agentic-security:security-fix-finding

Skill — fix a specific finding

Activates when the user is pointing at a specific finding from a prior scan and wants to remediate it. The deterministic toolchain (synthesize_fix → verify_fix → apply_fix) is the only correct path — do NOT use Edit directly.

When to fire

• User says "fix this one" / "how do I close X" / "patch the SQL injection at api/users.ts:42".
• User pastes a finding id (struct:src/api.js:42:SQL_Injection) or a stableId (abc12345) and asks for action.
• A prior turn produced a finding and the user said "ok do it".
• Conversation is in the middle of a triage flow.

What to do

1. Confirm the finding still exists. Read .agentic-security/last-scan.json and look up the finding by id / stableId / file+line. If it's gone, tell the user and stop — don't patch a finding that's already been resolved.

2. Decide appropriateness. Read the file around finding.line ± 30 via the Read tool. Is the canonical fix actually right here? If the surrounding code already validates upstream, or there's a custom sanitizer, or the file is a test fixture — STOP and report refused: <reason>. Don't proceed.

3. Route via MCP, not Edit. The deterministic path is:

   MCP synthesize_fix → MCP verify_fix → MCP apply_fix
   

   • synthesize_fix({ finding_id }) returns the stored replacement text. You do NOT modify it. You do NOT retype it.
   • verify_fix({ stable_id, files: {…} }) re-scans the patched file in memory and runs the project linter. Read the structured introduced[] array on failure (template-incomplete vs codebase- prior vs lint-failed — see agents/security-fixer.md for the decision tree).
   • apply_fix({ finding_id, confirm: true }) writes via fix-history.js with HMAC verification + reserved-path refusal + attempt-budget enforcement.

4. Batch mode. If the user wants to fix more than one finding, hand off to the security-fixer subagent with a list (≤ 10 findings per invocation per _CONFINEMENT.md). The subagent writes a PLAN.md to the scratchpad.

5. Run the project tests after apply. If the project has npm test, pytest, cargo test, or similar — invoke it via Bash. Surface pass/fail in the final report.

Don't

• Don't use Edit to apply security patches. The deterministic toolchain is the only path with HMAC + audit + budget + backup.
• Don't paraphrase the synthesized replacement. The bytes go through.
• Don't claim a fix is applied without a passing verify_fix.
• Don't grind through > 10 findings in one session — split into batches, write progress to AGENTS.md.

Canonical commands

• /fix --one <id> — patch a single finding (interactive)
• /fix --all --critical — batch by severity
• /fix --pr — bundle into a PR branch