1.2.0

• Multi-host: the roll call now covers every Agent Skills host, not just Claude Code.
  --agent picks Claude Code, Codex, Cursor, Gemini CLI, Copilot, OpenCode, Amp, Goose,
  Kiro, Windsurf, the universal ~/.agents/skills convention, a comma list, or all;
  with no flag the host is detected from CLAUDECODE, CODEX_SANDBOX, CURSOR_AGENT
  or GEMINI_CLI, and falls back to all. The report header names the host and every
  root it found, and under several hosts rows are grouped by who reads them.
• Project walk-up now finds .agents/skills (and .goose/, .kiro/, .windsurf/)
  alongside .claude/skills; --add-dir includes them too.
• Duplicate and overlap warnings only fire when one host would see both skills: the
  same skill installed for Claude Code and for Codex is a copy, not a conflict.
• Codex: skills switched off with [[skills.config]] enabled = false in
  ~/.codex/config.toml (or a project's .codex/config.toml) are reported as warnings.
  TOML is read with tomllib on 3.11+ and a small stdlib fallback on 3.10.
• Agent Skills spec checks: a name outside ^[a-z0-9]+(-[a-z0-9]+)*$ or over 64
  chars, and a description over 1024 chars, are warnings ("strict hosts skip it").
  --lint flags Claude-only frontmatter (context:) when other hosts read the skill.
• --audit: Codex's --dangerously-bypass-approvals-and-sandbox and the --yolo alias
  (Codex and Gemini CLI) count as skips-permissions; approval_policy = "never" is a
  review hit; openai.com, cursor.com, agentskills.io and skills.sh join the known hosts.
• JSON rows gain agent, where, visible_to and frontmatter_keys.
• SKILL.md rewritten for any host, with spec license, compatibility and metadata
  frontmatter; agents/openai.yaml added for Codex / ChatGPT UI metadata.
• 21 new tests (66 total), including a fallback-vs-tomllib equivalence check.

1.1.0

• --lint: description quality (too short, no "use when" cue, near-duplicate of a
  sibling, heavy) and a per-session context-cost estimate.
• --audit: content scan for hidden Unicode, pipe-to-shell, permission-bypass flags,
  injection phrasing, encoded blobs and unfamiliar hosts, with high / review / info
  severities. Dangerous strings inside test files are downgraded to review.
• --fix / --fix --apply: un-nest manifest-less plugin layouts and rename folders
  to match name. Dry run by default; never touches the plugin cache.
• --strict exit code and --skills-dir for checking a skills repo in CI.
• Scopes: project walk-up to the git root, --add-dir, plugin cache and synced
  plugins, and skills-directory plugins (.claude-plugin/plugin.json + skills/),
  which are now correctly recognised as valid rather than flagged as nested.
• Errors (will not register) and warnings (registers, but) are reported separately.
• 44-test suite and a CI matrix (Windows / macOS / Linux, Python 3.10 / 3.13) that
  also runs the tool on its own skill in strict mode.
• Docs corrected: Claude Code hot-reloads SKILL.md edits, not just adds and removes;
  --bare mode is the case without a watcher; /reload-plugins for plugin hooks.