[MirzaMerdovic]

Summary

The skill is now named doc-maintenance, and the plugin version is 0.5.0. The plugin name
stays haddock. The change also restates the docstring rule, directs Claude to read
documents that a name search does not find, and extends two eval cases.

Changes

• skills/doc-style/ is now skills/doc-maintenance/, and the frontmatter sets
  name: doc-maintenance. The README, CLAUDE.md, the primer, the plugin and marketplace
  descriptions, and the skill-loaded graders use the new name.
• .claude-plugin/plugin.json sets version 0.5.0. marketplace.json has no version
  field, so the manifest is the only source of the version.
• The docstring rule: a docstring documents the non-obvious caller-visible contract,
  including constraints, errors, and side effects. It does not repeat what the signature,
  the types, or the surrounding code show.
• Workflow step 2: a search for the changed names does not find a document that describes
  the behavior without naming an identifier. Claude reads the README section and the
  documentation page of the changed component even when the search finds no match.
• evals/historical-adr checks that Claude creates docs/adr/0005-*.md and that its
  Decision section names Redis.
• evals/restating-comment checks the retry behavior with an LLM grader that judges the
  behavior, not the code structure.
• The eval command in the README and CLAUDE.md adds --judge-model sonnet. The default
  judge, Haiku, answers without extended thinking. It failed correct retry code in every
  run.

Compatibility

• /haddock:doc-style and skills/doc-style/ no longer exist. A link to
  skills/doc-style from an install by hand breaks.
• An installed plugin gets 0.5.0 after claude plugin marketplace update throw-if-null
  and claude plugin update haddock@throw-if-null.

Validation

• claude plugin validate . passes with Claude Code 2.1.287 and 2.1.292, the version
  that CI pins. 2.1.292 gives one warning: CLAUDE.md at the plugin root is not loaded as
  project context. That file holds the repository instructions.

• In an isolated configuration directory, an install from the marketplace gives 0.5.0
  with skills/doc-maintenance/SKILL.md. The two update commands above change an
  installed 0.4.0 to 0.5.0.

• No reference to doc-style or 0.4.0 remains.

• Eval results:

  Model	Judge	Runs per case	Result
  Sonnet	Sonnet	1	Every grader passed except readme-consolidated. The skill loaded in 9 of 9 cases.
  Opus	Haiku	1	Every grader passed except retry-implemented (the Haiku judge defect) and comment-judgment in obsolete-comment.

• With the Sonnet judge, retry-implemented passed two correct implementations and
  failed two incorrect implementations, 3 of 3 votes each.

• The new ADR graders passed in 4 of 4 runs.

• After the step 2 change, Sonnet updated the README table in unrelated-cleanup in 5 of
  7 runs. Before the change, it did so in 3 of 9 runs.

• No shell script changed.

Known limitations

• Sonnet does not consolidate duplicated documents. It failed readme-consolidated in
  every run. Opus passed it in 7 of 7 runs. The grader stays strict, because two copies of
  a fact can diverge.
• With the Haiku judge, comment-judgment in obsolete-comment failed some Opus outputs
  that keep a valid constraint. Opus with the Sonnet judge is not tested.
• The duplicate-docs workspace is close to example 3 in examples/maintenance.md. The
  case does not show whether Claude follows the rule or copies the example.
• Each case ran 1 to 5 times per model.
• The evals ran from a copy of the working tree, not from the checkout.