Repository Guidelines

Project Structure & Inputs

Rust workspace: engine/ (analysis core), cli/ (safe-mcp-scan), server/ (safe-mcp-analyzer), schemas/ (JSON Schema helpers), safe-mcp/ (vendored corpus; keep gitignored), techniques/ (active specs), techniques_backup/ (deprioritized specs).
Technique metadata: canonical table safe-mcp/README.md, prioritized list safe-mcp/techniques/prioritized-techniques.md, mitigations under safe-mcp/mitigations/, per-technique guidance safe-mcp/techniques/<ID>/README.md. Prompts must include the per-technique README excerpt; evidence must cite which artifacts were read.

Toolchain: stable Rust (cargo fmt, cargo clippy --all-targets --all-features, cargo test --all).
CLI: cargo run -p cli -- --provider openai --model-name gpt-4o-mini T1001 --repo <path> --schema $(pwd)/schemas/technique.schema.json --json [--llm-review]. LLM review reuses the configured OpenAI model/key and is non-fatal on failure.
MCP server: cargo run -p server --bin safe-mcp-analyzer. Config (YAML/JSON) sets provider keys, allowlists, retries/timeouts, and path filters (include_*, exclude_*, max_file_bytes, docs allowed by default; 0 disables size checks).
Batch scans: ./run_scans.sh runs all specs under techniques/, continuing on failures and writing scan_outputs/*.json.

Track work in your chosen tracker; keep sequencing/dependencies noted. After every code change, record design decisions, prompts/model settings, filters used, and test commands. Mark status promptly.
Add new issues to your tracker; note ordering/dependencies when relevant.

rustfmt/clippy before commits; prefer explicit error types. IDs: techniques SAFE-T#### (schema enforces T#### inside specs), mitigations M-##, schema files kebab-case under schemas/.
Prompts carry file path/extension/lines plus README excerpt; findings must include evidence (path, line range, snippet) and state artifacts consulted. Avoid inventing findings; keep model temperature at 0.

Add unit tests beside modules (loader/validator/cache, README/prioritized/mitigation indices, chunking, prompts, adapters, aggregation/status, config). Use fixtures under tests/fixtures/; stub model adapters for deterministic outputs. Failures must be clear and user-facing.

DCO sign (git commit -s) with scoped subjects (e.g., engine: add retry adapter). PRs list commands run and note schema/config changes. Never commit safe-mcp/ mutations.

No secrets in git; read provider keys from env/config. Enforce provider allowlist/allow_remote_providers before making remote calls. Respect path filters when sending code to models; keep logs free of secrets and large payloads.