Forensic-deepdive

AI coding agents regularly fail not because the model is wrong but because the context they receive is wrong — file snippets without import resolution, symbol names without call-chain provenance, endpoints with no trace back to handlers. forensic-deepdive addresses this by parsing a codebase across nine languages, building a persistent embedded graph of its structural and historical relationships, and serving that graph through an MCP server that coding agents can query directly. The core artifacts are nodes for Files, Symbols, Modules, Commits, Authors, Endpoints, and DbTables, connected by edges for DEFINES, MEMBER_OF, IMPORTS, CALLS, EXTENDS, IMPLEMENTS, and TOUCHED_BY_COMMIT — a vocabulary precise enough for agents to answer questions like ‘what calls this endpoint’ or ‘who last changed this module and what else changed with it’.

The differentiating design choice is persistence. Most context-supply approaches reparse on every session; forensic-deepdive writes the graph to `/.deepdive/graph.lbug` and keeps it there. Queries return structured, pre-computed answers rather than ad-hoc file scans, which means agent sessions start with a complete architectural picture rather than rebuilding one from whatever fits in the context window. The five Markdown artifacts — the human-readable projection of the same graph — mean onboarding documentation is a by-product of the analysis, not a separate authoring task.

The tool fits teams running AI coding agents (via MCP) against large polyglot repositories where structural context is the bottleneck. It also fits teams that want durable, auto-generated code documentation without a dedicated docs workflow. Where it breaks: the graph reflects the state of the codebase at the time of the last analysis run, so rapidly changing repos produce stale structural data unless graph refresh is wired into CI. The vendor describes no hosted offering and no managed refresh service — operational discipline for graph freshness is entirely the team’s responsibility.