gitreins-tier2-judge-402

Answer

Wrote the solution to ~/gitreins-tier2-judge-402-solution.md.

Important caveat: this environment has no gitreins install or repo (<tool> is a broken pipx symlink; no .gitreins/, tasks.yaml, or verdict.json anywhere). So this is a reconstruction from the supplied context, not a replay against live artifacts. The commands are written to be directly runnable once GITREINS_ROOT points at a real checkout.

Summary of the diagnosis

Failure chain: task complete is an un-isolated atomic pipeline — tier1 → local gates → tier2 judge → write verdict.json → flip tasks.yaml → close row. An intermittent HTTP 402 from the paid tier‑2 judge aborts the command at step 3, so steps 4–6 never run: no verdict.json, task stuck in_progress.

Secondary bug: the offline cache key gitreins-tier2-verdict-402-infra never matches the actual class gitreins-tier2-judge-402 — an off‑by‑one in the lookup key, so the fallback always returns not_found.

The fix

1. Immediate (tick 719 precedent): re-run tier1+local gates with gitreins gates run --tier1 --local --json; only if they pass, synthesize an honest verdict.json with tier2.status="missing", flip tasks.yaml to complete, write guard_result.tier2="missing", and close the history row. The doc includes copy‑paste Python for the YAML/JSON edits (uses system pyyaml 6.0.3, confirmed available).
2. Durable: make tier‑2 best‑effort — bounded retries on {402,429,5xx}, then a non-raising fallback verdict, so the ledger write is always reached. Add verdict_cache_key() normalization so -judge-/-verdict-/-infra collapse to one key.

Verification

The doc ships a three-part assertion script (verdict honest + tier2: missing, ledger row complete with recorded guard result, history row closed) plus a GITREINS_JUDGE_STUB_STATUS=402 regression test that asserts exit=0 and a closed row even when the judge fails.

Key guardrail baked in: force-completion runs only after deterministic gates return 0; tier2: missing is recorded explicitly rather than forged as a pass.