` bare '403'/'401' match causes false auth_error on NDJSON timestamps
`2d560843`	#3310	bug(review): timeout path doesn't apply model cooldown — same slow models retried indefinitely
`8b7f4799`	—	fix(review): drop blocking cooldown sleep in subscriber + skip sync re-fires when cooled

Issues Closed

#3314 — north-mini-code-free review parse_error not applying persistent model cooldown → fixed by #3315
#3309 — auto-recover rebroadcast-blocked tasks → fixed by #3312
#3308 — find_auth_line_scan false auth_error on NDJSON timestamps → fixed by #3311
#3307 — review timeout path not applying model cooldown → fixed by #3310

Latest release: v0.80.17 (CI auto-tagged after all 4 PRs merged).

Operational Health

Volume (Last 24h)

Metric	Count
Dispatches	324
PRs created	98
Review decisions	87
Status changes	1181
Errors logged	58
Reroutes	21

High-volume, healthy throughput day.

Agent / Model Outcomes

Agent	Model	Outcome	Count
claude	sonnet	success	115
opencode	mimo-v2.5-free	success	21
codex	gpt-5.5	success	17
opencode	nemotron-3-ultra-free	success	17
opencode	north-mini-code-free	success	11
opencode	nemotron-3-ultra-free	failed	9
claude	sonnet	parse_error	8
opencode	deepseek-v4-flash-free	success	8
opencode	deepseek-v4-flash-free	failed	5
opencode	north-mini-code-free	parse_error	5
codex	gpt-5.3	failed	2
codex	gpt-5.2	failed	2

claude/sonnet dominates with 115 successes, confirming it's the only fully-healthy agent right now.

7 claude/sonnet agent parse_error outcomes ("claude invalid response") — these will benefit from the persistent cooldown fix in #3315 once the service upgrades to v0.80.17. The fix ensures InvalidResponse results in an escalating model cooldown rather than immediate retry.

codex/gpt-5.3: 2 more failures today ("not supported" error), bringing the total to ~19 since Jun 3. Fix requires human config change (see #3313).

Routing State

Active cooldowns:
  claude:haiku          22h51m (persisted)
  opencode              22h15m (billing_cycle_exhausted)
  opencode:deepseek-v4-flash-free  10h52m
  codex:gpt-5.3         9h
  minimax:opus          14h2m
  minimax               1d14h
  kimi                  14h20m
  kimi:haiku            1d13h
  minimax:haiku         6h59m
  codex                 ~42m (expires soon)

Effective routing pool right now:
  claude:sonnet ✓   claude:opus ✓
  codex:gpt-5.5 (in ~42m)

opencode billing cycle cooldown expires ~22h from now, restoring nemotron/north-mini/mimo routing. kimi resumes ~14h from now.

Task #3313 (codex gpt-5.3 permanently unavailable) shows ALL AGENTS COOLED in logs (remaining_secs=2779) — this resolves when codex agent cooldown expires in ~42 min, and claude can handle it immediately. The routing error log is expected and not a concern.

Blocked / Stuck Tasks

Task	Project	Status	Block Reason
internal:153015	bean	blocked	review agent rebroadcast escalated
internal:153022	bean	blocked	review agent rebroadcast escalated
2125	bean	blocked	review agent rebroadcast escalated
internal:153505	bean	blocked	review agent failure threshold exceeded
internal:153551	bean	blocked	CI failure limit (PR #2166 open)

The three "rebroadcast escalated" bean tasks will auto-recover on next tick after upgrading to v0.80.16+ (fix #3312). No manual intervention needed — just upgrade and restart.

Service Lag: CRITICAL ACTION NEEDED

Running: v0.80.13 — Released: v0.80.17 (4 versions behind, all from today)

All 4 missing versions ship critical fixes:

v0.80.14: review timeout now applies model cooldown
v0.80.15: find_auth_line_scan false 403/401 fixed
v0.80.16: rebroadcast-blocked tasks auto-recover
v0.80.17: review parse_error / InvalidResponse applies persistent model cooldown

brew update && brew upgrade orch
brew services restart orch

Open Issues

#	Summary	Action
#3313	codex gpt-5.3 permanently unavailable (19 wasted runs since Jun 3)	Human: remove `gpt-5.3` from codex model pool in `~/.orch/config.yml`
#3297	Service deployment lag (now 4 versions behind)	Run upgrade command above

Tomorrow's Priorities

UPGRADE SERVICE (v0.80.13 → v0.80.17) — auto-recovers 3 blocked bean tasks, stops parse_error cycling, deploys false auth_error fix
Remove gpt-5.3 from config — #3313 is waiting on a human config edit; 2 more wasted dispatches today
Monitor opencode recovery — billing cycle expires ~22h from now (Jun 14 ~21:00 UTC); expect routing pool to expand significantly

Prepared by Orch automation (internal:153781)

← All updates