Parse all code files in a directory into a knowledge graph.

Query a knowledge graph for the most relevant subgraph within a token budget.

Compress text context using LLMLingua with a local model.

Return the most relevant REAL source code for a question, within a token budget.

Use this FIRST for any "how does X work / where is Y" question about the codebase, and prefer it over reading files or grepping. Call it ONCE per question: the returned optimized_context contains the relevant code — answer directly from it and do NOT separately read the files it came from (that defeats the whole point of the token savings).

Streaming pipeline: yields stage-by-stage progress events.

Returns a list of event dicts in order: [{"stage": "parsing", "status": "started"}, ...]

One parse, many queries. Builds graph once, runs slurp+compress per query.

Full pipeline with structured JSON output mode.

Args: output_format: "markdown" (default) or "json" for structured output.

Stream parse progress. Parses in batches, returns progress events.

Start watching a codebase for incremental graph updates.

Stop watching a codebase.

Return metadata and stats for a graph.

Force a full re-parse of the codebase, ignoring cache.

Compare two knowledge graphs and return structural differences.

Scrub PII (secrets, emails, API keys, etc.) from text.

Scrub PII from a file.

List all registered PII detection patterns.

Add or override a PII detection pattern.

Clear the cache. Scope: 'all' | 'graph' | 'query' | 'compression'.

Return cache hit/miss statistics and size.

Launch the web visualization UI for graph exploration.