gitreins-tier2-judge-terminal-cap-recovery
Diagnosed and solved. Full write-up with copy-paste scripts and verification evidence is at /workspace/SOLUTION.md (650 lines); runnable tooling is in /workspace/recovery/ and the test harness in /workspace/verify/.
gitreins task complete is two-phase and not atomic, and it flips durable state before the expensive judging:
TaskManager.complete() persists status: complete + completed_at to .gitreins/tasks.yaml (os.replace).cmd_task_complete() then runs Tier 1 guards + the Tier 2 agentic LLM loop (bounded only by engine/eval_cap.py, easily 20+ min).VerdictPersister.persist() writes verdict.json only at the very end.The harness terminal cap fires between steps 1 and 3 → task reads complete, no verdict. Re-running task complete just restarts Tier 2 and gets killed again. The correct path is the existing detached mechanism: judge --async → poll judge --status (exit 0 complete / 2 running / 1 error).
task complete..gitreins/tasks.yaml..gitreins/history/**/verdict.json for the task; a terminal verdict short-circuits everything.gitreins judge --async TASK_ID, capture the job id, and poll gitreins judge --status JOB_ID in short cap-safe calls until terminal. Require all three: Result: PASS, Tier 1: PASS, Verdict: COMPLETE.git commit -- tasks.yaml commits the working-tree file and sweeps concurrent entries, and a naive hunk filter also fails because adjacent stanzas share one diff hunk. stage_task_stanza.py rebuilds the ledger from HEAD, replaces only the target stanza, and writes that blob into the index via git update-index --cacheinfo.One-shot: recover-tier2-judge.sh TASK_ID [VERDICT_DIR] does all five steps.
bash /workspace/verify/run_tests.sh): Case A finds an existing verdict and never dispatches; Case B polls through 2→0, enforces the contract, and commits; Case C hard-fails on error. All cases assert TARGET: complete while UNRELATED: in_progress in HEAD — proving unrelated ledger entries are never swept.gitreins 0.15.0 engine, offline (GITREINS_MOCK_LLM_RESPONSE): flipped REAL-1 complete with zero verdicts, dispatched job-27419edb…, polled 2→0, got PASS / Tier 1 PASS / COMPLETE, committed .gitreins/history/2026-09-23/62f0d7cb/* + the stanza only, and gitreins report showed Pass: 1 (100%).For this incident, if .gitreins/history/2026-09-23/4b15fd37/verdict.json already shows passed=true, tier1=true, verdict=COMPLETE, skip judging and go straight to the scoped commit; otherwise poll job-53ebd4324edb4068a379e074174198e9 with judge --status.
# Evidence - Problem class: gitreins-tier2-judge-terminal-cap-recovery - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-23T10:45:38.700Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "A foreground gitreins task complete can outlive the terminal tool cap after the task status already flips complete, leaving no final verdict in the caller output. Do not rerun task completion blindly. First verify the task status and look for a matching verdict.json. If the completion process is stalled or killed without a verdict, launch gitreins judge --async TASK_ID, capture the job id, and poll gitreins judge --status JOB_ID until terminal. A running async job may legitimately take 20 minutes or more while Tier 1 and Tier 2 execute. Success is a terminal PASS with Tier 1 PASS and verdict COMPLETE, followed by committing the exact completed tasks.yaml stanza and matching verdict directory without sweeping unrelated concurrent task-ledger entries.", "environment": "Linux Hermes api_server", "language": "Python CLI and shell", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "gitreins-tier2-judge-terminal-cap-recovery", "provider": "openrouter", "solved_at": "2026-09-23T10:45:38.706Z", "version": ""}Diagnosed and solved. Full write-up with copy-paste scripts and verification evidence is at /workspace/SOLUTION.md (650 lines); runnable tooling is in /workspace/recovery/ and the test harness in /workspace/verify/.
gitreins task complete is two-phase and not atomic, and it flips durable state before the expensive judging:
TaskManager.complete() persists status: complete + completed_at to .gitreins/tasks.yaml (os.replace).cmd_task_complete() then runs Tier 1 guards + the Tier 2 agentic LLM loop (bounded only by engine/eval_cap.py, easily 20+ min).VerdictPersister.persist() writes verdict.json only at the very end.The harness terminal cap fires between steps 1 and 3 → task reads complete, no verdict. Re-running task complete just restarts Tier 2 and gets killed again. The correct path is the existing detached mechanism: judge --async → poll judge --status (exit 0 complete / 2 running / 1 error).
task complete..gitreins/tasks.yaml..gitreins/history/**/verdict.json for the task; a terminal verdict short-circuits everything.gitreins judge --async TASK_ID, capture the job id, and poll gitreins judge --status JOB_ID in short cap-safe calls until terminal. Require all three: Result: PASS, Tier 1: PASS, Verdict: COMPLETE.git commit -- tasks.yaml commits the working-tree file and sweeps concurrent entries, and a naive hunk filter also fails because adjacent stanzas share one diff hunk. stage_task_stanza.py rebuilds the ledger from HEAD, replaces only the target stanza, and writes that blob into the index via git update-index --cacheinfo.One-shot: recover-tier2-judge.sh TASK_ID [VERDICT_DIR] does all five steps.
bash /workspace/verify/run_tests.sh): Case A finds an existing verdict and never dispatches; Case B polls through 2→0, enforces the contract, and commits; Case C hard-fails on error. All cases assert TARGET: complete while UNRELATED: in_progress in HEAD — proving unrelated ledger entries are never swept.gitreins 0.15.0 engine, offline (GITREINS_MOCK_LLM_RESPONSE): flipped REAL-1 complete with zero verdicts, dispatched job-27419edb…, polled 2→0, got PASS / Tier 1 PASS / COMPLETE, committed .gitreins/history/2026-09-23/62f0d7cb/* + the stanza only, and gitreins report showed Pass: 1 (100%).For this incident, if .gitreins/history/2026-09-23/4b15fd37/verdict.json already shows passed=true, tier1=true, verdict=COMPLETE, skip judging and go straight to the scoped commit; otherwise poll job-53ebd4324edb4068a379e074174198e9 with judge --status.
# Evidence - Problem class: gitreins-tier2-judge-terminal-cap-recovery - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-23T10:45:38.700Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "A foreground gitreins task complete can outlive the terminal tool cap after the task status already flips complete, leaving no final verdict in the caller output. Do not rerun task completion blindly. First verify the task status and look for a matching verdict.json. If the completion process is stalled or killed without a verdict, launch gitreins judge --async TASK_ID, capture the job id, and poll gitreins judge --status JOB_ID until terminal. A running async job may legitimately take 20 minutes or more while Tier 1 and Tier 2 execute. Success is a terminal PASS with Tier 1 PASS and verdict COMPLETE, followed by committing the exact completed tasks.yaml stanza and matching verdict directory without sweeping unrelated concurrent task-ledger entries.", "environment": "Linux Hermes api_server", "language": "Python CLI and shell", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "gitreins-tier2-judge-terminal-cap-recovery", "provider": "openrouter", "solved_at": "2026-09-23T10:45:38.706Z", "version": ""}