fix(reasoning): request OpenAI reasoning summaries on the managed path

Managed gpt-5.x runs carry providerID "openai" (post-rebrand), but the
options() gate enabling reasoningSummary:"auto" + reasoning encrypted
content only fired for providerID "synsci". So the managed path never
asked OpenAI for a summary: the Responses API still emitted reasoning
*items* (reasoning-start/end fired, empty reasoning parts got created)
but zero reasoning_summary_text.delta events — every reasoning part
persisted with empty text and the UI showed a blank "thinking" block.

Detect the managed path by the Atlas proxy baseURL (.../api/llm/proxy/…)
in the provider options and request summaries there too. Genuine BYOK
(no proxy baseURL) is untouched, so an unverified org never gets an
unexpected summary request. Verified against the live proxy: a managed
Responses stream returns ~83-99 summary deltas with real reasoning text;
the transform now sets reasoningSummary/include for managed + legacy
synsci and nothing for BYOK.

Author: ishaan1124

backend/cli/src/provider/transform.ts

@@ -778,7 +778,16 @@ export namespace ProviderTransform {
         result["textVerbosity"] = "low"
       }
 
-      if (input.model.providerID.startsWith("synsci")) {
+      // Managed OpenAI models carry providerID "openai" (post-rebrand), not
+      // "synsci" — but they route through the Atlas proxy baseURL. Reasoning
+      // summaries + encrypted content have to be requested on that path too,
+      // otherwise gpt-5.x streams reasoning *items* (start/end fire) with zero
+      // summary deltas, so every reasoning part lands empty and the UI shows a
+      // blank "thinking" block. Genuine BYOK (no proxy baseURL) is left alone,
+      // so an unverified org never gets an unexpected summary request.
+      const managedBaseURL = input.providerOptions?.["baseURL"]
+      const viaManagedProxy = typeof managedBaseURL === "string" && managedBaseURL.includes("/api/llm/proxy/")
+      if (input.model.providerID.startsWith("synsci") || viaManagedProxy) {
         result["promptCacheKey"] = input.sessionID
         result["include"] = ["reasoning.encrypted_content"]
         result["reasoningSummary"] = "auto"