feat(session): expose LLM response headers on assistant messages

packages/core/src/v1/session.ts

@@ -480,6 +480,10 @@ export const Assistant = Schema.Struct({
   structured: Schema.optional(Schema.Any),
   variant: Schema.optional(Schema.String),
   finish: Schema.optional(Schema.String),
+  // HTTP response headers from the LLM provider. Used by plugins (e.g. LiteLLM
+  // proxy) to read routing metadata such as x-litellm-model-api-base that
+  // identifies the actual model selected by an auto-router behind the proxy.
+  responseHeaders: Schema.optional(Schema.Record(Schema.String, Schema.String)),
 }).annotate({ identifier: "AssistantMessage" })
 export type Assistant = Omit<Types.DeepMutable<Schema.Schema.Type<typeof Assistant>>, "error"> & {
   error?: AssistantError

packages/llm/src/schema/events.ts

@@ -185,6 +185,10 @@ export const StepFinish = Schema.Struct({
   reason: FinishReason,
   usage: Schema.optional(Usage),
   providerMetadata: Schema.optional(ProviderMetadata),
+  // HTTP response headers from the LLM provider. Used by plugins (e.g. LiteLLM
+  // proxy) to read routing metadata such as x-litellm-model-api-base that
+  // identifies the actual model selected by an auto-router behind the proxy.
+  responseHeaders: Schema.optional(Schema.Record(Schema.String, Schema.String)),
 }).annotate({ identifier: "LLM.Event.StepFinish" })
 export type StepFinish = Schema.Schema.Type<typeof StepFinish>
 

packages/opencode/src/session/llm/ai-sdk.ts

@@ -104,6 +104,10 @@ export function toLLMEvents(
             reason: finishReason(event.finishReason),
             usage: usage(event.usage),
             providerMetadata: metadata,
+            // Capture HTTP response headers so plugins can read provider-specific
+            // routing metadata (e.g. x-litellm-model-api-base from a LiteLLM
+            // auto-router) that is only available in the response headers.
+            responseHeaders: event.response?.headers as Record<string, string> | undefined,
           }),
         ]
       })

packages/opencode/src/session/processor.ts

@@ -709,6 +709,10 @@ export const layer = Layer.effect(
             ctx.assistantMessage.finish = value.reason
             ctx.assistantMessage.cost += usage.cost
             ctx.assistantMessage.tokens = usage.tokens
+            // Store provider response headers so plugins can read routing
+            // metadata (e.g. x-litellm-model-api-base from a LiteLLM proxy).
+            if (value.responseHeaders && Object.keys(value.responseHeaders).length > 0)
+              ctx.assistantMessage.responseHeaders = value.responseHeaders
             yield* session.updatePart({
               id: PartID.ascending(),
               reason: value.reason,