Debate mode

Debate is Palabre's primary mode. Two agents answer each other on one subject with the context and previous responses.

From the TUI

palabre

Then:

/debate
/agents codex claude
/roles architect critic
Review this plan and resolve the important disagreements

Each response receives the subject, context, injected files, and history. The final summary organizes consensus, disagreements, proposed actions, and a short conclusion.

The response count is between 1 and 20. It is the total, not the count per agent. Palabre may stop after a complete round when explicit agreement is detected.

Direct commands

palabre codex-claude "Review this plan" -t 4
palabre run --subject "Review this plan" --agent-a codex --agent-b claude --turns 4

Add --no-early-stop, --no-summary, or --summary-agent <agent> as needed. --dry-run previews the session without calling an agent. The export uses .debate.md.

Roles, context and failures

/roles architect critic assigns temporary roles in agent order: a structured proposal, then a search for flaws. Roles: implementer, reviewer, architect, scout, critic, summarizer. --role-a and --role-b do not modify configuration; palabre agent-role <agent> <role> sets a persistent role.

The default limit is 4 responses, configurable with palabre config -t <n>; early stopping may produce fewer. To request the whole limit, use --no-early-stop. A positional subject must have several words; for one word, use -s "word".

palabre codex-claude "Review this plan" --context src docs -t 4

See Context and files. On agent failure (quota, timeout, empty reply), PALABRE attempts to preserve a partial export with an interruption section. With --checkpoint, palabre resume <session-id> resumes the next missing response or summary, without replaying an accepted answer. No automatic retry.

After Debate, Chat can continue from the subject and summary. To consult a conversation that already knows the project, use Relay.