---
name: agent-docs-verification
description: Use when AGENTS.md or agent-docs have been changed and you need post-change validation evidence such as diff cleanliness, routing checks, cycle checks, or summary of what was verified.
---

# Agent Docs Verification

## Overview

Use this skill after changing AGENTS or `agent-docs` to gather verification evidence and report exactly what was checked. It is the completion gate for documentation-tree changes.

## When to Use

- A doc refactor is complete and needs verification
- The user asks whether the change is really done
- You need to summarize what was validated versus what was not

## Workflow

1. Read `../agent-docs/references/verification-checklist.md`.
2. If the change edited this skill suite itself, also read `../agent-docs/references/skill-maintenance-checklist.md`.
3. Read `../agent-docs/references/severity-model.md` when the verification report includes findings or residual warnings.
4. Read `../agent-docs/references/report-format.md`.
5. Read `../agent-docs/references/behavior-check-template.md` when behavior checks are required.
6. Run the strongest repository validation that actually applies.
7. For repository AGENTS or `agent-docs` structure changes, run the validator at `../agent-docs/scripts/validate_agent_docs.py --repo-root <repo-root>`.
8. If the change edited this skill suite itself, also run `../agent-docs/scripts/validate_agent_docs.py --skill-suite-root <skills-root>`.
9. At minimum for docs-only work:
   - run `git diff --check`
10. When the change touched routing or structure, also re-check:
   - root reachability
   - stale references
   - direct cycles if relevant
   - duplicate authority or document-shape regressions when relevant
11. When the change touched repository-dependent claims, confirm those claims against current files, config, scripts, or command output rather than memory.
12. Treat the following as behavior-check-required changes:
   - root routing changes
   - child-doc reference changes
   - conflict-resolution changes
   - workflow changes that affect required follow-up skills such as `refactor -> consistency -> repo-alignment -> verification`
13. For behavior-check-required changes, run a small behavior check with 1 to 3 representative prompts and inspect whether the agent:
   - reads the root entrypoint before choosing a specialized workflow
   - follows the intended route and required follow-up checks
   - escalates duplicate-authority or routing ambiguity instead of guessing
   - states any unverified remainder explicitly
14. Record each representative prompt using `behavior-check-template.md`.
15. Treat the behavior check as failed if any representative prompt skips the root entrypoint, misses a required follow-up check, or resolves ambiguity by guessing.
16. When the change edited the `agent-docs` skill suite itself, also confirm:
   - each `agents/openai.yaml` still matches its `SKILL.md`
   - every shared reference still has a direct reference path from the relevant skill
   - no stale filenames, old skill names, or outdated terms remain after renames or splits
   - one concept still uses one term across `SKILL.md`, `agents/openai.yaml`, and `references/`
17. Report command evidence and exact scope using the verification shape from `report-format.md`:
   - what was verified
   - what was not verified
   - whether the claim is limited to docs structure only
   - whether representative behavior was checked or only static structure was reviewed
   - whether any remaining findings are `blocking`, `major`, `minor`, or `observation`

## Result Levels

- `fully verified`: required static checks passed and any required behavior check passed.
- `structurally verified only`: static checks passed and no behavior check was required for the change.
- `behavior check failed`: a behavior-check-required change was tested and at least one representative prompt failed the pass criteria.
- `verification incomplete`: required checks were skipped or the available evidence is not enough to classify the change as fully or structurally verified.

## Reporting Rules

- Do not say a change is complete without fresh evidence from the current turn.
- Name the commands that were run.
- Do not describe a change as fully verified if a required behavior check was skipped or failed.
- Name the applicable result level in the report.
- If only a minimum docs-only gate was available, say so explicitly.