5DVFGXociNnsbyPhzFe4iY6SisLEUE6eASWUf11Mn3KXffqU update

[ninja66-contributor]

No description provided.

[github-actions]

OpenRouter PR Judge

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

Score	Value
Overall	74
Real edit	76
Safety	95
Scope	90
Contract	98

Summary

This PR adds several mechanical changes to agent.py: a new _strip_mode_headers_from_content_diffs sanitization step that strips old/new mode header lines from content-bearing diffs; a TF-IDF-style doc-frequency boost in _rank_context_files (rare terms get +8, common stay at floor +3); an expanded identifier-stopword set with hook-prefix filtering; a type-aware self-check cue prepended to the self-check prompt; a soft-nudge mid-loop turn fired when many steps elapse without edits; an under-deliver detection path in the multishot wrapper that triggers attempt 2 with a tailored bootstrap when patch count is below estimated file count; and a small bump to the recently-observed-paths window. The solve signature, return shape, stdlib-only constraint, and DANGEROUS_PATTERNS are preserved.

Static Checks

• No static findings.

Judge Reasons

• Multiple distinct mechanical changes: patch sanitization, file-ranking score function, multishot underdeliver branch, and a new soft-nudge loop turn — not prompt-only.
• TF-IDF bonus is floor-clamped at the prior +3 so it cannot decrease ranking, a real algorithmic refinement rather than a cosmetic rename.
• solve(...) signature, return dict shape, env-var allowlist, and validator-owned boundaries appear unchanged.
• No new third-party imports, no new network endpoints, no sampling parameters added.
• DANGEROUS_PATTERNS list left intact; no weakening of safety filters.
• Some new prompt strings (type-aware cue, under-deliver bootstrap) include style/imitation guidance that edges toward judge-facing prose, but each is paired with real mechanical wiring (new branch / new gate), so not prompt-only.
• The under-deliver bootstrap text instructs the model to 'imitate the codebase's existing style' — borderline style-anchoring but reasonable since it triggers only on multi-file tasks where attempt 1 under-covered.

Risks

• goodhart: new _strip_mode_headers_from_content_diffs strips mode header lines specifically because 'the LLM judge dings agents for unrelated file mode churn' — this is explicit judge-shaping, though the effect (removing chmod-only noise) is also a legitimate cleanup.
• goodhart (minor): type-aware self-check cue and under-deliver bootstrap contain motivational/style language directed at shaping output for the judge; modest in scale and paired with mechanism.

Required Changes

• No required changes returned.