Exploration Skill

Intent

Build a reliable mental model before planning or patching.

Trigger

Use this skill when:

• •repository is unfamiliar
• •request references unknown modules
• •design decision needs structural context

Exploration Workflow

Step 1: Entry-point scan (breadth first)

Start from top-level anchors:

1. •runtime/package manifests
2. •app entrypoints
3. •build/test config
4. •major module directories

Suggested command sequence:

rg --files

Then inspect key anchors with read.

Step 2: Build dependency map

Identify:

• •upstream inputs (APIs, config, events)
• •core transforms (services, domain logic)
• •downstream outputs (CLI, HTTP, files, external tools)

Capture module edges in concise form:

MODULE_EDGE
- from: <module>
- to: <module>
- reason: "<import/call/data flow>"

Step 3: Locate critical paths

Find the hot path for the user request:

• •request entry
• •primary decision logic
• •persistence or side-effect boundary

Read deeply only in 2-4 critical files first.

Step 4: Control exploration depth

Depth limits:

• •initial scan: up to 50 files
• •deep dive: only files tied to hot path
• •stop broad scanning when repeated patterns emerge

If context remains unclear, list specific unknowns and continue focused search.

Step 5: Emit exploration outputs

ARCHITECTURE_MAP
- entrypoints:
  - <path>
- key_modules:
  - <module + role>
- data_flow:
  - "<A -> B -> C>"

UNKNOWNS
- "<unknown item>"
- "<unknown item>"

Heuristics

• •Prefer current implementation paths over legacy/dead code.
• •Use test files to infer expected behavior quickly.
• •Track naming conventions to infer boundary ownership.
• •Prioritize modules with high fan-in or fan-out.

Stop Conditions

• •Cannot identify real entrypoint after focused scan.
• •Request depends on generated/external code not present locally.
• •Tool-call budget exhausted without converging map.

Anti-Patterns (never)

• •Reading random files without hypothesis.
• •Deep-diving implementation before mapping boundaries.
• •Producing generic architecture summary not tied to paths.
• •Scanning entire repository for a local bugfix task.

Example

Input:

"Understand how verification state flows from tool results to final gate decision."

Expected flow:

1. •Locate extension entrypoints.
2. •Trace evidence ingestion path.
3. •Trace gate evaluation path.
4. •Return ARCHITECTURE_MAP and unresolved unknowns.