How to ask impact questions

How-to: ask impact questions

The question the README demo asks, as a repeatable workflow — verified live on cie itself (the uncut cast is in docs/demo).

The workflow

1. Find the symbol — search_symbol("resolve_backend") → node id, file, line range, signature.
2. Who calls it — callers("resolve_backend") → receiver- resolved call sites with confidence tags (13 callers for this symbol on cie itself).
3. Blast radius — affected_by("cie/config.py") → what depends on that file, multi-hop, with distances (30 hits: the CLI openers, _selected_backend, mcp_server's main…).
4. Which tests catch it — test_map("resolve_backend") → the pinning tests (7, with line numbers, after v0.1.4's direct-calls heuristic).

One plain-words question to an agent does all four if the client is registered — that's the whole demo.

Reading the envelope

Every tool returns the same shape:

{"ok": true, "tool": "callers", "results": [...], "truncated": false,
 "total": 13, "hint": null, "elapsed_ms": 12}

Empty results are honest, with a hint saying why — e.g. test_map with no match: "a test that only reaches its target through a helper is missed". confidence values: EXTRACTED (proved from the code) vs INFERRED (heuristic — verify before betting).

Gotchas

• affected_by takes a file path, not a symbol name (the per-file blast radius). callers takes the symbol.
• TESTS edges are heuristics, honestly not exhaustive — see Extraction pipeline.