[ninjaking66]

Adds dual-pass mid-loop rescue (55% + 78% wall-clock), compile-clean import preflight, signature-change audit, directory-companion preload, and issue-identifier path ranking — plus core.fileMode=false at solve start so incidental permission drift never reaches the scored diff. Removes three unused locals.

[github-actions]

OpenRouter PR Judge

Verdict: WARN
Model: anthropic/claude-opus-4.7
Threshold: 70

Score	Value
Overall	74
Real edit	78
Safety	92
Scope	85
Contract	95

Summary

Large but coherent miner-edit PR that adds several mechanical features to agent.py: (1) IDF-weighted path-ranking with tiered identifier-path boosts and stopword/generic-hook filtering, (2) selective region-preload for large context files, (3) extended same-directory companion-file detection across more languages and header/source pairs, (4) a dual-pass mid-loop rescue (early+late threshold) replacing the prior single-pass hail-mary, (5) a richer caller-audit / signature-change audit / import-resolution preflight wired into the self-check refinement turn, (6) a 'confident early-exit' that skips the self-check when all structural gates are clean, (7) stripping redundant mode-header lines from real-content diff blocks and disabling core.fileMode locally, and (8) lockfile/extra-skip additions to the context allowlist. solve() signature and return shape are preserved.

Static Checks

• agent.py has very few substantive changed lines.
• Large patch with 2057 changed lines; judge should inspect for churn.

Judge Reasons

• solve(repo_path, issue, model, api_base, api_key, ...) signature and return-dict shape preserved; multishot wrapper untouched at the contract level.
• Adds real mechanism changes (region preload, dual-pass mid-loop rescue, signature-change audit, import preflight, confident early-exit, mode-header sanitiser) rather than only prompt prose.
• DANGEROUS_PATTERNS untouched; no new third-party imports; stdlib only.
• No new env vars or sampling fields introduced; api_base/api_key path unchanged.
• 'Confident early-exit' is a behavioral change that could occasionally skip a self-check turn — modest impact, but it does change observable refinement behavior so similarity-evasion alone is not the motivation.
• _EDGECASE_GUARDRAIL list (existing in base) sanitises judge-influencing strings inside output patches — preserved, not weakened.
• Sets repo-local git core.fileMode=false to suppress incidental mode-only diff noise; scoped to the task repo via cwd.
• Some additions (e.g. broad _PY_KNOWN_THIRD_PARTY list, large stopword set) are list-data heavy but feed real logic, not cosmetic-only churn.

Risks

• goodhart: mode-header stripping and core.fileMode=false are explicitly motivated as 'reduce cursor-similarity distance' — this targets the similarity scorer, though it also produces cleaner real patches, so impact is mixed.
• scope-drift: very large +1825/-232 diff for one logical area; many helpers added (signature audit, import preflight, region preload) increase shared-infra surface that future forking miners inherit.
• cosmetic-copy: low risk — diff contains substantive new logic and control flow, not just renames/reorderings.

Required Changes

• No required changes returned.