Add routing_method experiment variable for A/B/C hydra experiment

[aashna]

Summary

Add copilotchat.autoModeRoutingMethod ECS flight variable to control which routing algorithm the auto-intent-service uses per-request. This enables A/B/C experiments for the HYDRA model router.

A/B/C Experiment Design

Path	ECS Value	Behavior
A	'' (empty, default)	Production automod — no router call
B	'binary'	Binary classifier v1 (current router)
C	'hydra'	HYDRA multi-head capability matching (new)

Fallback for B/C → automod (path A), signaled via fallback=true in the response.

Changes

• configurationService.ts: New AutoModeRoutingMethod experiment-based setting
• routerDecisionFetcher.ts:
  • Add routingMethod param to getRouterDecision()
  • Send routing_method in request body when set by ECS
  • Update RouterDecisionResponse with new server fields: routing_method, fallback, fallback_reason, hydra_scores, chosen_model, chosen_shortfall
  • Add routing method + fallback info to trace log and telemetry
• automodeService.ts:
  • Read AutoModeRoutingMethod exp var in per-turn routing call
  • Handle fallback=true responses (fall back to automod default selection)

Related

• Server-side: github/auto-intent-service branch aashnagarg/per-request-routing
• Proxy tests: github/copilot-api PR aashnagarg/routing-method-tests

[Copilot]

Pull request overview

Adds richer Auto Mode model-routing behavior (including HYDRA routing metadata) and wires post-summarization re-routing so the system can switch models at a compaction boundary.

Changes:

• Adds per-conversation routing (route once on first turn, reuse thereafter) plus related cache/telemetry in AutomodeService.
• Extends router decision shape to include HYDRA/fallback metadata and introduces an experiment/config key for routing method selection.
• Wires an optional post-summarization callback through the agent prompt/summarizer flow to allow re-routing after a summary is produced.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
src/platform/endpoint/node/test/automodeService.spec.ts	Adds unit tests for per-conversation routing vs per-turn routing behavior.
src/platform/endpoint/node/routerDecisionFetcher.ts	Extends router response typing/logging and adds an (intended) routing method parameter.
src/platform/endpoint/node/automodeService.ts	Implements per-conversation routing cache, adds routing method plumb-through, and introduces post-summarization re-routing API + telemetry.
src/platform/configuration/common/configurationService.ts	Adds experiment-based config keys for per-conversation routing and routing-method selection.
src/extension/prompts/node/agent/summarizedConversationHistory.tsx	Adds optional post-summary callback hook (and an exclude flag) to trigger re-routing after summarization.
src/extension/prompts/node/agent/agentPrompt.tsx	Plumbs resolveNextEndpoint into summarized history rendering.
src/extension/intents/node/agentIntent.ts	Wires post-summarization callback to call AutomodeService.resolveEndpointForSummarization() and passes it to foreground/background compaction.
src/extension/intents/node/editCodeIntent2.ts	Updates constructor injection / super-call ordering to pass IAutomodeService.
src/extension/intents/node/askAgentIntent.ts	Updates constructor injection / super-call ordering to pass IAutomodeService.
src/extension/intents/node/notebookEditorIntent.ts	Updates constructor injection / super-call ordering to pass IAutomodeService.

Comments suppressed due to low confidence (1)

src/platform/endpoint/node/automodeService.ts:503

• resolveEndpointForSummarization updates the cached endpoint, but does not update/clear conversationRoutedModel. If per-conversation routing is enabled, the next turn will still reuse the original first-turn model and can immediately undo the post-summarization switch. Update the cache entry so subsequent turns follow the post-summarization decision (e.g., set conversationRoutedModel to the new model, or clear it to force re-routing).

			this._autoModelCache.set(conversationId, {
				...entry,
				endpoint: autoEndpoint,
				lastSessionToken: token.session_token,
			});