llm-workflows.md

LLM workflow guide

Use CodeBadger in this order so requests stay bounded and results remain actionable:

1. Call get_backend_status before submitting builds. Respect recommended_max_concurrent_builds and page through CPG summaries only when needed.
2. Call generate_cpg once for the repository, or select an existing codebase_hash. Poll get_cpg_status until status is ready; stop when it is failed and use its error_code/error to recover.
3. Start with narrow browsing: list_methods, list_calls, get_type_definition, or get_cfg using a filename/name filter and a small limit. Check truncated and pagination fields before broadening.
4. Expand only after locating a relevant function or line. Use get_program_slice/get_variable_flow for local context, then taint or detector tools for a focused hypothesis.
5. For taint analysis, use find_taint_sources and find_taint_sinks first. find_taint_sinks is focused by default; use broad=true or explicit patterns when an exhaustive audit is intended.
6. Treat every result as bounded. Follow total_pages, available, returned, has_more, or truncated instead of assuming the first page is complete.
7. Rendered analysis tools default to compact output. Request detail="full" only when the bounded summary is insufficient; otherwise keep detail="compact" to reduce context usage. Native tool failures expose stable error-code prefixes such as VALIDATION_ERROR and QUERY_ERROR.

remove_cpg(delete_files=false) only releases the Joern worker and preserves the CPG. delete_files=true removes the CPG, copied source, and catalog record; use it only when permanent deletion is intended.