gitreins-judge-criterion-asserts-foreman-owned-state
Solution written to ~/solution-gitreins-judge-criterion.md. The referenced gitreins/hermes binaries in this sandbox are broken symlinks, so I reproduced the full lifecycle locally to verify the diagnosis and fix rather than asserting it.
Root cause: Criterion 7 (no push) asserted a repo state that the foreman's mandatory post-gate push makes false before the judge runs. It conflated a worker-process constraint ("the worker does not push", true and checkable at commit time) with a judge-time repo-state assertion ("the repository has no push"). Since the judge evaluates state after the foreman pushes, the criterion is unsatisfiable by construction.
Exact fix: 1. Rewrite criterion 7 as actor clause + accepted end state:
The WORKER does not push; the FOREMAN pushes after gates pass, so
origin/master == HEADandgit rev-list --count origin/master..HEAD == 0. 2. Re-board viagitreins task delete+task create(criteria 1–6 unchanged) +task complete --force; capture base SHA before dispatch. 3. Add a pre-dispatchlint_criteria.pythat fails closed on bare state negations (no push,no commit,uncommitted,clean tree,untracked) unless the criterion names the downstream actor/end state.
Verification (all executed):
- Lifecycle repro: OLD criterion 7 ('no push'): FAIL vs CORRECTED criterion 7: PASS, while the recorded worker-time origin/master..HEAD == 1 still proves the worker didn't push.
- Linter: rejects 7. no push (exit 1), accepts the corrected phrasing (exit 0).
- Secondary token-cap near-miss (+1.3% over 5.0M, used ≠ cap) documented with the rule: re-run unchanged first.
# Evidence - Problem class: gitreins-judge-criterion-asserts-foreman-owned-state - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-18T02:17:31.257Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "SYMPTOM: a board task's implementation was fully verified (6 of 7 criteria PASS with live evidence: exact 5 service_call edge pairs on a fresh corpus, 0 generated-stub endpoints, impact listing both dependents, additivity byte-identical at 440 non-service_call lines, cache-hit path restored then fast-pathed, regression test falsified two ways, all build/test gates green, exactly one commit with 'Addresses <id>' and the co-author trailer) yet tier2 returned FAIL - solely on criterion 7, which read 'no push'. The judge proved the violation honestly: git reflog show origin/master showed 'update by push' and HEAD == origin/master. The commit WAS pushed because the FOREMAN pushed it after the gates, as its tick doctrine requires ('a tick that ends with unpushed commits is NOT complete' + mandatory remote-parity verification). ROOT CAUSE: the criterion asserted a REPO STATE (no push) that the foreman's own mandatory post-gate push makes unsatisfiable at judge time. The worker obeyed the brief (it did not push); the criterion conflated 'the worker does not push' (a worker process constraint, true and checkable at commit time) with 'the repository has no push' (false the moment the foreman lands the work, which is the foreman's job). Because the judge evaluates the repo state AFTER the foreman pushes, the criterion is self-defeating by construction. FIX / RULE: never phrase a worker constraint or a work constraint as a judge-time repo-state assertion when another actor (the foreman, CI, a deploy step) is required to change that state as part of the same lifecycle. Write the acceptance state the judge will actually observe, e.g. 'the worker does not push; the foreman pushes after the gates pass, so origin/master == HEAD and git rev-list --count origin/master..HEAD == 0 (the accepted end state)'. Cheap detection: before dispatching, grep the criteria for state assertions that a downstream mandatory step invalidates ('no push', 'no commit', 'uncommitted', 'clean tree') and rewrite them as 'actor X does not do Y; actor Z does Y afterwards, so state S holds'. Remedy used here: gitreins task delete + task create with the same criteria 1-6 and the corrected criterion 7, then task complete --force. SECONDARY MEASUREMENT (same tick, class gitreins-tier2-input-token-cap-exceeded): tier2 died once at 'Input token budget (5.0M) exceeded (5.1M used)' = +1.3% over the 5M rung - the near-miss band, NOT starvation (used != cap). The unchanged re-run completed tier2 with per-criterion evidence and no cap line, confirming the corpus rule 'within ~1-3% of the cap on either side -> re-run unchanged FIRST'.", "environment": "Linux; <project> (Hilo, Rust workspace, 11 crates; JSONL foreman board; gitreins 0.13.x). Worker dispatched by a foreman via hermes chat with a brief that carried the gitreins acceptance criteria verbatim; the foreman owns the guard, the judge and the push.", "language": "rust", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "gitreins-judge-criterion-asserts-foreman-owned-state", "provider": "openrouter", "solved_at": "2026-09-18T02:17:31.257Z", "version": "gitreins 0.13.x"}Solution written to ~/solution-gitreins-judge-criterion.md. The referenced gitreins/hermes binaries in this sandbox are broken symlinks, so I reproduced the full lifecycle locally to verify the diagnosis and fix rather than asserting it.
Root cause: Criterion 7 (no push) asserted a repo state that the foreman's mandatory post-gate push makes false before the judge runs. It conflated a worker-process constraint ("the worker does not push", true and checkable at commit time) with a judge-time repo-state assertion ("the repository has no push"). Since the judge evaluates state after the foreman pushes, the criterion is unsatisfiable by construction.
Exact fix: 1. Rewrite criterion 7 as actor clause + accepted end state:
The WORKER does not push; the FOREMAN pushes after gates pass, so
origin/master == HEADandgit rev-list --count origin/master..HEAD == 0. 2. Re-board viagitreins task delete+task create(criteria 1–6 unchanged) +task complete --force; capture base SHA before dispatch. 3. Add a pre-dispatchlint_criteria.pythat fails closed on bare state negations (no push,no commit,uncommitted,clean tree,untracked) unless the criterion names the downstream actor/end state.
Verification (all executed):
- Lifecycle repro: OLD criterion 7 ('no push'): FAIL vs CORRECTED criterion 7: PASS, while the recorded worker-time origin/master..HEAD == 1 still proves the worker didn't push.
- Linter: rejects 7. no push (exit 1), accepts the corrected phrasing (exit 0).
- Secondary token-cap near-miss (+1.3% over 5.0M, used ≠ cap) documented with the rule: re-run unchanged first.
# Evidence - Problem class: gitreins-judge-criterion-asserts-foreman-owned-state - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-18T02:17:31.257Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "SYMPTOM: a board task's implementation was fully verified (6 of 7 criteria PASS with live evidence: exact 5 service_call edge pairs on a fresh corpus, 0 generated-stub endpoints, impact listing both dependents, additivity byte-identical at 440 non-service_call lines, cache-hit path restored then fast-pathed, regression test falsified two ways, all build/test gates green, exactly one commit with 'Addresses <id>' and the co-author trailer) yet tier2 returned FAIL - solely on criterion 7, which read 'no push'. The judge proved the violation honestly: git reflog show origin/master showed 'update by push' and HEAD == origin/master. The commit WAS pushed because the FOREMAN pushed it after the gates, as its tick doctrine requires ('a tick that ends with unpushed commits is NOT complete' + mandatory remote-parity verification). ROOT CAUSE: the criterion asserted a REPO STATE (no push) that the foreman's own mandatory post-gate push makes unsatisfiable at judge time. The worker obeyed the brief (it did not push); the criterion conflated 'the worker does not push' (a worker process constraint, true and checkable at commit time) with 'the repository has no push' (false the moment the foreman lands the work, which is the foreman's job). Because the judge evaluates the repo state AFTER the foreman pushes, the criterion is self-defeating by construction. FIX / RULE: never phrase a worker constraint or a work constraint as a judge-time repo-state assertion when another actor (the foreman, CI, a deploy step) is required to change that state as part of the same lifecycle. Write the acceptance state the judge will actually observe, e.g. 'the worker does not push; the foreman pushes after the gates pass, so origin/master == HEAD and git rev-list --count origin/master..HEAD == 0 (the accepted end state)'. Cheap detection: before dispatching, grep the criteria for state assertions that a downstream mandatory step invalidates ('no push', 'no commit', 'uncommitted', 'clean tree') and rewrite them as 'actor X does not do Y; actor Z does Y afterwards, so state S holds'. Remedy used here: gitreins task delete + task create with the same criteria 1-6 and the corrected criterion 7, then task complete --force. SECONDARY MEASUREMENT (same tick, class gitreins-tier2-input-token-cap-exceeded): tier2 died once at 'Input token budget (5.0M) exceeded (5.1M used)' = +1.3% over the 5M rung - the near-miss band, NOT starvation (used != cap). The unchanged re-run completed tier2 with per-criterion evidence and no cap line, confirming the corpus rule 'within ~1-3% of the cap on either side -> re-run unchanged FIRST'.", "environment": "Linux; <project> (Hilo, Rust workspace, 11 crates; JSONL foreman board; gitreins 0.13.x). Worker dispatched by a foreman via hermes chat with a brief that carried the gitreins acceptance criteria verbatim; the foreman owns the guard, the judge and the push.", "language": "rust", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "gitreins-judge-criterion-asserts-foreman-owned-state", "provider": "openrouter", "solved_at": "2026-09-18T02:17:31.257Z", "version": "gitreins 0.13.x"}