Research

Present the proposed axes to the user. Let them adjust, add, or remove before proceeding.

2. Phase 1 — Broad survey

Launch one subagent per axis in parallel in a single message. Each agent receives:

• The topic description
• Their specific research angle
• Codebase context if relevant (explore the codebase first, or pass key context from the conversation)
• Instructions to return structured findings

Agent prompt template:

Research the following topic from the perspective of your assigned angle.

## Topic
<topic description>

## Your Research Angle
<specific angle for this agent>

## Codebase Context
<relevant context about the project, tech stack, constraints>

## Output Format

Return your findings in this structure:

### Key Facts
What you found, with specifics. Cite sources where possible.

### Options Identified
Viable approaches within your angle. For each option: brief description, pros, cons.

### Risks
Downsides, gotchas, known problems.

### Open Questions
Things you couldn't fully answer or that need deeper investigation.

### Cross-References
Flag connections to other research angles:
- "This finding likely affects [other angle] because..."
- "The [other angle] researcher should also consider..."

Be thorough but concise. Focus on findings that would inform a decision, not exhaustive background.

The cross-references instruction is key — it primes each researcher to flag connections to other axes even though they can't communicate directly. The coordinator uses these in the gap analysis.

3. Gap analysis

After all Phase 1 agents complete, read all results and identify:

• Contradictions: Agent A says library X supports feature Y; Agent B says it doesn't
• Gaps: Nobody investigated topic Z, but multiple agents flagged it in cross-references
• Cross-cutting findings: A's finding about constraint P changes the viability of B's recommendation Q
• Unresolved questions: Things agents explicitly flagged as needing more investigation

Present the gap analysis to the user. Ask: should we do a follow-up phase, or is this sufficient?

4. Phase 2 — Targeted follow-up (optional)

If the user wants deeper investigation:

• Launch targeted subagents for each gap or contradiction
• Each gets the full Phase 1 findings plus their specific investigation target
• Fewer agents, more focused than Phase 1
• These agents may use web search or documentation lookup for specific questions

After Phase 2 completes, present new findings to the user. Do NOT offer a Phase 3 — two rounds is the limit. If significant unknowns remain, note them in the synthesis as open risks.

5. Synthesis

Write a research document covering:

# Research: <Topic>

**Date**: <YYYY-MM-DD>
**Research axes**: <list of angles investigated>
**Phases**: <1 or 2>

## Landscape Summary

What we now understand about the problem space. Establish the context
and constraints that shape the available options.

## Options

### Option A: <Name>

<Description>

**Pros:**
- ...

**Cons:**
- ...

**Fit for our case:** <brief assessment of how well this fits the specific project>

### Option B: <Name>
...

## Recommendation

The strongest path forward, with reasoning. Reference specific findings
that support this recommendation. If no single option is clearly best,
say so and explain what would tip the decision.

## Risks and Unknowns

What we still don't know, and whether it matters for the decision.
Distinguish between:
- **Risks we can mitigate** — known problems with known solutions
- **Unknowns that may not matter** — gaps in our knowledge that are unlikely to affect the outcome
- **Unknowns that could change the decision** — things that, if answered differently, would alter the recommendation

## Sources and References

Links, documentation, libraries, tools, and other resources discovered
during research. Group by relevance.

6. Output

@../../../../conventions/work-items.md

Detect the available system and confirm with the user.

• Work item: Create with title Research: <Topic> and the synthesis document as the body
• Markdown fallback: Write to ./plans/<topic-slug>-research.md

After saving, inform the user:

Research complete: <path or work item reference>
Next step: /decaf-planning:grill-me or /decaf-planning:write-a-prd

Example Usage

/decaf-planning:research Should we add real-time sync to Vestige?
/decaf-planning:research Evaluate options for replacing our REST API with GraphQL
/decaf-planning:research Migration path from SQLite to PostgreSQL for our access patterns