| name | graft-ai-governance-audit |
| description | Repository-specific audit workflow for Graft AI governance. Use when evaluating or changing AI tooling, MCP adoption, repository skills, ai-plan governance, Python helper scripts, environment inventory, or drift between AGENTS.md and AI workflow documents. |
Graft AI Governance Audit
Use this skill to evaluate or maintain Graft AI governance across MCP, skills, ai-plan, Python helpers, and
environment inventory.
Treat root AGENTS.md as the startup and execution-governance source of truth. This skill is a docs/automation audit
workflow; it does not replace repository startup, validation, closeout, or commit rules.
Workflow
- Complete the startup preflight from root
AGENTS.md.
- Classify the task as
docs/automation unless the requested change also modifies server, web, OpenAPI, or shared
runtime contracts.
- Read:
.ai/environment/tools.ai.yaml
ai-plan/design/governance/ai/AI工具与MCP接入治理规范.md
ai-plan/design/governance/ai/CodeGraph-MCP-辅助开发规范.md when CodeGraph is involved
ai-plan/design/governance/frontend/TDesign-MCP-辅助开发规范.md when TDesign or frontend component generation is involved
- Inspect concurrent work before edits:
git status --short
- keep ownership limited to AI governance docs,
.agents/skills/** governance files, and scripts/** audit helpers
- do not stage unrelated
server, web, OpenAPI, dashboard, topic recovery, or generated artifact changes
- Check live local tool state when relevant:
codex mcp list
codex mcp get <name>
codex mcp get context7
codex mcp get github
codex mcp get playwright
codex mcp get headroom when Headroom context compression or MCP tooling is involved
codegraph status only when CodeGraph index health matters
- Run the structure audit:
python3 scripts/validate_ai_governance.py
- If Python helper tests changed, run the focused tests:
python3 -m unittest discover -s scripts -p 'test_*.py'
- Report:
- startup receipt
- MCP adoption status
- skill coverage or drift
- workflow-level drift such as
graft-work-intake inventing rules outside document authority
- Python helper status
- validation commands and results
- any rejected tool and concrete reason
Guardrails
- Do not add MCP servers to
server/go.mod, web/package.json, CI, hooks, or runtime scripts.
- Do not commit personal MCP client configuration, tokens, database DSNs,
.codegraph/**, .ai/venv/**, or browser
artifacts.
- Do not introduce
memory MCP or another hidden recovery store; use ai-plan/public/**.
- Do not use GitHub or database MCP write capabilities unless a repository skill explicitly owns the permission,
workflow, and closeout evidence.
- Keep Context7, GitHub MCP, and Playwright MCP in user-level Codex MCP config only; do not commit client config.
- Use Context7 for current external library documentation, GitHub MCP for read-only PR/Actions context, and Playwright
MCP as browser exploration before
graft-web-browser-agent captures reproducible evidence.
- Keep new skills concise and prefer scripts for deterministic checks.
- If a thin workflow skill such as
graft-work-intake exists, verify that it routes through documented Work Intake
and Work Contract rules instead of creating a second intake authority.
Closeout Evidence
AI governance audit:
- task_class: docs/automation
- owned_scope: ai-plan/design/**, .agents/skills/**, scripts/**
- tools_checked: codex mcp list / validate_ai_governance / unittest
- mcp_changes: adopted | pilot | rejected | none
- validation: <commands and results>
- commit_scope: <confirmed owned paths>