Research Team

research-team

When to Use

• An unknown bug has no match in L1 memory — need external knowledge
• A library has released a new version and APIs may have changed
• You need to compare multiple solutions before committing to one
• Triggered by: "research this", "find a fix for", "investigate", "what's the best approach for", "look up", "research team", "autonomous research"

Do NOT use for

• Known bugs already documented in L1 (core/memory/L1/) — read L1 first
• Simple API lookups that a single WebSearch handles
• Tasks requiring code execution — hand off to autonomous-patching-loop after research
• See ingest-repo for structured code-level repo analysis

Team Roles

Researcher-A  → searches StackOverflow + GitHub Issues for the exact error
Researcher-B  → reads official docs / changelog for the library version in use
Researcher-C  → searches for alternative approaches / known workarounds
Synthesizer   → cross-references all 3 findings, ranks solutions by reliability

Research Loop Architecture

# Conceptual: multi-agent research pipeline (adapt to Claude Code subagents)

from dataclasses import dataclass
from typing import Callable

@dataclass
class ResearchQuery:
    topic: str
    context: str        # error message / file / line number
    library: str        # e.g. "react@18.3.0"
    max_sources: int = 10

def research_team(query: ResearchQuery, search_fn: Callable) -> dict:
    """
    Parallel research across 3 lanes, then synthesize.
    """
    # Lane A: Stack-style QA
    lane_a = search_fn(f"{query.topic} site:stackoverflow.com OR site:github.com/issues")

    # Lane B: Official docs + changelog
    lane_b = search_fn(f"{query.library} changelog deprecation {query.topic}")

    # Lane C: Alternative approaches
    lane_c = search_fn(f"alternative to {query.topic} {query.library} workaround")

    # Synthesizer: rank by recency + vote count + library version match
    all_results = lane_a + lane_b + lane_c
    ranked = sorted(
        all_results,
        key=lambda r: (r.get("votes", 0) + r.get("recency_score", 0)),
        reverse=True
    )[:query.max_sources]

    return {
        "query": query.topic,
        "sources": len(ranked),
        "top_solution": ranked[0] if ranked else None,
        "alternatives": ranked[1:3],
        "confidence": "high" if len(ranked) >= 5 else "medium",
    }

Claude Code Subagent Pattern

# ORCHESTRATOR prompt to spawn research-team
Task: Bug in [library]@[version]: [error message]

Spawn 3 parallel research agents:

Agent-A: Search StackOverflow and GitHub Issues for exact error.
  Query: "[error message] [library]"
  Output: top 3 matches with URL + date + vote count → write to .claude/signals/research-a.json

Agent-B: Read official docs/changelog for [library]@[version].
  Focus: breaking changes, deprecated APIs, migration guides.
  Output: relevant section → write to .claude/signals/research-b.json

Agent-C: Find alternative approaches for [topic] that avoid the error.
  Output: top 2 alternatives → write to .claude/signals/research-c.json

Synthesizer (after A+B+C complete):
  Read all 3 signal files.
  Rank solutions by: recency > vote count > version match.
  Write final report to .claude/signals/research-report.md
  Promote confirmed fix to L1: bash core/scripts/add-fact.sh "bug-fix" "<solution>" "high"

Output Report Format

# Research Report — [topic]
Date: [timestamp]  Sources scanned: [n]  Confidence: high|medium|low

## Recommended Fix
[Solution text + code snippet]
Source: [URL] ([date], [votes] votes)

## Why This Works
[1-2 sentence explanation tied to the library's behavior]

## Alternative Approaches
1. [Alt A] — [tradeoff]
2. [Alt B] — [tradeoff]

## Outdated / Rejected Solutions
- [Stack answer from 2020] — library API changed in v18, no longer applicable

## L1 Promotion
[ ] Promoted to core/memory/L1/ — bash core/scripts/add-fact.sh "bug-fix" "..." "high"

Integration with autonomous-patching-loop

research-team  →  finds solution
       │
       ▼
autonomous-patching-loop  →  applies fix on isolated branch
       │
       ▼
verify gate passes  →  merge + L1 promotion

Anti-Fake-Pass Checklist

• At least 3 sources consulted (not just 1 web search)
• Results filtered for library version match — old answers flagged as "may be outdated"
• Confidence level stated: high (≥5 sources), medium (3–4), low (<3)
• Rejected/outdated solutions documented — not silently dropped
• Confirmed fix promoted to L1 via add-fact.sh
• Report written to .claude/signals/research-report.md — not left in chat only
• Agent does NOT apply the fix directly — hands off to autonomous-patching-loop