Forensic Deepdive

forensic-deepdive

forensic-deepdive gives an AI coding agent forensic understanding of an unfamiliar codebase — a persistent knowledge graph, an MCP server, and five durable markdown artifacts, so your agent stops re-discovering the same architecture from scratch every session.

uv tool install forensic-deepdive && forensic extract /path/to/repo

Why this is different: every edge and every emitted claim carries a confidence tag — EXTRACTED (deterministic fact) / INFERRED (a heuristic resolved cleanly) / AMBIGUOUS (multiple candidates, shown, not guessed) — so your agent knows what to trust before it acts on it. Nothing else in this category tags confidence at all.

Why not GitNexus / CodeGraphContext / DeepWiki / Sourcegraph

GitNexus is the runaway leader in this category — but its PolyForm Noncommercial license locks every commercial user out. forensic-deepdive is the Apache-2.0 alternative: same persistent-graph-plus-MCP shape, plus honest confidence tags, git archaeology, persistent agent memory, and the 5 markdown artifacts as a fallback for any agent that doesn't speak MCP — with a license your employer's legal team will actually sign off on.

What it produces

forensic-deepdive analyzes any codebase (9 languages, polyglot) and produces:

1. A persistent embedded graph at <repo>/.deepdive/graph.lbug — File, Symbol, Module, Commit, Author, Endpoint, and DbTable nodes plus DEFINES, MEMBER_OF, IMPORTS, CALLS, EXTENDS, IMPLEMENTS, TOUCHED_BY_COMMIT, AUTHORED_BY, CO_CHANGES_WITH, and the cross-boundary HANDLES / CALLS_ENDPOINT / ROUTES_TO / INJECTS / PERSISTS_TO edges. Every edge carries a confidence tag (EXTRACTED / INFERRED / AMBIGUOUS) — no hidden heuristics. The single Endpoint join node unifies five cross-boundary protocols (HTTP, MCP tools, registry-dispatch, gRPC, messaging/AMQP), so a frontend call resolves to its backend handler across the stack as one ROUTES_TO edge.
2. An MCP server (forensic serve) exposing 9 composite tools (impact, context, archaeology, flow, query, record_insight, recall_insights, visualize, trace) consumable by Claude Code, Cursor, Codex, Continue, Cline, Windsurf — and any other MCP-aware agent.
3. Five durable markdown artifacts under <repo>/docs/codebase/, regenerated from the graph on every extract:
   • MAP.md — what's where, ranked by centrality.
   • HOTPATHS.md — the dependency hot spots, with a per-row confidence-mix column so you see exactly how cleanly each symbol resolves.
   • ARCHAEOLOGY.md — why the code looks the way it does (git history, top authors with %, bus factor, co-change clusters, defect proximity).
   • MENTAL_MODEL.md — the doc the original author would write to onboard a new hire.
   • AGENT_BRIEF.md — ≤5 KB of assertive Never/Always rules with per-rule confidence tags. Drop-in CLAUDE.md for any project.
4. Ten shims into the target repo — 4 editor rule files (CLAUDE.md, AGENTS.md, .cursor/rules/codebase.mdc, .continue/rules/codebase.md), 5 single-intent Claude skills under .claude/skills/codebase-{exploring,debugging,impact-analysis,refactoring,onboarding}/, and a .claude-plugin/plugin.json manifest. All write-if-absent — hand-edited files are never overwritten.

Extract also regenerates ARCHITECTURE.md — a system-level Mermaid view of the cross-boundary graph (ROUTES_TO / INJECTS / PERSISTS_TO, confidence-styled), a separate human-validation surface (not one of the five contract artifacts, exactly like forensic visualize and serve --ui). Regenerate it on its own with forensic diagram --repo <repo>. Use it to sanity-check the graph — a wrong edge there is a wrong edge everywhere.

Add --emit-vault to also write an Obsidian-friendly vault under <output>/vault/ — every artifact gets summary:/tags: frontmatter, cross-references become [[wikilinks]], and an INDEX.md MOC ties them together (with a .obsidian/ config). A local-first second brain for humans (graph view, backlinks) and agents (triage by summary: without opening files, a traversable index). Opt-in; off by default.

Status

v0.10.0 "The Upgrade Path" — an integrity release. v0.9 shipped correct, and then its findings run caught --refresh-shims being structurally unable to refresh half its targets, with 909 tests green over it. The cause wasn't the fix — it was that every test in the suite wrote into an empty directory, so only the first run had ever been tested. Everyone who isn't a new user lives on the untested path.

So 0.10 adds almost no surface. It makes the guarantees already made real: running Deepdive over a repo that already carries a previous release's output now provably converges (and provably leaves your hand-edited files alone), extract tells you when generated shims are stale instead of waiting for you to know the flag exists, and a stale examples/ fails CI rather than being caught by eye mid-release. New: forensic extract --timings, per-phase wall clock.

It also contains a cancelled feature, which is the more useful story. The planned performance work targeted PageRank. The profile required before touching it measured PageRank at 0.04 % of a large extract, and the already-batched store writes at 79–87 %. The optimization was disproven before it was written — see docs/findings/v0.10/PROFILE.md. Confirming a known technique had been applied had been mistaken for confirming it was fast.

The engine, the graph, the contract layer and the 5-artifact + 9-MCP-tool contract are unchanged from 0.9, and emitted content is byte-identical but for the version footer.

⚠ If you installed 0.8.0 or 0.9.0 and forensic serve stopped working — this is why. Those releases declare mcp>=1.27.1 with no upper bound. mcp 2.0.0 removed mcp.server.fastmcp, which the MCP server imports, so a fresh install of 0.8.0 or 0.9.0 today resolves to mcp 2.0.0 and fails on import. Our development lockfile pinned 1.27.1, so every local check stayed green and only the clean-environment wheel smoke in CI could see it. Fix: upgrade to 0.10.0 (which caps mcp<2), or pin mcp<2 yourself. Nothing else about those releases is affected.

What's proven, and what isn't — the honest-framing section moved to docs/findings/HONEST.md (unchanged content, DEC-119). Short version: proven usable with real agent auto-discovery; not proven to make autonomous issue resolution measurably faster end-to-end (hardware-gated). This is an assisted-analysis tool, and it is not overclaimed as more.

Quick start

# install from PyPI (puts `forensic` on PATH); or run ephemerally with uvx
uv tool install forensic-deepdive
forensic info            # banner + capability panel
forensic extract /path/to/repo

# …or from source for development:
git clone https://github.com/Dhevenddra/forensic-deepdive && cd forensic-deepdive
uv sync --all-extras

# what can it do? (banner + capability panel: artifacts, protocols, MCP tools, confidence legend)
uv run forensic info

# guided setup: analyze a repo, then wire it up as an MCP server
# (--yes takes every default: scriptable, and the one mode that needs no extra)
uv run forensic onboard --repo /path/to/repo

# run on any repo
uv run forensic extract /path/to/repo

# graph lands at <repo>/.deepdive/graph.lbug
# 5 markdown artifacts at <repo>/docs/codebase/
# 10 shims at <repo>/.claude/, .cursor/, .continue/, root

# trace a cross-stack feature slice (frontend call -> endpoint -> handler -> tail)
uv run forensic trace <symbol> --repo /path/to/repo

# interactive query REPL over one held-open store (needs the [interactive] extra)
# bare text = natural-language query (no LLM) · :cypher <q> = raw Cypher · :help · Ctrl-D exits
uv run forensic repl --repo /path/to/repo

# full-screen terminal graph browser — the loopback-free sibling of serve --ui
# 1/2/3 = Symbols/Files/Endpoints · type to filter · c/e/l = confidence/edge/language · Enter = context · i/f = impact/flow
uv run forensic browse --repo /path/to/repo

# the session shell: all of the above over ONE held-open graph, with history
# in-session: extract · query · trace · impact · flow · diagram · browse · onboard · serve
uv run deepdive --repo /path/to/repo

# query the graph as an MCP server (point it at the analyzed repo)
uv run forensic serve --repo /path/to/repo

# inspect every repo you've analyzed
uv run forensic list

Install from PyPI

Published as forensic-deepdive — no clone needed:

uv tool install forensic-deepdive        # puts `forensic` on PATH
forensic extract /path/to/repo

# …or run ephemerally, no install:
uvx forensic-deepdive extract /path/to/repo

Optional extras: uv tool install "forensic-deepdive[semantic]" (offline ONNX NL query), [interactive] (the forensic repl query console, the forensic browse TUI graph browser, and the deepdive session shell), [openapi] (YAML spec parsing), [graphiti] (temporal insight backend).

Upgrading from 0.8

Re-run extract with --refresh-shims once. The generated shims and skills under .claude/, .cursor/ and .continue/ are written only if absent, so a plain re-extract leaves your 0.8-era copies in place, and two of the skill files still cite internal decision IDs that 0.9 removed:

forensic extract /path/to/repo --force --refresh-shims

Only files Deepdive generated are rewritten. A shim you have hand-edited is never touched. pip install forensic-deepdive works too if you're not on uv.

Use it as an MCP server

forensic serve is a stdio MCP server exposing the 9 composite tools to any MCP-aware agent (Claude Code, Cursor, VS Code/Copilot, Codex, Continue, Cline, Windsurf). First build the graph once (forensic extract <repo>), then wire the server. Three ways, easiest first:

1. Claude Code plugin (self-hosted marketplace — no PyPI step):

/plugin marketplace add Dhevenddra/forensic-deepdive
/plugin install forensic-deepdive@dhevenddra

2. From the MCP Registry — indexed as io.github.Dhevenddra/forensic-deepdive, so registry-aware clients and discovery hubs (PulseMCP, MCPJungle, the VS Code @mcp index) can find and install it directly.

3. Manual config — generate a client snippet with forensic mcp-config, or paste:

{
  "mcpServers": {
    "forensic-deepdive": {
      "command": "uvx",
      "args": ["forensic-deepdive", "serve", "--repo", "."]
    }
  }
}

Per-client copy-paste blocks (Cursor, VS Code, Codex, the uvx-not-found GUI gotcha) are in docs/install.md.

The 9 supported languages

Python, C, Dart, Swift, TypeScript, JavaScript, Java, Go, Rust.

The 9 MCP tools

Tool descriptions are individually ≤200 tokens so the 9-tool envelope stays comfortably inside Anthropic's per-turn skill metadata budget.

The confidence taxonomy

Every edge and every emitted claim carries EXTRACTED / INFERRED / AMBIGUOUS:

• EXTRACTED — deterministic from AST or git log. Facts.
• INFERRED — a heuristic resolved cleanly (import-graph walk, receiver-type inference, single same-name candidate cross-file). High-trust but derived.
• AMBIGUOUS — multiple candidates surfaced; the resolver couldn't disambiguate. You see every candidate, not a silent guess.

HOTPATHS shows a per-row confidence-mix column so at a glance you can tell Logger (4 EXTRACTED + 1458 INFERRED — mostly clean) from ChatToolResponse (449 AMBIGUOUS — same-name cross-file collision).

Honest-mode (pure-static, zero LLM, zero network)

forensic extract works end-to-end with no ANTHROPIC_API_KEY, no OPENAI_API_KEY, no Ollama, no network. Graphiti is opt-in via the [graphiti] PyPI extra plus a 2-of-5 repo-size threshold (≥50 k LOC, ≥25 contributors, ≥18 mo old, ≥200 PRs/12 mo, ≥100 issues with discussion). The JsonlInsightStore is the always-available floor.

(The GitNexus/CodeGraphContext/DeepWiki/Sourcegraph comparison table is up top — it's the headline, not a footnote.)

Local development

git clone https://github.com/Dhevenddra/forensic-deepdive
cd forensic-deepdive
uv sync --all-extras
uv run forensic --version
uv run pytest -x          # 1000+ tests at v0.10.0
uv run ruff check src/ tests/
uv run forensic extract tests/fixtures/tiny_fixture

Read CLAUDE.md, DECISIONS.md (81 active DECs), and PROGRESS.md before making changes. This repo dogfoods its own pattern: every session starts with the protocol in CLAUDE.md, every architectural choice is captured as a DEC-N entry, and the artifact-name contract (MAP, HOTPATHS, ARCHAEOLOGY, MENTAL_MODEL, AGENT_BRIEF) is part of the public API.

Acknowledgments

• Aider (Paul Gauthier) for the PageRank-on-Tree-sitter repo-map pattern. Algorithm ported with attribution; we do not depend on aider as a package.
• Graphify (safishamsi) for the EXTRACTED / INFERRED / AMBIGUOUS confidence taxonomy. Productized in DEC-015 across every emitter.
• GitNexus (abhigyanpatwari) for the multi-repo registry pattern (~/.deepdive/registry.json, DEC-018), the composite-MCP-tool shape, and being the licensing wedge that makes this project's Apache-2.0 differentiation matter.
• Kuzu (now Apple-archived) for the embedded graph engine; LadybugDB for the live community fork that v0.2 ships against (DEC-013).
• Zep / getzep for Graphiti — the temporal knowledge graph that powers the above-threshold insight backend (DEC-019).
• Anthropic for the Skills format, Claude Code, and the MCP protocol that makes this whole product shape possible.
• Astral for uv and ruff.
• Repomix (yamadashy) for the original v0.1 flatten-the-repo pattern, now demoted to --legacy-repomix (DEC-017) but still available for legacy use cases.

Contributing

Contributions are welcome — see CONTRIBUTING.md for the dev setup, the verification gate, and the architectural invariants (the 5-artifact contract, the Endpoint keystone, the confidence taxonomy). By contributing you agree your work is licensed under Apache-2.0.

License

Apache-2.0. See LICENSE.

If you redistribute, modify, or build on this project, the Apache-2.0 terms apply: you must retain the copyright notice, the LICENSE text, and the NOTICE file, and state any changes you made (License §4). Attribution is required; the project is Copyright 2026 Dhevenddra (see NOTICE). The boilerplate header in the LICENSE appendix (Copyright [yyyy] [name of copyright owner]) is a template for applying the license to source files — it is not itself a requirement, and the LICENSE file is kept verbatim as the official Apache-2.0 text.