fix(openai-base): pass the streamed apply_patch diff to the app

.changeset/openai-streamed-apply-patch-diff.md

@@ -0,0 +1,7 @@
+---
+'@tanstack/openai-base': patch
+---
+
+Pass the full diff of a streamed `apply_patch` call to the app. The Responses API adds an `apply_patch_call` with `status: "in_progress"` and an empty `diff`, and streams the diff in before the item is done. The adapter ended the call when the item was added, so `execute` always got `diff: ""`, and an agent loop sent empty patches until it hit its iteration limit. A user-run call (`apply_patch`, `local_shell`, `shell`, `computer`) now starts when its item is added and ends, with its input, once the item is no longer in progress, as function calls already do.
+
+If the response completes while a user-run call is still in progress, the run now ends with a `RUN_ERROR` (code `incomplete-tool-call`) instead of running the tool with incomplete input.

packages/openai-base/src/adapters/responses-text.ts

@@ -1595,6 +1595,17 @@ export abstract class OpenAIBaseResponsesTextAdapter<
       return tracked
     }
 
+    /**
+     * The tool call chunks for a user-run item (`apply_patch_call`,
+     * `shell_call`, `local_shell_call`, `computer_call`). The first time a call
+     * is seen it emits `TOOL_CALL_START`. While the item is `in_progress` its
+     * input may be incomplete (an `apply_patch_call` has an empty diff), so
+     * `TOOL_CALL_END` waits until the item is no longer `in_progress`, and then
+     * carries the complete input. An ended call emits nothing.
+     */
+    // The call ids of user-run calls that are still `in_progress`.
+    const unendedUserCalls = new Set<string>()
+
     const userToolChunks = (
       item: unknown,
       outputIndex: number,
@@ -1609,21 +1620,21 @@ export abstract class OpenAIBaseResponsesTextAdapter<
         started: false,
       }
       toolCallMetadata.set(call.callId, tracked)
-      tracked.started = true
-      tracked.ended = true
       tracked.name = call.name
       streamedToolCall = true
       const timestamp = Date.now()
       const modelName = model || options.model
-      const metadata: OpenAIResponsesToolCallMetadata = {
-        openaiUserTool: call.name,
-        ...(call.itemId ? { itemId: call.itemId } : {}),
-        ...(call.maxOutputLength !== undefined
-          ? { maxOutputLength: call.maxOutputLength }
-          : {}),
-      }
-      return [
-        {
+      const chunks: Array<AdapterYieldChunk> = []
+      if (!tracked.started) {
+        tracked.started = true
+        const metadata: OpenAIResponsesToolCallMetadata = {
+          openaiUserTool: call.name,
+          ...(call.itemId ? { itemId: call.itemId } : {}),
+          ...(call.maxOutputLength !== undefined
+            ? { maxOutputLength: call.maxOutputLength }
+            : {}),
+        }
+        chunks.push({
           type: EventType.TOOL_CALL_START,
           toolCallId: call.callId,
           toolCallName: call.name,
@@ -1632,16 +1643,23 @@ export abstract class OpenAIBaseResponsesTextAdapter<
           timestamp,
           index: outputIndex,
           metadata,
-        },
-        {
-          type: EventType.TOOL_CALL_END,
-          toolCallId: call.callId,
-          toolCallName: call.name,
-          model: modelName,
-          timestamp,
-          input: call.input,
-        },
-      ]
+        })
+      }
+      if (isRecord(item) && item.status === 'in_progress') {
+        unendedUserCalls.add(call.callId)
+        return chunks
+      }
+      unendedUserCalls.delete(call.callId)
+      tracked.ended = true
+      chunks.push({
+        type: EventType.TOOL_CALL_END,
+        toolCallId: call.callId,
+        toolCallName: call.name,
+        model: modelName,
+        timestamp,
+        input: call.input,
+      })
+      return chunks
     }
 
     const emitReasoningDelta = function* (
@@ -2438,6 +2456,25 @@ export abstract class OpenAIBaseResponsesTextAdapter<
             hasEmittedTextMessageStart = false
           }
 
+          // A stateless stream can skip `output_item.done` and send an empty
+          // `output`. A user-run call that never left `in_progress` has
+          // incomplete input (an `apply_patch_call` has an empty diff), so
+          // the run fails instead of running the tool with it.
+          if (unendedUserCalls.size > 0) {
+            const message = `Response completed with user tool calls still in progress: ${[...unendedUserCalls].join(', ')}`
+            yield {
+              type: EventType.RUN_ERROR,
+              metadata: this.sourceMetadata(options.model),
+              model: model || options.model,
+              timestamp: Date.now(),
+              message,
+              code: 'incomplete-tool-call',
+              error: { message, code: 'incomplete-tool-call' },
+            }
+            runFinishedEmitted = true
+            return
+          }
+
           // Determine finish reason. Function-call output → tool_calls.
           // Otherwise surface incomplete_details.reason when present so
           // callers can distinguish length-limit / content-filter cutoffs

packages/openai-base/tests/responses-text.test.ts

@@ -4765,6 +4765,135 @@ describe('OpenAIBaseResponsesTextAdapter', () => {
       expect(finished).toMatchObject({ finishReason: 'tool_calls' })
     })
 
+    it('passes the streamed diff of an apply_patch_call to the app', async () => {
+      // The diff is empty when the item is added and streams in afterwards.
+      const diff = '@@\n-def fib\n+def fibonacci\n'
+      const item = (status: string, itemDiff: string) => ({
+        type: 'apply_patch_call',
+        id: 'apc_1',
+        call_id: 'call_patch',
+        status,
+        operation: { type: 'update_file', path: 'lib/fib.py', diff: itemDiff },
+      })
+      setupMockResponsesClient([
+        {
+          type: 'response.created',
+          response: {
+            id: 'resp-user-tool',
+            model: 'test-model',
+            status: 'in_progress',
+          },
+        },
+        {
+          type: 'response.output_item.added',
+          output_index: 0,
+          item: item('in_progress', ''),
+        },
+        {
+          type: 'response.apply_patch_call_operation_diff.delta',
+          output_index: 0,
+          item_id: 'apc_1',
+          delta: diff,
+        },
+        {
+          type: 'response.apply_patch_call_operation_diff.done',
+          output_index: 0,
+          item_id: 'apc_1',
+        },
+        {
+          type: 'response.output_item.done',
+          output_index: 0,
+          item: item('completed', diff),
+        },
+        {
+          type: 'response.completed',
+          response: {
+            id: 'resp-user-tool',
+            model: 'test-model',
+            status: 'completed',
+            output: [item('completed', diff)],
+            usage: { input_tokens: 1, output_tokens: 1, total_tokens: 2 },
+          },
+        },
+      ])
+      const adapter = new TestResponsesAdapter(testConfig, 'test-model')
+      const chunks: Array<AdapterYieldChunk> = []
+      for await (const chunk of adapter.chatStream({
+        threadId: 'thread-1',
+        logger: testLogger,
+        model: 'test-model',
+        messages: [{ role: 'user', content: 'edit the file' }],
+        tools: [applyPatchTool()],
+      })) {
+        chunks.push(chunk)
+      }
+
+      const starts = chunks.filter((chunk) => chunk.type === 'TOOL_CALL_START')
+      const ends = chunks.filter((chunk) => chunk.type === 'TOOL_CALL_END')
+      expect(starts).toHaveLength(1)
+      expect(ends).toHaveLength(1)
+      expect(ends[0]).toMatchObject({
+        toolCallId: 'call_patch',
+        input: {
+          operation: { type: 'update_file', path: 'lib/fib.py', diff },
+        },
+      })
+    })
+
+    it('fails the run when a user tool call is still in_progress at completion', async () => {
+      // A stateless stream can skip output_item.done and send an empty output.
+      setupMockResponsesClient([
+        {
+          type: 'response.created',
+          response: {
+            id: 'resp-open',
+            model: 'test-model',
+            status: 'in_progress',
+          },
+        },
+        {
+          type: 'response.output_item.added',
+          output_index: 0,
+          item: {
+            type: 'apply_patch_call',
+            id: 'apc_open',
+            call_id: 'call_open',
+            status: 'in_progress',
+            operation: { type: 'update_file', path: 'lib/fib.py', diff: '' },
+          },
+        },
+        {
+          type: 'response.completed',
+          response: {
+            id: 'resp-open',
+            model: 'test-model',
+            status: 'completed',
+            output: [],
+            usage: { input_tokens: 1, output_tokens: 1, total_tokens: 2 },
+          },
+        },
+      ])
+      const adapter = new TestResponsesAdapter(testConfig, 'test-model')
+      const chunks: Array<AdapterYieldChunk> = []
+      for await (const chunk of adapter.chatStream({
+        threadId: 'thread-1',
+        logger: testLogger,
+        model: 'test-model',
+        messages: [{ role: 'user', content: 'edit the file' }],
+        tools: [applyPatchTool()],
+      })) {
+        chunks.push(chunk)
+      }
+
+      const types = chunks.map((chunk) => chunk.type)
+      expect(types).not.toContain('TOOL_CALL_END')
+      expect(types).not.toContain('RUN_FINISHED')
+      expect(chunks.at(-1)).toMatchObject({
+        type: 'RUN_ERROR',
+        code: 'incomplete-tool-call',
+      })
+    })
+
     it('surfaces a local shell call and ignores a hosted container shell', async () => {
       const chunks = await collect([
         {

testing/e2e/src/routeTree.gen.ts

@@ -154,6 +154,7 @@ import { Route as ApiMoonshotUsageDetailsRouteImport } from './routes/api.moonsh
 import { Route as ApiMultimodalToolResultWireRouteImport } from './routes/api.multimodal-tool-result-wire'
 import { Route as ApiNeonWireRouteImport } from './routes/api.neon-wire'
 import { Route as ApiNonStreamingRunErrorRouteImport } from './routes/api.non-streaming-run-error'
+import { Route as ApiOpenaiApplyPatchWireRouteImport } from './routes/api.openai-apply-patch-wire'
 import { Route as ApiOpenaiCompletedResponseTextRouteImport } from './routes/api.openai-completed-response-text'
 import { Route as ApiOpenaiComputerUseWireRouteImport } from './routes/api.openai-computer-use-wire'
 import { Route as ApiOpenaiImage25ModelsRouteImport } from './routes/api.openai-image-2-5-models'
@@ -971,6 +972,11 @@ const ApiNonStreamingRunErrorRoute = ApiNonStreamingRunErrorRouteImport.update({
   path: '/api/non-streaming-run-error',
   getParentRoute: () => rootRouteImport,
 } as any)
+const ApiOpenaiApplyPatchWireRoute = ApiOpenaiApplyPatchWireRouteImport.update({
+  id: '/api/openai-apply-patch-wire',
+  path: '/api/openai-apply-patch-wire',
+  getParentRoute: () => rootRouteImport,
+} as any)
 const ApiOpenaiCompletedResponseTextRoute =
   ApiOpenaiCompletedResponseTextRouteImport.update({
     id: '/api/openai-completed-response-text',
@@ -1443,6 +1449,7 @@ export interface FileRoutesByFullPath {
   '/api/multimodal-tool-result-wire': typeof ApiMultimodalToolResultWireRoute
   '/api/neon-wire': typeof ApiNeonWireRoute
   '/api/non-streaming-run-error': typeof ApiNonStreamingRunErrorRoute
+  '/api/openai-apply-patch-wire': typeof ApiOpenaiApplyPatchWireRoute
   '/api/openai-completed-response-text': typeof ApiOpenaiCompletedResponseTextRoute
   '/api/openai-computer-use-wire': typeof ApiOpenaiComputerUseWireRoute
   '/api/openai-image-2-5-models': typeof ApiOpenaiImage25ModelsRoute
@@ -1651,6 +1658,7 @@ export interface FileRoutesByTo {
   '/api/multimodal-tool-result-wire': typeof ApiMultimodalToolResultWireRoute
   '/api/neon-wire': typeof ApiNeonWireRoute
   '/api/non-streaming-run-error': typeof ApiNonStreamingRunErrorRoute
+  '/api/openai-apply-patch-wire': typeof ApiOpenaiApplyPatchWireRoute
   '/api/openai-completed-response-text': typeof ApiOpenaiCompletedResponseTextRoute
   '/api/openai-computer-use-wire': typeof ApiOpenaiComputerUseWireRoute
   '/api/openai-image-2-5-models': typeof ApiOpenaiImage25ModelsRoute
@@ -1860,6 +1868,7 @@ export interface FileRoutesById {
   '/api/multimodal-tool-result-wire': typeof ApiMultimodalToolResultWireRoute
   '/api/neon-wire': typeof ApiNeonWireRoute
   '/api/non-streaming-run-error': typeof ApiNonStreamingRunErrorRoute
+  '/api/openai-apply-patch-wire': typeof ApiOpenaiApplyPatchWireRoute
   '/api/openai-completed-response-text': typeof ApiOpenaiCompletedResponseTextRoute
   '/api/openai-computer-use-wire': typeof ApiOpenaiComputerUseWireRoute
   '/api/openai-image-2-5-models': typeof ApiOpenaiImage25ModelsRoute
@@ -2070,6 +2079,7 @@ export interface FileRouteTypes {
     | '/api/multimodal-tool-result-wire'
     | '/api/neon-wire'
     | '/api/non-streaming-run-error'
+    | '/api/openai-apply-patch-wire'
     | '/api/openai-completed-response-text'
     | '/api/openai-computer-use-wire'
     | '/api/openai-image-2-5-models'
@@ -2278,6 +2288,7 @@ export interface FileRouteTypes {
     | '/api/multimodal-tool-result-wire'
     | '/api/neon-wire'
     | '/api/non-streaming-run-error'
+    | '/api/openai-apply-patch-wire'
     | '/api/openai-completed-response-text'
     | '/api/openai-computer-use-wire'
     | '/api/openai-image-2-5-models'
@@ -2486,6 +2497,7 @@ export interface FileRouteTypes {
     | '/api/multimodal-tool-result-wire'
     | '/api/neon-wire'
     | '/api/non-streaming-run-error'
+    | '/api/openai-apply-patch-wire'
     | '/api/openai-completed-response-text'
     | '/api/openai-computer-use-wire'
     | '/api/openai-image-2-5-models'
@@ -2695,6 +2707,7 @@ export interface RootRouteChildren {
   ApiMultimodalToolResultWireRoute: typeof ApiMultimodalToolResultWireRoute
   ApiNeonWireRoute: typeof ApiNeonWireRoute
   ApiNonStreamingRunErrorRoute: typeof ApiNonStreamingRunErrorRoute
+  ApiOpenaiApplyPatchWireRoute: typeof ApiOpenaiApplyPatchWireRoute
   ApiOpenaiCompletedResponseTextRoute: typeof ApiOpenaiCompletedResponseTextRoute
   ApiOpenaiComputerUseWireRoute: typeof ApiOpenaiComputerUseWireRoute
   ApiOpenaiImage25ModelsRoute: typeof ApiOpenaiImage25ModelsRoute
@@ -3771,6 +3784,13 @@ declare module '@tanstack/react-router' {
       preLoaderRoute: typeof ApiNonStreamingRunErrorRouteImport
       parentRoute: typeof rootRouteImport
     }
+    '/api/openai-apply-patch-wire': {
+      id: '/api/openai-apply-patch-wire'
+      path: '/api/openai-apply-patch-wire'
+      fullPath: '/api/openai-apply-patch-wire'
+      preLoaderRoute: typeof ApiOpenaiApplyPatchWireRouteImport
+      parentRoute: typeof rootRouteImport
+    }
     '/api/openai-completed-response-text': {
       id: '/api/openai-completed-response-text'
       path: '/api/openai-completed-response-text'
@@ -4409,6 +4429,7 @@ const rootRouteChildren: RootRouteChildren = {
   ApiMultimodalToolResultWireRoute: ApiMultimodalToolResultWireRoute,
   ApiNeonWireRoute: ApiNeonWireRoute,
   ApiNonStreamingRunErrorRoute: ApiNonStreamingRunErrorRoute,
+  ApiOpenaiApplyPatchWireRoute: ApiOpenaiApplyPatchWireRoute,
   ApiOpenaiCompletedResponseTextRoute: ApiOpenaiCompletedResponseTextRoute,
   ApiOpenaiComputerUseWireRoute: ApiOpenaiComputerUseWireRoute,
   ApiOpenaiImage25ModelsRoute: ApiOpenaiImage25ModelsRoute,