@juspay/neurolink 9.65.2 → 9.66.0

package/dist/proxy/openaiFormat.js

@@ -0,0 +1,800 @@
+/**
+ * OpenAI Chat Completions API format conversion layer.
+ *
+ * Provides a request parser (OpenAI -> NeuroLink), a response serializer
+ * (NeuroLink -> OpenAI), a streaming SSE state machine, and an error
+ * envelope helper.  Together they allow NeuroLink to act as a
+ * drop-in OpenAI API proxy.
+ *
+ * Reference: https://platform.openai.com/docs/api-reference/chat/create
+ */
+import { jsonSchema, tool } from "../utils/tool.js";
+import { randomBytes } from "crypto";
+import { normalizeJsonSchemaObject } from "../utils/schemaConversion.js";
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+/** Generate a unique chat completion ID in the OpenAI format. */
+export function generateChatCompletionId() {
+    return `chatcmpl-${randomBytes(12).toString("base64url").slice(0, 24)}`;
+}
+/** Generate an OpenAI-format tool call ID (`call_` + random chars). */
+export function generateOpenAIToolCallId() {
+    return `call_${randomBytes(12).toString("base64url").slice(0, 24)}`;
+}
+// ---------------------------------------------------------------------------
+// Request parser: OpenAI -> NeuroLink internal format
+// ---------------------------------------------------------------------------
+/**
+ * Parse an incoming OpenAI Chat Completions request into an intermediate
+ * representation consumable by NeuroLink's generate/stream pipeline.
+ *
+ * Handles:
+ * - System prompt extraction from system messages
+ * - Message flattening (text + image parts)
+ * - Tool definition conversion
+ * - tool_choice mapping
+ * - top_p, temperature, max_tokens pass-through
+ */
+export function parseOpenAIRequest(body) {
+    // --- system prompt ---
+    const systemParts = [];
+    for (const msg of body.messages) {
+        if (msg.role === "system") {
+            systemParts.push(msg.content);
+        }
+    }
+    const systemPrompt = systemParts.length > 0 ? systemParts.join("\n\n") : undefined;
+    // --- messages ---
+    // Find the index of the last user message so we can distinguish the
+    // current turn from history.  Images from historical messages are kept
+    // inline as text references; only images from the latest user message
+    // are extracted into the top-level `images` array.
+    const conversationMessages = [];
+    const images = [];
+    let lastUserPrompt = "";
+    let lastUserMsgIdx = -1;
+    for (let i = body.messages.length - 1; i >= 0; i--) {
+        if (body.messages[i].role === "user") {
+            lastUserMsgIdx = i;
+            break;
+        }
+    }
+    // NOTE: This loop intentionally does NOT use MessageBuilder because the proxy
+    // layer translates between OpenAI's wire format and NeuroLink's internal
+    // representation. MessageBuilder is for SDK-side message construction from
+    // user inputs (files, images, etc.).
+    for (let msgIdx = 0; msgIdx < body.messages.length; msgIdx++) {
+        const msg = body.messages[msgIdx];
+        const isLatestUserMsg = msgIdx === lastUserMsgIdx;
+        if (msg.role === "system") {
+            // System messages are already extracted above; skip them in the
+            // conversation history to avoid duplication.
+            continue;
+        }
+        if (msg.role === "user") {
+            if (typeof msg.content === "string") {
+                conversationMessages.push({ role: msg.role, content: msg.content });
+                lastUserPrompt = msg.content;
+            }
+            else if (Array.isArray(msg.content)) {
+                const textParts = [];
+                for (const part of msg.content) {
+                    if (part.type === "text") {
+                        textParts.push(part.text);
+                    }
+                    else if (part.type === "image_url") {
+                        if (isLatestUserMsg) {
+                            images.push(part.image_url.url);
+                        }
+                        else {
+                            textParts.push("[image]");
+                        }
+                    }
+                }
+                const combined = textParts.join("\n");
+                conversationMessages.push({ role: msg.role, content: combined });
+                lastUserPrompt = combined;
+            }
+        }
+        else if (msg.role === "assistant") {
+            const textParts = [];
+            if (msg.content) {
+                textParts.push(msg.content);
+            }
+            if (msg.tool_calls) {
+                for (const tc of msg.tool_calls) {
+                    textParts.push(`[tool_use:${tc.id}:${tc.function.name}] ${tc.function.arguments}`);
+                }
+            }
+            const combined = textParts.join("\n");
+            conversationMessages.push({ role: msg.role, content: combined });
+        }
+        else if (msg.role === "tool") {
+            conversationMessages.push({
+                role: "user",
+                content: `[tool_result:${msg.tool_call_id}] ${msg.content}`,
+            });
+        }
+    }
+    // --- tools ---
+    const tools = {};
+    if (body.tools) {
+        for (const t of body.tools) {
+            tools[t.function.name] = tool({
+                description: t.function.description ?? "",
+                inputSchema: jsonSchema(normalizeJsonSchemaObject(t.function.parameters ?? { type: "object" })),
+            });
+        }
+    }
+    // --- tool_choice ---
+    let toolChoice;
+    let toolChoiceName;
+    if (body.tool_choice) {
+        if (typeof body.tool_choice === "string") {
+            switch (body.tool_choice) {
+                case "auto":
+                    toolChoice = "auto";
+                    break;
+                case "required":
+                    toolChoice = "required";
+                    break;
+                case "none":
+                    toolChoice = "none";
+                    break;
+            }
+        }
+        else if (typeof body.tool_choice === "object" &&
+            body.tool_choice.type === "function") {
+            toolChoice = "required";
+            toolChoiceName = body.tool_choice.function.name;
+        }
+    }
+    // --- stop sequences ---
+    let stopSequences;
+    if (body.stop) {
+        stopSequences = Array.isArray(body.stop) ? body.stop : [body.stop];
+    }
+    // --- response format ---
+    let responseFormat;
+    if (body.response_format) {
+        responseFormat = {
+            type: body.response_format.type,
+            jsonSchema: body.response_format.json_schema,
+        };
+    }
+    return {
+        model: body.model,
+        maxTokens: body.max_tokens ?? body.max_completion_tokens ?? 4096,
+        temperature: body.temperature,
+        topP: body.top_p,
+        systemPrompt,
+        stream: body.stream === true,
+        prompt: lastUserPrompt,
+        images,
+        conversationMessages,
+        tools,
+        toolChoice,
+        toolChoiceName,
+        stopSequences,
+        responseFormat,
+    };
+}
+// ---------------------------------------------------------------------------
+// Response serializer: NeuroLink result -> OpenAI response
+// ---------------------------------------------------------------------------
+/**
+ * Map NeuroLink finish-reason strings to OpenAI finish_reason values.
+ */
+function mapFinishReason(finishReason) {
+    switch (finishReason) {
+        case "stop":
+        case "end_turn":
+            return "stop";
+        case "length":
+        case "max_tokens":
+            return "length";
+        case "tool-calls":
+        case "tool_use":
+            return "tool_calls";
+        case "content_filter":
+        case "safety":
+            return "content_filter";
+        default:
+            return finishReason ? "stop" : null;
+    }
+}
+/**
+ * Serialize a NeuroLink GenerateResult into an OpenAI Chat Completions response.
+ */
+export function serializeOpenAIResponse(result, requestModel) {
+    const inferredFinishReason = result.toolCalls &&
+        result.toolCalls.length > 0 &&
+        (!result.finishReason || result.finishReason === "stop")
+        ? "tool_calls"
+        : result.finishReason;
+    // Build tool_calls array if present
+    let toolCalls;
+    if (result.toolCalls && result.toolCalls.length > 0) {
+        toolCalls = result.toolCalls.map((tc) => ({
+            id: tc.toolCallId || generateOpenAIToolCallId(),
+            type: "function",
+            function: {
+                name: tc.toolName,
+                arguments: JSON.stringify(tc.args ?? {}),
+            },
+        }));
+    }
+    // Content is null when only tool calls are present (no text content)
+    const content = toolCalls && !result.content ? null : result.content || "";
+    return {
+        id: generateChatCompletionId(),
+        object: "chat.completion",
+        created: Math.floor(Date.now() / 1000),
+        model: result.model ?? requestModel,
+        choices: [
+            {
+                index: 0,
+                message: {
+                    role: "assistant",
+                    content,
+                    ...(toolCalls ? { tool_calls: toolCalls } : {}),
+                },
+                finish_reason: mapFinishReason(inferredFinishReason),
+            },
+        ],
+        usage: {
+            prompt_tokens: result.usage?.input ?? 0,
+            completion_tokens: result.usage?.output ?? 0,
+            total_tokens: result.usage?.total ?? 0,
+        },
+    };
+}
+// ---------------------------------------------------------------------------
+// Error envelope
+// ---------------------------------------------------------------------------
+/** Map HTTP status codes to OpenAI error types. */
+function errorTypeFromStatus(status) {
+    switch (status) {
+        case 400:
+            return "invalid_request_error";
+        case 401:
+            return "authentication_error";
+        case 403:
+            return "permission_error";
+        case 404:
+            return "not_found_error";
+        case 429:
+            return "rate_limit_error";
+        default:
+            return status >= 500 ? "server_error" : "invalid_request_error";
+    }
+}
+/**
+ * Build an OpenAI-compatible error envelope.
+ */
+export function buildOpenAIError(status, message) {
+    return {
+        error: {
+            message,
+            type: errorTypeFromStatus(status),
+            code: null,
+        },
+    };
+}
+// ---------------------------------------------------------------------------
+// SSE helpers
+// ---------------------------------------------------------------------------
+/**
+ * Format a single OpenAI SSE frame.
+ * OpenAI uses only `data:` lines (no `event:` prefix unlike Claude).
+ */
+export function formatOpenAISSE(data) {
+    return `data: ${JSON.stringify(data)}\n\n`;
+}
+// ---------------------------------------------------------------------------
+// Streaming SSE state machine
+// ---------------------------------------------------------------------------
+/**
+ * Stateful SSE serializer that emits a well-formed OpenAI streaming response.
+ *
+ * Tracks lifecycle state (`idle` -> `streaming` -> `done`) and the current
+ * tool call index for multi-tool streaming.
+ *
+ * Usage:
+ * ```ts
+ * const sse = new OpenAIStreamSerializer(requestModel);
+ *
+ * // Start the stream
+ * yield* sse.start();
+ *
+ * // Text deltas
+ * for await (const chunk of textStream) {
+ *   yield* sse.pushDelta(chunk);
+ * }
+ *
+ * // Tool use
+ * yield* sse.pushToolUse(toolId, toolName, toolInput);
+ *
+ * // Finalize
+ * yield* sse.finish("stop", usage);
+ * ```
+ */
+export class OpenAIStreamSerializer {
+    state = "idle";
+    id;
+    model;
+    started = false;
+    toolCallIndex = -1;
+    constructor(model) {
+        this.id = generateChatCompletionId();
+        this.model = model;
+    }
+    /** Current lifecycle state (exposed for testing). */
+    getState() {
+        return this.state;
+    }
+    // -----------------------------------------------------------------------
+    // Internal helpers
+    // -----------------------------------------------------------------------
+    makeChunk(delta, finishReason = null, usage) {
+        return {
+            id: this.id,
+            object: "chat.completion.chunk",
+            created: Math.floor(Date.now() / 1000),
+            model: this.model,
+            choices: [
+                {
+                    index: 0,
+                    delta,
+                    finish_reason: finishReason,
+                },
+            ],
+            ...(usage ? { usage } : {}),
+        };
+    }
+    // -----------------------------------------------------------------------
+    // Public API
+    // -----------------------------------------------------------------------
+    /**
+     * Emit the opening frame with `role: "assistant"`.
+     */
+    *start() {
+        if (this.state !== "idle") {
+            return;
+        }
+        this.started = true;
+        this.state = "streaming";
+        yield formatOpenAISSE(this.makeChunk({ role: "assistant" }));
+    }
+    /**
+     * Push a text content delta.
+     */
+    *pushDelta(text) {
+        if (this.state === "done" || this.state === "error") {
+            return;
+        }
+        if (!this.started) {
+            yield* this.start();
+        }
+        yield formatOpenAISSE(this.makeChunk({ content: text }));
+    }
+    /**
+     * Push the start of a tool call (id, name, empty arguments).
+     */
+    *pushToolCallStart(id, name) {
+        if (this.state === "done" || this.state === "error") {
+            return;
+        }
+        if (!this.started) {
+            yield* this.start();
+        }
+        this.toolCallIndex += 1;
+        yield formatOpenAISSE(this.makeChunk({
+            tool_calls: [
+                {
+                    index: this.toolCallIndex,
+                    id,
+                    type: "function",
+                    function: { name, arguments: "" },
+                },
+            ],
+        }));
+    }
+    /**
+     * Push an arguments delta for the current tool call.
+     */
+    *pushToolCallArgDelta(index, argsChunk) {
+        if (this.state === "done" || this.state === "error") {
+            return;
+        }
+        yield formatOpenAISSE(this.makeChunk({
+            tool_calls: [
+                {
+                    index,
+                    function: { arguments: argsChunk },
+                },
+            ],
+        }));
+    }
+    /**
+     * Push a complete tool use: emits tool call start followed by chunked
+     * argument deltas (~100 chars per chunk).
+     */
+    *pushToolUse(id, name, input) {
+        if (this.state === "done" || this.state === "error") {
+            return;
+        }
+        yield* this.pushToolCallStart(id, name);
+        const jsonStr = JSON.stringify(input ?? {});
+        const CHUNK_SIZE = 100;
+        const currentIndex = this.toolCallIndex;
+        for (let i = 0; i < jsonStr.length; i += CHUNK_SIZE) {
+            const chunk = jsonStr.slice(i, i + CHUNK_SIZE);
+            yield* this.pushToolCallArgDelta(currentIndex, chunk);
+        }
+        // If the input was empty, still emit at least one delta
+        if (jsonStr.length === 0) {
+            yield* this.pushToolCallArgDelta(currentIndex, "{}");
+        }
+    }
+    /**
+     * Finalize the stream: emit finish_reason chunk, then `data: [DONE]`.
+     */
+    *finish(finishReason, usage) {
+        if (this.state === "idle") {
+            yield* this.start();
+        }
+        if (this.state === "done" || this.state === "error") {
+            return;
+        }
+        const mappedReason = mapFinishReason(finishReason) ?? "stop";
+        const usagePayload = usage
+            ? {
+                prompt_tokens: usage.input,
+                completion_tokens: usage.output,
+                total_tokens: usage.total,
+            }
+            : undefined;
+        yield formatOpenAISSE(this.makeChunk({}, mappedReason, usagePayload));
+        yield "data: [DONE]\n\n";
+        this.state = "done";
+    }
+    /**
+     * Emit an error event.  Transitions to terminal ERROR state.
+     */
+    *emitError(message) {
+        this.state = "error";
+        yield formatOpenAISSE({
+            error: { message, type: "server_error" },
+        });
+    }
+}
+// ---------------------------------------------------------------------------
+// OpenAI <-> Claude (Anthropic) format bridge
+// ---------------------------------------------------------------------------
+/**
+ * Convert an OpenAI Chat Completions request to a Claude Messages API request.
+ *
+ * Used by the OpenAI proxy endpoint to internally loopback requests targeting
+ * Claude models through the proxy's native /v1/messages passthrough path,
+ * so they benefit from OAuth account rotation, retry, SSE interception, etc.
+ */
+export function convertOpenAIToClaudeRequest(openai) {
+    // --- system messages ---
+    const systemMessages = openai.messages.filter((m) => m.role === "system");
+    const system = systemMessages.length > 0
+        ? systemMessages.map((m) => ({ type: "text", text: m.content }))
+        : undefined;
+    // --- conversation messages ---
+    const messages = [];
+    for (const msg of openai.messages) {
+        if (msg.role === "system") {
+            continue;
+        }
+        if (msg.role === "user") {
+            if (typeof msg.content === "string") {
+                messages.push({ role: "user", content: msg.content });
+            }
+            else if (Array.isArray(msg.content)) {
+                const blocks = msg.content.map((part) => {
+                    if (part.type === "text") {
+                        return { type: "text", text: part.text };
+                    }
+                    if (part.type === "image_url") {
+                        return {
+                            type: "image",
+                            source: { type: "url", url: part.image_url.url },
+                        };
+                    }
+                    return { type: "text", text: "" };
+                });
+                messages.push({ role: "user", content: blocks });
+            }
+        }
+        else if (msg.role === "assistant") {
+            const blocks = [];
+            if (msg.content) {
+                blocks.push({ type: "text", text: msg.content });
+            }
+            if (msg.tool_calls) {
+                for (const tc of msg.tool_calls) {
+                    let input;
+                    try {
+                        input = tc.function.arguments
+                            ? JSON.parse(tc.function.arguments)
+                            : {};
+                    }
+                    catch {
+                        input = {};
+                    }
+                    blocks.push({
+                        type: "tool_use",
+                        id: tc.id,
+                        name: tc.function.name,
+                        input,
+                    });
+                }
+            }
+            messages.push({
+                role: "assistant",
+                content: blocks.length === 1 && blocks[0].type === "text"
+                    ? blocks[0].text
+                    : blocks,
+            });
+        }
+        else if (msg.role === "tool") {
+            messages.push({
+                role: "user",
+                content: [
+                    {
+                        type: "tool_result",
+                        tool_use_id: msg.tool_call_id,
+                        content: msg.content,
+                    },
+                ],
+            });
+        }
+    }
+    // --- tools ---
+    const tools = openai.tools?.map((t) => ({
+        name: t.function.name,
+        description: t.function.description || "",
+        input_schema: t.function.parameters,
+    }));
+    // --- tool_choice ---
+    let tool_choice;
+    if (openai.tool_choice === "auto") {
+        tool_choice = { type: "auto" };
+    }
+    else if (openai.tool_choice === "required") {
+        tool_choice = { type: "any" };
+    }
+    else if (openai.tool_choice === "none") {
+        tool_choice = { type: "none" };
+    }
+    else if (typeof openai.tool_choice === "object" &&
+        openai.tool_choice.type === "function") {
+        tool_choice = { type: "tool", name: openai.tool_choice.function.name };
+    }
+    const result = {
+        model: openai.model,
+        messages,
+        max_tokens: openai.max_tokens ?? openai.max_completion_tokens ?? 4096,
+        stream: openai.stream ?? false,
+    };
+    if (system) {
+        result.system = system;
+    }
+    if (openai.temperature !== undefined) {
+        result.temperature = openai.temperature;
+    }
+    if (openai.top_p !== undefined) {
+        result.top_p = openai.top_p;
+    }
+    if (tools && tools.length > 0) {
+        result.tools = tools;
+    }
+    if (tool_choice) {
+        result.tool_choice = tool_choice;
+    }
+    if (openai.stop) {
+        result.stop_sequences = Array.isArray(openai.stop)
+            ? openai.stop
+            : [openai.stop];
+    }
+    return result;
+}
+/**
+ * Convert a non-streaming Claude Messages response to an OpenAI Chat
+ * Completions response by bridging through {@link InternalResult}.
+ */
+export function convertClaudeToOpenAIResponse(claude, requestModel) {
+    const content = claude.content
+        .filter((b) => b.type === "text")
+        .map((b) => b.text)
+        .join("");
+    const toolCalls = claude.content
+        .filter((b) => b.type === "tool_use")
+        .map((b) => {
+        const tu = b;
+        return {
+            toolCallId: tu.id,
+            toolName: tu.name,
+            args: tu.input ?? {},
+        };
+    });
+    const internal = {
+        content,
+        model: claude.model,
+        finishReason: claude.stop_reason ?? "end_turn",
+        usage: {
+            input: claude.usage.input_tokens,
+            output: claude.usage.output_tokens,
+            total: claude.usage.input_tokens + claude.usage.output_tokens,
+        },
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+    };
+    return serializeOpenAIResponse(internal, requestModel);
+}
+/**
+ * Create a TransformStream that parses Claude Messages API SSE events from
+ * the upstream response and re-emits them as OpenAI Chat Completions SSE
+ * frames.
+ *
+ * Handles the canonical Claude SSE event types:
+ *  - message_start       -> emits the opening `role: "assistant"` chunk
+ *  - content_block_start -> text block: no-op; tool_use block: emit tool call start
+ *  - content_block_delta -> text_delta: emit content delta;
+ *                           input_json_delta: emit tool call argument delta
+ *  - content_block_stop  -> no-op
+ *  - message_delta       -> captures stop_reason and output token usage
+ *  - message_stop        -> emits the final `finish_reason` chunk + `[DONE]`
+ */
+export function createClaudeToOpenAIStreamTransform(requestModel) {
+    const serializer = new OpenAIStreamSerializer(requestModel);
+    const encoder = new TextEncoder();
+    const decoder = new TextDecoder();
+    let buffer = "";
+    // Track per-content-block state so we can map Claude's indexed block
+    // stream to OpenAI's flat delta stream.
+    const blockState = new Map();
+    let stopReason;
+    let usage;
+    let inputTokens = 0;
+    let nextToolCallIndex = 0;
+    let finished = false;
+    const emit = (controller, gen) => {
+        for (const frame of gen) {
+            controller.enqueue(encoder.encode(frame));
+        }
+    };
+    const handleEvent = (eventName, dataStr, controller) => {
+        let data;
+        try {
+            data = JSON.parse(dataStr);
+        }
+        catch {
+            return;
+        }
+        switch (eventName) {
+            case "message_start": {
+                const message = (data.message ?? {});
+                if (message.usage?.input_tokens !== undefined) {
+                    inputTokens = message.usage.input_tokens;
+                }
+                emit(controller, serializer.start());
+                return;
+            }
+            case "content_block_start": {
+                const index = typeof data.index === "number" ? data.index : 0;
+                const block = (data.content_block ?? {});
+                if (block.type === "tool_use") {
+                    const toolCallIndex = nextToolCallIndex++;
+                    blockState.set(index, { kind: "tool_use", toolCallIndex });
+                    emit(controller, serializer.pushToolCallStart(block.id ?? "", block.name ?? ""));
+                }
+                else if (block.type === "text") {
+                    blockState.set(index, { kind: "text" });
+                }
+                else if (block.type === "thinking") {
+                    blockState.set(index, { kind: "thinking" });
+                }
+                else {
+                    blockState.set(index, { kind: "other" });
+                }
+                return;
+            }
+            case "content_block_delta": {
+                const index = typeof data.index === "number" ? data.index : 0;
+                const delta = (data.delta ?? {});
+                const state = blockState.get(index);
+                if (!state) {
+                    return;
+                }
+                if (delta.type === "text_delta" && state.kind === "text") {
+                    emit(controller, serializer.pushDelta(delta.text ?? ""));
+                }
+                else if (delta.type === "input_json_delta" &&
+                    state.kind === "tool_use" &&
+                    state.toolCallIndex !== undefined) {
+                    emit(controller, serializer.pushToolCallArgDelta(state.toolCallIndex, delta.partial_json ?? ""));
+                }
+                // thinking_delta is intentionally dropped — OpenAI has no equivalent.
+                return;
+            }
+            case "content_block_stop": {
+                // No-op: OpenAI stream has no per-block close event.
+                return;
+            }
+            case "message_delta": {
+                const delta = (data.delta ?? {});
+                if (delta.stop_reason) {
+                    stopReason = delta.stop_reason;
+                }
+                const u = (data.usage ?? {});
+                if (u.output_tokens !== undefined) {
+                    usage = {
+                        input: inputTokens,
+                        output: u.output_tokens,
+                        total: inputTokens + u.output_tokens,
+                    };
+                }
+                return;
+            }
+            case "message_stop": {
+                if (!finished) {
+                    finished = true;
+                    emit(controller, serializer.finish(stopReason, usage));
+                }
+                return;
+            }
+            default:
+                // ping, error, and unknown events are ignored.
+                return;
+        }
+    };
+    // Parse any complete SSE events present in `buffer`, mutating it as events
+    // are consumed. Shared between the streaming `transform` and the terminal
+    // `flush` (after the decoder is drained) so trailing events aren't lost.
+    const drainBufferedEvents = (controller) => {
+        // Claude SSE events are separated by blank lines (`\n\n`).
+        let sepIdx = buffer.indexOf("\n\n");
+        while (sepIdx !== -1) {
+            const rawEvent = buffer.slice(0, sepIdx);
+            buffer = buffer.slice(sepIdx + 2);
+            let eventName = "";
+            const dataLines = [];
+            for (const line of rawEvent.split("\n")) {
+                if (line.startsWith("event:")) {
+                    eventName = line.slice(6).trim();
+                }
+                else if (line.startsWith("data:")) {
+                    dataLines.push(line.slice(5).trim());
+                }
+            }
+            if (eventName && dataLines.length > 0) {
+                handleEvent(eventName, dataLines.join("\n"), controller);
+            }
+            sepIdx = buffer.indexOf("\n\n");
+        }
+    };
+    return new TransformStream({
+        transform(chunk, controller) {
+            buffer += decoder.decode(chunk, { stream: true });
+            drainBufferedEvents(controller);
+        },
+        flush(controller) {
+            // Drain any bytes still held inside the TextDecoder, then re-run the
+            // event parser so a complete trailing event that arrived without a
+            // closing `\n\n` is not silently lost.
+            buffer += decoder.decode();
+            drainBufferedEvents(controller);
+            // If the upstream closed without a message_stop, still finalize.
+            if (!finished) {
+                finished = true;
+                emit(controller, serializer.finish(stopReason, usage));
+            }
+        },
+    });
+}