| name | explore |
| description | Deep autonomous codebase exploration — traces behaviors end-to-end, creates boundaries by concern, builds navigable depth with sub-boundaries, enforces quality via cartographer_check_depth. Use when the user says explore, map, understand, or analyze deeply. |
| model | opus |
| effort | high |
Codebase Exploration
Verify tools work: call cartographer_get_summary. If it fails, stop and
tell the user the Cartographer MCP tools are not connected.
Set the project root: cartographer_set_project.
Save a snapshot: cartographer_snapshot with label "pre-exploration".
Checklist
Copy this checklist. Track progress. Do not skip items.
Exploration Progress:
- [ ] Orient: read project, understand what it does
- [ ] Behaviors: trace key behaviors end-to-end as slices (as many as the system warrants)
- [ ] Boundaries: cluster entities by concern, record boundaries
- [ ] Depth: create sub-boundaries (check_depth must pass)
- [ ] Perspectives: create at least one focused perspective
- [ ] Synthesis: report findings to user
How to Explore
Behaviors first. Identify what the system DOES, not what files it has.
Trace each behavior from entry point to completion. Record entities as
encountered: actors, capabilities, state, side effects, invariants, failure
points. Record each path as a behavior slice.
Boundaries emerge from behaviors. After tracing flows, cluster entities
that participate in the same behaviors. Name by concern ("authentication",
"payment processing"), NEVER by directory ("src", "lib", "components").
Depth is mandatory. Every boundary with more than 3 entities must have
sub-boundaries. Read the source for the key classes/functions inside each
boundary. Create a child boundary. Record its internal capabilities, state,
invariants as children.
Perspectives for navigation. Create at least one perspective for the
most important concern. If cartographer_create_perspective is unavailable,
skip gracefully — sub-boundaries and slices provide depth regardless.
Before Finishing
Call cartographer_check_depth. Fix every issue it reports. Call it again.
Repeat until passed: true. Do NOT report to the user until it passes.
Bad exploration (avoid)
- 50 entities, all flat, no sub-boundaries, no slices — a parts list
- Boundaries named "Service", "UI", "Plugin" — directory mirrors
- Descriptions like "Express middleware handler" — engineer jargon
- No behavior flows — structure without stories
Good exploration
- 30 entities with 2+ depth levels and 5 behavior flows — a navigable map
- Boundaries named "Payment processing", "Real-time sync" — concerns
- Descriptions like "Validates cart before checkout" — readable by anyone
- Stories: "When a user checks out, this chain fires"
