Auto mode: use the router's chosen_model instead of candidate_models[0]

[aashna]

Auto mode was picking the wrong model. When resolving the auto pseudo-model, we took candidate_models[0] from the router response and ignored chosen_model entirely. chosen_model is the router's actual pick after any server-side re-ranking (the Cost Sorting experiment being the one that bit us). The integrator contract is explicit about this:

Use chosen_model for the upcoming chat call, and use candidate_models as the ordered fallback list.

— auto-intent-service integrator docs. The Copilot CLI already does this; VS Code was the odd one out.

This showed up as two separate-looking bugs that are really the same root cause:

1. We sent a different model than the router chose. In the report, /models/session/intent returned GPT-5.4 mini but the chat completion actually went to GPT-5.3 Codex.
2. The "Routed to …" label didn't match the response footer. The explainability label read the router's pick (GPT-5.4 mini) while the credits footer showed what actually served the request (GPT-5.3 Codex). Once selection reads chosen_model, the label, the request, and the footer all line up.

There's a knock-on effect too: because the Cost Sorting experiment only re-ranks chosen_model, anything reading candidate_models[0] was effectively invisible to it. That matches what we saw — the experiment moved the CLI but not VS Code.

What changed

In _tryRouterSelection (automodeService.ts):

• Resolve chosen_model against the client's known endpoints first.
• Only fall back to walking candidate_models when chosen_model is missing or we don't have an endpoint for it.
• Emit chosen_model as candidateModel in the automode.routerModelSelection telemetry, so the "router pick vs actual model" comparison is actually meaningful.

Why the tests didn't catch it

Every mocked router response in the spec set chosen_model === candidate_models[0], so reading the wrong field gave the right answer by accident. The new tests decouple them:

• should prefer chosen_model over candidate_models[0] — resolved endpoint follows chosen_model when the two differ.
• should surface chosen_model in the routing decision the UI displays — the decision the "Routed to" label consumes matches the served endpoint (guards bug gulp-symdest does not preserve links on electron #2).
• should emit candidateModel from chosen_model, not candidate_models[0] — pins the telemetry field.
• Reworked the old iterate all candidate_models test into should fall back to candidate_models when chosen_model has no endpoint so it actually exercises the fallback path.

I confirmed the new tests fail on the pre-fix code (resolving gpt-5.3-codex instead of gpt-5.4-mini) and pass after the fix.

Test plan

• vitest run automodeService.spec.ts — 37 passing
• Verified the decoupled tests fail on the old code, reproducing the reported GPT-5.4 mini → GPT-5.3 Codex mismatch

[Copilot]

Pull request overview

This PR fixes VS Code Copilot “Auto mode” routing to honor the router’s authoritative chosen_model (including post-ranking/experiment adjustments) instead of incorrectly using candidate_models[0], aligning behavior with the auto-intent-service integrator contract and the Copilot CLI.

Changes:

• Update router selection logic to resolve chosen_model against knownEndpoints, falling back to the ordered candidate_models list only when needed.
• Update routing telemetry (automode.routerModelSelection) so candidateModel reflects the router’s chosen_model.
• Add regression tests that decouple chosen_model from candidate_models[0] and validate both selection and telemetry.

Show a summary per file

File	Description
extensions/copilot/src/platform/endpoint/node/automodeService.ts	Prefer router chosen_model for endpoint resolution; fall back to candidate_models when chosen_model can’t be resolved; emit telemetry using the router’s authoritative pick.
extensions/copilot/src/platform/endpoint/node/test/automodeService.spec.ts	Add/adjust tests to ensure selection prefers chosen_model and telemetry reports candidateModel from chosen_model, plus a real fallback-path test.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Low