Run a Codex and Claude Code debate

Palabre can make Codex CLI and Claude Code discuss a technical decision before you move to implementation. Each agent receives the same subject and context, then answers the previous analysis.

Use this workflow to challenge an architecture, review a migration plan, compare two strategies, or uncover risks missed by the first proposal.

Prepare Codex and Claude Code

Install and authenticate both CLIs through their usual mechanisms. Palabre does not replace their accounts or subscriptions.

Then verify availability:

codex --version
claude --version
palabre config --sync-agents
palabre doctor

The following command should report the preset as available:

palabre presets --json

Start the debate

The codex-claude preset makes Codex speak first and Claude Code second:

palabre codex-claude "Compare these architectures and recommend the more robust option" -t 4

The reverse variant makes Claude Code speak first:

palabre claude-codex "Review this migration plan and resolve the disagreements" -t 4

Position controls speaking order. Roles still come from your configuration unless you override them for the session.

Assign complementary roles

A pair is more useful when the agents do not receive exactly the same mission. For example:

palabre run \
  --subject "Should this service be split in two?" \
  --agent-a codex \
  --agent-b claude \
  --role-a architect \
  --role-b critic \
  --turns 4

Codex then structures an architecture proposal while Claude Code looks for fragile assumptions and unaddressed consequences.

Add project context

Use --files for a strict and reproducible selection:

palabre codex-claude "Review this caching strategy" \
  --files src/cache.ts src/config.ts \
  -t 4

Use --context to scan a folder tolerantly:

palabre codex-claude "Propose an incremental migration" \
  --context src docs/architecture \
  -t 4

Palabre marks files as untrusted data inside prompts. Size limits and exclusions are documented in Context and files.

Read the summary

After the transcript, the summary agent organizes:

• consensus points;
• disagreements and uncertainty;
• proposed actions;
• a short conclusion.

The session is exported under .palabre/ as a .debate.md file. This export preserves the opposing arguments instead of keeping only the conclusion.

Choose Debate or Ask

Use Debate when agents should answer each other's objections. Use Ask when you first want two independent answers without mutual influence.

For a broader workflow overview, see Use multiple AI agents from the terminal and Debate mode.