research-scout

name	research-scout
description	Read bookmarks from .claude/bookmarks.json, fetch each URL, analyze content, and propose orchestrator improvements.
disable-model-invocation	false
allowed-tools	WebFetch, WebSearch, Read, Write, Edit, Bash
user-invocable	true

Research Scout — Bookmark Analysis for Orchestrator Improvements

Read URLs from .claude/bookmarks.json, fetch their content, analyze it, and propose concrete improvements to the agent orchestrator.

Phase 1: Load Bookmarks & State

1. Read .claude/bookmarks.json to get the list of URLs
2. Read .claude/research-scout-state.json to see which URLs have already been processed
3. Filter out:
   • Already processed URLs (present in state file)
   • URLs with added date older than 7 days
4. Report how many new URLs to process

Phase 2: Fetch & Analyze Each URL

For each unprocessed URL (max 10 per run):

1. Fetch the content using WebFetch:

   • For X/Twitter URLs (x.com, twitter.com): convert to api.vxtwitter.com first
   • For all other URLs: fetch directly
   • If fetch fails, try WebSearch with the URL as query

2. Analyze the content looking for ideas that could improve:

   • Memory — state persistence, context optimization, cross-session knowledge
   • Router — task routing algorithms, cost optimization, complexity heuristics
   • Agents — new roles, better prompting patterns, coordination improvements
   • Skills — new tool ideas, workflow patterns, error recovery
   • Tools — API integrations, plugins, MCP extensions

3. Score relevance (0-1):

   • Applicable: relevant to AI orchestration? (> 0.5 required)
   • Novel: adds something we don't already have?
   • Actionable: can be implemented in our codebase?

4. Propose improvements with:

   • Target component (memory/router/agents/skills/tools)
   • Title and description
   • Files to modify
   • Expected benefit

Phase 3: Update State

After processing all URLs, update .claude/research-scout-state.json:

# Read current state
cat .claude/research-scout-state.json

Then use Edit or Write to add each processed URL to the processed dict:

{
  "processed": {
    "https://example.com/article": {
      "processed_at": "2026-03-08T12:00:00Z",
      "summary": "Article about multi-agent routing patterns",
      "improvements": ["Adaptive routing based on task history"]
    }
  },
  "last_run": "2026-03-08T12:00:00Z"
}

Phase 4: Report & PR

Print a summary report:

RESEARCH SCOUT REPORT
=====================
Bookmarks:  X total, Y new, Z skipped (old/processed)
Processed:  N URLs

Improvements Found:
  [memory]  Title — description (files: ...)
  [router]  Title — description (files: ...)
  [agents]  Title — description (files: ...)

State: updated .claude/research-scout-state.json

If improvements were found, create a PR:

1. Create a branch named research-scout/YYYY-MM-DD
2. Commit the updated state file and findings
3. Create a PR with the findings as the body:

BRANCH="research-scout/$(date +%Y-%m-%d)"
git checkout -b "$BRANCH"
git add .claude/research-scout-state.json .claude/bookmarks.json .claude/research-scout-findings.md
git commit -m "research-scout: findings from $(date +%Y-%m-%d)"
git push -u origin "$BRANCH"
gh pr create --title "research-scout: findings $(date +%Y-%m-%d)" --body "$(cat .claude/research-scout-findings.md)"

If no improvements were found, just commit state updates to the current branch.

Rules

• Max 10 URLs per run — quality over quantity
• Max 5 improvements per URL — only the most relevant
• Skip irrelevant content — if a URL has nothing useful for the orchestrator, mark it as processed with empty improvements and move on
• Don't modify code — only propose improvements, don't implement them
• Always create a PR when findings exist — never push findings directly to main
• All output in English