suggest-fixes

Suggest Fixes for SkillGuard Findings

Reads SkillGuard scan findings and generates concrete code fixes. Can either suggest fixes or directly apply them with user approval.

Usage

1. Generate a JSON report:

   npx skillguard <target> --json report.json
   

2. Ask to fix the findings. Read the report and the source files to generate patches.

Fix Strategies by Dimension

Prompt Injection

• Remove injection phrases from prompts
• Use structured prompt templates instead of string concatenation
• Remove zero-width characters with text sanitization

Permission Escalation

• Remove dangerouslyDisableSandbox — run in sandbox
• Replace sudo with user-level operations
• Change chmod 777 to chmod 755 or 644
• Add allowed-tools to SKILL.md frontmatter

Data Exfiltration

• Isolate env var reads from network requests
• Add URL allowlists for outbound HTTP
• Remove direct .env file reading — use injected env vars

Destructive Operations

• Add path allowlist validation before rm -rf
• Replace git push --force with --force-with-lease
• Add confirmation prompts before destructive ops

Supply Chain

• Pin dependency versions: package@1.2.3 instead of package
• Replace curl | bash with download-then-review-then-execute
• Add @sha256: digests to Docker FROM statements

Code Security

• Replace shell=True with shell=False and argument lists
• Replace eval() with JSON.parse() or safe alternatives
• Use parameterized SQL queries instead of f-strings

Credential Leaks

• Move hardcoded secrets to environment variables
• Add .env to .gitignore
• Replace hardcoded API keys with process.env.KEY_NAME

Least Privilege

• Add allowed-tools frontmatter listing only needed tools
• Split skills with shell+network+write into focused sub-skills

License Compliance

• Check dependency license compatibility
• Add license declarations to SKILL.md frontmatter

Resource Abuse

• Add maxIterations or timeout to loops
• Set reasonable retry limits with exponential backoff
• Add recursion depth limits

Workflow

1. Read the JSON report to identify findings
2. Read each affected source file
3. For each finding, propose a minimal, targeted fix
4. Apply fixes with Edit tool (with user approval)
5. Re-run npx skillguard <target> to verify the fix reduced the risk score