geml-code-graph

Build, view, update, and navigate a project's call graph as GEML codemap documents. Use when asked to see/update/build a project's code graph or codemap (看下/更新下 code-graph), when asked "who calls X" /

SKILL.md

GitHub

Code-graph navigation (codemap profile)

This is the in-repo variant of integrations/*/skills/geml-code-graph/SKILL.md: paths are repository-relative and it adds the local node geml-parser/dist/geml.js codemap … form; the plugin copies are byte-identical to each other and ship with the agent host plugins (geml skill install installs only the geml skill).

The call graph lives as text documents, not a database (profile: spec/profiles/geml-codemap/geml-codemap-profile.md): one GEML document per container (module / dir / file), each with ONE meta (module, src, entry, resolution-default), empty-body code blocks per method, and up to three CSV edge tables — #calls (out), #called-by (in), #unresolved (blind spots). The build's verify has checked that every edge reference resolves.

The moves

# 1. resolve a name — where does a symbol live
node -e "console.log(JSON.stringify(require('./.geml/codemap/_index/name-lookup.json')['hashtableFind'],null,1))"
#   → [{"anchor":"c:hashtable.c#hashtableFind(…)","doc":"hashtable.c.geml","id":"hashtableFind"}, …]
#   Multiple entries = real ambiguity (e.g. a .c definition and a .h inline) — inspect each.

# 2. container overview — the module's surface, one glance
head -8 .geml/codemap/hashtable.c.geml          # meta: entry = the externally-called methods

# 3. open the method block (src= tells you exactly where the code is)
geml get .geml/codemap/hashtable.c.geml '#hashtableFind'

# 4. forward: what it calls (grep your method's rows; follow doc.geml#id refs)
geml get .geml/codemap/hashtable.c.geml '#calls'

# 5. reverse: who calls it (aggregated, with file:line sites)
geml get .geml/codemap/hashtable.c.geml '#called-by'

A reference is #id (same document) or sibling.geml#id (that document, that block) — geml get it the same way. index.geml holds the repo-level view: app entries in its meta, #modules / #module-edges aggregate tables.

Reading the tables

Line Meaning
#calls row, empty confidence resolved at the document's resolution-default, high confidence
#calls row kind=candidate dispatch ambiguity: one of several implementations, right after its main call row. Treat the SET as the answer, never just the first
#calls row confidence medium/low the extractor is less sure — say so when reporting
#unresolved rows (hidden table) calls the extractor could NOT resolve — blind spots, not evidence of absence; fall back to grep when one matters
#called-by absent for a method no resolved callers. Under resolution-default = heuristic that means little; under cpg it is strong (but pointer/dynamic dispatch still lands in #unresolved)

Symbol classes: .accessor (bean get/set/is leaves — the graph view hides them by default, tables keep them) · .leaf (calls nothing, only called — usually skippable when tracing logic) · .test (test territory) · .flow-entry (critical-flow start).

"看下/更新下 X 项目的 code-graph" — the end-to-end move

The toolkit ships inside the @geml/geml package: geml codemap … (without a global install: npx -y @geml/geml codemap …; inside the geml repo: node geml-parser/dist/geml.js codemap …).

Dispatch first — generation is slow, the conversation must not block on it

Indexers take real time (scip: seconds–minutes; Joern on a repo: minutes). Pick the executor BEFORE starting:

  • Codemap exists, user wants to look → inline, seconds: serve --background + open the browser. No subagent.
  • Update asked and _index/refresh.json exists → no subagent either: geml codemap refresh <dir> --background (detached process, costs the conversation nothing). Open the CURRENT graph immediately — serve renders live, so when the refresh lands, F5 shows it; say exactly that.
  • geml files must be (re)generated agentically — first build, no recipe recorded, adapters change, or a refresh failed → hand the WHOLE generation to ONE subagent (Agent tool; run_in_background: true so the user can keep working). Its prompt must be self-contained: project root; detect the languages per the table below (never ask); the exact indexer + geml codemap build --history + geml codemap verify commands; verify MUST exit 0; write _index/refresh.json with the exact commands used; return container/method/entry counts, verify result, and any language gaps. The MAIN conversation does the last mile itself when the subagent reports: serve --background, open the browser (if an older codemap was already on screen, telling the user to F5 is the whole move).
  1. Have a codemap? <proj>/.geml/codemap/index.geml exists → skip to step 4 (view) or step 3 (update was asked). An older codemap//graph/ tree from before the rename is not special: regenerate into .geml/codemap/ (one build; carry the *.gemlhistory sidecars over first if they matter) and remove the old directory.

  2. Detect the language(s) — NEVER ask the user. (Steps 2–3 are the generation work — per Dispatch above they normally run inside the subagent.) Judge from manifests first, then source-file counts (Glob/ls). Multiple languages with real code (≥ a handful of files each) → one build with REPEATED --adapter groups; the codemap merges them (Java+TS validated).

    Signal Indexer → adapter
    tsconfig.json / mostly .ts .tsx .js npx --yes @sourcegraph/scip-typescript index --output index.scip (run IN the target repo/subproject) → --adapter scip --raw index.scip
    React / JSX (.tsx .jsx) same scip route, verified tier: <Child /> render edges, custom-hook calls, and useReducer(reducer, …) wiring all resolve high — arrow components (const Foo = () =>) included. Indirect dispatch is absent, not #unresolved: callback-prop calls (onToggle(…)), dispatch()→reducer case handling, and context-injected functions ride scip locals/members and leave NO edge — grep when one matters. Also invisible: memo()/forwardRef()-wrapped components (const = call, inner fn is a local) and module-scope render(<App />) callers
    Cargo.toml / .rs rust-analyzer scip . --output rust.scip (run IN the crate/workspace root; missing → rustup component add rust-analyzer or the rust-analyzer GitHub releases page) → --adapter scip --raw rust.scip. Precise tier: rust-analyzer-resolved, cross-file/cross-crate calls included; calls into std/external crates land in #unresolved
    pom.xml / build.gradle / .java Joern (locate per Locating Joern below; JDK required): GEML_SRC=<abs-src> GEML_OUT=<abs-raw> GEML_LANG=JAVASRC joern --script <pkg>/codemap/joern-export.sc → --adapter joern --raw <raw>. GEML_LANG takes Joern's --language names, UPPERCASE — lowercase javasrc fails with "No CPG generator exists"
    .c / .h same Joern route, GEML_LANG=NEWC (valkey-validated)
    .py / go.mod / .kt Joern frontends, GEML_LANG=PYTHONSRC etc. (usable tier — SAY SO in your report)
    only a code-review-graph graph.db --db <graph.db> (heuristic tier — say so)
    none of the above report honestly which languages are unsupported; do not guess

    .vue / .svelte SFCs: covered — use the AUTO build (geml codemap build --root <proj>), not the manual per-indexer route. It virtualizes each SFC project (Volar / svelte2tsx, fetched hermetically via npx) into shadow TS with line-map sidecars, runs one scip pass over shadows + the project's real TS/JS, and attributes every symbol back to the original file and line. Template event handlers surface as edges from a synthetic <Component>.template node (@click="save" → #App-template, #save; mustapi-validated across three Vue apps, 85/85 SFCs). Honest residuals — say them when reporting: component-TAG usage (<Child/>) is not a call edge; Nuxt auto-imports (unimported ref, auto-registered components) don't resolve, so those references drop; top-level <script setup> calls, including computed(() => …) bodies, drop exactly like module-level calls in plain TS; a failed virtualization falls back to plain TS indexing and says so.

    Vendored source trees explode the job list — next.js's packages/next/src/compiled/ carries ~140 checked-in package.json bundles, each becoming its own scip job. Prune them at build time: geml codemap build --root <proj> --exclude "src/compiled/**" (repeatable; the exclusion also keeps their symbols out of the graph).

    Locating Joern — never hardcode a path. Resolve it fresh on each run, in this order: (1) joern on PATH — if joern --version works, use it; (2) else read ~/.claude/skills/geml-code-graph/config.json ({"joern": "<launcher-or-dir>"}) and pass it as geml codemap build … --joern <path> (or export GEML_JOERN); (3) else ASK the user for the joern-cli location (Windows: the folder unzipped from joern-cli.zip; macOS/Linux: the joern-install.sh install dir), WRITE it into that JSON file, then reuse it. <path> may be the launcher itself or the directory holding it (joern.bat on Windows, joern on unix). Ask at most once per machine — after that the JSON answers. Mirrors the CLI's own --joern / GEML_JOERN resolution.

  3. Build + verify (also the "更新" path — builds are deterministic, only changed documents are rewritten):

    geml codemap build --adapter scip --raw index.scip --root <proj> \
         --out <proj>/.geml/codemap --history      # --container module|dir|file: match
                                             # the layout (default dir; flat C repo → file)
    geml codemap verify <proj>/.geml/codemap       # MUST exit 0 before showing anyone
    

    First successful build: record the recipe so refresh (and the commit hook) can replay it — write <proj>/.geml/codemap/_index/refresh.json with the EXACT commands you ran:

    { "root": "..",
      "steps": ["npx --yes @sourcegraph/scip-typescript index --output index.scip",
                "geml codemap build --adapter scip --raw index.scip --root . --out .geml/codemap --history",
                "geml codemap verify .geml/codemap"] }
    

    From then on, "更新下" = geml codemap refresh <proj>/.geml/codemap (skips itself when git HEAD hasn't moved; log at _index/refresh.log).

  4. View — finish with the browser OPEN, not with instructions.

    geml codemap serve <proj>/.geml/codemap --background   # detached: SURVIVES the agent session;
                                                     # http://localhost:8140, pages render live
                                                     # from .geml — rebuild + F5, never stale.
                                                     # already-running port → reused, not stacked.
    geml codemap serve <proj>/.geml/codemap --stop         # stop it (pid: .geml/codemap/_index/serve.pid)
    geml codemap render <proj>/.geml/codemap               # serverless alternative: bake .html next to
                                                     # each doc; open file:///…/.geml/codemap/index.html
    

    Always --background (a viewer must not die with the session). Then open it for the user: Windows start "" <url> (or Start-Process <url>), macOS open <url>, Linux xdg-open <url>. Port taken by something else → pick another (--port), open that one.

index.html is the module overview; clicking a module opens its page inside the graph area (nested view). Method pages: click = callee chain, ⊕ on an entry = full caller chain, breadcrumb walks back up.

Keep it in sync on every commit (optional per-project hook)

With the recipe recorded (step 3), a Claude Code PostToolUse hook makes any git commit Claude runs in that project refresh the codemap in the BACKGROUND (never blocks the commit; non-commit commands exit instantly; projects without refresh.json are silently skipped). Add to the project's .claude/settings.json:

{ "hooks": { "PostToolUse": [ { "matcher": "Bash", "hooks": [
  { "type": "command", "command": "geml codemap refresh .geml/codemap --hook --commit" }
] } ] } }

(.geml/codemap = the codemap dir relative to the project root; use an absolute path if the hook cwd differs.) With --commit, the refreshed documents land as their own follow-up commit — chore(codemap): refresh for <sha>, codemap dir only — so the next push carries code + graph together. It is loop-safe (the follow-up commit changes no source file, so the refresh it triggers skips) and it stands down when HEAD moved during the refresh or a merge is in progress. Drop --commit to keep the old behavior: refreshed files stay in the working tree for you to include in a later commit.

Between commits (editing-time sync), geml codemap serve <dir> --watch re-runs the recipe after 30s of quiet whenever an indexed source file changes — pages render live, so a browser reload shows the new graph.

Add --history [-m msg] to build to snapshot changed documents into .gemlhistory sidecars — then geml history get .geml/codemap/<doc>.geml shows the graph's evolution and geml revert .geml/codemap/<doc>.geml '#method' --rev -1 rolls one method's edges back. Language maturity tiers and the smoke-test gate: docs/design/specs/geml-codemap/DESIGN-geml-code-graph.md §3.4. An MCP wrapper exists (geml mcp --root <dir>, which serves the four read-only geml_codemap_* tools next to the document tools when the root holds a graph); the CLI path works without it.