explore

name	explore
description	Deep codebase exploration to answer specific questions
model	haiku
argument-hint	<question>

YOU ARE EXECUTING THE /explore SKILL. The user triggered this skill. Follow ALL instructions below step by step. Do NOT treat this as a freeform conversation - execute the skill workflow.

Read-only exploration task.

Ultra Think Strategy

Ultra think before answering:

• After exploration results: reflect on completeness, identify gaps
• Before answering: consider multiple interpretations, ensure accuracy
• Cross-reference findings from different sources

---

1. PARSE QUESTION

Extract from $ARGUMENTS:

• Key terms and concepts to search
• Whether backend, frontend, or both
• Whether external documentation is needed

---

1.5 CLARIFY IF AMBIGUOUS

Use AskUserQuestion if the question could have multiple interpretations:

• "How does X work?" -> Which layer? What aspect?
• "Where is X handled?" -> At which level?
• "Find X" -> Find what exactly? (usage/definition/all)

If question is clear and specific, skip to section 2.

---

2. EXPLORE (PARALLEL)

Launch focused agents in a single message (parallel execution). There is no upper limit on agent count: one agent per distinct angle, scaling up with the question's breadth. The lists below are a starting menu, not a quota.

Agent-count principle

• You decide the count from the actual scope. Assess how many distinct concerns / areas / file-clusters the question really spans, and launch one focused agent for each. Do not anchor to a fixed number or to file counts.
• No ceiling, real floor. Use at least 2 agents per domain the question touches (1 only for a genuinely trivial single-file lookup). Above that floor there is no maximum.
• One distinct focus per agent. Split by angle so two agents never run the same search.
• No batching. Launch them all in one message.

Backend (one per distinct angle, no maximum):

1. explore-codebase: "Find [keywords] definitions and core logic in backend/internal/domain/ and backend/internal/application/"
2. explore-codebase: "Find [keywords] implementations and usage in backend/internal/infrastructure/ and backend/internal/presentation/"
3. explore-codebase: "Find similar patterns to [keywords] in backend/ for comparison"

Frontend (one per distinct angle, no maximum):

1. explore-codebase: "Find [keywords] components and hooks in frontend/src/components/ and frontend/src/hooks/"
2. explore-codebase: "Find [keywords] types, API calls, and pages in frontend/src/types/, frontend/src/lib/, and frontend/src/app/"
3. explore-codebase: "Find similar patterns to [keywords] in frontend/src/ for comparison"

Supporting (each fans out the same way, one per distinct concern, no maximum):

• Database: explore-db - "[env] - [tables/schema question]" (one per table cluster / env)
• Library: explore-docs - "[library] [feature]" (one per library / feature)
• External: websearch - "[topic]" (one per topic)

For simple single-file lookups, 1 agent per domain is sufficient.

---

2.5 POST-EXPLORATION CHECK

After agents return, verify coverage:

1. Full code path traced? Can I trace the complete flow? If gaps -> launch targeted explore-codebase
2. Similar patterns identified? Do I have a reference implementation? If not -> launch explore-codebase
3. Data model complete? If not -> launch additional parallel explore-db agents
4. Library docs sufficient? If not -> launch additional parallel explore-docs agents

---

3. ANSWER

Provide comprehensive response:

## Answer

[Direct answer to the question]

## Evidence

### Code Found
- `path/to/file.ts:XX` - [description]
- `path/to/file.go:YY` - [description]

### Patterns Identified
- [Pattern 1]: [explanation]

### Documentation (if explore-docs used)
- [Library]: [relevant info]

### External Sources (if websearch used)
- [Title](URL) - [what we learned]

## Recommendations (if applicable)
- [Suggestion based on findings]

---

Rules

• PARALLEL - launch relevant agents in ONE message
• CITE - always reference file:line
• READ-ONLY - do not modify anything
• THOROUGH - gather complete context before answering