Deep call-path queries (e.g. "trace HTTP insert → nativelog write") return no result — sparse C++ call graph + agent doesn't use path_between

[gangtao]

Symptom

Asking the chat agent a call-path question like "Trace the call path from HTTP insert to nativelog write" (one of the empty-state suggested questions) returns "I could not find a documented call-path…" even though the endpoint entities exist. The agent runs ~9 search_entities keyword calls and gives up.

Root cause (two compounding factors)

1. The agent treats it as keyword search, not graph traversal.
The trace shows 9 search_entities calls and zero path_between / neighbors. "Trace the call path from A to B" is inherently a path_between(A, B) question — find the two endpoint nodes, walk the calls edges. The agent never attempts a traversal, so it never reconstructs a path.

2. Even with the right tool, the path isn't in the graph.
Measured on proton-enterprise@v3.3.1:

• Call graph is sparse: only ~38% of function/method nodes (22,715 / 60,467) have any outgoing calls edge.
• InterpreterInsertQuery (the INSERT entry point) has 0 outgoing calls edges — a leaf in the extracted graph, so nothing to trace from it.
• proton is ingested code-only (AST). C++ AST call extraction misses virtual dispatch, templates, and callbacks — exactly how ClickHouse/proton dispatches between layers via abstract interfaces (IStorage/IInterpreter, indirect_call is only 851 edges vs 46k direct calls).
• The real path (HTTP handler → parser → interpreter → storage → WAL/nativelog) is many layers deep, beyond path_between's MAX_PATH_DEPTH = 6 cap (src/tpk/tools.py:52).

The individual entities are findable (InterpreterInsertQuery, NativeLog ×92, HTTPHandler ×22) — browsable in Explorer — but the end-to-end chain connecting them can't be reconstructed from the current data.

Options

A. Small — nudge the agent to use path_between/neighbors for trace questions.
Improve the agent system prompt / tool descriptions so "trace"/"call path"/"how does X reach Y"/"what calls X" questions drive neighbors (single-hop: what calls X / what X calls) and path_between (endpoint-to-endpoint). Makes single/few-hop call questions work reliably. Does not fix deep cross-layer traces — the agent can't invent edges that aren't extracted. Low effort, real value for the common "what calls X?" case.

B. Large — richer call-graph extraction.
Either a semantic (LLM) extraction pass on proton (currently code-only), or a proper C++ call-graph analyzer (e.g. clang-based) that resolves virtual dispatch/templates. This is the real fix for deep-path questions, but a substantial effort. Possibly also raise/parameterize MAX_PATH_DEPTH for deep codebases.

Suggested near-term action

Do A (small agent-prompt improvement) and set expectations in the UI: consider replacing the "Trace the call path from HTTP insert to nativelog write" suggested question with one the current graph can answer well (e.g. "What calls CheckpointCoordinator::trigger?"), keeping deep-trace as a target for B.

Context

Found while debugging chat answers. Related fixes already merged this cycle: stale bare-name node keys re-ingested; repo-filter now accepts bare name or name@ref (#14). This issue is specifically about call-path reconstruction depth/coverage and agent tool selection, which those did not address.