| name | endor-fetch-and-search-call-graph |
| description | Fetch project call graph artifacts, decode zstd payloads into searchable node and edge files, and run safe direct-edge or multi-hop path searches on the customer PackageVersion plane. Use for static symbol paths (e.g. app code to a dependency API)—not for Finding/CVE reachability triage (hand off to endor-reachability-provenance), scan finding lists (endor-retrieve-scan-results), or natural-language function summaries (VectorStoreQuery; separate flow). |
Fetch and Search Call Graph
Purpose
Turn raw call graph storage into searchable, join-friendly JSON (decoded_callables.json, decoded_edges.json when using --decode-zstd) and run deterministic node/edge/path queries on the customer tenant graph (keyed by PackageVersion).
Use this skill for static call-graph facts: symbol presence, direct edges, and transitive paths. Do not treat graph edges as proof of runtime execution.
Plane gate (read first)
| User ask | Use | Do not use |
|---|
“Does our code call library X?” (e.g. APIClient → httpx) | This skill | VectorStoreQuery, finding reachability |
| “Is the vulnerable function reachable for this Finding?” | uv run endor-reachability-context → endor-reachability-provenance | Call-graph search alone |
| “List findings / scans for a project” | endor-retrieve-scan-results | Call graph |
| “Semantic function summary / NL code search” | client.VectorStore + VectorStoreQuery (tenant store; scope by project meta.name) | CallGraphData.decode |
Vocabulary: Call-graph reachability = static edges in decoded JSON. Finding reachability = finding tags, oss vuln metadata, stitched customer + oss planes — different skill.
Ordering (escalation ladder)
Do not skip to ad-hoc CallGraphData.decode on the first PackageVersion row.
-
Credentials — uv run --env-file .env (or equivalent) for endorlabs.Client.
-
Fetch (required first) — endor-project-retrieval-bundle with --callgraph-export and --decode-zstd:
uv run --env-file .env endor-agent-context \
--tenant "<tenant>" \
--namespace "<namespace>" \
--project "<project-uuid-or-name>" \
--output-dir .endorlabs-context/workspace/projects \
--callgraph-export \
--decode-zstd
-
Discover paths — read context_manifest.json → artifacts.callgraph_export. Nested callgraph_export/callgraph_export_manifest.json lists decoded file paths.
-
Reuse only when present — if a prior bundle exists under workspace/projects/, read its manifest first; do not re-fetch unless stale or force requested.
-
Single-PV spot-check (last resort) — client.CallGraphData.decode(package_version) only after PackageVersion selection. See facade-helpers.md.
-
Search — uv run endor-callgraph-search with --callables, --edges, and patterns; use --path-from / --path-to for multi-hop BFS. Follow path search protocol.
-
Reason — join on method_id; record direct edges and transitive chains when wrappers sit in between.
Stop condition: If artifacts.callgraph_export is null or every decode attempt returns NotFoundError for PVs tried within cap → report “no CallGraphData for project” with PV count tried. Do not fall back to tenant-wide unscoped search.
PackageVersion selection
When using programmatic decode (step 5 above):
- Resolve
Project first; use namespace=project.namespace on PackageVersion.list_by_project.
- Prefer PVs with
spec.call_graph_available=true, ordered refs/heads/main / main first, then other source_code_reference.ref values.
- Use
resolve_package_version_with_callgraph or build_callgraph_pv_inventory — when no row has call_graph_available, the inventory message explains listed refs (do not guess).
- Try
CallGraphData.decode(pv) on ordered candidates; NotFoundError → next PV.
- Record
pv_uuid, source_ref, and pv_label in the output template.
Pass 3 of endor-agent-context records which PVs exported graphs in the export manifest — prefer those paths over guessing.
Path search protocol
Thin wrappers (e.g. get() → _request() → _request_with_retry() → httpx.Client.request()) often yield zero direct edges between the user-named source and target. That is normal.
- Existence — confirm both ends appear in
decoded_callables.json (--node-pattern or grep URIs).
- Direct edge —
endor-callgraph-search with --source-pattern and --target-pattern.
- If
edge_hits_total is 0 — run endor-callgraph-search --path-from … --path-to … or import find_call_graph_path. See call-graph-format-and-search.md.
- Distinguish return-type nodes (
.../httpx._models/Response) from call targets (httpx._client/Client.request()).
- Collision — if one source fans out to many same-name targets, list the fanout set.
Example (portable patterns — substitute your module names):
uv run endor-callgraph-search \
--callables "<decoded_callables.json>" \
--edges "<decoded_edges.json>" \
--path-from "APIClient" \
--path-from "get" \
--path-to "Client.request" \
--max-depth 6 \
--out .endorlabs-context/workspace/runs/callgraph-search/path.json
Why this skill exists
Raw call graph payloads are not directly search-friendly. Safe retrieval starts from decoded files and uses method IDs for joins. URI substring search alone misses multi-hop chains.
Library and CLI entrypoints
endorlabs.workflows.callgraph.run_callgraph_export — enumerates PVs and writes call graph exports (Pass 3 via --callgraph-export).endorlabs.workflows.callgraph.find_call_graph_path — BFS on decoded JSON.endorlabs.workflows.callgraph.resolve_package_version_with_callgraph — first decodable PV with call_graph_available (main ref first).endorlabs.workflows.callgraph.build_callgraph_pv_inventory — diagnostic when no call-graph PV exists for the project context.client.CallGraphData.decode(package_version) — decoded shape (CallGraphDecoded: summary, callables, edges, envelope).endor-callgraph-search — direct-edge and --path-from / --path-to search on local JSON.endor-callgraph-path — live project → decode → path search.endor-vector-query — list/probe/query tenant vector stores (separate from call-graph export).- Bundle orchestration:
uv run endor-agent-context ... --callgraph-export (see endor-project-retrieval-bundle).
Inputs
tenant — endorlabs.Client tenant / auth.namespace — project list scope; use the project’s tenant_meta.namespace for PV lists and decode.project — 24-hex project UUID or meta.name (repo URL).- Decode —
--decode-zstd when you need text JSON for search.
- Search CLI — paths to
--callables and --edges; repeatable patterns; --path-from / --path-to for BFS path mode.
Outputs
- Directory under
.endorlabs-context/workspace/projects/ (bundle) or workspace/runs/callgraph-search/ (ad-hoc) with callgraph_export_manifest.json when using export.
- Decoded
decoded_*.json when --decode-zstd is set.
context_manifest.json at bundle root: artifacts.callgraph_export locates export output (null if Pass 3 was not run).
Bounds
- Cap
PackageVersion enumeration (--callgraph-max-pages, --callgraph-page-size on export). Do not assume full-tenant graph dumps.
- Pass 2 default
--pv-limit may hydrate only a subset — read multipass-llm-contract.md before inferring repo-wide posture from a partial bundle.
Output template
Use this structure in the final result:
## Call graph result
- Project: <uuid> namespace=<project.namespace>
- PV used: <uuid> <label> (or export manifest row)
- Source: <uri or search pattern>
- Target: <uri or search pattern>
- Direct edges: <n> (`edge_hits_total` or ID join)
- Transitive path: <full URI chain> or none within depth N
- Artifacts: <paths to decoded_callables.json, decoded_edges.json, search JSON>
- Ground truth (optional): <local file:line if repo checkout available>
## Notes
- Static graph only — not runtime execution proof.
- Finding/CVE reachability: hand off to endor-reachability-provenance if needed.
Example use cases
- Verify whether function A reaches library B (including through private helpers).
- List nodes mentioning a dependency (
httpx, requests, etc.).
- Find edges where source and target belong to different subsystem patterns.
- Extract same-name target fanout for one source symbol.
Safety checks
- Prefer decoded artifacts over raw envelope JSON.
- Confirm symbol existence before claiming reachability.
- Zero direct edges ≠ no path — run BFS before concluding.
- Use method IDs for final joins, not URI text alone.
- Treat static graph edges as graph facts, not runtime execution proof.
Related skills
Documentation hops
