@code-yeongyu/senpi-ai 2026.10.25-4 → 2026.10.26-2

package/dist/api/anthropic-messages.js

@@ -5,6 +5,7 @@ import { readProviderDiagnostic } from "../provider-diagnostic.js";
 import { registerSessionResourceCleanup } from "../session-resources.js";
 import { isVideoMimeType } from "../types.js";
 import { combineAbortSignals } from "../utils/abort-signals.js";
+import { pushAssistantProgress } from "../utils/assistant-capture.js";
 import { claudeCodeVersionTooOldHint, getClaudeCodeVersion, isClaudeCodeVersionTooOldError, recoverClaudeCodeVersion, } from "../utils/claude-code-version.js";
 import { splitDeferredTools } from "../utils/deferred-tools.js";
 import { appendAssistantMessageDiagnostic } from "../utils/diagnostics.js";
@@ -920,7 +921,7 @@ async function* iterateAnthropicEvents(response, signal) {
     }
 }
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = resolveTranscript(context, supportsMidConvoSystemMessages(model));
     const currentTools = getCurrentTools(normalizedContext.messages);
     (async () => {
@@ -1248,7 +1249,12 @@ export const stream = (model, context, options) => {
                             index: event.index,
                         };
                         output.content.push(block);
-                        // Native blocks are represented in output.content but have no dedicated stream event variant.
+                        pushAssistantProgress(stream, {
+                            type: "message_block_state",
+                            contentIndex: output.content.length - 1,
+                            phase: "insert",
+                            block,
+                        });
                     }
                 }
                 else if (event.type === "content_block_delta") {
@@ -1304,6 +1310,11 @@ export const stream = (model, context, options) => {
                         if (block && block.type === "thinking") {
                             block.thinkingSignature = block.thinkingSignature || "";
                             block.thinkingSignature += event.delta.signature;
+                            pushAssistantProgress(stream, {
+                                type: "message_block_metadata",
+                                contentIndex: index,
+                                thinkingSignatureDelta: event.delta.signature,
+                            });
                         }
                     }
                 }
@@ -1346,6 +1357,12 @@ export const stream = (model, context, options) => {
                             if (partialJson !== undefined && isRecord(block.raw)) {
                                 block.raw = { ...block.raw, input: parseStreamingJson(partialJson) };
                             }
+                            pushAssistantProgress(stream, {
+                                type: "message_block_state",
+                                contentIndex: index,
+                                phase: "end",
+                                block,
+                            });
                         }
                     }
                 }

package/dist/api/azure-openai-responses.js

@@ -44,7 +44,7 @@ function formatAzureOpenAIError(error) {
  * Generate function for Azure OpenAI Responses API
  */
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = resolveTranscript(context, model.compat?.supportsMidConvoSystemMessages);
     // Start async processing
     (async () => {

package/dist/api/bedrock-converse-stream.js

@@ -3,6 +3,7 @@ import { NodeHttpHandler } from "@smithy/node-http-handler";
 import { HttpProxyAgent } from "http-proxy-agent";
 import { HttpsProxyAgent } from "https-proxy-agent";
 import { calculateCost } from "../models.js";
+import { pushAssistantProgress } from "../utils/assistant-capture.js";
 import { appendAssistantMessageDiagnostic } from "../utils/diagnostics.js";
 import { normalizeProviderError } from "../utils/error-body.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
@@ -24,7 +25,7 @@ const EMPTY_TEXT_PLACEHOLDER = "<empty>";
 /** Matches the placeholder the Anthropic API path uses for redacted thinking. */
 const REDACTED_THINKING_PLACEHOLDER = "[Reasoning redacted]";
 export const stream = (model, context, options = {}) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     // Bedrock has no mid-conversation system messages; fold them into the leading prompt.
     const normalizedContext = collapseSystemMessages(context);
     (async () => {
@@ -500,12 +501,23 @@ function handleContentBlockDelta(event, blocks, output, stream) {
             if (delta.reasoningContent.signature && !thinkingBlock.redacted) {
                 thinkingBlock.thinkingSignature =
                     (thinkingBlock.thinkingSignature || "") + delta.reasoningContent.signature;
+                pushAssistantProgress(stream, {
+                    type: "message_block_metadata",
+                    contentIndex: thinkingIndex,
+                    thinkingSignatureDelta: delta.reasoningContent.signature,
+                });
             }
             if (delta.reasoningContent.redactedContent?.length) {
                 // Encrypted reasoning from non-Anthropic models on Bedrock (e.g. OpenAI GPT-5.6).
                 // The payload is opaque, so keep it verbatim in `thinkingSignature` the way the
                 // Anthropic path stores redacted thinking, and replay it on the next turn.
                 if (!thinkingBlock.redacted) {
+                    pushAssistantProgress(stream, {
+                        type: "message_block_metadata",
+                        contentIndex: thinkingIndex,
+                        redacted: true,
+                        ...(thinkingBlock.thinkingSignature ? { thinkingSignature: "" } : {}),
+                    });
                     thinkingBlock.redacted = true;
                     thinkingBlock.thinkingSignature = "";
                     thinkingBlock.thinking += REDACTED_THINKING_PLACEHOLDER;

package/dist/api/cursor-agent.js

@@ -475,7 +475,7 @@ export function mapH2TransportError(error, baseUrl) {
     return error;
 }
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const requestView = toCursorRequestView(context);
     (async () => {
         const output = {
@@ -3061,7 +3061,7 @@ export function processInteractionUpdate(update, output, stream, state, usageSta
                 target[kStreamingLastParseLen] = nextBuffer.length;
             }
             const idx = output.content.indexOf(target);
-            stream.push({ type: "toolcall_delta", contentIndex: idx, delta: chunk, partial: output });
+            stream.push({ type: "toolcall_delta", contentIndex: idx, delta: chunk, partial: output }, { parsedThrough: target[kStreamingLastParseLen] ?? 0 });
         }
     }
     else if (updateCase === "toolCallCompleted") {

package/dist/api/devin-agent/stream-state.js

@@ -6,6 +6,7 @@
  * or closes them. This module owns the open-block state so the adapter emits
  * senpi's start/delta/end triples in the right order.
  */
+import { pushAssistantProgress, pushAssistantSignature } from "../../utils/assistant-capture.js";
 import { parseStreamingJson } from "../../utils/json-parse.js";
 import { StopReason } from "./gen/cascade_pb.js";
 export function createDevinStreamState() {
@@ -41,6 +42,7 @@ function appendThinking(message, output, events, state) {
     if (block?.type !== "thinking")
         return;
     block.thinking += message.deltaThinking;
+    pushAssistantSignature(events, state.thinkingIndex, block, message.deltaSignature);
     if (message.deltaSignature)
         block.thinkingSignature = message.deltaSignature;
     events.push({
@@ -84,6 +86,12 @@ function appendToolCall(call, output, events, state) {
     const block = output.content[entry.contentIndex];
     if (block?.type !== "toolCall")
         return;
+    if (call.name && call.name !== block.name)
+        pushAssistantProgress(events, {
+            type: "message_block_metadata",
+            contentIndex: entry.contentIndex,
+            name: call.name,
+        });
     if (call.name)
         block.name = call.name;
     if (!call.argumentsJson)

package/dist/api/devin-agent.js

@@ -20,7 +20,7 @@ import { AssignModelRequestSchema, AssignModelResponseSchema, GetChatMessageRequ
 import { applyDevinResponse, createDevinStreamState } from "./devin-agent/stream-state.js";
 import { buildDevinChatRequest, buildDevinRouterPrompt, DEVIN_ASSIGN_MODEL_PATH, DEVIN_CHAT_HEADERS, DEVIN_CHAT_MESSAGE_PATH, DEVIN_DEFAULT_BASE_URL, DEVIN_USER_JWT_PATH, decodeDevinFrames, devinCliMetadata, encodeDevinRequestFrame, postDevinUnary, readDevinTrailerError, } from "./devin-agent/wire.js";
 export const stream = (model, context, options) => {
-    const events = new AssistantMessageEventStream();
+    const events = new AssistantMessageEventStream({ captureBlocks: true });
     void run(model, context, events, options);
     return events;
 };

package/dist/api/google-generative-ai.js

@@ -1,5 +1,6 @@
 import { GoogleGenAI, } from "@google/genai";
 import { calculateCost, clampThinkingLevel } from "../models.js";
+import { pushAssistantCompleteBlock, pushAssistantSignature } from "../utils/assistant-capture.js";
 import { formatProviderError, normalizeProviderError } from "../utils/error-body.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { providerHeadersToRecord } from "../utils/headers.js";
@@ -13,7 +14,7 @@ import { applyExtraBody, buildBaseOptions, GOOGLE_RESERVED_BODY_KEYS } from "./s
 // Counter for generating unique tool call IDs
 let toolCallCounter = 0;
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = collapseSystemMessages(context);
     (async () => {
         const output = {
@@ -50,10 +51,13 @@ export const stream = (model, context, options) => {
             const googleStream = await retryGoogleRequest(() => client.models.generateContentStream(params), options);
             stream.push({ type: "start", partial: output });
             let currentBlock = null;
+            let currentIndex = 0;
             let groundingMetadataEmitted = false;
             let urlContextMetadataEmitted = false;
             const blocks = output.content;
             const blockIndex = () => blocks.length - 1;
+            // blockIndex() drifts past the open block once a native part is appended; capture keeps the block's own index.
+            const pushOpen = (event) => stream.push(event, { contentIndex: currentIndex });
             for await (const chunk of googleStream) {
                 await options?.onProviderStreamEvent?.(chunk, model);
                 // @google/genai documents GenerateContentResponse.responseId as an output-only field
@@ -71,7 +75,7 @@ export const stream = (model, context, options) => {
                                 (!isThinking && currentBlock.type !== "text")) {
                                 if (currentBlock) {
                                     if (currentBlock.type === "text") {
-                                        stream.push({
+                                        pushOpen({
                                             type: "text_end",
                                             contentIndex: blocks.length - 1,
                                             content: currentBlock.text,
@@ -79,7 +83,7 @@ export const stream = (model, context, options) => {
                                         });
                                     }
                                     else {
-                                        stream.push({
+                                        pushOpen({
                                             type: "thinking_end",
                                             contentIndex: blockIndex(),
                                             content: currentBlock.thinking,
@@ -90,18 +94,21 @@ export const stream = (model, context, options) => {
                                 if (isThinking) {
                                     currentBlock = { type: "thinking", thinking: "", thinkingSignature: undefined };
                                     output.content.push(currentBlock);
+                                    currentIndex = blockIndex();
                                     stream.push({ type: "thinking_start", contentIndex: blockIndex(), partial: output });
                                 }
                                 else {
                                     currentBlock = { type: "text", text: "" };
                                     output.content.push(currentBlock);
+                                    currentIndex = blockIndex();
                                     stream.push({ type: "text_start", contentIndex: blockIndex(), partial: output });
                                 }
                             }
                             if (currentBlock.type === "thinking") {
                                 currentBlock.thinking += part.text;
+                                pushAssistantSignature(stream, currentIndex, currentBlock, part.thoughtSignature);
                                 currentBlock.thinkingSignature = retainThoughtSignature(currentBlock.thinkingSignature, part.thoughtSignature);
-                                stream.push({
+                                pushOpen({
                                     type: "thinking_delta",
                                     contentIndex: blockIndex(),
                                     delta: part.text,
@@ -110,8 +117,9 @@ export const stream = (model, context, options) => {
                             }
                             else {
                                 currentBlock.text += part.text;
+                                pushAssistantSignature(stream, currentIndex, currentBlock, part.thoughtSignature);
                                 currentBlock.textSignature = retainThoughtSignature(currentBlock.textSignature, part.thoughtSignature);
-                                stream.push({
+                                pushOpen({
                                     type: "text_delta",
                                     contentIndex: blockIndex(),
                                     delta: part.text,
@@ -123,7 +131,7 @@ export const stream = (model, context, options) => {
                             handledAsKnownPart = true;
                             if (currentBlock) {
                                 if (currentBlock.type === "text") {
-                                    stream.push({
+                                    pushOpen({
                                         type: "text_end",
                                         contentIndex: blockIndex(),
                                         content: currentBlock.text,
@@ -131,7 +139,7 @@ export const stream = (model, context, options) => {
                                     });
                                 }
                                 else {
-                                    stream.push({
+                                    pushOpen({
                                         type: "thinking_end",
                                         contentIndex: blockIndex(),
                                         content: currentBlock.thinking,
@@ -155,16 +163,18 @@ export const stream = (model, context, options) => {
                             };
                             output.content.push(toolCall);
                             stream.push({ type: "toolcall_start", contentIndex: blockIndex(), partial: output });
+                            // The adapter never parses its own spelling: the complete parsed seed stays authoritative.
                             stream.push({
                                 type: "toolcall_delta",
                                 contentIndex: blockIndex(),
                                 delta: JSON.stringify(toolCall.arguments),
                                 partial: output,
-                            });
+                            }, { parsedThrough: 0 });
                             stream.push({ type: "toolcall_end", contentIndex: blockIndex(), toolCall, partial: output });
                         }
                         if (!handledAsKnownPart) {
                             output.content.push(toProviderNativeContent(part));
+                            pushAssistantCompleteBlock(stream, blockIndex(), blocks[blockIndex()]);
                         }
                     }
                 }
@@ -174,6 +184,7 @@ export const stream = (model, context, options) => {
                         subtype: "groundingMetadata",
                         raw: candidate.groundingMetadata,
                     });
+                    pushAssistantCompleteBlock(stream, blockIndex(), blocks[blockIndex()]);
                     groundingMetadataEmitted = true;
                 }
                 if (candidate?.urlContextMetadata && !urlContextMetadataEmitted) {
@@ -182,6 +193,7 @@ export const stream = (model, context, options) => {
                         subtype: "urlContextMetadata",
                         raw: candidate.urlContextMetadata,
                     });
+                    pushAssistantCompleteBlock(stream, blockIndex(), blocks[blockIndex()]);
                     urlContextMetadataEmitted = true;
                 }
                 if (candidate?.finishReason) {
@@ -213,7 +225,7 @@ export const stream = (model, context, options) => {
             }
             if (currentBlock) {
                 if (currentBlock.type === "text") {
-                    stream.push({
+                    pushOpen({
                         type: "text_end",
                         contentIndex: blockIndex(),
                         content: currentBlock.text,
@@ -221,7 +233,7 @@ export const stream = (model, context, options) => {
                     });
                 }
                 else {
-                    stream.push({
+                    pushOpen({
                         type: "thinking_end",
                         contentIndex: blockIndex(),
                         content: currentBlock.thinking,

package/dist/api/google-vertex.js

@@ -1,5 +1,6 @@
 import { GoogleGenAI, ResourceScope, } from "@google/genai";
 import { calculateCost, clampThinkingLevel } from "../models.js";
+import { pushAssistantCompleteBlock, pushAssistantSignature } from "../utils/assistant-capture.js";
 import { formatProviderError, normalizeProviderError } from "../utils/error-body.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { providerHeadersToRecord } from "../utils/headers.js";
@@ -17,7 +18,7 @@ const GCP_VERTEX_CREDENTIALS_MARKER = "gcp-vertex-credentials";
 // Counter for generating unique tool call IDs
 let toolCallCounter = 0;
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = collapseSystemMessages(context);
     (async () => {
         const output = {
@@ -55,10 +56,13 @@ export const stream = (model, context, options) => {
             const googleStream = await retryGoogleRequest(() => client.models.generateContentStream(params), options);
             stream.push({ type: "start", partial: output });
             let currentBlock = null;
+            let currentIndex = 0;
             let groundingMetadataEmitted = false;
             let urlContextMetadataEmitted = false;
             const blocks = output.content;
             const blockIndex = () => blocks.length - 1;
+            // blockIndex() drifts past the open block once a native part is appended; capture keeps the block's own index.
+            const pushOpen = (event) => stream.push(event, { contentIndex: currentIndex });
             for await (const chunk of googleStream) {
                 await options?.onProviderStreamEvent?.(chunk, model);
                 // Vertex uses the same @google/genai GenerateContentResponse type as Gemini.
@@ -76,7 +80,7 @@ export const stream = (model, context, options) => {
                                 (!isThinking && currentBlock.type !== "text")) {
                                 if (currentBlock) {
                                     if (currentBlock.type === "text") {
-                                        stream.push({
+                                        pushOpen({
                                             type: "text_end",
                                             contentIndex: blocks.length - 1,
                                             content: currentBlock.text,
@@ -84,7 +88,7 @@ export const stream = (model, context, options) => {
                                         });
                                     }
                                     else {
-                                        stream.push({
+                                        pushOpen({
                                             type: "thinking_end",
                                             contentIndex: blockIndex(),
                                             content: currentBlock.thinking,
@@ -95,18 +99,21 @@ export const stream = (model, context, options) => {
                                 if (isThinking) {
                                     currentBlock = { type: "thinking", thinking: "", thinkingSignature: undefined };
                                     output.content.push(currentBlock);
+                                    currentIndex = blockIndex();
                                     stream.push({ type: "thinking_start", contentIndex: blockIndex(), partial: output });
                                 }
                                 else {
                                     currentBlock = { type: "text", text: "" };
                                     output.content.push(currentBlock);
+                                    currentIndex = blockIndex();
                                     stream.push({ type: "text_start", contentIndex: blockIndex(), partial: output });
                                 }
                             }
                             if (currentBlock.type === "thinking") {
                                 currentBlock.thinking += part.text;
+                                pushAssistantSignature(stream, currentIndex, currentBlock, part.thoughtSignature);
                                 currentBlock.thinkingSignature = retainThoughtSignature(currentBlock.thinkingSignature, part.thoughtSignature);
-                                stream.push({
+                                pushOpen({
                                     type: "thinking_delta",
                                     contentIndex: blockIndex(),
                                     delta: part.text,
@@ -115,8 +122,9 @@ export const stream = (model, context, options) => {
                             }
                             else {
                                 currentBlock.text += part.text;
+                                pushAssistantSignature(stream, currentIndex, currentBlock, part.thoughtSignature);
                                 currentBlock.textSignature = retainThoughtSignature(currentBlock.textSignature, part.thoughtSignature);
-                                stream.push({
+                                pushOpen({
                                     type: "text_delta",
                                     contentIndex: blockIndex(),
                                     delta: part.text,
@@ -128,7 +136,7 @@ export const stream = (model, context, options) => {
                             handledAsKnownPart = true;
                             if (currentBlock) {
                                 if (currentBlock.type === "text") {
-                                    stream.push({
+                                    pushOpen({
                                         type: "text_end",
                                         contentIndex: blockIndex(),
                                         content: currentBlock.text,
@@ -136,7 +144,7 @@ export const stream = (model, context, options) => {
                                     });
                                 }
                                 else {
-                                    stream.push({
+                                    pushOpen({
                                         type: "thinking_end",
                                         contentIndex: blockIndex(),
                                         content: currentBlock.thinking,
@@ -159,16 +167,18 @@ export const stream = (model, context, options) => {
                             };
                             output.content.push(toolCall);
                             stream.push({ type: "toolcall_start", contentIndex: blockIndex(), partial: output });
+                            // The adapter never parses its own spelling: the complete parsed seed stays authoritative.
                             stream.push({
                                 type: "toolcall_delta",
                                 contentIndex: blockIndex(),
                                 delta: JSON.stringify(toolCall.arguments),
                                 partial: output,
-                            });
+                            }, { parsedThrough: 0 });
                             stream.push({ type: "toolcall_end", contentIndex: blockIndex(), toolCall, partial: output });
                         }
                         if (!handledAsKnownPart) {
                             output.content.push(toProviderNativeContent(part));
+                            pushAssistantCompleteBlock(stream, blockIndex(), blocks[blockIndex()]);
                         }
                     }
                 }
@@ -178,6 +188,7 @@ export const stream = (model, context, options) => {
                         subtype: "groundingMetadata",
                         raw: candidate.groundingMetadata,
                     });
+                    pushAssistantCompleteBlock(stream, blockIndex(), blocks[blockIndex()]);
                     groundingMetadataEmitted = true;
                 }
                 if (candidate?.urlContextMetadata && !urlContextMetadataEmitted) {
@@ -186,6 +197,7 @@ export const stream = (model, context, options) => {
                         subtype: "urlContextMetadata",
                         raw: candidate.urlContextMetadata,
                     });
+                    pushAssistantCompleteBlock(stream, blockIndex(), blocks[blockIndex()]);
                     urlContextMetadataEmitted = true;
                 }
                 if (candidate?.finishReason) {
@@ -217,7 +229,7 @@ export const stream = (model, context, options) => {
             }
             if (currentBlock) {
                 if (currentBlock.type === "text") {
-                    stream.push({
+                    pushOpen({
                         type: "text_end",
                         contentIndex: blockIndex(),
                         content: currentBlock.text,
@@ -225,7 +237,7 @@ export const stream = (model, context, options) => {
                     });
                 }
                 else {
-                    stream.push({
+                    pushOpen({
                         type: "thinking_end",
                         contentIndex: blockIndex(),
                         content: currentBlock.thinking,

package/dist/api/lazy.js

@@ -1,3 +1,4 @@
+import { assistantEvents, pushAssistantProgress, setAssistantReader } from "../utils/assistant-capture.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { isOAuthRefreshUnavailableError } from "../utils/oauth-refresh-error.js";
 import { describeReplacedInstall } from "../utils/provider-failure-description.js";
@@ -39,9 +40,13 @@ function hasResult(source) {
 class LazyAssistantMessageEventStream extends AssistantMessageEventStream {
     setCancellationHandler(handler) {
         this.cancellationHandler = handler;
+        // The forwarder carries records, so this wrapper is carrier-aware and keeps its cancellation for them.
+        setAssistantReader(this, (project) => this.withCancellation(this.mapIterator(project)));
     }
     [Symbol.asyncIterator]() {
-        const iterator = super[Symbol.asyncIterator]();
+        return this.withCancellation(super[Symbol.asyncIterator]());
+    }
+    withCancellation(iterator) {
         return {
             next: () => iterator.next(),
             return: async () => {
@@ -68,7 +73,8 @@ async function forwardStream(target, source, iterator) {
 export function lazyStream(model, setup) {
     const outer = new LazyAssistantMessageEventStream();
     const inner = setup();
-    const iterator = inner.then((source) => source[Symbol.asyncIterator]());
+    // Progress crosses as it is read, so the forwarder loop runs exactly the turns it runs for events alone.
+    const iterator = inner.then((source) => assistantEvents(source, (progress) => pushAssistantProgress(outer, progress))[Symbol.asyncIterator]());
     let cancellation;
     outer.setCancellationHandler(() => {
         cancellation ??= iterator.then(async (sourceIterator) => {

package/dist/api/mistral-conversations.js

@@ -16,7 +16,7 @@ const MAX_MISTRAL_ERROR_BODY_CHARS = 4000;
  * Stream responses from the native Mistral Chat Completions endpoint.
  */
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = resolveTranscript(context, model.compat?.supportsMidConvoSystemMessages);
     (async () => {
         const output = createOutput(model);

package/dist/api/openai-codex-responses.js

@@ -142,7 +142,7 @@ function copyBytesToArrayBuffer(bytes) {
 // Main Stream Function
 // ============================================================================
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = resolveTranscript(context, model.compat?.supportsMidConvoSystemMessages);
     (async () => {
         const output = {

package/dist/api/openai-completions.js

@@ -318,7 +318,7 @@ function resolveCacheRetention(cacheRetention, env) {
     return "short";
 }
 export const stream = (model, context, options) => {
-    const stream = new AssistantMessageEventStream();
+    const stream = new AssistantMessageEventStream({ captureBlocks: true });
     const normalizedContext = resolveTranscript(context, getCompat(model).supportsMidConvoSystemMessages);
     (async () => {
         const output = {
@@ -589,19 +589,24 @@ export const stream = (model, context, options) => {
                 for (const block of blocks) {
                     const contentIndex = getContentIndex(block);
                     if (block.type === "text") {
-                        stream.push({ type: "text_start", contentIndex, partial: output });
+                        stream.push({ type: "text_start", contentIndex, partial: output }, { offset: 0 });
                         for (const delta of deferredTextDeltas) {
                             stream.push({ type: "text_delta", contentIndex, delta, partial: output });
                         }
                     }
                     else if (block.type === "thinking") {
-                        stream.push({ type: "thinking_start", contentIndex, partial: output });
+                        stream.push({ type: "thinking_start", contentIndex, partial: output }, { offset: 0 });
                         for (const delta of deferredThinkingDeltas) {
                             stream.push({ type: "thinking_delta", contentIndex, delta, partial: output });
                         }
                     }
                     else {
-                        stream.push({ type: "toolcall_start", contentIndex, partial: output });
+                        const rawArgumentPrefix = (deferredToolCallDeltas.get(block) ?? []).join("");
+                        stream.push({ type: "toolcall_start", contentIndex, partial: output }, {
+                            rawArgumentPrefix,
+                            parsedThrough: rawArgumentPrefix.length,
+                            offset: 0,
+                        });
                         for (const delta of deferredToolCallDeltas.get(block) ?? []) {
                             stream.push({ type: "toolcall_delta", contentIndex, delta, partial: output });
                         }

package/dist/api/openai-responses-shared.js

@@ -1,5 +1,6 @@
 import { CONTEXT_PROVENANCE_FIELD, contextProvenanceFingerprint, getContextProvenance, } from "../context-provenance.js";
 import { calculateCost, supportsConfigurationUpdate } from "../models.js";
+import { enableAssistantCapture, pushAssistantProgress } from "../utils/assistant-capture.js";
 import { splitDeferredTools } from "../utils/deferred-tools.js";
 import { shortHash } from "../utils/hash.js";
 import { parseStreamingJson } from "../utils/json-parse.js";
@@ -581,6 +582,7 @@ function reconcileNativeImageGenerationCall(item) {
     };
 }
 export async function processResponsesStream(openaiStream, output, stream, model, options) {
+    enableAssistantCapture(stream);
     let sawTerminalResponseEvent = false;
     let nativeImageBase64Chars = 0;
     const outputSlots = new Map();
@@ -644,7 +646,7 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 contentIndex: output.content.length - 1,
             };
             outputSlots.set(outputIndex, slot);
-            stream.push({ type: "toolcall_start", contentIndex: slot.contentIndex, partial: output });
+            stream.push({ type: "toolcall_start", contentIndex: slot.contentIndex, partial: output }, { rawArgumentPrefix: block.partialJson });
             return slot;
         }
         if (isResponseCustomToolCallItem(item)) {
@@ -686,6 +688,12 @@ export async function processResponsesStream(openaiStream, output, stream, model
             reconcileNativeImageSlot(outputIndex, slot, imageItem);
         output.content.push(block);
         outputSlots.set(outputIndex, slot);
+        pushAssistantProgress(stream, {
+            type: "message_block_state",
+            contentIndex: slot.contentIndex,
+            phase: "insert",
+            block,
+        });
         return slot;
     };
     const getOrCreateSlot = (outputIndex, item) => {
@@ -732,10 +740,16 @@ export async function processResponsesStream(openaiStream, output, stream, model
             if (!imageItem)
                 continue;
             const existingSlot = getSlot(outputIndex, "providerNative");
+            const slot = existingSlot ?? createSlot(outputIndex, outputItem);
             if (existingSlot)
                 reconcileNativeImageSlot(outputIndex, existingSlot, imageItem);
-            else
-                createSlot(outputIndex, outputItem);
+            if (slot)
+                pushAssistantProgress(stream, {
+                    type: "message_block_state",
+                    contentIndex: slot.contentIndex,
+                    phase: "end",
+                    block: slot.block,
+                });
             outputSlots.delete(outputIndex);
         }
     };
@@ -897,6 +911,14 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 if (delta.length > 0)
                     pushToolCallDelta(slot, delta);
             }
+            else {
+                pushAssistantProgress(stream, {
+                    type: "message_tool_arguments_reset",
+                    contentIndex: slot.contentIndex,
+                    rawArgumentPrefix: event.arguments,
+                    parsedThrough: event.arguments.length,
+                });
+            }
         }
         else if (event.type === "response.custom_tool_call_input.delta") {
             const slot = getSlot(event.output_index, "toolCall");
@@ -917,6 +939,12 @@ export async function processResponsesStream(openaiStream, output, stream, model
             const imageItem = readNativeImageGenerationCall(item);
             if (imageItem && slot?.type === "providerNative") {
                 reconcileNativeImageSlot(event.output_index, slot, imageItem);
+                pushAssistantProgress(stream, {
+                    type: "message_block_state",
+                    contentIndex: slot.contentIndex,
+                    phase: "end",
+                    block: slot.block,
+                });
                 finalizedNativeImageOutputIndexes.add(event.output_index);
                 outputSlots.delete(event.output_index);
             }
@@ -990,6 +1018,12 @@ export async function processResponsesStream(openaiStream, output, stream, model
             else if (slot?.type === "providerNative") {
                 slot.block.subtype = item.type;
                 slot.block.raw = item;
+                pushAssistantProgress(stream, {
+                    type: "message_block_state",
+                    contentIndex: slot.contentIndex,
+                    phase: "end",
+                    block: slot.block,
+                });
                 outputSlots.delete(event.output_index);
             }
         }