When an AI agent edits a function, today's tools hand it every dependency the code structurally touches — dead code, deprecated fallbacks, paths that haven't executed in years. The agent burns tokens on all of it, and sometimes optimizes for the wrong one.
The fix
Run the app, collect OpenTelemetry traces, and map each trace back to the exact function in the parsed code tree. Now every function and every call edge carries an execution count — the graph stops being a guess and starts being a record of what actually happened.
What that unlocks
- Hot pathGive the agent only the dependencies that actually ran recently, ranked by frequency.
- Real payloadsInstead of guessing what a dynamically-typed argument looks like, show the real JSON that passed through it.
- Latency attributionPoint at the specific functions responsible for slow requests.
The hard part
The thesis of the project: correlating a runtime span back to a static function. You have byte offsets from the parser and file/line info from the trace. Making those line up reliably — through async, decorators, and transpiled code — is the real work.
The pitch
"AST graphs are a map. I added traffic data."
Status
In active development — heading to GitHub soon. Check back, or reach out if you want a look before it's public.