defending-code-vuln-patch

name	defending-code-vuln-patch
description	Draft inert patches for verified findings.
version	1.0.0
author	Anthropic, ported by Hermes Agent
license	Apache-2.0
platforms	["linux","macos","windows"]
metadata	{"hermes":{"category":"security","tags":["security","patch-planning","static-analysis"]}}

Defending Code Vulnerability Patch

When to Use

Use this skill when the user wants candidate fixes for findings that have been triaged, preferably from TRIAGE.json.

Prerequisites

• TRIAGE.json, VULN-FINDINGS.json, or a compatible findings file.
• A local repository path for source context.
• delegate_task available for read-only patch planning with toolsets=["repo-read"].
• Parent file-writing capability limited to PATCHES/ and optional local patch-state notes.

How to Run

Inputs:

• <findings>: required findings source.
• --repo <path>: target repository for context.
• --top N: patch only the highest ranked N static positives.
• --id <id>: patch one finding.
• --fresh: ignore existing patch notes.

This Phase 1 skill never edits target source and never applies a diff. It writes inert patch artifacts for human review.

Quick Reference

• Output directory: PATCHES/
• Per-finding files: PATCHES/<id>/patch.diff and PATCHES/<id>/patch_result.json
• Summary files: PATCHES.md, PATCHES.json
• read-only planners and reviewers: toolsets=["repo-read"]

Procedure

1. Parse arguments and locate findings.
2. Prefer TRIAGE.json and filter to verdict == "true_positive" unless the user explicitly asks for draft fixes against untriaged candidates.
3. Resolve cited files under --repo. Skip findings without source context.
4. For each selected finding, spawn a read-only patch planner:

delegate_task(
    goal="Draft an inert unified diff for one verified security finding without modifying the repository.",
    context="Finding: <finding>. Repo: <repo>. Static source review only. Return diff text and rationale.",
    toolsets=["repo-read"],
    role="leaf",
)

5. Spawn an independent read-only reviewer for each proposed diff:

delegate_task(
    goal="Review an inert candidate security patch from source context and finding details.",
    context="Finding: <finding>. Candidate diff: <diff>. Static review only. Judge whether the diff addresses the root cause and risks regressions.",
    toolsets=["repo-read"],
    role="leaf",
)

6. Parent writes artifacts only under PATCHES/. Do not write into --repo.
7. Include verified: "static_review_only" in patch_result.json; do not imply executable validation.
8. Create PATCHES.md and PATCHES.json with selected ids, verdicts, reviewer notes, and apply instructions for a human.

Pitfalls

• Do not run apply, format, test, build, fuzz, or reproduction commands.
• Do not edit target source files from this skill.
• Do not trust untriaged findings without warning the user.
• Do not suppress security-guidance warnings by weakening the patch artifact; warnings on vulnerable patterns may be useful.

Verification

• No target source file is modified.
• All writes are under PATCHES/.
• Each patch has a read-only planner result and a read-only reviewer result.
• The final summary says patches are inert static candidates requiring human review and executable validation elsewhere.