code-understand gives an agent an on-demand architecture map of a codebase it does not know yet. It runs the obsidian-wiki code-understand command, which ranks the load-bearing symbols and files and cites each one by file and line.
The agent then reads only what the map points at, which keeps orientation, impact analysis and caller lookups short. It uses CodeGraph when installed, an existing Graphify graph next, and a built-in extractor that needs nothing else.
What You Get
- A ranked list of load-bearing symbols and files with file and line citations
- Impact analysis for a diff, so you see the blast radius before finishing a change
- Guidance on which entries are structural facts and which are text matches to confirm
- A fallback path that works with no extra install
Common Use Cases
- Orient yourself in a new repository before the first edit
- Answer who else calls this function before changing its signature
- Rank the files to read before a large refactor
- Check what a set of changed files could break
What Makes It Unique
It treats the map as a pointer to the code, and labels how strong each piece of evidence is: real call-graph edges from CodeGraph versus text matches from the builtin backend. It also degrades quietly, falling back to ordinary grep and read when the command is missing.
Under The Hood
The skill is instructions for the agent plus one command, obsidian-wiki code-understand, which scans the project with a regex extractor and ripgrep, or with CodeGraph or a Graphify graph when present. The agent runs shell commands and reads project files. CodeGraph keeps a disposable, git-ignored index folder in the project, so the skill can write files there, and it tells the agent never to put that output in the wiki. It may suggest installing CodeGraph with npm, which downloads a package from the npm registry, and the skill tells the agent to ask you first. Otherwise it makes no web requests.