Analyze code for LLM workflow patterns using Gemini.

* Calculate adaptive batch size based on file complexity. * Larger files → smaller batches, smaller files → larger batches. * This improves reliabil

* Clear the graph completely (used when cache is cleared before reanalysis)

* Merge labels for edges with the same source→target pair. * E.g., two edges "Event Type → callback" with labels "connected" and "stop" * become one

* Deduplicate locations by (line, type) key * Keeps first occurrence of each unique location

* Detect ALL AI services/APIs in the content (not just the first one) * Returns array of detected service names

* Determine node type from function info * Only 3 types: step, llm, decision * - llm: Function makes LLM API calls * - step: Everything else (decis

* Extract function name from a C/C++ declarator chain. * C function names are nested inside function_declarator, possibly * wrapped in pointer_decla

* Convert function name to a readable label (temporary until LLM provides real labels)

* Get all file paths under a node

* Extract the property/method name from a member-like node. * Different languages use different field names for the property.

* Generate octagon points for a component * Prioritizes width over height (wider than tall)

* Hash content (raw)

* Initialize graph after file picker closes (for cached data)

* Invalidate a single file and its cross-file edges