[codex] fix responses-to-chat tool stream fidelity (#378)

* fix responses chat tool stream fidelity

* fix responses tool call follow-up review comments

src/server/transformers/openai/chat/helpers.ts

@@ -18,6 +18,18 @@ function asRecord(value: unknown): Record<string, unknown> | null {
   return isRecord(value) ? value : null;
 }
 
+function cloneJson<T>(value: T): T {
+  if (Array.isArray(value)) {
+    return value.map((item) => cloneJson(item)) as T;
+  }
+  if (isRecord(value)) {
+    return Object.fromEntries(
+      Object.entries(value).map(([key, item]) => [key, cloneJson(item)]),
+    ) as T;
+  }
+  return value;
+}
+
 function toStringArray(value: unknown): string[] {
   if (!Array.isArray(value)) return [];
   return value
@@ -375,11 +387,19 @@ function extractCitationsFromPayload(
 }
 
 export function extractChatUsageDetails(payload: Record<string, unknown>): OpenAiChatUsageDetails | undefined {
-  const usage = asRecord(payload.usage);
+  const usage = asRecord(payload.usage) ?? asRecord(asRecord(payload.response)?.usage);
   if (!usage) return undefined;
 
-  const promptDetails = toNumericRecord(usage.prompt_tokens_details ?? usage.promptTokensDetails);
-  const completionDetails = toNumericRecord(usage.completion_tokens_details ?? usage.completionTokensDetails);
+  const promptDetails = toNumericRecord(
+    usage.prompt_tokens_details
+    ?? usage.promptTokensDetails
+    ?? usage.input_tokens_details,
+  );
+  const completionDetails = toNumericRecord(
+    usage.completion_tokens_details
+    ?? usage.completionTokensDetails
+    ?? usage.output_tokens_details,
+  );
   if (!promptDetails && !completionDetails) return undefined;
 
   return {
@@ -416,9 +436,25 @@ export function mergeChatUsageDetails(
 }
 
 export function extractUsagePayload(payload: Record<string, unknown>): Record<string, unknown> | undefined {
-  const usage = asRecord(payload.usage);
+  const usage = asRecord(payload.usage) ?? asRecord(asRecord(payload.response)?.usage);
   if (!usage) return undefined;
-  return usage;
+  const mappedUsage = cloneJson(usage);
+
+  const promptTokens = mappedUsage.prompt_tokens ?? mappedUsage.input_tokens;
+  const completionTokens = mappedUsage.completion_tokens ?? mappedUsage.output_tokens;
+  if (promptTokens !== undefined) mappedUsage.prompt_tokens = promptTokens;
+  if (completionTokens !== undefined) mappedUsage.completion_tokens = completionTokens;
+  delete mappedUsage.input_tokens;
+  delete mappedUsage.output_tokens;
+
+  const promptDetails = mappedUsage.prompt_tokens_details ?? mappedUsage.input_tokens_details;
+  const completionDetails = mappedUsage.completion_tokens_details ?? mappedUsage.output_tokens_details;
+  if (promptDetails !== undefined) mappedUsage.prompt_tokens_details = cloneJson(promptDetails);
+  if (completionDetails !== undefined) mappedUsage.completion_tokens_details = cloneJson(completionDetails);
+  delete mappedUsage.input_tokens_details;
+  delete mappedUsage.output_tokens_details;
+
+  return mappedUsage;
 }
 
 export function extractChatResponseExtras(payload: unknown): Pick<

src/server/transformers/openai/chat/index.test.ts

@@ -372,6 +372,37 @@ describe('openAiChatTransformer.outbound', () => {
 });
 
 describe('openAiChatTransformer.stream', () => {
+  it('serializes assistant starter chunks for tool-first responses streams', () => {
+    const context = openAiChatTransformer.createStreamContext('gpt-5');
+    const event = openAiChatTransformer.transformStreamEvent({
+      type: 'response.created',
+      response: {
+        id: 'resp-tool-start',
+        model: 'gpt-5',
+        created_at: 1706000000,
+        status: 'in_progress',
+        output: [],
+      },
+    }, context, 'gpt-5');
+
+    const payloads = parseSsePayloads(
+      openAiChatTransformer.serializeStreamEvent(event, context, createClaudeDownstreamContext()),
+    );
+
+    expect(payloads[0]).toMatchObject({
+      id: 'resp-tool-start',
+      model: 'gpt-5',
+      choices: [{
+        index: 0,
+        delta: {
+          role: 'assistant',
+          content: '',
+        },
+        finish_reason: null,
+      }],
+    });
+  });
+
   it('preserves annotations, citations, and usage payload on serialized stream chunks', () => {
     const context = openAiChatTransformer.createStreamContext('gpt-5');
     const event = openAiChatTransformer.transformStreamEvent({
@@ -416,6 +447,40 @@ describe('openAiChatTransformer.stream', () => {
     ]);
   });
 
+  it('maps nested responses usage into chat-completions usage chunks', () => {
+    const context = openAiChatTransformer.createStreamContext('gpt-5');
+    const event = openAiChatTransformer.transformStreamEvent({
+      type: 'response.completed',
+      response: {
+        id: 'resp-usage',
+        model: 'gpt-5',
+        status: 'completed',
+        output: [],
+        usage: {
+          input_tokens: 11,
+          output_tokens: 7,
+          total_tokens: 18,
+          input_tokens_details: { cached_tokens: 3 },
+          output_tokens_details: { reasoning_tokens: 2 },
+        },
+      },
+    }, context, 'gpt-5');
+
+    const payloads = parseSsePayloads(
+      openAiChatTransformer.serializeStreamEvent(event, context, createClaudeDownstreamContext()),
+    );
+
+    expect(payloads[0]).toMatchObject({
+      usage: {
+        prompt_tokens: 11,
+        completion_tokens: 7,
+        total_tokens: 18,
+        prompt_tokens_details: { cached_tokens: 3 },
+        completion_tokens_details: { reasoning_tokens: 2 },
+      },
+    });
+  });
+
   it('serializes multi-choice stream chunks without collapsing choice-specific deltas', () => {
     const context = openAiChatTransformer.createStreamContext('gpt-5');
     const event = openAiChatTransformer.transformStreamEvent({