AGENTS.md git:20260908.0c1891fA

Immutable. This exact content is served forever at /api/v1/blob/8cbfaa5094732ea3.

# AGENTS.md — accessibility-skills

| Agent | Type | Command | Companion |
|-------|------|---------|-----------|
| a11y-planner | planner | `/a11y-planner` | a11y-critic |
| a11y-critic | critic | `/a11y-critic` | a11y-planner, a11y-test |
| a11y-test | tester | `/a11y-test` | a11y-critic |
| perspective-audit | auditor | `/perspective-audit` | a11y-planner, a11y-critic |
| maintain-accessibility-skills | maintainer | `/maintain-accessibility-skills` | — |

## Lifecycle

The critic serves at **two checkpoints**, with optional perspective-audit escalation at each:

```
plan → critique plan → [perspective audit] → revise → implement → test → critique implementation → [perspective audit] → fix → re-test
```

- `a11y-planner` is the pre-implementation design surface (step 1).
- `a11y-critic` reviews the plan for gaps (step 2) and the implementation after testing (step 6).
- `perspective-audit` provides deep review from 7 access perspectives when escalated by the planner or critic (steps 2b and 6b).
- `a11y-test` runs automated Playwright keyboard tests and axe-core scans (step 5).
- All four lifecycle skills are shipped from this repository.

## Benchmark Team Agents

Agents for running, scoring, documenting, and maintaining the eval suite.

| Agent | Type | Model | Tools | Job |
|-------|------|-------|-------|-----|
| bench-runner | executor | sonnet | Bash, Read | Execute benchmarks (Claude API, Codex/OpenAI, Ollama, Gemini/other hosted adapters when present), run scoring scripts |
| bench-reporter | writer | sonnet | all | Update BENCHMARK.md, README.md, EVAL-GAPS-PLAN.md with results |
| bench-reviewer | reviewer | sonnet | read-only | Audit fixture/rubric quality, verify scoring accuracy |
| fixture-builder | builder | sonnet | all | Create/enrich fixture triplets (.md, .metadata.yaml, .rubric.yaml) |

See `.claude/teams/README.md` for team workflows.

## Commits and pushes

- Authorized implementation work includes making focused commits and pushing the task's feature branch to its configured remote after relevant checks pass. Do not ask for separate confirmation solely because multiple files changed. Honor explicit hold, no-commit, or no-push instructions.
- Stage only reviewed files for the current task and preserve unrelated changes. Ask only when the destination is ambiguous, the operation would rewrite or delete shared history, or the action would expand the approved scope.