sherlock-plugins

Sherlock V2

Deep research agent for Claude Code. Decomposes complex questions into a dependency graph, researches them in parallel, validates every finding, and produces cited reports.

Install

# Add the marketplace
/plugin marketplace add trilogy-group/cc-skills-sherlock

# Install the plugin
/plugin install sherlock@sherlock-plugins

Or for local development:

claude --plugin-dir /path/to/cc-skills-sherlock

Usage

# Start a new research session
/sherlock "What are the best neighborhoods in Austin for families under $600k?"

# Resume a previous session
/sherlock --resume

# List all sessions
/sherlock --list

# Regenerate report from existing research
/sherlock --report <session-id>

# Re-export updated report to Google Docs/Sheets
/sherlock --update-export <session-id>

How It Works

1. REFINE — 2-3 clarifying questions to sharpen your goal
2. PLAN — Decomposes into a beads dependency graph with real dependencies (bd dep add)
3. EXECUTE — Dispatches parallel researcher subagents (batch size from config)
4. VERIFY — Creates validation beads for every claim, dispatches validators in parallel
5. REPORT — Synthesizes into a cited markdown report + CSV with full provenance

What's New in V2.1

• Config-driven execution — ~/.sherlock/config.yaml controls batch size, model, bead budget, depth limit, and validation mode. No more hardcoded values.
• Real dependency graph — Beads are wired with bd dep add, bd ready finds dispatchable work. Parent beads close when children resolve.
• Full validation mode — Every claim gets a validation bead, dispatched in parallel. Not just spot-checks.
• Structured JSON findings — Researchers write JSON, not freeform text. Conductor parses reliably.
• Retry on source failure — Researchers try alternative sources when URLs return 403/404.
• Cross-source contradiction detection — Conflicting data is flagged and surfaced in the report.
• Incremental CSV — Data rows written after each batch, not just at the end.
• Bead-level provenance — Every CSV row and report claim traces to a specific bead ID.
• Permission cleanup — Stale domain-specific WebFetch(domain:...) entries are removed automatically.
• Project-scoped permissions — Permissions live in .claude/settings.local.json per project.
• Google Workspace update-export — --update-export pushes corrected reports without recreating.
• Cost warning for Opus researchers — Config warns before running expensive researcher models.

Monitoring & Steering Research

Reviewing beads/tasks in progress

• Progress display — After each batch, shows completed beads (with source domains and bead IDs), active researchers, and queued beads (from bd ready)
• Press t in Claude Code to toggle the task list view
• Type "summary" mid-session for a synthesis of all findings so far
• Type "threads" to see research threads with completion percentages

Modifying or refining beads mid-session

• "focus on X" — Create new beads for X, deprioritize others
• "stop looking at Y" — Close Y-related beads
• "also check Z" — Add new research beads (within budget)
• "pause" / "report" / "quit" — Control flow

What Makes It Different

• Parallel research via beads dependency graph — configurable batch size (default 4)
• Every claim validated — full validation creates verification beads for all findings
• Structured provenance — every CSV row and report claim traces to a bead ID
• Cross-source contradiction detection — conflicting data is flagged, not hidden
• Trust Summary in every report — honest accounting of validation results
• Resumable sessions — walk away and come back days later
• Steerable — redirect research mid-session
• Config-driven — batch size, model, budget, validation mode all configurable

Prerequisites

• beads CLI (brew install beads) — auto-installed on first run, verified working
• Claude Code with Opus model (conductor) and configurable researcher model
• gogcli (brew install gogcli) — optional, enables Google Workspace export

Additional Tools

gogcli — Google Workspace Integration

gogcli is an optional CLI that connects Sherlock to Google Workspace. When configured, Sherlock can export finished reports directly to Google Docs and structured data to Google Sheets — no copy-pasting or file uploads needed.