doc-steward

doc-steward

Keep agent-facing repository documentation predictable: put each fact at the right residency, altitude, and volatility level; evaluate it deterministically where possible; and gate every write.

Run commands from this skill directory. Every scripts/, references/, and agents/ path below is relative to this file.

Route the request

Choose exactly one starting mode. Never cross from a read-only mode into a write mode without explicit user intent.

\* EVALUATE changes no audited document. Pass --history only when the user explicitly wants .doc-steward/history.jsonl updated.

DEFINE

1. Classify the target:

   python3 scripts/checks/tier_assess.py <repo-root> --json

Precedence is --tier > explicit --config > auto-detection. Unknown offline signals round down. An explicit --config may also set exclude_paths to take target-relative subtrees out of audit scope — frozen archives and vendored doc trees, where every finding is unfixable because the material is not editable. Report what was excluded; a narrowed scope that goes unstated reads as a clean audit.

1. Read references/standard-core.md for the three rulers, tier, profile, and

taxonomy. Use references/rule-catalog.md to distinguish spec-required rules from house opinion.

1. Open references/templates.md only when the user requests required-document

checklists or skeletons.

DEFINE is complete when the requested standard, classification, and rule status have been presented with no mutation. Stop unless the user also requested an audit.

EVALUATE

1. Run the deterministic audit for the exact target:

   python3 scripts/checks/doc_lint.py --target <repo-root> --json

Use --tier, --config, --fail-on, or --history only when the request calls for them. Explicit YAML config needs PyYAML from requirements.txt; no-config evaluation remains stdlib-only.

1. Run deep inspectors only when explicitly requested. Dispatch the applicable

read-only checklists in parallel when the harness supports it, otherwise apply the same checklists sequentially:

• agents/inspector-structure.md — residency, structure, duplication, no-ops.
• agents/inspector-taxonomy.md — altitude and cross-tool wiring.
• agents/inspector-staleness.md — volatility and implementation drift.
• agents/inspector-design.md — DESIGN rules when the frontend profile fires.

1. Render the deterministic result as a fenced DOC-STEWARD REPORT. Follow

references/rubric.md for verdicts and finding format. Put deep-inspector findings in a separate, unscored judgment appendix; never alter the deterministic composite with them.

EVALUATE is complete only when:

• the report was freshly generated for the exact canonical target;
• every unavailable or failed checker appears under skipped;
• every deterministic finding includes its catalog severity and remedy;
• every judgment finding cites file:line and passes the quote-gate;
• the final output states target, tier/profile, dimensions, grade, findings,

skipped checks, whether history was enabled, and — when exclude_paths was configured — the excluded subtrees and the fact that they narrow the document corpus only; and

• no audited document changed.

ENFORCE

Open references/apply-workflow.md and follow it completely. The essential sequence is:

1. Save a fresh default EVALUATE report outside the target worktree.
2. Confirm the target is on an existing non-default feature branch with a fully

clean worktree. The script does not create or switch branches.

1. Preview the exact dispositions without --apply. Use --scaffold and

--link-map only for exact user-requested paths or mappings.

1. Show the previewed write set and obtain explicit approval.
2. Repeat the same command with --apply.
3. Inspect the complete diff and run target validation plus git diff --check.

The classifier and LOW-RISK-AUTO allowlist live in references/do-dont-table.md. ENFORCE never blind-deletes or overwrites a present scaffold target. The script never stages, commits, pushes, or opens a pull request; perform those repository actions only under separate user authorization.

ENFORCE is complete only when every finding has a disposition, preflight and verification succeeded, and the exact changed paths and remaining escalations have been reported. If rollback verification fails, stop and report the target for manual inspection.

LEARN

Open references/learning-sink.md and capture exactly one selected finding. Learning is never implied by EVALUATE or ENFORCE. Keep the noop sink unless the user explicitly supplies a trusted, reviewed adapter; capturing a lesson never changes repository documentation.

LEARN is complete when the sink returns a success or safe rejection receipt and no unapproved write-back occurred.

Read on demand

Invariants

• Read-only by default; writes require explicit mode and intent.
• scripts/lib/rules.py is the rule catalog's single source of truth; regenerate

references/rule-catalog.md with scripts/gen_rule_catalog.py.

• Quote every judgment finding at file:line; hypotheses are not findings.
• This package must pass its own deterministic audit, resolve every pointer, and

keep this entrypoint within the dogfood line budget.