Context graph

A map of the project, made without AI and kept up to date, that agents read instead of opening dozens of files, so they use less context.

Turning it on

In a repository where you have not decided yet, the Project window offers Generate the context graph?; the graph button of its header does the same at any time (generate, open the report, regenerate, turn off). The answer is saved in .atelier/config.json.

While it is on, a thread of its own watches the files (on each save and every 10 seconds), reads again only those that changed and rebuilds the graph, as a background task shown in the status bar.

What is read

• JavaScript and TypeScript;
• Rust: functions, types, methods, imports, modules, calls, inheritance;
• the headings and links of Markdown files.

Each link is EXTRACTED (read in the code) or INFERRED (deduced from a name that is unique in the project). Files that rely on each other form groups.

Files

In .atelier/graph/:

File	Content
GRAPH_REPORT.md	The summary to read first: central items, groups, and an estimate of the share of tokens saved. Written in English, for AI agents.
graph.json	The graph, in node-link format
cache.json	What each file said, so that nothing is read again after a restart

The folder is ignored by git (.atelier/.gitignore): the graph regenerates, there is nothing to share. Remove the line to commit it.

For AI agents

Agents query it through the graph_overview, graph_query, graph_node and graph_path MCP tools, whose answers stop at a token budget. The commit command also uses it to group files (see Commits by AI).