Review Agentic Workflows

1. Produce a security-first review of workflow changes.
2. Compile workflows with validation and security scanners.
3. Flag suspicious changes that weaken protections.
4. Use run history (logs/audit) when available to find optimization opportunities.

Self-contained setup (do not assume environment is ready)

Step 0) Verify CLI availability

Run from the repository root:

Review workflow

1) Scope the review

Run this scope check in the review step:

If source .md files changed, treat generated .lock.yml drift as part of the review.

2) Compile with validation + security tools

For changed workflows, run strict compilation with validators:

If gh aw extension is unavailable but local binary exists:

Fail review on compilation errors or High/Critical security findings unless explicitly justified.

3) Enforce security best practices

Require and verify:

• least-privilege permissions: (no write-all without explicit rationale)
• pinned third-party actions by full commit SHA
• safe handling of untrusted GitHub event data (no direct template injection into shell)
• explicit safe-outputs limits (max, constrained event/action sets)
• no broadening of network/tool access without justification
• no integrity downgrades (for example lower min-integrity)

4) Detect suspicious weakening changes

Treat these as suspicious until proven safe:

• permission expansion (especially new write scopes or global writes)
• relaxed security controls (strict: false, reduced guardrails, disabled scans)
• larger write blast radius (safe-outputs limits removed or sharply increased)
• reduced provenance controls (unpinning actions, mutable refs)
• wider external access (new unrestricted network domains/ecosystems)
• prompt or script edits that reintroduce command/template injection risk

Use targeted diffs and call out before/after impact.

5) Audit history and optimize (when run data exists)

If workflow run IDs/URLs are available, audit them:

Look for optimization opportunities:

• high token/cost usage
• repeated retries/tool failures
• long-running steps or bottleneck jobs
• unnecessary MCP/tool invocations
• firewall denials causing retries or wasted turns

Recommend minimal, safe optimizations that keep or improve security posture.

Review output contract

Return findings in three sections:

1. Security regressions (must-fix) — high-confidence weakening changes.
2. Validation/scanner results — compile and tool outcomes.
3. Optimization opportunities — optional improvements backed by logs/audit evidence.

Each finding should include severity, file(s), rationale, and a concrete remediation direction.