feat: Stream reasoning events as agent thoughts (#182)

We weren't showing any reasoning/thinking before, but this also allows
for streaming in the reasoning deltas if the backend supports it.

Author: benbrandt

src/CodexAcpClient.ts

@@ -480,7 +480,7 @@ export class CodexAcpClient {
             input: input,
             approvalPolicy: agentMode.approvalPolicy,
             sandboxPolicy: agentMode.sandboxPolicy,
-            summary: disableSummary ? "none" : null,
+            summary: disableSummary ? "none" : "auto",
             effort: effort,
             model: modelId.model,
             serviceTier: serviceTier,

src/CodexEventHandler.ts

@@ -20,6 +20,9 @@ import type {
     ItemStartedNotification,
     ThreadItem,
     ModelReroutedNotification,
+    ReasoningSummaryPartAddedNotification,
+    ReasoningSummaryTextDeltaNotification,
+    ReasoningTextDeltaNotification,
     ThreadGoalClearedNotification,
     ThreadGoalUpdatedNotification,
     ThreadTokenUsageUpdatedNotification,
@@ -60,6 +63,7 @@ export class CodexEventHandler {
     private readonly activeGuardianApprovalReviews = new Set<string>();
     private readonly activeImageGenerationItems = new Set<string>();
     private readonly emittedImageViewItems = new Set<string>();
+    private readonly seenReasoningDeltaItemIds = new Set<string>();
 
     constructor(connection: acp.AgentSideConnection, sessionState: SessionState) {
         this.connection = connection;
@@ -145,6 +149,12 @@ export class CodexEventHandler {
                 return this.handleGuardianApprovalReviewCompleted(notification.params);
             case "thread/compacted":
                 return this.createContextCompactedEvent();
+            case "item/reasoning/summaryTextDelta":
+                return this.createReasoningDeltaEvent(notification.params);
+            case "item/reasoning/textDelta":
+                return this.createReasoningDeltaEvent(notification.params);
+            case "item/reasoning/summaryPartAdded":
+                return this.createReasoningSectionBreakEvent(notification.params);
             case "model/rerouted":
                 return this.createModelReroutedEvent(notification.params);
             case "fuzzyFileSearch/sessionUpdated":
@@ -159,9 +169,6 @@ export class CodexEventHandler {
             case "command/exec/outputDelta":
             case "hook/started":
             case "hook/completed":
-            case "item/reasoning/summaryTextDelta":
-            case "item/reasoning/summaryPartAdded":
-            case "item/reasoning/textDelta":
             case "turn/diff/updated":
             case "item/commandExecution/terminalInteraction":
             case "item/fileChange/outputDelta":
@@ -290,6 +297,28 @@ export class CodexEventHandler {
         };
     }
 
+    private createReasoningDeltaEvent(
+        event: ReasoningSummaryTextDeltaNotification | ReasoningTextDeltaNotification
+    ): UpdateSessionEvent {
+        this.seenReasoningDeltaItemIds.add(event.itemId);
+        return this.createAgentThoughtEvent(event.delta);
+    }
+
+    private createReasoningSectionBreakEvent(event: ReasoningSummaryPartAddedNotification): UpdateSessionEvent {
+        this.seenReasoningDeltaItemIds.add(event.itemId);
+        return this.createAgentThoughtEvent("\n\n");
+    }
+
+    private createAgentThoughtEvent(text: string): UpdateSessionEvent {
+        return {
+            sessionUpdate: "agent_thought_chunk",
+            content: {
+                type: "text",
+                text,
+            }
+        };
+    }
+
     private async createItemEvent(event: ItemStartedNotification): Promise<UpdateSessionEvent | null> {
         switch (event.item.type) {
             case "fileChange":
@@ -351,15 +380,10 @@ export class CodexEventHandler {
                 }
                 return createImageGenerationUpdate(event.item);
             case "reasoning":
-                const summary = event.item.summary[0];
-                if (!summary) return null;
-                return {
-                    sessionUpdate: "agent_thought_chunk",
-                    content: {
-                        type: "text",
-                        text: summary
-                    }
+                if (this.seenReasoningDeltaItemIds.delete(event.item.id)) {
+                    return null;
                 }
+                return this.createCompletedReasoningEvent(event.item);
             case "webSearch":
                 return createWebSearchCompleteUpdate(event.item);
             case "collabAgentToolCall":
@@ -376,6 +400,15 @@ export class CodexEventHandler {
         }
     }
 
+    private createCompletedReasoningEvent(item: ThreadItem & { type: "reasoning" }): UpdateSessionEvent | null {
+        const parts = item.summary.length > 0 ? item.summary : item.content;
+        const text = parts.filter(part => part.length > 0).join("\n\n");
+        if (text.length === 0) {
+            return null;
+        }
+        return this.createAgentThoughtEvent(text);
+    }
+
     private createExitedReviewModeEvent(item: ThreadItem & { type: "exitedReviewMode" }): UpdateSessionEvent | null {
         const text = item.review.trim();
         if (text.length === 0) {

src/__tests__/CodexACPAgent/CodexAcpClient.test.ts

@@ -1179,22 +1179,22 @@ describe('ACP server test', { timeout: 40_000 }, () => {
         };
     }
 
-    it ('should disable resasoning.summary if key authorization is used', async () => {
+    it ('should disable reasoning.summary if key authorization is used', async () => {
         const { mockFixture, turnStartSpy } = setupPromptFixture({ account: { type: "apiKey" } });
 
         await mockFixture.getCodexAcpAgent().prompt({ sessionId: "id", prompt: [{ type: "text", text: "test" }] });
 
         expect(turnStartSpy).toHaveBeenCalledWith(expect.objectContaining({ summary: "none" }));
     });
 
-    it ('should not disable resasoning.summary by default', async () => {
+    it ('should enable reasoning.summary by default', async () => {
         const { mockFixture, turnStartSpy } = setupPromptFixture({
             account: { type: "chatgpt", email: "test@example.com", planType: "pro" },
         });
 
         await mockFixture.getCodexAcpAgent().prompt({ sessionId: "id", prompt: [{ type: "text", text: "test" }] });
 
-        expect(turnStartSpy).toHaveBeenCalledWith(expect.objectContaining({ summary: null }));
+        expect(turnStartSpy).toHaveBeenCalledWith(expect.objectContaining({ summary: "auto" }));
     });
 
     it ('should disable reasoning.summary when model lacks reasoning', async () => {
@@ -1208,7 +1208,7 @@ describe('ACP server test', { timeout: 40_000 }, () => {
         expect(turnStartSpy).toHaveBeenCalledWith(expect.objectContaining({ summary: "none" }));
     });
 
-    it ('should not disable reasoning.summary when model supports reasoning', async () => {
+    it ('should enable reasoning.summary when model supports reasoning', async () => {
         const { mockFixture, turnStartSpy } = setupPromptFixture({
             account: { type: "chatgpt", email: "test@example.com", planType: "pro" },
             supportedReasoningEfforts: [
@@ -1219,7 +1219,7 @@ describe('ACP server test', { timeout: 40_000 }, () => {
 
         await mockFixture.getCodexAcpAgent().prompt({ sessionId: "id", prompt: [{ type: "text", text: "test" }] });
 
-        expect(turnStartSpy).toHaveBeenCalledWith(expect.objectContaining({ summary: null }));
+        expect(turnStartSpy).toHaveBeenCalledWith(expect.objectContaining({ summary: "auto" }));
     });
 
     it ('should reject prompt with images when model does not support image input', async () => {

src/__tests__/CodexACPAgent/data/reasoning-completed-parts.json

@@ -0,0 +1,15 @@
+{
+  "method": "sessionUpdate",
+  "args": [
+    {
+      "sessionId": "test-session-id",
+      "update": {
+        "sessionUpdate": "agent_thought_chunk",
+        "content": {
+          "type": "text",
+          "text": "First summary\n\nSecond summary"
+        }
+      }
+    }
+  ]
+}

src/__tests__/CodexACPAgent/data/reasoning-deltas-and-section-break.json

@@ -0,0 +1,45 @@
+{
+  "method": "sessionUpdate",
+  "args": [
+    {
+      "sessionId": "test-session-id",
+      "update": {
+        "sessionUpdate": "agent_thought_chunk",
+        "content": {
+          "type": "text",
+          "text": "First thought"
+        }
+      }
+    }
+  ]
+}
+{
+  "method": "sessionUpdate",
+  "args": [
+    {
+      "sessionId": "test-session-id",
+      "update": {
+        "sessionUpdate": "agent_thought_chunk",
+        "content": {
+          "type": "text",
+          "text": "\n\n"
+        }
+      }
+    }
+  ]
+}
+{
+  "method": "sessionUpdate",
+  "args": [
+    {
+      "sessionId": "test-session-id",
+      "update": {
+        "sessionUpdate": "agent_thought_chunk",
+        "content": {
+          "type": "text",
+          "text": "Raw reasoning detail"
+        }
+      }
+    }
+  ]
+}

src/__tests__/CodexACPAgent/data/send-attachments-turn-start.json

@@ -45,7 +45,7 @@
       "excludeTmpdirEnvVar": false,
       "excludeSlashTmp": false
     },
-    "summary": null,
+    "summary": "auto",
     "effort": "effort",
     "model": "model",
     "serviceTier": null

src/__tests__/CodexACPAgent/reasoning-events.test.ts

@@ -0,0 +1,105 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import type { ServerNotification } from "../../app-server";
+import type { SessionState } from "../../CodexAcpServer";
+import { AgentMode } from "../../AgentMode";
+import {
+    createCodexMockTestFixture,
+    createTestSessionState,
+    setupPromptAndSendNotifications,
+    type CodexMockTestFixture
+} from "../acp-test-utils";
+
+describe("CodexEventHandler - reasoning events", () => {
+    let mockFixture: CodexMockTestFixture;
+    const sessionId = "test-session-id";
+
+    beforeEach(() => {
+        mockFixture = createCodexMockTestFixture();
+        vi.clearAllMocks();
+    });
+
+    const sessionState: SessionState = createTestSessionState({
+        sessionId,
+        currentModelId: "model-id[effort]",
+        agentMode: AgentMode.DEFAULT_AGENT_MODE
+    });
+
+    it("streams reasoning deltas and section breaks without duplicating the completed item", async () => {
+        const notifications: ServerNotification[] = [
+            {
+                method: "item/reasoning/summaryTextDelta",
+                params: {
+                    threadId: sessionId,
+                    turnId: "turn-1",
+                    itemId: "reasoning-1",
+                    summaryIndex: 0,
+                    delta: "First thought",
+                },
+            },
+            {
+                method: "item/reasoning/summaryPartAdded",
+                params: {
+                    threadId: sessionId,
+                    turnId: "turn-1",
+                    itemId: "reasoning-1",
+                    summaryIndex: 1,
+                },
+            },
+            {
+                method: "item/reasoning/textDelta",
+                params: {
+                    threadId: sessionId,
+                    turnId: "turn-1",
+                    itemId: "reasoning-1",
+                    contentIndex: 0,
+                    delta: "Raw reasoning detail",
+                },
+            },
+            {
+                method: "item/completed",
+                params: {
+                    threadId: sessionId,
+                    turnId: "turn-1",
+                    completedAtMs: 0,
+                    item: {
+                        type: "reasoning",
+                        id: "reasoning-1",
+                        summary: ["Completed summary should not duplicate"],
+                        content: ["Completed content should not duplicate"],
+                    },
+                },
+            },
+        ];
+
+        await setupPromptAndSendNotifications(mockFixture, sessionId, sessionState, notifications);
+
+        await expect(mockFixture.getAcpConnectionDump([])).toMatchFileSnapshot(
+            "data/reasoning-deltas-and-section-break.json"
+        );
+    });
+
+    it("emits all completed reasoning parts when no deltas streamed", async () => {
+        const notifications: ServerNotification[] = [
+            {
+                method: "item/completed",
+                params: {
+                    threadId: sessionId,
+                    turnId: "turn-1",
+                    completedAtMs: 0,
+                    item: {
+                        type: "reasoning",
+                        id: "reasoning-2",
+                        summary: ["First summary", "Second summary"],
+                        content: ["Raw content fallback"],
+                    },
+                },
+            },
+        ];
+
+        await setupPromptAndSendNotifications(mockFixture, sessionId, sessionState, notifications);
+
+        await expect(mockFixture.getAcpConnectionDump([])).toMatchFileSnapshot(
+            "data/reasoning-completed-parts.json"
+        );
+    });
+});