[antiscj007]

improve harness

[github-actions]

OpenRouter PR Judge

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

Score	Value
Overall	72
Real edit	75
Safety	90
Scope	85
Contract	95

Summary

This PR makes several mechanical additions to agent.py: a sed-to-python rewriter in extract_commands, new patch sanitizers (mode-metadata lines, lockfile diffs, build cache directories, .map files), a TypeScript compilation refinement gate (MAX_TS_COMPILE_TURNS) using tsc --noEmit, a thin-patch budget nudge when most time is spent on read-only steps, language-aware directory sibling preloading, hardcoded-secret/credential advisory in the self-check prompt, acceptance-criteria checklist injection, functional-area gap detection (CI/Docker), a second coverage nudge if patch unchanged, large-file skip in preload, and a sed-failure observation hint. Plus disabling git core.fileMode.

Static Checks

• No static findings.

Judge Reasons

• Adds real mechanical changes: command rewriting, new refinement gate (tsc), new sanitizers, new budget-pressure nudge, edit-vs-read step tracking
• Preserves solve() signature and return shape; multishot wrapper unchanged
• No new third-party imports, no new env vars, no new sampling parameters
• DANGEROUS_PATTERNS not weakened
• Inference still routed through validator-supplied api_base/api_key
• Hardcoded-pattern advisory targets generated patch content, not validator-side scoring
• MAX_TOTAL_REFINEMENT_TURNS still capped; MAX_COVERAGE_NUDGES raised to 2 with patch-unchanged guard
• The acceptance-criteria checklist insertion into self-check prompt is prompt-tuning paired with real mechanical changes elsewhere

Risks

• scope-drift: hardcoded patterns like curDay look benchmark-specific and resemble memorization of expected validator inputs (minor)
• goodhart: the new _strip_lockfile_diffs and _strip_mode_metadata_lines could be patch-shaping toward Cursor-baseline style rather than fixing the issue, though stripping lockfiles is a defensible cleanup
• obfuscation: minor — some new helpers use single-letter locals (m, fm, fm2) but readability is overall preserved

Required Changes

• No required changes returned.