/deep-research
Launch deep research across Claude and Codex with cross-pollination
> /plugin marketplace add hamelsmu/research-council > /plugin install research-council@research-council
How it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/deep-research
Context preview
What this command does when you run it.
Launch deep research across Claude and Codex with cross-pollination
Command definition
deep-research.mddescription: "Launch deep research across Claude and Codex with cross-pollination"
argument-hint: "[--test] <research topic>"
allowed-tools:
- Bash
- Read
- Write
- Edit
- Glob
- Grep
- WebSearch
- WebFetch
Run the setup script to validate prerequisites and create the research workspace:
set -o noglob; bash "${CLAUDE_PLUGIN_ROOT}/scripts/setup-research.sh" $ARGUMENTS; _rc=$?; set +o noglob; (exit $_rc)If setup fails (non-zero exit code), help the user fix the issue (install missing CLIs, set API keys, etc.) and DO NOT proceed further.
If setup succeeds (exit code 0), read the state file to get the research details:
cat .claude/deep-research.local.md
Extract the research_id, models, and workspace path (`research/<research_id>/`) from the state file. Tell the user:
1. The research council is now active with 2 AI agents (Claude, Codex) 2. They can monitor progress with: `tail -f research/<research_id>/progress.log` 3. When all agents finish their research and cross-pollination refinement, you will synthesize the final report
Then **finish your response** — the Stop hook will automatically launch all research agents. You do not need to run any additional commands.
When you are given the synthesis prompt (after the research agents complete), read ALL refined reports carefully and write a comprehensive synthesis to the specified path. Structure it as:
1. **Executive Summary** — the most important findings across all investigations 2. **Key Findings** — organized by THEME (not by source agent), with the strongest evidence from all reports 3. **Areas of Consensus** — where both agents agree, with combined evidence 4. **Areas of Disagreement** — where agents differed, with analysis of why 5. **Novel Insights** — unique findings from the cross-pollination refinement round 6. **Open Questions** — what remains uncertain even after two independent investigations 7. **Sources** — comprehensive, deduplicated list from all reports 8. **Methodology** — brief note on the multi-agent research process used
Read more
description: "Launch deep research across Claude and Codex with cross-pollination" argument-hint: "[--test] <research topic>" allowed-tools: - Bash - Read - Write - Edit - Glob - Grep - WebSearch - WebFetch
Run the setup script to validate prerequisites and create the research workspace:
set -o noglob; bash "${CLAUDE_PLUGIN_ROOT}/scripts/setup-research.sh" $ARGUMENTS; _rc=$?; set +o noglob; (exit $_rc)If setup fails (non-zero exit code), help the user fix the issue (install missing CLIs, set API keys, etc.) and DO NOT proceed further.
If setup succeeds (exit code 0), read the state file to get the research details:
cat .claude/deep-research.local.md
Extract the research_id, models, and workspace path (`research/<research_id>/`) from the state file. Tell the user:
1. The research council is now active with 2 AI agents (Claude, Codex) 2. They can monitor progress with: `tail -f research/<research_id>/progress.log` 3. When all agents finish their research and cross-pollination refinement, you will synthesize the final report
Then **finish your response** — the Stop hook will automatically launch all research agents. You do not need to run any additional commands.
When you are given the synthesis prompt (after the research agents complete), read ALL refined reports carefully and write a comprehensive synthesis to the specified path. Structure it as:
1. **Executive Summary** — the most important findings across all investigations 2. **Key Findings** — organized by THEME (not by source agent), with the strongest evidence from all reports 3. **Areas of Consensus** — where both agents agree, with combined evidence 4. **Areas of Disagreement** — where agents differed, with analysis of why 5. **Novel Insights** — unique findings from the cross-pollination refinement round 6. **Open Questions** — what remains uncertain even after two independent investigations 7. **Sources** — comprehensive, deduplicated list from all reports 8. **Methodology** — brief note on the multi-agent research process used
A Claude Code plugin that orchestrates deep research across two coding agents — Claude Code and Codex CLI (OpenAI) — running in parallel with cross-pollination refinement and final synthesis.

