[Critical] Judge inherits user project .claude settings (no cwd)

[slaFFik]

Problem

src/server/judge.ts:217 spawns claude with --setting-sources project but no cwd. The judge process inherits the server's working directory — typically the user's project root — and walks up applying their .claude/settings.json, hooks, and project-scoped tool overrides.

Why it matters

• User hooks intercept every judge call, potentially logging or mutating prompt/response.
• Project permission rules leak into a process that is supposed to use no tools.
• A user's .claude/settings.json defaultModel could override the judge's --model haiku.

This is silent score contamination — judge results will quietly differ between projects for reasons that have nothing to do with the variants under test.

Suggested fix

• Set cwd on the judge spawn to a known-clean directory (the run's <run>/ folder, or os.tmpdir() for full isolation).
• Consider passing --setting-sources user for the judge instead of project.

Source

Counselors review (all three agents flagged this).