• Security Vulnerabilities — SQL injection, hardcoded secrets, authentication bypasses, XSS/CSRF
• Correctness — Does the code match the stated purpose? Logical errors?
• Code Quality — Bugs, error handling, race conditions, edge cases
• Static Analysis — Reviews code changes directly from git diff

Review method: Static code review (reads diff output, does NOT execute code or run tests)

Rule: Verifier must cite specific files, line numbers, and code snippets in verdict.

jobs:
  verify:
    runs-on: ubuntu-latest
    env:
      PROOF_AGENT_PROVIDER_BASE_URL: https://api.anthropic.com
      PROOF_AGENT_PROVIDER_TYPE: anthropic
      PROOF_AGENT_PROVIDER_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
      PROOF_AGENT_MODEL: claude-sonnet-4-20250514
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: AndreaGriffiths11/proof-agent@main
        with:
          github-token: ${{ secrets.GITHUB_TOKEN }}

jobs:
  verify:
    runs-on: ubuntu-latest
    env:
      PROOF_AGENT_PROVIDER_TYPE: azure
      PROOF_AGENT_PROVIDER_BASE_URL: https://mycompany.openai.azure.com
      PROOF_AGENT_PROVIDER_API_KEY: ${{ secrets.AZURE_OPENAI_KEY }}
      PROOF_AGENT_MODEL: gpt-4-turbo
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: AndreaGriffiths11/proof-agent@main
        with:
          github-token: ${{ secrets.GITHUB_TOKEN }}

jobs:
  verify:
    runs-on: ubuntu-latest
    env:
      PROOF_AGENT_PROVIDER_BASE_URL: http://localhost:11434/v1
      PROOF_AGENT_MODEL: qwen2.5-coder:3b  # Good balance of size/performance
      # Also works: qwen2.5:0.5b, gemma2:2b, deepseek-coder:6.7b
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: AndreaGriffiths11/proof-agent@main
        with:
          github-token: ${{ secrets.GITHUB_TOKEN }}

Supported providers: Anthropic, Azure OpenAI, OpenAI-compatible endpoints (OpenAI, Ollama, vLLM, etc.).

Cost optimization: Use a different model for the verifier step with PROOF_AGENT_VERIFIER_MODEL.

⚠️ Prompt Optimization: Proof Agent works with smaller models when prompts are concise and task-focused. Minimum 1B parameters for basic detection, 2B+ parameters recommended for reliable production use. Very complex instructions may reduce effectiveness regardless of model size.

- uses: AndreaGriffiths11/proof-agent@main
  with:
    # GitHub token (use built-in GITHUB_TOKEN)
    github-token: ${{ secrets.GITHUB_TOKEN }}
    
    # Optional token used for Copilot SDK requests
    # Falls back to github-token when omitted
    copilot-token: ${{ secrets.COPILOT_TOKEN }}
    
    # Git ref to compare against
    base-ref: origin/main
    
    # Force verification even when thresholds are not met
    force: false
    
    # Block PR merge if FAIL
    block-on-fail: true
    
    # Post verdict as PR comment
    post-comment: true
    
    # Comment format: collapse (default), summary, or full
    comment-mode: collapse
    
    # Max comment length in characters
    max-comment-length: 2000

Comment modes:

• collapse — First paragraph visible, rest in expandable section
• summary — Verdict + key findings only
• full — Everything visible (truncates at max-comment-length)

Customize thresholds and patterns:

thresholds:
  min_files_changed: 3
  always_verify:
    - "**/*auth*"
    - "**/*secret*"
    - "**/*permission*"
    - "**/Dockerfile"
    - "**/*.env*"
  never_verify:
    - "**/.gitignore"

retry:
  max_attempts: 3
  escalate_on_max: true

Auto-verify when:

• ≥3 files changed
• ANY file matches: *auth*, *secret*, *permission*, Dockerfile, *.env*
• You explicitly force verification with force: true or bash scripts/verify.sh <base-ref> --force

Skip for:

• Formatting-only changes
• .gitignore changes

When thresholds are not met, the action exits successfully and reports SKIP in its action output/comment.

1. Worker agent generates src/auth.py, tests/test_auth.py, updates requirements.txt
2. Proof Agent detects: 3+ files changed + *auth* pattern → triggers verification
3. Verifier agent spawns, receives:
   • Original request
   • Files changed (list)
   • Approach taken (git diff output)
4. Verifier reviews the code changes:
   • Scans src/auth.py for hardcoded secrets, SQL injection, auth bypasses
   • Checks tests/test_auth.py for edge case coverage
   • Reviews requirements.txt for suspicious dependencies
5. Verifier finds:
   • Hardcoded API key in src/auth.py:42 (API_KEY = "sk-1234...")
   • SQL query uses string interpolation (injection risk)
   • Missing input validation on username parameter
6. Verdict: FAIL — Security issues (hardcoded secret + SQL injection + missing validation)
7. Proof Agent posts comment, blocks merge
8. Developer fixes issues → pushes new commit → re-verifies → PASS

Check workflow permissions:

permissions:
  contents: read
  pull-requests: write
  copilot-requests: write  # ← Required for Copilot SDK requests

Verify Copilot requests are enabled:

• The organization must allow Copilot CLI usage from GitHub Actions.
• If your repository or organization requires a dedicated token, pass copilot-token.

Check workflow permissions:

permissions:
  pull-requests: write  # ← Required for posting comments
  contents: read
  copilot-requests: write

For private repos, use secrets.GITHUB_TOKEN (already has correct permissions).

Proof Agent skips if <3 files changed AND no sensitive files detected.

To force verification:

• Change 3+ files, OR
• Touch a sensitive file: *auth*, *secret*, Dockerfile, *.env*

Single-agent limitations:

• Same model that made the mistake will rationalize it
• Confirmation bias in self-review
• No incentive to find flaws

Adversarial separation:

• Verifier has no stake in worker's success
• Forced to provide evidence
• Different prompts catch different issues

Real-world analogy:

• Code review (separate developer)
• Security audit (external team)
• Peer review (different researcher)