Deep Research

Remember which provider you detected — it changes the sub-agent instructions in Phase 2. You can reuse this initial search result as part of Phase 2.

PHASE 1: PLAN

Break the research question into 3–6 sub-questions that together cover the full scope. Each sub-question should target a distinct angle:

• Core facts / definitions
• Current state / recent developments
• Competing viewpoints or alternatives
• Evidence quality / data sources
• Practical implications / applications
• Risks, limitations, or criticisms

Present the plan to the user:

I'll research "<topic>" by investigating these sub-questions:

1. <sub-question>
2. <sub-question>
3. <sub-question>
...

Mode: [Deep / Quick]

Shall I proceed, or would you like to adjust the sub-questions?

Wait for user confirmation before proceeding.

PHASE 2: RESEARCH

Deep Mode (parallel sub-agents)

For each sub-question, spawn a background sub-agent:

sessions_spawn background:true timeout:120

Give each sub-agent this prompt (fill in <sub-question>):

Research the following question thoroughly:
"<sub-question>"

Instructions:
1. Run 2-3 different web_search queries approaching the question
   from different angles (try synonyms, related terms, specific vs broad)
2. For each search:
   - Standard search: web_fetch the 3 most relevant results
   - Perplexity Sonar: DO NOT web_fetch — extract findings directly
     from the synthesized search response
3. Extract key findings with source attribution
4. Note when sources disagree or data is uncertain

Return your findings in this EXACT format:

## Findings
- [Finding 1] (Source: <url>)
- [Finding 2] (Source: <url>)
- [Finding 3] (Source: <url>)

## Sources
1. <title> — <url> — <brief description>
2. <title> — <url> — <brief description>

Be thorough. Prefer recent sources (last 1-2 years). Note disagreements.

After spawning, inform the user: "Researching N sub-questions in parallel..." Monitor sessions and collect results once finished.

Quick Mode (serial)

Do the research yourself sequentially:

For each sub-question:

1. Run 2 web_search queries from different angles
2. Standard search: web_fetch the top 2-3 results from each search Perplexity Sonar: Skip web_fetch — extract findings and source URLs directly from the synthesized search response
3. Record findings with source URLs
4. Move to the next sub-question

PHASE 3: SYNTHESIZE

3a. Collect & Parse

Gather all findings from sub-agents (or your own serial research). Parse each finding into: {claim, source_url, source_title, sub_question}.

3b. Deduplicate Citations

Build a global reference list. Rules:

• Same URL = same reference number everywhere in the report
• Assign [1], [2], [3]... in order of first appearance
• Strip tracking parameters from URLs before comparing (utm_source, utm_medium, fbclid, etc.)

3c. Cross-Reference

Across all sub-question results:

• Consensus: Claims supported by 2+ independent sources
• Contradictions: Claims where sources disagree — note both sides
• Gaps: Sub-questions with thin or no results
• Confidence: Rate each major finding (high / medium / low) based on source quality and agreement

PHASE 4: DELIVER

4a. Write the Report

Use this format exactly:

# <Research Topic>

**Date:** YYYY-MM-DD
**Status:** Complete | Partial (if gaps exist)
**Sub-questions researched:** N

---

## Executive Summary

<2-3 paragraphs synthesizing all key findings. Lead with the most important
conclusions. Mention confidence levels for major claims.>

## Detailed Findings

### <Sub-topic 1>
- Finding with inline citation [1]
- Finding with citation [2]
- <Note any disagreements between sources>

### <Sub-topic 2>
- Finding with citation [3]
- Finding with citation [1] (reuse if same source)

### <Sub-topic N>
...

## Cross-cutting Analysis

### Consensus
- <Claims well-supported across multiple sources>

### Contradictions
- <Where sources disagree, with citations for each side>

### Gaps & Limitations
- <What couldn't be determined, and why>

### Confidence Assessment
| Finding | Confidence | Basis |
|---------|-----------|-------|
| <key claim> | High/Med/Low | <why> |

## Recommendations / Next Steps
<If applicable: what to investigate further, actions to take>

## References
[1] <Title> — <URL>
[2] <Title> — <URL>
...

4b. Save to Workspace

write path:"research/<topic-slug>/research.md"

Where <topic-slug> is the topic in lowercase, spaces replaced with hyphens, special characters removed. Example: "AI Safety in 2026" → ai-safety-in-2026.

4c. Share via Messaging (optional)

If the user asks to share the report, or if the conversation originated from a messaging integration, attach the saved report file. Use whichever tool is available:

• Slack: slack action:"sendMessage" to:"<channel>" content:"Research report: <topic>" mediaUrl:"file://research/<topic-slug>/research.md"
• Discord: discord action:"sendMessage" to:"<channel>" content:"Research report: <topic>" mediaUrl:"file://research/<topic-slug>/research.md"
• iMessage: imsg send --to "<recipient>" --text "Research report: <topic>" --file research/<topic-slug>/research.md
• BlueBubbles: bluebubbles action:"sendAttachment" path:"research/<topic-slug>/research.md" caption:"Research report: <topic>"

Rules:

• Only share if the user requests it or the conversation context implies it (e.g., user asked from a Slack channel)
• If the messaging tool isn't available, skip silently — just deliver via the saved file and chat summary
• Include a brief caption with the topic name, not the full summary

Slack Threading

When the conversation originates from a Slack channel, use threading to keep the channel clean:

1. All intermediate messages go in a thread of the original user message. This includes: the research plan (Phase 1), progress updates ("Researching N sub-questions..."), sub-agent results, and the detailed findings. Use thread_ts set to the original message timestamp.

2. Only the final deliverables go to the main channel AND the thread. Specifically:

   • The executive summary text
   • The research file attachment

   To post in both the channel and the thread, send the executive summary and file to the channel (no thread_ts), and also post them in the thread (thread_ts set to the original message).

Example thread flow:

#channel
├─ User: "Research solid-state batteries for EVs"
│  ├─ [thread] Bot: "I'll investigate these 5 sub-questions... proceed?"
│  ├─ [thread] User: "Go ahead"
│  ├─ [thread] Bot: "Researching 5 sub-questions in parallel..."
│  ├─ [thread] Bot: "All sub-agents complete. Synthesizing..."
│  ├─ [thread] Bot: Executive summary + report file
│  └─ [thread] Bot: "Would you like me to go deeper on any section?"
├─ Bot: Executive summary + report file  ← also in main channel

4d. Present to User

Show the Executive Summary directly in chat, then:

Full report saved to: research/<topic-slug>/research.md

I found N sources across M sub-questions.

Would you like me to:
- Go deeper on any section?
- Research additional angles?
- Share the report to a channel or contact?
- Export in a different format?

ERROR HANDLING

• Sub-agent timeout: Note which sub-question has incomplete data. Include in "Gaps" section. Offer to retry.
• No search results: Try broader/alternative queries. If still nothing, note the gap — never fabricate sources or findings.
• web_fetch failures: Skip the URL, try the next result.
• All sub-agents fail: Fall back to quick mode automatically.

RULES

1. Every claim needs a citation. No unsourced assertions in findings.
2. Never invent URLs or sources. If you can't find it, say so.
3. Prefer recent sources — within the last 1-2 years when possible.
4. Be honest about uncertainty. "Low confidence" is better than false certainty.
5. Keep the executive summary readable — no jargon walls, no citation spam. Save detailed citations for the findings sections.
6. Respect the user's time. If quick mode is more appropriate, suggest it.
7. No external API keys required. Use only built-in tools: web_search, web_fetch, sessions_spawn, read, write.