LLM Integration Code Review Patterns

Stack-specific rules loaded by dh:code-reviewer when prompt files, model selection logic, or evaluation harness code are detected.

Read Review principles before applying these checks; it defines authority, applicability, evidence, and blocking criteria. Verify model identifiers, request parameters, and error semantics against the provider and SDK version the target actually uses.

Prompt Hygiene

• Keep instruction authority distinguishable from user/retrieved content. Trace where untrusted text can influence instructions, tool arguments, or consequential actions.
• Check that user-controlled content is treated as data at its intended boundary. String interpolation, escaping, role separation, or templating alone does not establish resistance to prompt injection.
• Review prompt composition for accidental instruction mixing, missing delimiters, or lost context using concrete inputs; do not report every f-string as an injection defect.
• Recommend dedicated prompt resources when versioning, reuse, or review clarity benefits; do not impose a line-count rule on inline strings.

Model Selection