fix: complete span telemetry on shim and exhaustion terminals

The prompt-shim early return and the maxSteps-exhausted return now attach the same llm.usage.* and gen_ai.* root-span attribute pairs as the other generateText terminals; streamText's teardown block already covers its own paths. The usage-only stream sentinel reads service_tier through the typed StreamChunkExtras surface instead of inline casts. Refresh the agency example model.

Author: jddunn

src/api/generateText.ts

@@ -1657,6 +1657,13 @@ export async function generateText(opts: GenerateTextOptions): Promise<GenerateT
           totalTokens: (totalUsage.totalTokens ?? 0) + loopResult.totalTokens,
         };
         metricUsage = shimUsage;
+        // Dual-emit the same root-span attribute pairs as the native
+        // terminals below — the shim early-return is a first-class chat
+        // terminal, not a bypass (spec batch-1 residual).
+        span?.setAttribute('agentos.api.finish_reason', loopResult.finishReason);
+        span?.setAttribute('agentos.api.tool_calls', loopResult.toolCalls.length);
+        attachUsageAttributes(span, shimUsage);
+        attachGenAiAttributes(span, { providerName: resolved.providerId, operationName: 'chat', requestModel: resolved.modelId, responseModel: lastResponseModelId, usage: shimUsage, ...(opts.serviceTier !== undefined ? { requestServiceTier: opts.serviceTier } : {}), ...(lastServiceTier !== undefined ? { responseServiceTier: lastServiceTier } : {}) });
         fireLlmUsageObserver({
           provider: resolved.providerId,
           model: resolved.modelId,
@@ -2081,6 +2088,7 @@ export async function generateText(opts: GenerateTextOptions): Promise<GenerateT
       span?.setAttribute('agentos.api.finish_reason', 'tool-calls');
       span?.setAttribute('agentos.api.tool_calls', allToolCalls.length);
       attachUsageAttributes(span, totalUsage);
+      attachGenAiAttributes(span, { providerName: resolved.providerId, operationName: 'chat', requestModel: resolved.modelId, responseModel: lastResponseModelId, usage: totalUsage, ...(opts.serviceTier !== undefined ? { requestServiceTier: opts.serviceTier } : {}), ...(lastServiceTier !== undefined ? { responseServiceTier: lastServiceTier } : {}) });
       fireLlmUsageObserver({
         provider: resolved.providerId,
         model: resolved.modelId,

src/api/hitl.ts

@@ -13,7 +13,7 @@
  *
  * // Auto-approve everything (useful in tests and CI environments)
  * const testAgency = agency({
- *   agents: { worker: { provider: 'openai', model: 'gpt-4o-mini' } },
+ *   agents: { worker: { provider: 'openai', model: 'gpt-5.5' } },
  *   hitl: {
  *     approvals: { beforeTool: ['delete-file'] },
  *     handler: hitl.autoApprove(),

src/core/llm/providers/implementations/OpenAIProvider.ts

@@ -115,7 +115,7 @@ namespace OpenAIAPITypes {
     finish_reason: string | null;
     logprobs?: unknown;
   }
-  export interface ChatCompletionStreamResponse {
+  export interface ChatCompletionStreamResponse extends StreamChunkExtras {
     id: string;
     object: string;
     created: number;
@@ -1252,8 +1252,8 @@ export class OpenAIProvider implements IProvider {
                 choices: [],
                 isFinal: true,
                 usage: usageOnlyUsage,
-                ...(typeof (apiChunk as { service_tier?: string }).service_tier === 'string'
-                  ? { serviceTier: (apiChunk as { service_tier?: string }).service_tier }
+                ...(typeof apiChunk.service_tier === 'string'
+                  ? { serviceTier: apiChunk.service_tier }
                   : {}),
             };
         }