How to Enable and Use GitHub HydraFusion in Copilot CLI

01

Before you start

You need GitHub Copilot access and GitHub Copilot CLI. GitHub's September 4 launch announcement says the HydraFusion research preview is available on all Copilot plans through the CLI experimental path.

Because this is experimental functionality, update the CLI and recheck the current model picker before assuming launch-day commands or labels will remain unchanged.

02

1. Enable HydraFusion in Copilot CLI

1. Run <code>/update</code> to install the latest Copilot CLI version.
2. Run <code>/experimental on</code>.
3. Run <code>/model</code>.
4. Select <code>HydraFusion (Research Preview)</code>.

You do not install HydraFusion as a separate package or model. Once selected, you continue using Copilot CLI normally while HydraFusion decides which internal workflow to run.

03

2. Understand what happens after selection

Single sends the task to one model. Cascade starts with an efficient model and may escalate to a stronger model. Critique has one model draft, an isolated model family review it without tools, and the drafting model revise once.

HydraFusion chooses the route automatically. A longer response time can therefore mean additional review or escalation rather than a stalled CLI session.

04

3. Start with focused repository tasks

• a focused bug fix
• a code-generation task with clear acceptance criteria
• debugging a failing test
• a repository-level change that benefits from independent review

GitHub currently recommends substantial, well-scoped first-turn tasks for the preview. Avoid judging the router from one prompt because different tasks can take different execution paths.

05

4. Track quality, latency and usage together

GitHub bills HydraFusion according to tokens consumed by the underlying models at their standard rates. Compound workflows can invoke more than one model, so evaluate total usage rather than assuming one visible request equals one inference call.

A practical comparison is to run similar tasks with HydraFusion and a manually chosen baseline, then record correctness, manual intervention, latency and usage. Benchmark savings are useful launch evidence but are not a guarantee for your codebase.

06

5. If HydraFusion is not showing

1. Run <code>/update</code> again.
2. Confirm <code>/experimental on</code> is enabled.
3. Reopen <code>/model</code>.
4. Verify you are in GitHub Copilot CLI rather than another Copilot surface.
5. Check whether GitHub has changed the preview since the launch announcement.

Research-preview interfaces can move quickly. Missing menu entries should be treated first as a version or rollout question, not as proof that the product has been discontinued.

07

6. Keep normal engineering controls

HydraFusion's critic is isolated and tool-less, and GitHub describes fail-safe patch application when a workflow is cancelled or fails validation. Those controls do not replace code review or repository safeguards.