Org-wide OpenCode model-pool outage: all 3 provider families down since 2026-07-25 — human credential/budget action required

[seonghobae]

Summary

Every opencode-review model-pool run across the org has ended OPENCODE_MODEL_POOL_OUTCOME: exhausted since ~2026-07-25T00:30Z (~34h). The last real model review completed 2026-07-24T23:42:52Z (run 30134503832, 43.5m). Later "success" runs (e.g. 30193838586 at 08:01Z, 30196130362 for #579) passed only via pre-existing same-head approvals — their pools were also exhausted. Approved-but-unreviewed PRs (e.g. #618 at head c4a9023) cannot merge until this is fixed.

Per-provider signature (human action needed)

Provider family	Failure class	Evidence	Action
github-models (majority of pool)	authentication-or-permission (~5s fatal per candidate)	runs 30153577309 (att. 21), 30195391225, 30196418659, 30196803994 (naruon#1081: 24× auth class)	STRIX_GITHUB_MODELS_TOKEN secret (PAT, models: read) is rejected — likely expired. Rotate/renew the org/repo secret. Fallback github.token is also failing, so check GitHub Models org entitlement/billing too.
openai gpt-5.6-luna	quota-or-budget	same runs; persists >34h across a daily reset boundary	Hard budget cap reached — raise budget or wait for the billing-cycle reset.
openrouter (deepseek-v3.2, qwen3-coder)	credit-exhausted (HTTP 402)	same runs; root cause documented in merged #621: runaway runs 30120972549 (~3h × 102 pool cycles) and 30122360773 (700+ attempts into 402) burned the entire org credit	Top up OpenRouter credits. Spend guards from #621 now bound future burn.

Org-wide confirmation (2026-07-26 09:29–09:44Z batch, all identical signature)

kaefa#70, html4tree#178, pg-erd-cloud#596, clearfolio#161, bandscope#645, newsdom-api#401, naruon#1081, fast-mlsirm#259 — all failure, all provider classes as above.

What was already done autonomously

• PR Handle stacked PR retarget and merge-queue trigger gaps #618 adversarial review: no PR-caused defects; every other required check green at head c4a9023; stale auto-merge (from 07-25) cleared via scheduler dispatch 30195149878.
• Same-head OpenCode re-dispatched twice (runs 30195391225, 30196418659) per the "opencode-review failures are infrastructure → re-run only" rule; both exhausted.
• No timing/gate changes were made (fixed rule: review timing untouched; no required-check weakening).

Self-healing after fix

No code change needed. Scheduled scans (:00/:30 repo, :15/:45 org) re-dispatch stale OpenCode reviews automatically once providers work again; approved heads then auto-merge. For an immediate retry: gh api repos/ContextualWisdomLab/.github/dispatches -f event_type=merge-scheduler -F 'client_payload[pr_number]=618' -F 'client_payload[trigger_reviews]=true' -F 'client_payload[enable_auto_merge]=true' (note: omitting the boolean fields makes GitHub expressions coerce them to false — see gotcha below).

Scheduler payload gotcha (worth a follow-up docs/test PR)

pr-review-merge-scheduler.yml maps client_payload.trigger_reviews != false → GitHub expressions coerce absent (null) to false via numeric coercion, so null != false is false. A bare client_payload[pr_number]=N dispatch runs with TRIGGER_REVIEWS/ENABLE_AUTO_MERGE/UPDATE_BRANCHES: false (observed: run 30195265329 → decision block; run 30195382302 with explicit trigger_reviews=true → review_dispatch).