Skip to content

Roundhouse

Roundhouse runs your prompt across multiple LLM models at the same time. Each model plans independently — no model sees another’s output. You review, annotate, and critique between phases. A synthesizer then combines the best ideas into a single unified plan and drops it into chat.

The result is higher-quality plans that blend the genuine strengths of different models, with a human checkpoint before synthesis runs.

Type /roundhouse to open the model picker, select your planners with Tab, and press Enter to launch.

/roundhouse

You can mix any combination of providers — cloud and local. A primary model (your currently active model) handles synthesis.

Roundhouse runs as a full-screen experience inside the TUI:

  • Left panel (65%) — streams output from each model live as they plan
  • Right panel (35%) — phase navigator, model list with status icons, running cost, and annotation count

Press j/k to switch between models in the left panel while they’re running.

The original prompt is shown dim and italic below the Roundhouse header throughout all phases, so you always know what you asked.

All selected models receive the same prompt and plan independently. Status icons in the right panel show which models are still running (), completed (), or failed.

After planning completes, Roundhouse pauses for your review. At this gate you have four options:

KeyAction
cProceed to critique phase — models critique each other’s plans
sSkip to synthesis — go straight to the final plan
aAnnotate — add guidance before synthesis runs
qCancel roundhouse

If you chose c, each model reviews the others’ plans and identifies weaknesses, missing edge cases, or better alternatives. Another review gate follows.

Pressing a at any gate lets you type freeform guidance — for example, “Use Claude’s database approach but GPT-4o’s error handling”. Annotations are injected into the next phase’s prompt as a User Guidance section. Models are instructed to treat your guidance as higher priority than their own judgment.

Your primary model receives all plans (and any annotations) and synthesizes a single unified implementation plan. This is streamed live in the left panel.

On completion, the synthesized plan is:

  • Inserted as an Assistant message in chat for immediate follow-up
  • Saved to .caboose/roundhouse/<YYYY-MM-DD>-<slug>.md in your project

Press Esc or Ctrl+C at any time to cancel Roundhouse from any phase. Slash commands are disabled while Roundhouse is active.

  • Use Roundhouse for high-stakes architectural decisions where different models may have genuinely different strengths.
  • Annotations are the most powerful part — even a brief note like “keep it simple” or “prioritize backwards compatibility” significantly shapes the synthesis.
  • Roundhouse output files are gitignored by default.
  • For quick tasks, a single model is faster. Roundhouse is best when the planning quality matters more than speed.