Skip to content

5DeX1wy4rnt8BpdALoq77fByMZxuVaZkKoiKwC6Z1ncuCz2d update - #1571

Open
ninja66-contributor wants to merge 1 commit into
unarbos:mainfrom
ninja66-contributor:miner/baihu003-1778777233
Open

ninja66-contributor wants to merge 1 commit into
unarbos:mainfrom
ninja66-contributor:miner/baihu003-1778777233

Conversation

@ninja66-contributor

Copy link
Copy Markdown

No description provided.

@github-actions

Copy link
Copy Markdown

OpenRouter PR Judge

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

Score Value
Overall 72
Real edit 70
Safety 95
Scope 90
Contract 100

Summary

The PR adds a new soft-nudge mid-loop turn (fires once at step>=6 / elapsed>=90s when no patch yet, before the hail-mary threshold), lowers the mid-loop hail-mary fraction from 0.55 to 0.50, expands an identifier stopword set and adds a generic-hook-prefix filter to avoid spurious identifier path-boosts for things like useState/handleClick, prepends a type-aware cue to the self-check prompt, and widens the recently-observed-paths log window from 30 to 80. These are mechanical changes (new prompt turn, new gating constants, new filter set), not pure prompt churn or rename-only.

Static Checks

  • No static findings.

Judge Reasons

  • Adds a real new mechanism: soft-nudge turn with its own counter, threshold, and gating, wired into the solve loop before the existing hail-mary path.
  • Adds a genuine filter (_HOOK_GENERIC_PREFIXES_LOWER) that changes _issue_identifier_path_boost behavior, reducing false-positive ranking boosts from generic React/JS hook names.
  • Self-check prompt now conditionally prepends a type-aware cue derived from issue keywords; this is prompt content but is paired with the new soft-nudge mechanism and filter changes, not prompt-only.
  • solve() signature, return shape, stdlib-only, DANGEROUS_PATTERNS, and validator-owned boundaries are unchanged.
  • No new env vars, no new hosts, no sampling fields, no obfuscation, no removal of MINER-EDITABLE/VALIDATOR CONTRACT markers.

Risks

  • goodhart: the expanded _IDENTIFIER_STOPWORDS list is large and tuned-looking (lots of React/JS-specific tokens); could be benchmark-shaped, but it has a plausible mechanical justification (avoid generic boosts) so warn-level only.
  • goodhart: type-aware cue in self-check prompt is judge-adjacent prose, but it instructs the inner model about correctness rather than addressing the judge directly.

Required Changes

  • No required changes returned.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant