Skill Sunset v0.3.0

Skill Sunset v0.3.0

Skill Sunset 0.3.0 adds a privacy-preserving way to check whether a Skill is likely to activate for the requests it claims to support. The audit remains local, deterministic, read-only, and independent of any AI API.

Highlights

• Generate activation-checklist.md alongside the existing audit reports.
• Compare expected request paraphrases with one nearby negative case for each generic Skill.
• Keep missing runtime load evidence explicitly UNKNOWN; it never becomes evidence for RETIRE.
• Normalize activation-report paths across macOS, Linux, and Windows.
• Add four redacted real-world findings with public reproducible examples.
• Stabilize timing-boundary tests without weakening the existing 25% regression threshold.

Try it

npx skill-sunset@latest audit --codex --open

For Claude Code:

npx skill-sunset@latest audit --claude --open

The generated report bundle now includes an activation checklist. It uses only local Skill metadata and synthetic paraphrases provided by the report generator; it does not read chat transcripts, provider credentials, or model usage history.

Evidence boundary

An activation checklist is a local inspection aid, not proof that a runtime loaded or ignored a Skill. A missing observation remains UNKNOWN, and no activation result authorizes deletion. Findings marked TEST still require controlled validation before any cleanup.

See the real-world cases and changelog for details.