Defensively review code or structure artifacts by reading architecture and data flow before reading full source. Use structural signals to identify vulnerabilities, logic flaws, parser/scanner risks, CI/CD risks, and LLM/agent tool-boundary risks.
When to Use
- security review, vulnerability review, SAST triage, AST, structure map, call graph, Source/Sink, taint flow
- Vulnerability Review, Security Review, Structure Map, AST Report, Dependencies, Complexity
- Reviewing owned or explicitly authorized code, design docs, scan results, or generated structure maps
- Hardening parsers, scanners, CI/CD tools, file walkers, and agent/tool-call boundaries against malformed or adversarial input
Safety Scope
- Keep the work defensive: review, risk explanation, safe verification ideas, and minimal fixes.
- Do not provide unauthorized testing, intrusion, persistence, evasion, credential theft, weaponized PoC, or destructive external steps.
- If exploitability is uncertain, place the item in Hypotheses rather than Findings.
- If code changes are requested, keep them minimal and verify with existing tests or a focused local check.
Inputs to Prefer
Use provided structure artifacts first, then inspect the smallest code/config range needed.
| Input | Examples |
|---|---|
| Structure | AST summaries, class/function lists, Mermaid graphs, JSON reports |
| Data flow | Source, Propagator, Sanitizer, Sink, taint flow |
| Call graph | caller -> callee edges, boundary-crossing calls, fallback paths |
| Complexity | large functions, branch concentration, exception density |
| Scope/state | variable scope, global state, shared mutable state, auth boundaries |
| Static results | secret scan, dependency scan, lint/type/SAST warnings |
| Context | README, design notes, CI workflow, deployment config |
Structure Map Build
If no equivalent artifact exists, before entering findings review, create a minimal read-only structure map. The procedure (7 steps) and the minimum map contract (entry_points / files / symbols / imports / call_edges / complexity / sources / sinks / sanitizers / scan_limits) are references/structure-map.md. If it cannot be generated, record the blocker and limits in Structure Map Summary.
Review Flow
- Gate the target: assume review is allowed for current workspace/user-provided artifacts; ask only if ownership or authorization is unclear.
- Classify the target: web app, CLI, library, scanner, CI/CD tool, infrastructure definition, data processor, extension, or agent tool.
- Read an existing structure map, or build the minimal map from the contract above before full source reading.
- Model trust boundaries: who controls which input, where it flows, what authority the Sink has, and what data is sensitive.
- Trace Source -> Propagator -> Sanitizer -> Sink. Look for missing validation, wrong order, branch gaps, implicit deserialization, and type-conversion surprises.
- Review call graph paths that bypass authn/authz, validation, initialization, feature flags, or normal handlers.
- Review complexity and logic risks: giant functions, broad try/catch, deep nesting, fallback behavior, shared state, TOCTOU, and cache races.
- For parsers/scanners/file walkers, prioritize crash, unbounded recursion, symlink loops, huge/deep ASTs, encoding edge cases, dependency resolution failures, and detection bypass.
- For CI/CD, review workflow inputs, artifact/cache trust, token permissions, PR boundaries, release gates, and secret exposure.
- For LLM/agent code, review prompt injection paths, external document trust, tool-call authorization, path safety, command policy, and log redaction.
Evidence Rules
- Findings require structural signal, reachability, impact, and a defensive verification or fix direction.
- Confidence is High, Medium, or Low. Avoid certainty language unless verified.
- Keep unsupported ideas in Hypotheses with the missing evidence and next check.
- Reference local files with precise paths and line links when possible. Do not paste full source files.
Output Format
Return in the order Findings -> Structure Map Summary -> Hypotheses -> Code to Inspect -> Recommended Fix Plan -> Verification Summary. All table schemas and the Recommended Fix Plan body are references/output-format.md. If there are zero Findings, state that explicitly at the beginning.
Thanks to / References
Inspired by these structure-first review resources. Treat them as references, not required dependencies.