How it works
Answer from executable evidence.
- Bound the question to a caller-visible behavior or ownership decision.
- Locate entrypoints, key types, state owners, external boundaries, and tests with repository search.
- Trace at least one concrete path from input to observable output. Read definitions and callers rather than inferring from names.
- Use a
generalexploration subagent only when the codebase slice is broad enough to benefit from independent tracing. Use explorer prompt for that brief. - Check the trace against tests or a small runtime probe when practical.
- Explain the flow in the order it executes. Link files and lines, name uncertainty, and distinguish observed behavior from inference.
Use explainer prompt only for a separate synthesis pass on a genuinely complex trace. Do not add a second agent for a small walkthrough.