The code graph

Six tables in SQL, walked with a recursive CTE. No graph database, and none needed.

The six tables

Table	One row is	Key columns
graph_files	A file the reading covered	path, language, loc, symbol count
graph_modules	A module	name, path
graph_packages	A package grouping modules	name
graph_symbols	A declaration	path, qname, name, kind, line, end_line, visibility
graph_imports	One file importing another	source path, target path
graph_calls	One symbol calling another	caller, callee

Every row is scoped by kb_id and project_id, so a re-analysis cannot mix two readings into one answer.

What it can answer

These are the questions no embedding encodes, which is why the graph exists alongside retrieval.

Who calls this? graph_calls, by callee name.

Who imports this file? graph_imports, by target — one hop.

What reaches this file, and how far away is it? A recursive CTE over graph_imports, returning each file with its distance in hops. This is the blast radius, and distance is the useful part: the file one hop away breaks first.

Does a test reach it? The same walk, asking whether anything under a test path appears in the set. A file nothing tests is a file a change to it is unverified by the suite, and that is worth saying out loud.

What does this file declare? graph_symbols, by path — which is also the symbol outline in the source viewer, and the table the lexical half of retrieval matches names against.

What it cannot answer, honestly

Call edges are resolved only where the callee is unambiguous

A method called through a variable, a callback passed as an argument, anything dispatched dynamically — those edges are not recorded, because recording a guess would be worse than recording nothing.

So find_callers returning nothing is not proof of no callers, and it says so in its own output rather than letting a reader infer certainty that is not there.

Import edges are more complete than call edges, which is why blast radius is built on imports.

Reach, and what it is for

reach_weight ranks how much a file matters by how much depends on it: capped transitive reach, plus five per written page that describes it. It lives in the base rather than in an app, so ranking by reach never requires opening one.

A file is high risk to edit at a reach weight of 25 or more, or at 8 or more with no test reaching it.

Where you meet it

• The studio, under any file in the source viewer: what it imports, who imports it, what reaches it and how far, whether a test covers it, and which written pages describe it.
• before_edit over MCP, which assembles all of it into the answer an agent wants before it changes something.
• What changed, which compares two readings.