explore

Explore the codebase in three layers, stopping at the first one that answers the question. Do not skip ahead to reading full files "to be safe" -- each layer is a decision point, not a formality.

Layer 1 -- file index. Glob/find only, zero file reads. Get the shape: directory layout, naming conventions, where the relevant area likely lives. Stop here if the task is now obvious (e.g. "add a field to this one struct" and you can see exactly which file that is).

Layer 2 -- entry points and signatures. Grep for the symbols, exports, and call sites that matter to the task. You're building a map of what calls what, not reading implementations yet. Stop here if you now know which functions/files the change touches and roughly how they fit together.

Layer 3 -- targeted reads. Read with offset/limit only for the files directly relevant to the upcoming change -- never a full-repo read sweep. If a file is large, read the section around the relevant symbol, not the whole thing.

After each layer, explicitly decide: enough context to proceed, or one more layer? Don't default to Layer 3 out of caution -- most tasks resolve at Layer 1 or 2.

Return a structured summary to the caller: what you found, which files/symbols matter, and your recommendation -- not the raw contents you read along the way. The raw reads stay in this forked context; only the summary crosses back.

Before returning, cache the summary so the next session doesn't re-derive it (best-effort -- if this fails in any way, skip it silently and return your summary as normal; it is a cache, never a deliverable):

deadeye notes-append explore <<'EOF'
<the question you were asked>
- <a file or symbol that matters, and why>
- recommendation: <one line>
EOF

Ten lines maximum. If deadeye isn't on PATH, retry once with ~/.deadeye/bin/deadeye; if that also fails, move on.