the docs say file edits can survive a discarded draft. flagging those edits separately in the final diff would save reviewers from reconstructing which attempt changed what.

Hydrafusion was my most interesting feature on CoPilot day. I'll try it but ... what's with that Ai voice?

Why is it picking models that I have disabled? For example Claude Opus 5, GPT-5.6 sol. Instead of cheaper and better Sonnet 5.5 or Opus 5.5 and GPT-6.1 Sol? (Same issue with 'auto' mode)

can the app show which model handled each stage? that trace would help separate a bad first draft from a critique step that failed to catch it.

Multi-model routing is useful when the path to the result stays visible. In Cascade and Critique, can users see which model drafted, why escalation occurred, and how the draft changed after critique? That would make outcomes easier to evaluate.

wait does it actually escalate or just loop the same two models

HydraFusion is basically model teamwork without the group chat. Route → critique → revise → one final answer. Pretty clever way to hide the orchestration complexity.

How does Critique mode bill? Drafter + critic + one revision. Is that 2 premium requests, 3, or one HydraFusion rate?

Multi-model routing is promising when each step has a clear role. It would be useful to evaluate the trade-offs in latency, cost, and revision quality, and show how escalation thresholds are calibrated on real tasks.

The underappreciated part is 'escalate' - tiering intelligence by task difficulty is the real product shift here. Curious whether the escalation signal comes from the critiquer model's confidence or from aggregated user edits over time.

isnt that ultra bad for getting cache hits which are model specific

Multi-model routing inside the editor is the part I've been waiting for — most of my Copilot time isn't "write more code," it's "this draft is almost right, now challenge it." Curious how often HydraFusion escalates vs sticking with the first model, and whether teams end up pinning routes per repo so juniors don't burn tokens on critique loops by accident.

I'd benchmark HydraFusion against the best single model on accepted changes per dollar, not just answer quality. A critic that catches one expensive bug can pay for the extra passes; three models agreeing on the same wrong assumption cannot.

Copilots integration with HydraFusion is a good sign for code summarization. The real test is whether it can handle codebases with weak or missing comments.

HydraFusion is an important step in runtime model orchestration. The assurance boundary is equally important: MULTI-MODEL REVIEW ≠ VERIFIED EFFECT Draft → critique → revise can improve solution quality. But it does not by itself establish authorization, execution, or the resulting system state. The evidence chain continues: RESULT → AUTHORIZATION → EXECUTION → OBSERVATION → VERIFIED EFFECT That distinction will become increasingly important as agentic systems move from generating answers to taking real-world actions.

Multi-model routing is where production AI actually lives. Our team ships client builds using different models for different tasks in the same workflow. HydraFusion making this native to Copilot is the right call.

The transparency and progress updates are the right seam for a multi-model workflow. A compact workflow receipt would make it more useful: which path ran, where escalation happened, and what the critic changed before the final answer.

how does it pick which model to use for a given task

Doesn't work with BYOK?

The interesting bit is the escalation policy, not just the model menu. I’d want traces showing why a task moved from draft to critique to human review.

One thing to watch: if HydraFusion discards a draft, file edits it already made aren't automatically undone. GitHub says to review those changes before committing.

router over models, that's a weekend

lovely pipeline, it'll escalate straight to the human anyway

Routing a task across draft, critique, and escalate models is useful only with a policy that picks the path. Log which model won, why it escalated, and who approved the final result.

wait does the routing burn more copilot quota than just picking one model?

Imma have to give this a try.

Interesting

Please I'm finding it difficult to receive TOTP on my registered phone number on GitHub.

I can see your acquisition by Microsoft worked as planned - you've really taken on Microsoft's testing and reliability - same old status page

HydraFusion now sits in the Copilot picker and routes a task across models to draft, critique, revise, or escalate into one result. I will not put a production change on that route until a named person owns the final accept.

Hydrafusion doesn’t support BOYK models.

Having models critique each other before handing you the answer is pretty slick

Models reviewing models is starting to feel pretty normal now.

A two-line fix probably doesn't need a whole panel discussion. Keeping simple edits quick would be just as useful as getting the hard ones right.