Add evals for add-dart-lint-validation-rule

[reidbaker-agent]

adding evals for the add-dart-lint-validation-rule skill.

This pr wasa blocked by #189

Evaluation Results

Evaluations were run to ensure that the updated SKILL.md instructions correctly guide agents in creating custom validation rules, specifically focusing on correctly documenting CLI flags in README.md without hallucinating prefixes.

Eval 1: RequireSpecificMetadataRule

• Goal: Agent creates a rule named RequireSpecificMetadataRule that verifies metadata has a required_version field.
• Expected README.md Flag:--[no-]require-specific-metadata
• Result:PASS.
• Notes: The agent successfully created the rule, documented it in RULES.md, added passing tests using an in-memory SkillContext, and perfectly adhered to the naming requirements for the flag in README.md without hallucinating a check- prefix.

Eval 2: FileExistenceRule

• Goal: Agent creates a disabled-by-default rule ensuring an info.txt file exists, using temp directories for testing.
• Expected README.md Flag:--[no-]file-existence
• Result:PASS.
• Notes: The agent successfully created the rule, configured it to be disabled by default, wrote temp directory tests, and exactly matched the file-existence flag pattern in README.md.

Conclusion

The TDD-driven refinements to the SKILL.md file (adding explicit CRITICAL FORMATTING and CRITICAL NAMING guidelines) successfully resolved the previous issues where agents hallucinated CLI flags. The subagents can now consistently author and document new lint rules cleanly.

[gemini-code-assist]

Code Review

This pull request introduces a comprehensive evaluation framework for AI agent skills within the dart_skills_lint package. It adds a new run-evals skill, defines per-skill and cross-skill evaluation rubrics (including a shared code quality rubric), and implements a structural consistency test for all evals.json files. Feedback on the newly added test highlights potential issues with brittle path resolution when running tests from different directories, as well as overly strict schema validation that could make future updates fragile.

[reidbaker]

/gemini-review