codeflash-internal/tiles/codeflash-internal-rules/rules/optimization-patterns.md

Optimization Patterns

Router Dispatch

• Shared routers in core/shared/ dispatch by the language field on the request schema
• Lazy imports inside the endpoint body to avoid circular dependencies:

  if data.language in ("javascript", "typescript"):
      from core.languages.js_ts.optimizer import optimize_javascript  # noqa: PLC0415
      return await optimize_javascript(request, data)
  

• Default language is Python

Context Extraction

• BaseOptimizerContext in optimizer_context.py handles prompt management and code extraction
• Two context types: SingleOptimizerContext (single-file) and MultiOptimizerContext (multi-file)
• extract_code_and_explanation_from_llm_res() parses LLM markdown response into code blocks
• parse_and_generate_candidate_schema() converts extracted code to OptimizeResponseItemSchema

Model Distribution

• get_model_distribution() in optimizer_config.py splits calls between OpenAI and Anthropic
• Formula: claude_calls = (total - 1) // 2, gpt_calls = total - claude_calls
• MAX_OPTIMIZER_CALLS = 6, MAX_OPTIMIZER_LP_CALLS = 7
• Claude gets fewer calls as it's more expensive

Postprocessing

• postprocess.py handles deduplication and validation of optimization candidates
• Deduplication: normalize with ast.parse() + ast.dump(), skip duplicates with identical AST
• Equality check: compare optimized code to original to skip no-ops
• Uses libcst for all code transformations (preserves formatting)

Prompt Templates

• Prompts stored as .md files alongside their modules
• Rendered with Jinja2 (e.g., build_prompt() in testgen)
• System and user prompts are constructed per-context type