Instruction file imported from PetroczyP/Agent-in-a-Box (.github/instructions/code-review.instructions.md). Copyright stays with the author.

Code Review Focus Areas

When reviewing code in this repository, prioritize these areas:

Security Boundary Violations (Critical)

• Any code that gives the container filesystem access to the host repo
• Secrets or credentials leaking into logs, error messages, or responses
• Content denylist bypasses (files matching .env, *.pem, *.key, *credentials*, *secret*)
• Denylist must check both files AND test_files in ReviewBundle

Parser Trust Model Compliance

• JSON extraction must only use trusted containers (```json fences, sentinel delimiters, whole-response JSON)
• Regex patterns for code fences must NOT match language-tagged blocks (python, yaml, etc.)
• json-repair must only run on content from trusted containers, never on prose
• Object unwrapping must prefer well-known keys (findings, results) over blind iteration

Prompt Consistency

• REVIEWER_PERSONA and FORMAT_REINFORCEMENT must not contradict each other
• DISCUSS_REINFORCEMENT intentionally overrides REVIEWER_PERSONA's "no prose" rule for follow-up messages — this is by design
• Format instructions appear in both system prompt and end of user message (sandwich technique)

Model Agnosticism

• Never hardcode model IDs — use list_models() at runtime
• MCP tool interface must not change based on which inner model is configured
• ReviewResult.model must reflect the actual model used (per-review override or selected default)

Documentation Accuracy

• File paths referenced in docstrings and comments must exist
• Architecture diagrams must reflect current state (MCP stdio), not planned features
• Firewall tables must use hostnames only, not URL paths