Changelog

All notable changes to this project will be documented in this file.

[1.6.1] — May 2026

Fixed

• Narrowed B1 (Credential Access) trigger: ~/.config/ alone no longer escalates to CRITICAL — only triggers when combined with credential/token/key/secret patterns. Generic app config paths now correctly classified as B5 (WARNING), preventing false DO NOT INSTALL verdicts.
• Added warning-clearing rule: warnings investigated and found to be false positives no longer count toward the verdict. A skill with all warnings cleared and no CRITICALs now correctly receives APPEARS SAFE.

Added

• Explicit verdict label table in SKILL.md — exact emoji+phrase required (DO NOT INSTALL / PROCEED WITH CAUTION / APPEARS SAFE)
• Eval scenarios (evals/) covering all three audit modes, all check categories, and prompt injection resistance — 97% score, 1.28× uplift

[1.6.0] — May 2026

Changed

• Refactored three workflows into a shared Audit Procedure section — each mode now only specifies its unique setup steps
• Consolidated three inline transparency notices into a single lookup table
• Collapsed Known Risks and Self-Audit Limitation into one inline note
• Removed Common Install Methods Reference section (duplicated Workflow 1 resolution table)
• Collapsed speculative resolution rows (npm, pip, marketplace) into a single fallback entry
• Tightened Fetch Safety Boundary to explicitly cover credential values in fetched content (W007)
• SKILL.md reduced from ~300 lines to 131 lines

Fixed

• Declared references/ files as docs in tile.json so Tessl bundler includes them
• Renamed .tileignore to .tesslignore (deprecation fix)
• Excluded test-fixtures/ from published bundle via .tesslignore
• Fixed allowed-tools format (Read,WebFetch,Glob → Read WebFetch Glob)
• Added tile.json and publish workflow (tile was not previously auto-publishing)

Added

• references/index.md as docs entrypoint linking security-checks.md and report-format.md
• Known risks documentation for Tessl security findings W007, W011, W012

[1.0.0] — April 2026

Added

• Initial release of skill-safety-auditor
• Three audit modes: Pre-download (Mode 1), Downloaded not installed (Mode 2), Already installed (Mode 3)
• 14 security checks across 4 categories: Frontmatter, Script Content, Prompt Injection, Source Provenance
• Structured report template with severity tiers: Critical / Warning / Info
• Bundled reference files: security-checks.md and report-format.md
• Support for all common skill install methods: GitHub URLs, git clone, npx, npm, pip, curl/wget, marketplace names, .skill files