graft-security-remediation

name	graft-security-remediation
description	Repository-specific workflow for triaging, remediating, validating, and closing Graft GitHub Code Scanning and Dependabot alerts, including branch hygiene, scoped fixes, commit/push/PR flow, and post-push security recheck.

Graft Security Remediation

Use this skill when a task asks to inspect, classify, remediate, validate, or recheck:

• GitHub security/code-scanning
• GitHub security/dependabot
• related GitHub Advanced Security findings surfaced through those alert APIs

Treat root AGENTS.md as the startup, authority, validation, commit, and closeout source of truth. This skill adds a security-specific workflow plus reusable inventory assets; it does not create a second release path.

Workflow

1. Complete the startup preflight from root AGENTS.md and classify the task as server, web, cross-boundary, or docs/automation.

2. Read the required subdomain AGENTS.md before edits, then inspect the current worktree with git status --short.

3. Build one full open-alert inventory before planning or fixing:

   python3 .agents/skills/graft-security-remediation/scripts/inventory_open_alerts.py \
     --repo GeWuYou/Graft \
     --kind all \
     --output - > /tmp/graft-security-alerts.json
   

   Read references/inventory.md when you need the normalized schema, gh examples, or the offline import path.

4. Classify every alert as exactly one of:

   • real-vulnerability
   • false-positive
   • needs-investigation
   • already-remediated-locally
   • blocked

5. Publish a concise remediation plan before edits:

   • in-scope alerts
   • classification evidence
   • intended authority-first repair
   • validation commands
   • branch / commit / recheck path

6. Remediate with the smallest authority-correct fix:

   • verify source -> sink for code scanning alerts
   • verify direct vs indirect usage and patched version reality for dependency alerts
   • do not add suppressions, aliases, or compatibility layers unless root AGENTS.md exception rules are satisfied

7. Run the smallest correct validation for the touched scope:

   • server: focused go test, then go run ./cmd/graft validate backend --stage lint
   • web: bun run check
   • cross-boundary: validate both sides

8. Recheck after local validation. Read references/recheck.md when you need the standard evidence checklist or post-push expectations.

Guardrails

• Do not start with a partial subset of alerts when a full inventory is still missing.
• Do not claim false-positive or already-remediated-locally without code-level evidence from the current branch.
• Do not invent remote write steps as the default path. Push, PR creation, alert dismissal, and any GitHub write action require explicit user request or an owning repository skill that authorizes that write path.
• Keep the owned scope limited to the confirmed remediation slice; do not bundle unrelated worktree changes.

Closeout Evidence

Security remediation evidence:
- inventory_source: gh-api | local-json | mixed
- alerts_reviewed: <count by kind>
- disposition_summary: real / false-positive / stale / blocked
- authority_repair: <canonical layer changed or no-change>
- validation: <commands and results>
- recheck_status: local-only | pushed-awaiting-github | github-confirmed