team-attention/hoyeon/deep-research

/deep-research <research question or topic>
/deep-research --auto <research question or topic>

# Session dir init
SESSION_ID="[CLAUDE_SESSION_ID from UserPromptSubmit hook]"
RESEARCH_DIR="$HOME/.hoyeon/$SESSION_ID/research"
mkdir -p "$RESEARCH_DIR"
echo "RESEARCH_DIR=$RESEARCH_DIR"
 
# Gemini check
command -v gemini && echo "GEMINI_AVAILABLE=true" || echo "GEMINI_AVAILABLE=false"
 
# Chromux check — resolve path literally, remember the output
CX=$(command -v chromux 2>/dev/null || echo "") && [ -n "$CX" ] && echo "CHROMUX=$CX" || (npx @team-attention/chromux help >/dev/null 2>&1 && echo "CHROMUX=npx @team-attention/chromux" || echo "CHROMUX=MISSING")

/path/to/chromux launch default --headless 2>/dev/null || true

cat > "$RESEARCH_DIR/plan.md" << 'PLAN_EOF'
# Research Plan
 
## Research Question
[original query]
 
## Complexity Tier
[tier] — [reason]
 
## Research Angles
[numbered list: 3-6 angles]
 
## Agent Assignments
[for each angle: agent type (WebSearch or Browser), objective, seed queries or target URLs]
 
## Gemini Status
[available/unavailable] — will research: [full query]
 
## Browser Agent Targets
[URLs/sites identified as needing browser-based extraction, with rationale]
 
## Expected Output Format
[report structure]
PLAN_EOF
echo "Plan saved to $RESEARCH_DIR/plan.md"

AGENT [N] (WebSearch): [Angle Name]
OBJECTIVE: [One clear sentence]
SEARCH TERRITORY: [What to investigate]
DO NOT OVERLAP WITH: [Other agents' territories]
SEED QUERIES:
  1. [Short, broad query - 2-3 words]
  2. [Medium specificity - 3-5 words]
  3. [Narrow/specific follow-up]
PREFERRED SOURCES: [docs/papers/news/blogs/github/forums]
OUTPUT FILE: $RESEARCH_DIR/agent-[N]-findings.md

BROWSER AGENT [N]: [Angle Name]
OBJECTIVE: [One clear sentence]
TARGET URLS: [Specific URLs to visit]
EXTRACT: [What specific data/content to extract from each URL]
OUTPUT FILE: $RESEARCH_DIR/browser-[N]-findings.md
RATIONALE: [Why browser is needed — e.g., "Reddit uses infinite scroll / JS-rendered comments"]

Bash(run_in_background=true):
  /absolute/path/to/script/gemini-research.sh "<full research query>" "$HOME/.hoyeon/SESSION_ID/research" 300

You are Research Agent [N], a focused investigator. ultrathink before each search.
 
ASSIGNMENT:
- Topic: [original research question]
- Your angle: [angle name]
- Objective: [specific objective]
- Search territory: [what to search]
- Stay away from: [other agents' territories]
- Preferred sources: [source types]
 
SEARCH STRATEGY — Start Wide, Then Narrow:
1. Begin with SHORT, BROAD queries (2-3 words). Evaluate what's available.
2. Based on initial results, form more specific follow-up queries.
3. Go deeper on the most promising leads.
4. Aim for [N] total search queries (per complexity tier).
 
For each search cycle:
- Use WebSearch with a focused query
- Evaluate the results. Ask: Is this relevant? Is this from a credible source?
  Does this add new information?
- For the best 2-3 results, use WebFetch to extract full content.
  Include a focused question about what to extract.
- Take detailed notes with EXACT source URLs for every claim.
 
CREDIBILITY RANKING:
- Tier 1 (HIGH): Official docs, peer-reviewed papers, government sites,
  primary sources (company blogs, SEC filings)
- Tier 2 (MEDIUM): Established media (Reuters, Bloomberg, TechCrunch),
  well-known technical blogs, conference talks
- Tier 3 (LOW): Personal blogs, forums, social media, SEO content farms
  -> Use Tier 3 only to corroborate Tier 1-2 findings, never as sole source.
 
WRITE YOUR FINDINGS to the file: [RESEARCH_DIR]/agent-[N]-findings.md
 
Use this exact structure:
 
# Agent [N]: [Angle Name]
## Search Queries Used
1. "[query]" -> [number] relevant results
2. ...
 
## Key Findings
### [Sub-topic A]
- [Factual claim] -- Source: [URL] (Credibility: HIGH/MED/LOW)
- [Factual claim] -- Source: [URL] (Credibility: HIGH/MED/LOW)
 
### [Sub-topic B]
...
 
## Source Registry
| # | URL | Title | Type | Credibility | Date |
|---|-----|-------|------|-------------|------|
| 1 | ... | ...   | docs | HIGH        | 2025 |
 
## Gaps & Uncertainties
- [What you couldn't find or verify]
- [Where sources contradicted each other]
 
## Unexpected Discoveries
- [Anything surprising or tangential but valuable]
 
RULES:
- NEVER fabricate sources or URLs. If you can't find it, say so.
- NEVER search for things outside your assigned territory.
- If you discover something critical outside your territory, note it
  in "Unexpected Discoveries" for the lead agent to handle.
- Prefer sources from the last 12 months unless historical context needed.
- For Korea-relevant topics, search in BOTH English and Korean.

You are a Browser Research Agent. Your task is to extract specific information from web
pages using a real Chrome browser via chromux.
 
## Chromux Setup
 
Resolve chromux: The chromux path is [INLINE LITERAL CHROMUX PATH — e.g., /usr/local/bin/chromux].
Chrome is already running in headless mode.
 
Generate your session ID:

Remember the output literally (e.g., ab12 → your session ID is exp-ab12). Inline both the
chromux path and session ID in every subsequent Bash call.
 
## Research Objective
 
Topic: [original research question]
Your angle: [angle name]
Objective: [one clear sentence — what information to find]
 
## Target URLs
 
Visit these URLs in order:
1. [URL 1] — Extract: [specific data points, e.g., "top comments discussing X", "version numbers", "benchmark table"]
2. [URL 2] — Extract: [specific data points]
[additional URLs if needed]
 
## Extraction Instructions