Explore

You are exploring: {topic} — specifically the "{facet}" angle.

## Investigation goals
{2-3 specific questions this agent should answer}

## How to investigate
- Read source files, trace execution paths, check test coverage
- Look at imports/exports to understand dependency graphs
- Check for patterns, anti-patterns, inconsistencies
- Note anything surprising or potentially problematic

## Report format
Return a structured markdown report with:
1. **Summary** (2-3 sentences)
2. **Findings** (detailed, with file:line references and code snippets)
3. **Diagram** (mermaid diagram showing relationships/flow)
4. **Observations** (patterns, risks, opportunities, surprises)

Be thorough. Include file paths and line numbers for every claim.
Do NOT write any files — just return your findings as text.

.agents/explorations/YYYY-MM-DD/<topic-slug>/

# {Topic} — Exploration Overview

> Explored on YYYY-MM-DD

## TL;DR

{3-5 bullet points with the most important findings}

## Facets Explored

| Facet | Key Finding | Doc |
|-------|-------------|-----|
| {facet-1} | {one-liner} | [{facet-1}.exp.md](./{facet-1}.exp.md) |
| {facet-2} | {one-liner} | [{facet-2}.exp.md](./{facet-2}.exp.md) |
| ... | ... | ... |

## Cross-Cutting Themes

{Patterns and insights that emerged across multiple facets}

## Risks & Opportunities

{Things that look fragile, inconsistent, or ripe for improvement}

## Open Questions

{Questions that came up but weren't fully answered — threads to pull later}

# {Topic} — {Facet Name}

> Part of [{Topic} exploration](./overview.exp.md) | YYYY-MM-DD

## Summary

{2-3 sentence overview of this facet}

## Findings

### {Finding 1}

{Detailed explanation with code references}

```language
// file/path:line — relevant code snippet

{diagram showing relationships, data flow, or structure}

### 5. Present results

After writing all files, give the user a concise summary:
- Link to `overview.exp.md`
- Top 3 most important/surprising findings
- Any open questions or recommended next steps

## Guidelines

**Go deep, not wide.** Each agent should trace code paths to their endpoints, not just grep for keywords. Read implementations, understand control flow, check edge cases.

**Every claim needs evidence.** File paths, line numbers, code snippets. No hand-waving.

**Mermaid diagrams are mandatory** for each facet doc. They should illuminate structure or flow that's hard to see from code alone.

**Quote every edge label and compile every block.** Write `A -.->|"read(2)"| B`: unquoted labels fail
the lexer when they contain `(` or `|`, and quoting always parses. Compile with
`mmdc -i block.mmd -o /tmp/block.svg`, after
`export PUPPETEER_EXECUTABLE_PATH="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"` —
`mmdc` drives a headless browser and finds none by default, so without that variable it reports a
launch failure that reads like "mermaid is unavailable". A grep check cannot see a lexer error.

**Name files descriptively.** Use kebab-case names that describe the facet: `error-handling.exp.md`, `data-flow.exp.md`, `test-coverage.exp.md`, not `part-1.exp.md`.

**Date folders use today's date** in `YYYY-MM-DD` format. Each exploration gets its own topic subfolder, so multiple explorations on the same day never collide.

**Agent count:** Aim for 3-6 parallel agents. Fewer than 3 isn't deep enough. More than 6 risks shallow coverage per agent.