[Patrick-Erichsen]

When a bundled SkillSpector child fails, the stored worker diagnostic omits which declared skill failed and what its structured report says. A current Solaris job ran 319.7 seconds and ended with a non-timeout exit 2 after 20.6 seconds, leaving only the exit/termination packet.

Record the child's zero-based index in the existing sorted root list, root/completion counts, budgets, and allowlisted terminal statuses, reasons and analyzer counts. Precreate report files with mode 0600 for the existing restricted reader. Preserve sequential execution, budgets, accepted exit-1 findings, retry behavior and scanner/judge gates. Paths, arbitrary hashes, findings, provider errors and scanner streams remain excluded.

Validation:

• Real worker orchestration with controlled scanner subprocesses: absent context → root 1 of 2 after one completion, with a fixed failure reason; failure/judge state is unchanged and privacy canaries are excluded.
• 96 focused controls and the full coverage run pass: 8,366 tests passed, five configured skips. Static and TypeScript/build gates pass. The existing deadline regression now checks private empty output after timeout and preserves rejection/no judge handoff; the first failed assertion/run is retained.
• The installed Linux SkillSpector CLI preserves private output and writes valid JSON on an unchanged original root offline. Independent restricted P0–P3 review covers all five changed files with zero findings.

This proves diagnostic behavior. Solaris's model-enabled failure cause and publication remain unresolved; the related issue stays open. Related: #4043. The privacy contract from #4062 is preserved.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
clawhub	Ready	Preview	Oct 30, 2026 8:58am UTC

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.