codemap

A static analyzer that turns a Python package's source into a queryable code graph. It reads source only — no runtime import — so it works on any package and stays decoupled from the code it analyzes. One canonical, deterministic graph store → many renders: API surface, dependency/architecture audit, RAG chunks, an Obsidian vault, mermaid diagrams, change-set review, and a SCIP index for interop with Sourcegraph / Glean and other precise-code-intelligence tools.

Status: 🟢 M0–M20 implemented + research track (R1/R2) — schema 0.13, 765 tests with no failures on Python 3.11–3.14 (in CI: the full suite including the dogfood pass, a determinism check, a wheel smoke test, and ctags/SCIP interop against the real CLIs), warm serve surface with 31 ops (28 exposed as MCP tools), and SCIP export. See DESIGN.md (product design & v1 boundaries), BACKLOG.md (roadmap), and research/ (tool landscape).

Why it exists

Docs describe code and CLIs call code; without a parsed map of the code both are done blind. codemap builds that map as facts: modules, classes, functions, the public API surface, import/inherit/ export edges, best-effort call edges, registry-family implements links, string-key column dataflow, and per-call argument contracts — then answers questions over it.

Design principles: source-only (static ast/griffe, never imports the target), deterministic (canonical sorted JSON, no timestamps — diffable; byte-stable on the fast tier, see the caveat below), CLI-AI-first (JSON by default, stable exit codes), honest (approximations are labeled, not hidden).

Point questions and whole-graph questions

Most code-intelligence tools answer point questions — you name a symbol and they walk outward a few steps. Where is this defined? Who calls it? What breaks if I change it? The answer's cost scales with the neighbourhood, not the repository, which is why an index or a vector store can serve it. The field is good at this, and several tools are faster at it than codemap.

Whole-graph questions have no starting symbol, because the property being asked about belongs to the graph and to no node in it:

Is there a dependency cycle anywhere? · Which module is most expensive to change? · Does the code still respect the layering I intended? · Where has behaviour concentrated into one class?

A cycle is invisible from inside every file that participates in it — each one looks perfectly reasonable alone. You cannot seed the question, and there is no partial answer. codemap report architecture computes all of them in one pass; on the dogfood target (88 modules, 715 import edges) that is:

layers         core 9 · data 12 · indicators 16 · analysis 42 · visualization 7 · cli 1
               analysis → core 38 edges · indicators → core 22 · data → core 13
violation    ⚠ analysis ↔ core             — one backward edge, written inside a function
cycle          pipeline → cache → pipeline  — the classic Python import-order landmine
lazy cycles    40 more, closed only by a function-local import — not import-time failures,
               still mutual coupling: neither module can be extracted without the other
coupling       core.logging_config  Ca 96   — a breaking change here reaches 96 modules
concentration  ZoneVisualizer 35 methods, worst function CC 66 / MI 12.5

That single layer violation is worth its line: it is reached only through an import written inside a function, so until R1-C29 codemap could not see it and reported the architecture as clean. A gate you can walk around by making the import lazy is not a gate.

Then codemap check turns the shape you want into a CI gate, so description and intent cannot drift apart.

Four rival tools have now been measured hands-on on the same benchmark scope — three of them on byte-identical input verified by content hash, the fourth (graphlens) on a near-identical staging that predates the harness — including the field's most-adopted tool at 68k stars. All four answer the point questions. None answers these on a surface a caller can reach — nor the related "how does each caller actually pass its arguments" (call_contract) — because a slice of source is the wrong shape for the answer.

That sentence used to end at "None answers these", and a second pass over the 68k-star tool made it narrower: its importable library does carry a cycle finder, reachable from neither its CLI nor its MCP tools and called nowhere in its own source. Measured on the same tree it reports 136 cycles; codemap reports 1. Scoring both against the truth set — every intra-package import, function-local ones included — is less flattering than that sounds:

	reported	of the 41 real ones	precision	recall
codemap, that morning	1	1	100%	2.4%
the peer's library API	136	13	10%	32%
codemap, after R1-C29	41	41	100%	100%

Theirs over-reports because it walks name-resolved call edges — a dict.get becomes a call into an unrelated class. Ours under-reported because its import map was module-level only, and it phrased that as "import graph is acyclic" — a property claim over a partial map, which is the worse of the two errors even though it was the smaller one. Both halves of that came from #11, filed off a different target and fixed the same day (R1-C29). The lesson is sharper than the win: a whole-graph answer inherits every flaw of the graph it is computed from, including the edges that graph never read — and the tool cannot be the judge of its own recall.

→ docs/whole-graph-questions.md — the full argument, every number above reproduced from one run, and the honest limits.

Install

Python 3.11+ — the range is measured on 3.11–3.14 in CI, not assumed (docs/ci.md).

The distribution is codmap; everything else is codemap. codemap was already taken on PyPI, so you install codmap — and then the command, the import and this repository are all spelled codemap, as they always were.

pip install codmap             # then: codemap build ./yourpkg

# optional: MCP server (`codemap serve --mcp`)
pip install 'codmap[mcp]'

# optional: SCIP export (`codemap export scip`)
pip install 'codmap[scip]'

Or straight from source: pip install git+https://github.com/kogriv/codemap.

Working on codemap itself, from a clone:

uv venv && uv pip install -e '.[mcp,scip]'    # or: pip install -e '.[mcp,scip]'

Dependencies: griffe (structure), networkx (query backend), jedi (deep call resolution). Optional extras: mcp (Model Context Protocol server), scip (protobuf, for SCIP export).

Quickstart

# build the canonical graph of a package (its directory name is the package name every id starts with)
codemap build ./yourpkg -o graph.json

# repo-scoped: add consumers (tests/examples) + docs for blast-radius/impact
codemap build ./yourpkg --deep --mode full --consumer ./tests --docs ./docs -o graph.json

# deep tier: union N independent samples (one build misses a real edge ~1 time in 4)
codemap build ./yourpkg --deep --repeat 3 -o graph.json

# ask about a symbol (JSON by default; --format text for humans)
codemap query analyze_zones --graph graph.json

# reports over the graph
codemap report architecture   --graph graph.json   # layers, coupling, god-objects, cycles
codemap report dependencies   --graph graph.json
codemap report dead-code      --graph graph.json
codemap report impact --symbol MyClass --graph graph.json   # who references it, and which
                                                            # flows it lands on — at which step
codemap report api-surface --graph graph.json --format json   # every kind has a json form

# change-set review straight from a diff → risk-sorted dossier
git diff | codemap review - --graph graph.json

# exports (see docs/export.md)
codemap export rag     --graph graph.json -o chunks.jsonl
codemap export mermaid --graph graph.json --mkind class
codemap export vault   --graph graph.json -o vault/
codemap export scip    --graph graph.json -o index.scip   # SCIP index (needs [scip] extra)
codemap export ctags   --graph graph.json -o tags         # universal-ctags tags file

# semantic (concept) search via an opt-in adapter → codemap symbols (needs the tool + opt-in)
codemap semantic "detect swing pivots" --build ./pkg --root pkg

# token-budgeted context pack — most relevant graph slice under N tokens (ranked; --seed to focus)
codemap pack --graph graph.json --budget 2000 --seed analyze_zones

# warm resident process — JSON requests over stdin/stdout (31 ops)
codemap serve --graph graph.json --source-root .

# …or expose the same surface as MCP tools for an AI-agent host (needs [mcp] extra)
codemap serve --graph graph.json --source-root . --mcp

# keep it current while you work: source → incremental rebuild → the warm server reloads
codemap watch ./pkg -o graph.json &
codemap serve --graph graph.json --watch

What it answers

• Structure & API — public surface, signatures, docstrings, deprecation.
• Dependencies & architecture — import cycles, layers + direction/violations, coupling (Ca/Ce/instability), god-objects & call-hubs, per-function complexity (cyclomatic / MI) blended with structural coupling.
• Impact / blast radius — who uses X, across the whole repo (core + tests + docs).
• Change review — a diff → the symbols it touches, their callers, signature-change surface, touched columns, cross-root consumers, risk rank.
• Dispatch seams — registry/factory families and the Protocol each impl satisfies.
• Dataflow — producers/consumers of a string-keyed DataFrame column.
• Semantic search (opt-in) — a concept query routed to an external adapter, with each fuzzy hit resolved to the exact codemap symbol at its location (codemap semantic). See docs/integrations.md.
• Context pack — the most relevant slice of the graph under a token budget, ranked by importance or by relevance to seed symbols (codemap pack --budget N [--seed X]). See docs/pack.md.
• Interop — export the graph as a SCIP index (definitions + symbol info + inherits/implements relationships) so Sourcegraph, Glean and other SCIP consumers can drive go-to-definition, symbol search and type hierarchy over it. See docs/export.md.

How it compares

codemap is the precise structural leg for index-free AI agents — it complements embeddings-RAG and Repomix-style packing rather than competing with them. Its bet is to be the best deterministic, diffable, provenance-aware graph in that slot, and to interoperate outward (SCIP, ctags) instead of locking the graph away.

A code graph an agent can trust: source-only, deterministic, diffable — no index to go stale, no LSP to provision.

The research track measures this against the field hands-on, on a shared benchmark scope. The positioning doc is the publication layer — the narrative and the numbers behind the claims above; comparison.md is the coverage matrix that backs them.

Where the comparison actually lands, after four hands-on cards: peers are faster, cover more languages, and several are easier to install. What none of them answers is the whole-graph class — architecture, layers, cycles, coupling — and per-call argument contracts. That is the substance behind "complements rather than competes", written out in docs/whole-graph-questions.md.

Honesty is part of the bet: the call graph is a measured lower bound, not a guess. docs/accuracy.md reports it — 100% precision / 100% decidable-recall on a hand-labeled suite, an openly-stated ~60% recall against all true edges (the price of Python's dynamism), and a grep-vs-graph proof that the graph is ~2× cheaper than grep for impact on unique names, tens of × on polymorphic ones, and no cheaper for locating a symbol.

One caveat on "deterministic"

The fast tier is byte-stable: two builds of an unchanged tree produce identical files, and CI compares the bytes on every push. The deep tier (--deep, jedi) is not. Ten deep builds of an unchanged tree here produced two distinct artifacts — differing in two per-symbol call counters out of 2133 nodes, no edges — and on a larger external tree one build in seven lost one real call edge of 9524. The cause is jedi's per-script execution budget: an inference that runs out of it returns nothing, and nothing is recorded as unresolved.

So a deep graph is one sample of a slightly fuzzy function, not a function of its input. Every deep build says so in its own diagnostics, and codemap diff says so when both sides are deep — read a delta of a few call edges as possible tool noise before reading it as a change in the code. Anything that must be reproducible byte-for-byte — a gate, a two-release comparison — belongs on the fast tier.

How many times to build. A consumer measured 175 full deep builds of one tree and found a real edge present in 75 % of them; eight builds here agreed (5 of 8, the same edge, and nothing else unstable among 13 675). So one build misses such an edge about 1 time in 4, two builds 1 in 16, three 1 in 64. codemap build --deep --repeat N builds N samples — each in a fresh interpreter, the regime the share was measured in; in-process repeats come in correlated streaks — and unions them: an edge seen in fewer than N runs carries extras.seen, provenance.samples records N and how many edges varied, and the note on the graph says both. Refused out loud on the fast tier (byte-stable already). With --incremental and watch, N becomes a property of the chain and every place the chain samples — base, fallback, each tick's recompute — samples N times; measured on twenty real commits, that is what removes the chain's misses, and a periodic full rebuild does not.

Measurement: gaps/deep_tier_nondeterminism_2026-09-02.md, gaps/deep_tier_union_by_repeat_2026-09-04.md.

Dogfooding

codemap is validated end-to-end against a real external package. Place a target repo as a sibling and run the full flow against its package — e.g. codemap build ../bquant/bquant (point at the package directory that holds __init__.py, not the repo root) — treating codemap purely as a third-party tool. The gaps/ directory records those dogfood runs: each is a pre-registered set of hypotheses, a run on the live graph, findings, and the milestone that closed them.

Documentation

• DESIGN.md — product design, the query catalog, v1 boundaries.
• docs/whole-graph-questions.md — start here for what codemap is for. Point questions vs whole-graph questions, why the field answers only the first class, the five questions codemap answers with a real run behind each, and the honest limits.
• docs/export.md — export recipes: RAG, mermaid, Obsidian vault, SCIP + ctags interop.
• docs/accuracy.md — measured call-graph accuracy, the honest static ceiling, and the grep-vs-graph value proof (both harnesses guarded in CI).
• docs/architecture-contracts.md — declare the intended architecture in codemap.toml and enforce it with codemap check (CI gate; codemap dogfoods its own).
• docs/api-diff.md — codemap diff two snapshots for added/removed/changed symbols and API breaking-change detection (release gate + review --base).
• docs/integrations.md — the opt-in router/adapter layer over external tools (codemap route / codemap semantic); license policy; adding an integration.
• docs/dead-code.md — graded dead-code candidates (high/medium/low + provenance reason) with a [dead_code] whitelist and --min-confidence filter.
• docs/pack.md — codemap pack: PageRank ranking + token-budgeted context slice for AI agents (global importance or seed-focused relevance).
• docs/attribute-edges.md — accesses edges: who reads/writes a class field, honest field-level impact (accessors; unknown vs none).
• docs/incremental.md — codemap build --incremental: recompute only changed modules (~12× faster on --deep), byte-identical on the fast tier; plus the automatic codemap watch + serve --watch loop (save → answerable in ~8 s on a real package at the defaults, of which 4.3 s is the rebuild).
• docs/test-mapping.md — codemap tests <symbol>: which tests exercise a symbol, as runnable pytest node ids, with the measured distance cutoff and an unknown that never pretends to be "untested".
• docs/hard-python.md — what the extractor does with metaclasses, dynamic classes, star imports, quoted annotations, .pyi stubs and symlinked trees; and the conditions where it warns instead of answering.
• docs/provenance.md — the provenance block: which tool, which tier, which input tree built a graph; what stays in the sidecar; the schema-mismatch warning and diff's comparability check.
• docs/flat-layout.md — flat module directories (sibling imports, no __init__.py): labelled resolution="flat" edges, and the empty-import-graph warning that stops a vacuous graph from reading as a clean one.
• research/blog/ — the build-story series: field notes on building codemap and measuring it against rival tools (EN + RU). See the section below.
• BACKLOG.md — milestones M0–M18, the research track (R1), and deferred work.
• gaps/ — dogfood runs, coverage analysis, the living axis register.
• research/ — survey of adjacent code-analysis tools and how codemap relates to each (integrate / wrap / learn); source of the R1 capability roadmap. See research/positioning.md for the publication-layer narrative and research/comparison.md for the hands-on coverage matrix.

Writing — the build-story series

Field notes on building codemap, and on measuring it honestly against the nearest rival tools. Published here in the repo; every post exists in English and Russian. Index: research/blog/.

#	Post
0	A code graph an agent can trust — what codemap is, the bet it makes, and the honest limits.	EN · RU
1	The competitor wasn't broken. We were. — I nearly published that a rival's impact analysis was broken. The bug was my PATH.	EN · RU
2	The one that does more — and why that's fine. — a 1.7 GB hybrid rival that proved the thesis instead of threatening it.	EN · RU
3	The competitor that does less — and that's why I take it. — the emptiest coverage row was the most useful find. For its license, not its features.	EN · RU
4	My determinism test went red. The tool was fine. — the input was moving under it, and the artifact could not say so. How the graph learned to name what built it.	EN · RU
5	A month of dogfooding. Then one more repository found seven bugs in two days. — eleven pre-registered axes, all asked of one tree. What was missing was not an angle but a shape.	EN · RU
6	I measured the 68,000-star competitor. It was faster than mine. That wasn't the finding. — four rivals, and the same two columns empty in all of them.	EN · RU

New here? Read 1 → 0 → 2 → 3 → 4 → 5 → 6. Every number in every post reproduces from a tool card or the comparison hub — measurements, not verdict.

License

MIT.