HydraFusion (policy max): router returns models outside the policy universe, then silently falls back to gpt-5.6-luna #5097

Describe the bug

With HydraFusion selected (policy max), most turns fail at the routing step and silently continue on a lower-tier model. The UI only shows:

HydraFusion routing failed; continuing with gpt-5.6-luna

The session log records the reason:

{
  "type": "session.fusion_route_failed",
  "syntheticModel": "hydrafusion",
  "policy": "max",
  "reason": "invalid_response",
  "errorMessage": "fusion routing returned an invalid response: fusion route references a model outside the selected policy universe: gpt-6.1-sol",
  "fallbackModel": "gpt-5.6-luna"
}

In one session over two days there were 10 routing failures against 3 successful routes:

Both rejected models are enabled for the account. The model list the runtime publishes to VS Code in the same session lists gpt-6.1-sol (and claude-sonnet-5.5) with "policyState": "enabled". The server-side router (routeSource: "capi_plan") therefore seems to work with a different model set than the runtime's "policy universe" for max. I couldn't find out how the runtime builds that set.

There are two problems here:

1. Router and runtime disagree on the allowed models. The router keeps choosing models the runtime then rejects, so most turns never run as HydraFusion.
2. The fallback is a silent capability downgrade. When the route is rejected, the turn runs on gpt-5.6-luna instead. A user who picked HydraFusion with the max policy for a demanding task (planning, in my case) ends up on a much cheaper model, with only a transient warning.

A related observation: the routes that did succeed under max also only used gpt-5.6-luna (primary/fallback/follow-up) and gpt-5.6-terra (secondary) with the Critique pattern. Higher-tier models such as claude-opus-5.5, which is enabled for the account, were never chosen. Those resolved routes also report "policyVersion": null and "modelUniverseVersion": null, which may be relevant to the mismatch.

Affected version

Copilot runtime 1.0.92-4.unstable.r37397886721.gad270aa (bundled @github/copilot-sdk-linux-x64 1.0.17-unstable.37397886721.gad270aa)

Steps to reproduce the behavior

1. In VS Code agent mode (Copilot SDK agent host), select HydraFusion as the model. The routing policy is max.
2. Send a normal multi-step coding/planning request.
3. The warning "HydraFusion routing failed; continuing with gpt-5.6-luna" appears for most turns.
4. The session's events.jsonl contains session.fusion_route_failed events with reason: "invalid_response" and "fusion route references a model outside the selected policy universe: ".

This happened consistently over two days. I haven't tested other policies.

Expected behavior

• The router only returns models that are inside the runtime's policy universe for the selected policy. Models the account has enabled (such as gpt-6.1-sol) should either be in the max universe or not be routed to.
• If a route is invalid anyway, the fallback should not silently downgrade capability. Either fall back to a model of comparable tier for the selected policy, or fail the turn with a clear, retryable error so the user can pick a model.
• Ideally, the UI makes it obvious that the turn did not run as HydraFusion.

Additional context

• VS Code 1.141.0, Windows client with a WSL (Ubuntu) remote; the agent host runs in WSL (Linux x86_64).
• GitHub Enterprise Cloud account with data residency (*.ghe.com host). I don't know whether the router's model set differs for those tenants.
• Possibly related: HydraFusion is offered in the picker to users whose runtime rejects every turn microsoft/vscode#339959 (HydraFusion offered although the runtime rejects it: a different symptom, but the same client/runtime model-policy mismatch) and For HydraFusion, please err on the side of higher intelligence OR provide an ... app#4569 (preference for higher-capability routing).
• I can share more session.fusion_* event fields if that helps.