Daily Review — 2026-06-18

2026-06-18

What Shipped (Last 24h)

An exceptionally productive day: 8 commits merged and 10+ issues closed. Almost all were targeted bug fixes to the runner, parser, router, jobs scheduler, and worktree subsystems.

Commit	PR	Description
`92c98607`	#3334	fix(router): skip agent dispatch for repos with active GitHub Actions billing failures
`cbf54563`	#3333	fix(parser): add missing 'fixed' status alias to normalize_status
`7cdcf359`	#3329	fix(jobs): quarantine per-project load_jobs failures instead of aborting scheduler tick
`a77d8f83`	—	test(worktree): cover origin-anchored worktree creation end-to-end
`e678d400`	—	fix(worktree): anchor new worktrees on origin/<default_branch>
`735d0b26`	#3325	fix(router): use alias instead of full API model ID as default route
`28faa4da`	#3327	fix: skip empty model success resets
`0234c4f0`	#3323	docs(posts): daily review 2026-06-18 (earlier draft)

Closed Issues (Today)

Issue	Description
#3332	ops: GitHub Actions billing failures on gabrielkoerich/bean
#3330	bug(parser): normalize_status missing 'fixed' alias
#3326	bug(jobs): duplicate job id in one project aborts scheduler tick
#3325	bug(router): LLM router pool uses full API model ID for codex alias
#3324	bug(runner): record_agent_success with None model writes incomplete KV keys
#3321	bug(codex): app-server event stream lag classified as silence
#3318	bug(runner): billing_cycle_exhausted for github-copilot triggers wrong cooldown
#3314	bug(runner): opencode/north-mini-code-free review parse_error
#3309	enhancement(sync): auto-recover tasks blocked by 'review recovery timeout'
#3308	bug(runner): find_auth_line_scan bare '403'/'401' match causes false positives
#3307	bug(review): timeout path doesn't apply model cooldown
#3302	bug(runner): 'You have exceeded your monthly quota' (GitHub Copilot) not classified
#3301	bug(runner): opencode 'Upstream idle timeout exceeded' classified as silence
#3300	bug(parser): normalize_status missing 'verified' and 'alert'

Notable fixes:

#3334 / #3329 combo: The scheduler can now survive one bad project's jobs config without aborting the entire tick. Paired with the billing-failure router skip, gabrielkoerich/bean GitHub Actions failures no longer cascade to orch routing.
#3325: The LLM router was storing the full API model ID (e.g. gpt-5.4-codex) instead of the configured alias, causing model-level cooldown lookups to miss. Fixed.
#3327: Success resets with None model were overwriting failure counts for all models on an agent. Fixed.
Worktree anchoring (e678d400 + a77d8f83): New worktrees now anchor on origin/<default_branch> instead of local HEAD, preventing stale base refs on long-running services.

Operational Health

Volume (Last 24h)

Metric	Count
Status changes	257
Dispatches	80
Branch deletes	74
Pushes	66
Routed	35
Review starts	32
PRs created	31
Reroutes	6
Errors	12
Push/worktree recoveries	2

High throughput day — 80 dispatches, 31 PRs, 32 reviews.

Agent / Model Outcomes (Last 24h)

Agent	Model	Outcome	Count
claude	sonnet	success	23
kimi	opus	success	9
opencode	deepseek-v4-flash-free	success	7
codex	gpt-5.4	success	4
codex	gpt-5.5	success	4
opencode	nemotron-3-ultra-free	success	3
minimax	opus	rate_limit	3
minimax	sonnet	rate_limit	2
claude	sonnet	failed	2
claude	sonnet	(no outcome recorded)	2
opencode	mimo-v2.5-free	success	2
minimax	opus	success	1
claude	haiku	success	1
claude	haiku	failed	1
claude	haiku	blocked	1
claude	sonnet	rate_limit	1
kimi	opus	failed	1

Effective pool: claude/sonnet is the workhorse (23 successes). Codex is back and healthy with gpt-5.4 and gpt-5.5 (8 combined successes — the dead gpt-5.2/5.3 models are behind us). Kimi/opus is running well (9 successes). Minimax is degraded with 5 rate-limit hits and an agent-level cooldown.

Active Cooldowns

Key	Remaining	Reason
minimax	~1d22h	persisted (rate limits)
minimax:haiku	~2h59m	persisted

No other agent or model cooldowns are active. Routing is healthy across all other agents.

Routing Health

Clean logs — no AllAgentsCooledError, no watchdog triggers, no GitHub API connectivity failures. Sync ticks are completing in 1.7–3.0 s. The LLM router is making accurate routing decisions (this task itself was correctly classified as claude/medium).

Blocked / Stuck Tasks

Task	Project	Status	Tries	Block Reason
#3331	gabrielkoerich/orch	blocked	3	Service v0.80.20 lags latest v0.80.24 — upgrade needed
#3317	gabrielkoerich/orch	blocked	3	codex `gpt-5.2` in config — waiting on human config edit
#3313	gabrielkoerich/orch	blocked	8	codex `gpt-5.3` in config — waiting on human config edit

Note: issue #3331 title says "lags v0.80.22" but the engine logs show latest_version=0.80.24. The gap is 4 patch versions, not 2. The fixes in #3327, #3329, #3333, #3334 have all been merged and should be in one of those releases.

Service Version Gap

Running: v0.80.20 · Latest: v0.80.24

Four versions behind. Today's critical fixes (scheduler job quarantine, router billing-failure skip, normalize_status aliases, model success reset) are merged but not yet deployed. Recommend:

brew update && brew upgrade orch
brew services restart orch
orch -V

Priorities for Tomorrow

Upgrade service to v0.80.24 — four versions behind; today's critical fixes are sitting undeployed. Run the post-push cycle from CLAUDE.md.
Config edit: remove gpt-5.2 and gpt-5.3 from codex model pool — #3313 and #3317 remain blocked until ~/.orch/config.yml is updated. The generic cooldown system works, but the models keep re-entering the pool when cooldowns expire.
Monitor minimax recovery — agent-level cooldown expires in ~2 days. Verify it re-routes cleanly when it clears.
Watch 2 unrecorded claude/sonnet outcomes — two runs have no outcome in task_runs. Could be the runner crash path or a timing edge case. Monitor whether the pattern recurs.

Prepared by Orch automation (internal:154084)

← All updates