how

The deliverable is a mental model: an explanation pitched at a senior engineer onboarding onto the subsystem, stopping short of annotated source. Every claim in it is traced to the implementation.

1. Frame the question

Take the question from the invocation, or from the subject under discussion this session. When its scope is ambiguous, state your reading in one line and proceed.

Cut the question into angles, each a distinct slice of the subsystem: one angle for a single module or function, two to four for a subsystem spanning files or services. When in doubt, take fewer.

Done when every angle names a file or symbol to start from, and every part of the question sits inside one of them.

2. Explore

Dispatch one explorer per angle in a single message (subagent_type: Explore, model: sonnet, search breadth very thorough). Each prompt carries the path to references/explorer-brief.md, the question, and that explorer's angle.

Done when every return traces its angle to code, a file path behind each component and each flow step. Re-dispatch an explorer that came back on names alone.

3. Explain

Synthesise the findings into one mental model. Where explorers overlap, merge; where they contradict, settle it by reading the code yourself. Write per /writing-for-humans, and: