v0.15.0 — "Fable Ultra"

Our biggest release ever: 42 commits, 378 files changed, ~18.5k insertions
since v0.14.1. The headline is a full plugin architecture for distributing
lint rules as PyPI packages, alongside the removal of the built-in LLM fix
engine, a new dead-file detection rule, and a long list of autofix,
correctness, security, and performance fixes.

Highlights

Plugin architecture (#312, #311, #313, #314)

Rules can now be distributed as ordinary pip-installable packages via
skillsaw.plugins entry points — reviewable, versioned, and auditable,
unlike repo-local custom rule files.

• Rule plugins: any package exposing a skillsaw.plugins entry point
  contributes rules, namespaced and attributed (source: plugin:<name>
  in JSON output).
• Extension points: plugins can register custom repo types
  (SKILLSAW_REPO_TYPES) and lint-tree contributors
  (SKILLSAW_TREE_CONTRIBUTORS), pulling new file types into the parse
  tree and content linting.
• Plugin CLIs: skillsaw <name> dispatches to a plugin's
  skillsaw-<name> console script — registered plugins only; arbitrary
  executables on PATH are never run.
• Management: new skillsaw plugins command, --no-plugins flag, and
  plugins: {disable: [...]} config block. A plugin that fails to load
  reports a plugin-load-error violation instead of aborting the lint.
• GitHub Action: new plugins input installs plugins in CI (#314).
• Docs: a full authoring guide (docs/plugins.md), a scaffold example
  under examples/plugins/, and a skillsaw-create-plugin skill.

⚠️ Breaking: LLM fix engine removed (#310)

The built-in LLM autofix path is gone in favor of coding agents (see the
skillsaw-fix skill and per-rule skillsaw explain guidance):

• Removed flags: skillsaw fix --llm/--ai, --model, --max-iterations,
  --all, -y/--yes, --workers, --patch-file, --apply-patch, and
  the hidden skillsaw lint --fix/--llm/--dry-run.
• Removed pip extras: skillsaw[llm], [vertexai], [bedrock]
  (installing them now fails).
• Removed rule API: AutofixConfidence.LLM, llm_fix_prompt,
  llm_fix_frontmatter — custom rules and plugins using them must update.
• The llm: config section is now an ignored unknown-key warning;
  SKILLSAW_MODEL is ignored.

⚠️ Breaking: GitHub Action defaults to --no-custom-rules (#323)

For CI safety on untrusted PRs, the Action no longer executes repo-local
.skillsaw-custom.py rules by default. Nothing fails — custom rules
silently stop running. Set no-custom-rules: false to opt back in, or
migrate to a rule plugin. The CLI default is unchanged, and custom-rule
loading now prints a clear one-line colored warning (#359).

New rule: agentskill-unreferenced-files (#347)

Detects dead files in skill directories — files never referenced from
SKILL.md or anything it transitively references. Warning severity,
enabled: auto on skill/plugin/marketplace repos, with an exclude
config parameter. Reachability is thorough: markdown links and text
mentions (case-insensitive), directory mentions (./fonts,
assets/fonts, fonts/), and Python import chains — both in .py
helpers and inside fenced code blocks — are all followed transitively.

Spec compliance & validation

• metadata.pluginRoot honored when resolving marketplace plugin
  sources, per the Claude Code marketplace spec — bare source names now
  resolve, with a root-relative fallback for marketplaces that mix styles;
  traversing/absolute pluginRoots are rejected (#355).
• marketplace.json source validation per the current spec: duplicate
  plugin names, ../absolute paths, typed-source required fields, unknown
  source types (#338).
• Frontmatter hooks scanned: hooks-dangerous and hooks-prohibited
  now inspect hooks: declared in skill/agent frontmatter, not just
  hooks JSON files (#346).
• Hooks & MCP validation synced with current Claude Code docs —
  exec-form args, MCP timeout, MessageDisplay event, ws
  transport (#336).
• Digit-leading skill names (e.g. 1password) now valid per the
  agentskills.io spec (#337).

Autofix reliability

• SAFE autofix corruption fixed: path truncation from backtrackable
  lookarounds, and duplicate name: keys from non-converging
  frontmatter fixes (#325).
• BOM and CRLF line endings preserved byte-exactly across autofix (#332).
• agentskill-name autofix no longer folds inline YAML comments into the
  rename manifest; comments are preserved on the rewritten line (#352).
• content-broken-internal-reference resolves percent-encoded links and
  never emits unparseable destinations (#350).

Rule accuracy

• content-embedded-secrets: entropy gating + placeholder allowlist
  kills documentation-placeholder false positives (password = "changeme", <your-api-key>, ${VAR} templates). Structured
  high-confidence tokens (AKIA…, ghp_…, JWTs, private keys) are always
  reported. Configurable via entropy-threshold and
  additional-placeholders (#354).
• **-glob exclude semantics fixed: **/templates/** now matches a
  top-level templates/ directory, matching gitignore semantics — a
  strict superset of the old behavior, applied to global, per-rule, and
  agentskill-unreferenced-files excludes (#353).
• agentskill-description max length is configurable via
  max_length (#348).
• Four small bug fixes from the #322 sweep: custom-rule abstract-class
  guard, @~/ imports, stacked disable-next-line, docs crash (#351).

Security & performance

• ReDoS protection: config-driven banned-references regexes run under a
  wall-clock timeout (regex-timeout, default 2s) (#330).
• promptfoo file:// references contained to the repo root — traversal,
  absolute paths, and symlink escapes are rejected (#349).
• Fixed O(n²) behavior in markdown_doc._ContentMap.locate (#328).
• Benchmarks vs 0.14.1: synthetic suite at parity; plugin discovery adds
  no measurable startup cost (--version/--help never touch entry
  points).

Docs & DX

• "How to fix" guidance added to every builtin rule page — also surfaced
  through skillsaw explain, feeding agent-driven fix loops (#309).
• SARIF output docs, exit-code reference, configuration reference, clean
  skillsaw add scaffolds, stale pin cleanup (#326).
• CLI internals: __main__.py god module split into a cli/ package
  (#308); block hierarchy extracted to skillsaw.blocks (#307);
  single-source rule registration (#311).

Upgrade notes — what may change on your repo

• agentskill-unreferenced-files may fire new warnings (auto-enabled on
  skill-type repos).
• Expanded marketplace-json-valid and frontmatter-hooks scanning may
  surface new violations; plugins under metadata.pluginRoot are now
  discovered and linted.
• A top-level templates/ directory is now excluded by default; digit-
  leading skill names and placeholder-shaped "secrets" no longer flagged.
• skillsaw plugins now lists installed plugin packages (exit 0) instead
  of linting the plugins/ directory — update CI that relied on the old
  behavior.
• Action users: custom rules require no-custom-rules: false (see above).

Known issues

• The secrets entropy gate trades false positives for some false
  negatives on generic key = "value" assignments (e.g. long digit-only
  values, values containing allowlisted words like token); structured
  token detection is unaffected.