Maintaining agent docs

Maintaining Agent Docs

Keep repository guidance accurate without turning provider entrypoints into competing rulebooks. Treat AGENTS.md as the shared operating contract, provider files as selected adapters, README.md as human orientation, and llms.txt as an optional docs-site index.

Procedure

1. Scan the repository before proposing content. Find root and nested instruction files, manifests, lockfiles, build definitions, CI configuration, and durable docs. Record dirty target files and preserve nearest-scope rules.

2. Extract commands and constraints from repository evidence. Cite each claimed command or convention by file and line. Mark an unknown as TODO: verify in the proposal; never invent an executable command.

3. Select only the provider profiles the user or repository requires. Load provider-shims.md for current prerequisites, behavior, and tests. Keep provider facts out of shared policy.

4. Propose changes inside uniquely named managed regions:

   <!-- BEGIN maintaining-agent-docs:root -->
   ...generated content...
   <!-- END maintaining-agent-docs:root -->
   

   Never nest regions, reuse an identifier in one file, or rewrite text outside a region. For unmanaged files, propose an insertion point instead of a full replacement.

5. Run the package validator in read-only mode:

   python skills/maintaining-agent-docs/scripts/validate.py --check --strict .
   # Equivalent framework entrypoint:
   polymind validate-agent-docs --check --strict .
   

   Add --diff to preview duplicate-block removals. Use --check-remote only after network access is selected and approved.

6. Show a unified diff and rollback method for every proposed mutation. Get explicit approval for the exact file list and diff.

7. Back up or otherwise make existing targets recoverable, apply only the approved changes, rerun the validator, and report evidence.

Validation contract

The validator is non-mutating. It checks real YAML frontmatter, marker ordering, nesting and uniqueness, normalized policy duplication with a configurable threshold, local-link containment, optional remote links, redacted secret findings with explicit allowlists, provider import rules, nested scopes, and basic instruction completeness.

• --check performs the default read-only check.
• --diff shows advisory diffs without writing.
• --strict turns warnings into a failing exit code.
• --secret-allowlist FILE accepts SHA-256 digests of known fixture values, one digest per line. Never place raw secrets in an allowlist.

Hard guardrails

• Preserve human-authored content and nested instruction scopes.
• Keep shared policy in one canonical file; adapters may add runtime-only facts.
• Reject local links that are absolute, traverse above the repository, or resolve outside it through a symlink.
• Redact possible secrets in every diagnostic.
• Generate llms.txt only for a published documentation site.
• Separate checking from writing. Never interpret validator success as approval.

Load detailed references only when their file type is in scope:

• agents-md.md for canonical and nested instructions.
• provider-shims.md for selected provider adapters.
• readme.md for human-facing documentation.
• llms-txt.md for an approved docs-site index.