@trigger.dev/sdk 4.4.5 → 4.5.0-rc.0

package/dist/esm/v3/chat-server.js

@@ -0,0 +1,734 @@
+/**
+ * Server-side helpers for the `chat.agent` head-start flow — a
+ * customer's warm process (Next.js route handler, Express, etc.)
+ * gets the conversation moving while the heavy chat.agent run boots
+ * in parallel. Mid-turn, ownership of the durable stream hands over
+ * to the agent.
+ *
+ * The `chat.headStart({ agentId, run })` entry point returns a
+ * Next.js-style POST handler. Inside the customer's `run` callback
+ * they call `streamText` themselves, spreading
+ * `chat.toStreamTextOptions({ tools })` to inherit handover wiring.
+ * The handler runs `streamText` step 1 in the customer's process
+ * while the chat.agent run boots in parallel; on `tool-calls` the
+ * agent run picks up tool execution and continues, on pure-text the
+ * agent run exits clean without an LLM call.
+ *
+ * Two-layer naming: customer-facing surface is "head start"
+ * (describes the *benefit* — fast first-turn TTFC). The internal
+ * protocol still uses "handover" (describes the *mechanism* — the
+ * conversation hands off mid-turn from the warm process to the
+ * agent). Customers see `chat.headStart`, `HeadStartSession`, etc.
+ * The wire format and run-loop locals stay on `handover` /
+ * `handover-prepare` / `handover-skip`.
+ *
+ * Cooperative ordering only — handler stops writing to `session.out`
+ * before sending the `handover` chunk on `session.in`. No S2 fencing.
+ *
+ * ⚠️ HARD CONSTRAINT — bundle isolation
+ *
+ * This module is the customer-facing boundary for the route handler.
+ * The whole TTFC win comes from the customer's process being
+ * lightweight while the heavy agent run boots in parallel. **The
+ * route-handler bundle must not include heavy tool execute deps**:
+ * E2B, puppeteer/playwright, native bindings, the trigger SDK
+ * runtime, turndown, image processing libs, anything that pulls
+ * weight or pulls `node:` builtins.
+ *
+ * "Schema-only" tools must live in a module that imports only `ai`
+ * (for `tool()`) and `zod`. The agent task module imports those
+ * schemas and adds execute fns elsewhere — that's where the heavy
+ * deps live, and it's never reached by the route handler bundle.
+ *
+ * Runtime "strip executes" helpers (anything that takes a tool
+ * catalog with executes and removes them) DO NOT solve this. The
+ * import chain is resolved at bundle/build time, so importing the
+ * full catalog drags every dep in regardless of what the SDK does
+ * with the value at runtime.
+ *
+ * IMPORTANT (internal): this module must NOT import from `./ai.ts`.
+ * `ai.ts` statically imports `agentSkillsRuntime` (which uses `node:`
+ * builtins unfit for some serverless runtimes) and the heavy task
+ * runtime. Allowed imports: `./ai-shared.js`, `./chat-client.js`,
+ * `@trigger.dev/core/v3` (api client), `ai` (types + lightweight
+ * helpers like `stepCountIs` / `convertToModelMessages`).
+ */
+import { SessionStreamInstance, TRIGGER_CONTROL_SUBTYPE, apiClientManager, } from "@trigger.dev/core/v3";
+import { convertToModelMessages, generateId as generateAssistantMessageId, stepCountIs, } from "ai";
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+export const chat = {
+    /**
+     * Returns a Next.js-style POST handler for the chat.agent
+     * head-start flow. Customer mounts it as
+     * `export const { POST } = chat.headStart({...})` (or
+     * `export const POST = chat.headStart({...})`).
+     *
+     * Pair with the browser transport's `headStart: "/api/chat"`
+     * option so the first message of a brand-new chat lands here
+     * before the agent run boots.
+     */
+    headStart(opts) {
+        return async (req) => {
+            const session = await openHandoverSession({
+                req,
+                agentId: opts.agentId,
+                idleTimeoutInSeconds: opts.idleTimeoutInSeconds,
+            });
+            const helper = {
+                toStreamTextOptions(spreadOpts) {
+                    return session.buildStreamTextOptions(spreadOpts);
+                },
+                session: session.handle,
+            };
+            const result = await opts.run({
+                messages: session.uiMessages,
+                signal: session.combinedSignal,
+                chat: helper,
+            });
+            return session.handle.handoverResponse(result);
+        };
+    },
+    /**
+     * Lower-level primitive for power users who want to call
+     * `streamText` themselves outside the `run` callback shape — custom
+     * transforms, non-AI-SDK code paths, or manual control over the
+     * response. Same wiring `chat.headStart` builds on internally.
+     */
+    openSession(opts) {
+        return openHandoverSession(opts).then((s) => s.handle);
+    },
+    /**
+     * Wrap a Web Fetch handler — `(req: Request) => Promise<Response>` —
+     * as a Node `http` listener — `(req: IncomingMessage, res: ServerResponse) => Promise<void>`.
+     *
+     * Use this to mount `chat.headStart` (or any other Web Fetch
+     * handler) inside Node-only frameworks like Express, Fastify, Koa,
+     * or raw `node:http`. Web-native frameworks (Next.js App Router,
+     * Hono, SvelteKit, Remix, Workers, Bun, Deno, etc.) don't need
+     * this — they pass `Request` objects directly.
+     *
+     * Streams the response body chunk-by-chunk to the Node response,
+     * so the `chat.headStart` SSE chunks reach the browser as they
+     * arrive (no buffering). Aborts the underlying handler if the
+     * client closes the connection.
+     *
+     * Type-only import of `node:http` types — no runtime dep on `node:http`,
+     * so this stays safe to bundle into edge / Workers builds (the
+     * function just won't be called there).
+     *
+     * @example
+     * ```ts
+     * import express from "express";
+     * import { chat } from "@trigger.dev/sdk/chat-server";
+     *
+     * const handler = chat.headStart({
+     *   agentId: "my-chat",
+     *   run: async ({ chat: helper }) => streamText({ ... }),
+     * });
+     *
+     * const app = express();
+     * app.post("/api/chat", chat.toNodeListener(handler));
+     * ```
+     */
+    toNodeListener,
+};
+async function openHandoverSession(opts) {
+    const wirePayload = (await opts.req.json());
+    const chatId = wirePayload.chatId;
+    if (!chatId) {
+        throw new Error("[chat.handover] request body missing `chatId`");
+    }
+    // Slim wire — head-start ships full history via `headStartMessages` (not
+    // `message`/`messages`) because the route handler runs on the customer's
+    // own HTTP endpoint and isn't subject to the 512 KiB `/in/append` cap.
+    // The full UIMessage[] flows through `wirePayload` into the auto-trigger
+    // `basePayload` below, where the agent run boot consumes it on first turn.
+    const uiMessages = (wirePayload.headStartMessages ?? []);
+    // `convertToModelMessages` is async — resolve once up front so the
+    // synchronous `toStreamTextOptions` builder can hand back a fully
+    // formed object. AI SDK's `streamText` validates `messages` as a
+    // `ModelMessage[]` synchronously and rejects a Promise.
+    const modelMessages = await convertToModelMessages(uiMessages);
+    const apiClient = resolveApiClient();
+    const idleTimeoutInSeconds = opts.idleTimeoutInSeconds ?? 60;
+    // Create the session and trigger the chat.agent's `handover-prepare`
+    // run atomically. `createSession` is idempotent on `(env, externalId
+    // = chatId)` and the auto-triggered run uses `triggerConfig.
+    // basePayload` as the wire payload — so a single round-trip both
+    // ensures the session exists and starts the agent booting with the
+    // right trigger.
+    //
+    // Awaited intentionally: subsequent writes to `session.out` (the
+    // tee from the customer's `streamText` to S2) need the session to
+    // exist, and the handover signal at end-of-step-1 needs the agent
+    // run to be there to consume it. The added latency (~one round trip
+    // to the control plane) is bounded; the agent's compute boot still
+    // overlaps with LLM TTFB.
+    const created = await apiClient.createSession({
+        type: "chat.agent",
+        externalId: chatId,
+        taskIdentifier: opts.agentId,
+        triggerConfig: {
+            basePayload: {
+                ...wirePayload,
+                chatId,
+                trigger: "handover-prepare",
+                idleTimeoutInSeconds,
+            },
+            idleTimeoutInSeconds,
+        },
+    });
+    const sessionPublicAccessToken = created.publicAccessToken;
+    // Combined abort signal: request lifecycle OR an internal timeout
+    // mirroring the agent's idle wait so a hung handler doesn't sit
+    // forever.
+    const abortController = new AbortController();
+    const requestAbort = opts.req.signal;
+    if (requestAbort) {
+        if (requestAbort.aborted)
+            abortController.abort();
+        else
+            requestAbort.addEventListener("abort", () => abortController.abort(), { once: true });
+    }
+    const idleTimer = setTimeout(() => abortController.abort(new Error("chat.handover: idle timeout")), idleTimeoutInSeconds * 1000);
+    const buildStreamTextOptions = (spreadOpts) => {
+        // The customer spreads this object into their `streamText` call
+        // and then adds `model`, `system`, etc. on top. We set the four
+        // keys handover correctness depends on:
+        //
+        //   - `messages`: the wire payload's UIMessages, converted
+        //     (Promise resolved upfront so the spread is synchronous)
+        //   - `tools`: customer's schema-only tool set
+        //   - `stopWhen`: `stepCountIs(1)` — step 1 only. Agent run picks
+        //     up tool execution and step 2+ after the handover signal.
+        //   - `abortSignal`: combined request-lifecycle + idle timeout
+        //
+        // The customer's `StreamTextResult` exposes `finishReason` and
+        // `response.messages` directly, so we don't need to install an
+        // `onStepFinish` capture hook — we read those off the result in
+        // `handoverWhenDone`.
+        return {
+            messages: modelMessages,
+            tools: spreadOpts?.tools,
+            stopWhen: stepCountIs(1),
+            abortSignal: abortController.signal,
+        };
+    };
+    // Tee a UIMessage stream into session.out via S2 direct-write,
+    // batched. `SessionStreamInstance` calls `initializeSessionStream`
+    // once to fetch S2 credentials, then pipes via `StreamsWriterV2`'s
+    // `BatchTransform` — one S2 append per ~200ms of chunks instead of
+    // one HTTP round-trip per UIMessageChunk.
+    let sessionWriter = null;
+    const tee = (stream) => {
+        const [a, b] = stream.tee();
+        sessionWriter = new SessionStreamInstance({
+            apiClient,
+            baseUrl: apiClient.baseUrl,
+            sessionId: chatId, // Sessions are addressable by externalId (chatId).
+            io: "out",
+            source: b,
+            signal: abortController.signal,
+        });
+        return a;
+    };
+    /** Wait for the teed S2 writer to drain. Called before signaling handover. */
+    const flushSessionWriter = async () => {
+        if (!sessionWriter)
+            return;
+        try {
+            await sessionWriter.wait();
+        }
+        catch {
+            // Drop write errors — the customer's response stream is the
+            // source of truth for what the user sees. Durability/resume
+            // best-effort.
+        }
+    };
+    const handover = async (args) => {
+        const chunk = {
+            kind: "handover",
+            partialAssistantMessage: args.partialAssistantMessage,
+            messageId: args.messageId,
+            isFinal: args.isFinal,
+        };
+        await apiClient.appendToSessionStream(chatId, "in", JSON.stringify(chunk));
+    };
+    /**
+     * Sent only on dispatch error (handler aborted before producing a
+     * `finishReason`). Normal pure-text and tool-call finishes go
+     * through `handover()` with the appropriate `isFinal` flag.
+     */
+    const handoverSkip = async () => {
+        const chunk = { kind: "handover-skip" };
+        await apiClient.appendToSessionStream(chatId, "in", JSON.stringify(chunk));
+    };
+    // A stable assistant messageId for this turn. The customer's
+    // `toUIMessageStream` is configured to emit its `start` chunk with
+    // this id, the handover signal carries it to the agent, and the
+    // agent's post-handover `toUIMessageStream` reuses it — so all
+    // chunks (customer's step 1 + agent's step 2) merge into one
+    // assistant message on the browser side.
+    const turnMessageId = generateAssistantMessageId();
+    let resolveDecision;
+    const decisionPromise = new Promise((resolve) => {
+        resolveDecision = resolve;
+    });
+    const handoverWhenDone = async (result) => {
+        // Owns idle-timer cleanup via the finally below, so both the
+        // sugar (`handoverResponse`) and the escape-hatch
+        // (`chat.openSession()` → `handle.handoverWhenDone(...)`) clean up
+        // the timer the same way.
+        try {
+            // `result.finishReason` is a Promise<FinishReason> on the AI SDK
+            // result. Wait for the stream to settle, then dispatch.
+            const finishReason = await result.finishReason;
+            // Drain the S2 tee so any in-flight handler writes (last
+            // `tool-input-available` parts, the synthetic `finish-step` for
+            // pure-text) are visible before the agent reads from session.out
+            // / session.in. Cooperative ordering — agent doesn't read past
+            // these unless we've finished writing them.
+            await flushSessionWriter();
+            const responseMessages = (await result.response).messages;
+            if (finishReason === "tool-calls") {
+                // Reshape pending tool-calls into AI SDK's tool-approval round
+                // so the agent's `streamText` resumes by executing them
+                // before the step-2 LLM call.
+                const reshaped = reshapeForHandoverResume(responseMessages);
+                await handover({
+                    partialAssistantMessage: reshaped,
+                    messageId: turnMessageId,
+                    isFinal: false,
+                });
+            }
+            else {
+                // Pure-text (or any non-tool-calls) finish — customer's step 1
+                // IS the final response. The agent runs the turn-loop hooks
+                // (`onChatStart`, `onTurnStart`, `onTurnComplete`, etc.) using
+                // this partial as the response, but skips the LLM call. That
+                // way persistence (`onTurnComplete` writing to DB), self-
+                // review, and any post-turn work all fire normally.
+                await handover({
+                    partialAssistantMessage: responseMessages,
+                    messageId: turnMessageId,
+                    isFinal: true,
+                });
+            }
+            resolveDecision({ kind: "handover" });
+        }
+        catch (err) {
+            // Dispatch failed before we could send the handover signal.
+            // Tell the agent to exit clean (no hooks fire) and close the
+            // response stream so it doesn't hang waiting for agent chunks.
+            resolveDecision({ kind: "handover-skip" });
+            try {
+                await handoverSkip();
+            }
+            catch {
+                // best-effort
+            }
+            throw err;
+        }
+        finally {
+            clearTimeout(idleTimer);
+        }
+    };
+    /**
+     * Build a single ReadableStream that:
+     *   1. Forwards the customer's `streamText` chunks (step 1) directly
+     *      to the response — same low-latency path as before.
+     *   2. After step 1 ends and the dispatch decision lands:
+     *      - `handover-skip`: closes the response immediately. The agent
+     *        run exits without writing more chunks.
+     *      - `handover`: subscribes to `session.out` from the sequence
+     *        ID where the customer's tee left off, forwarding the agent
+     *        run's chunks (tool-output-available, step 2 LLM text,
+     *        `finish-step`, etc.) until `trigger:turn-complete`.
+     *
+     * The browser sees one continuous SSE response per first turn, just
+     * like a normal `streamText` would produce.
+     */
+    const stitchHandoverStream = (customerBranch) => {
+        return new ReadableStream({
+            async start(controller) {
+                try {
+                    // Phase 1: forward customer's chunks.
+                    const reader = customerBranch.getReader();
+                    try {
+                        while (true) {
+                            const { done, value } = await reader.read();
+                            if (done)
+                                break;
+                            controller.enqueue(value);
+                        }
+                    }
+                    finally {
+                        reader.releaseLock();
+                    }
+                    // Phase 2a: wait for handoverWhenDone to decide.
+                    const decision = await decisionPromise;
+                    if (decision.kind === "handover-skip") {
+                        controller.close();
+                        return;
+                    }
+                    // Phase 2b: agent is taking over. Resume from session.out
+                    // starting AFTER the customer tee's last write, so we don't
+                    // re-emit chunks the browser already saw.
+                    const writeResult = sessionWriter
+                        ? await sessionWriter.wait().catch(() => undefined)
+                        : undefined;
+                    const customerLastEventId = writeResult?.lastEventId;
+                    // Capture the latest S2 event id seen on session.out via
+                    // `onPart`. After the stream closes we emit it to the
+                    // browser as a `trigger:session-state` control chunk so the
+                    // transport can hydrate `state.lastEventId` for turn 2's
+                    // subscribe — without it, turn 2 reads session.out from the
+                    // start and replays turn 1 to the user.
+                    //
+                    // The agent's `turn-complete` control record is now header-
+                    // form on S2 (see `client-protocol.mdx`), so the
+                    // `for await (const chunk of agentStream)` loop below NEVER
+                    // sees it as a data chunk — `subscribeToSessionStream` routes
+                    // it to `onControl`. Use that to know when to stop and
+                    // synthesise the data-chunk shape the browser bridge still
+                    // expects (this HTTP response stream is NOT S2 and keeps the
+                    // legacy chunk shape for the customer-server-to-browser hop).
+                    let latestEventId;
+                    let turnComplete = false;
+                    // Dedicated abort signal for this agent subscription. Aborted
+                    // from `onControl` the moment turn-complete fires so the
+                    // `for await` loop below exits immediately instead of blocking
+                    // until S2's long-poll closes (~60s). Combined with the outer
+                    // `abortController.signal` via `AbortSignal.any` so a request-
+                    // wide abort still tears the subscription down.
+                    const subscriptionAbort = new AbortController();
+                    const agentStream = await apiClient.subscribeToSessionStream(chatId, "out", {
+                        ...(customerLastEventId != null
+                            ? { lastEventId: customerLastEventId }
+                            : {}),
+                        signal: AbortSignal.any([abortController.signal, subscriptionAbort.signal]),
+                        onPart: (part) => {
+                            if (part.id)
+                                latestEventId = part.id;
+                        },
+                        onControl: (event) => {
+                            if (event.subtype === TRIGGER_CONTROL_SUBTYPE.TURN_COMPLETE) {
+                                turnComplete = true;
+                                // Synthesise the data-chunk shape for the browser
+                                // bridge. The customer-server-to-browser response is
+                                // not S2; it keeps the legacy chunk shape so the
+                                // browser's transport can recognise turn-complete the
+                                // same way it always has.
+                                controller.enqueue({
+                                    type: "trigger:turn-complete",
+                                });
+                                // Stop the SSE read now. Without this the `for await`
+                                // can't see the control event (control records are
+                                // never enqueued into the data stream) and would idle
+                                // until S2's long-poll timeout closes the connection.
+                                subscriptionAbort.abort();
+                            }
+                        },
+                    });
+                    try {
+                        for await (const chunk of agentStream) {
+                            // Data records only — control records are routed via
+                            // `onControl` above and trigger the subscription abort.
+                            controller.enqueue(chunk);
+                            if (turnComplete)
+                                break;
+                        }
+                    }
+                    catch (err) {
+                        // AbortError from `subscriptionAbort` is the expected exit
+                        // path once turn-complete fires; surface anything else.
+                        const isAbort = err instanceof Error && err.name === "AbortError";
+                        if (!isAbort || !turnComplete)
+                            throw err;
+                    }
+                    // Final control chunk: hand the browser transport the
+                    // `lastEventId` it should use for the next turn's
+                    // session.out subscribe. Filtered out before reaching the
+                    // AI SDK on the browser side.
+                    if (latestEventId != null) {
+                        controller.enqueue({
+                            type: "trigger:session-state",
+                            lastEventId: latestEventId,
+                        });
+                    }
+                    controller.close();
+                }
+                catch (err) {
+                    controller.error(err);
+                }
+            },
+            cancel() {
+                // Browser closed the connection. Trigger the abort so any
+                // pending session.out subscription stops too.
+                abortController.abort();
+            },
+        });
+    };
+    const handoverResponse = (result) => {
+        // `generateMessageId` makes the customer's `start` chunk carry
+        // `turnMessageId`, so the browser-side AI SDK keys the assistant
+        // message by it. The agent's post-handover stream emits chunks
+        // with the same id (passed via the handover signal) — both sides
+        // merge into one message on the browser.
+        const teed = tee(result.toUIMessageStream({
+            generateMessageId: () => turnMessageId,
+        }));
+        // `handoverWhenDone` re-throws on dispatch failure for visibility,
+        // but the recovery (resolveDecision + handoverSkip) has already run
+        // by then and `stitchHandoverStream` closes the response cleanly via
+        // `decisionPromise`. The user-facing path is fine; we only suppress
+        // the unhandled-rejection so processes started with
+        // `--unhandled-rejections=throw` don't crash on what is effectively
+        // a logged failure with no further action to take.
+        // (Idle-timer cleanup lives inside `handoverWhenDone` itself.)
+        void handoverWhenDone(result).catch(() => { });
+        const stitched = stitchHandoverStream(teed);
+        // Encode UIMessageChunks as SSE for the AI SDK transport on the
+        // browser. AI SDK's `toUIMessageStreamResponse()` does this same
+        // thing internally; replicate the format here so we don't have
+        // to bridge through the SDK's response helper.
+        const encoder = new TextEncoder();
+        const sseStream = stitched.pipeThrough(new TransformStream({
+            transform(chunk, controller) {
+                controller.enqueue(encoder.encode(`data: ${JSON.stringify(chunk)}\n\n`));
+            },
+        }));
+        return new Response(sseStream, {
+            headers: {
+                "Content-Type": "text/event-stream",
+                "X-Vercel-AI-UI-Message-Stream": "v1",
+                "Cache-Control": "no-cache, no-transform",
+                Connection: "keep-alive",
+                // Browser transport reads these to hydrate session state
+                // for subsequent (non-handover) turns. Once the browser has
+                // the PAT it talks directly to `session.in` / `session.out`
+                // without going back through the handler.
+                "X-Trigger-Chat-Id": chatId,
+                "X-Trigger-Chat-Access-Token": sessionPublicAccessToken,
+            },
+        });
+    };
+    const handle = {
+        chatId,
+        tee,
+        handoverWhenDone,
+        handoverResponse,
+        handover,
+        handoverSkip,
+    };
+    return {
+        uiMessages,
+        combinedSignal: abortController.signal,
+        handle,
+        buildStreamTextOptions,
+    };
+}
+function resolveApiClient() {
+    // Reuse the SDK's standard apiClientManager so customers configure
+    // base URL + secret key the same way as for `tasks.trigger(...)`.
+    const client = apiClientManager.clientOrThrow();
+    return client;
+}
+/** @internal — exposed via `chat.toNodeListener`. */
+function toNodeListener(webHandler) {
+    return async function nodeListener(req, res) {
+        const abort = new AbortController();
+        res.on("close", () => abort.abort());
+        try {
+            const url = `http://${req.headers.host ?? "localhost"}${req.url ?? "/"}`;
+            const method = req.method ?? "GET";
+            const hasBody = method !== "GET" && method !== "HEAD";
+            // Read full body upfront. Chat wire payloads are small (sub-KB
+            // typically) so accumulating avoids the duplex-stream ceremony
+            // some Node versions need for streaming request bodies into
+            // a Web Request.
+            let body;
+            if (hasBody) {
+                const chunks = [];
+                for await (const chunk of req) {
+                    chunks.push(chunk);
+                }
+                if (chunks.length > 0) {
+                    let total = 0;
+                    for (const c of chunks)
+                        total += c.length;
+                    const merged = new Uint8Array(total);
+                    let offset = 0;
+                    for (const c of chunks) {
+                        merged.set(c, offset);
+                        offset += c.length;
+                    }
+                    body = merged.buffer.slice(merged.byteOffset, merged.byteOffset + merged.byteLength);
+                }
+            }
+            // Flatten Node header values: arrays → comma-joined (per RFC 7230 §3.2.2).
+            const webHeaders = new Headers();
+            for (const [name, value] of Object.entries(req.headers)) {
+                if (value == null)
+                    continue;
+                if (Array.isArray(value)) {
+                    for (const v of value)
+                        webHeaders.append(name, v);
+                }
+                else {
+                    webHeaders.set(name, value);
+                }
+            }
+            const webReq = new Request(url, {
+                method,
+                headers: webHeaders,
+                body,
+                signal: abort.signal,
+            });
+            const webRes = await webHandler(webReq);
+            res.statusCode = webRes.status;
+            // `Headers.forEach` exposes the value comma-joined for multi-valued
+            // headers, which `setHeader` accepts. Set-Cookie is handled separately
+            // via `getSetCookie()` to preserve multiple values.
+            webRes.headers.forEach((value, key) => {
+                if (key.toLowerCase() === "set-cookie")
+                    return;
+                res.setHeader(key, value);
+            });
+            const setCookies = typeof webRes.headers.getSetCookie === "function"
+                ? webRes.headers.getSetCookie()
+                : [];
+            if (setCookies.length > 0) {
+                res.setHeader("set-cookie", setCookies);
+            }
+            if (!webRes.body) {
+                res.end();
+                return;
+            }
+            // Pipe the Web Response body to the Node response. On client
+            // disconnect (`abort.signal`), cancel the reader so a pending
+            // `read()` rejects and we exit the loop instead of blocking on
+            // a stream that will never produce more chunks.
+            const reader = webRes.body.getReader();
+            const onAbort = () => {
+                reader.cancel(abort.signal.reason).catch(() => { });
+            };
+            if (abort.signal.aborted)
+                onAbort();
+            else
+                abort.signal.addEventListener("abort", onAbort, { once: true });
+            try {
+                while (true) {
+                    const { done, value } = await reader.read();
+                    if (done)
+                        break;
+                    res.write(value);
+                }
+            }
+            catch {
+                // Reader was cancelled (client disconnect). Silently end.
+            }
+            finally {
+                abort.signal.removeEventListener("abort", onAbort);
+            }
+            res.end();
+        }
+        catch (err) {
+            if (!res.headersSent) {
+                res.statusCode = 500;
+                res.setHeader("content-type", "text/plain; charset=utf-8");
+                res.end(err instanceof Error ? err.message : "Internal error");
+            }
+            else {
+                res.end();
+            }
+        }
+    };
+}
+/**
+ * Reshape a step-1 partial so the agent's `streamText` resumes by
+ * executing pending tool-calls before the next LLM call.
+ *
+ * When the customer's handler runs `streamText` with schema-only tools
+ * (no `execute` fns) and `stopWhen: stepCountIs(1)`, the LLM emits
+ * tool-calls but AI SDK can't execute them — the partial we ship is
+ * `[{ assistant: text + tool-call }]`. Splicing that as-is onto the
+ * agent's accumulator and calling `streamText` throws
+ * `MissingToolResultsError` synchronously inside
+ * `convertToLanguageModelPrompt`.
+ *
+ * AI SDK's documented escape hatch for "external party decides what
+ * to do with a tool-call, then SDK executes" is the tool-approval
+ * round. By appending a `tool-approval-request` part to the assistant
+ * message and a trailing `tool` message with a matching
+ * `tool-approval-response { approved: true }`, AI SDK:
+ *   1. Suppresses `MissingToolResultsError` for approved tool-calls
+ *      (`convert-to-language-model-prompt.ts:135-144`).
+ *   2. Hits its initial-tool-execution branch
+ *      (`stream-text.ts:1342-1486`) on the next `streamText` call,
+ *      runs the agent-side `execute` fns, and synthesizes
+ *      `tool-result` parts before the step-2 LLM call.
+ *
+ * If the customer's tools already had `execute` fns (rare for the
+ * handover use case but valid), the partial already contains a
+ * `tool-result` per tool-call — we leave those alone and only inject
+ * approvals for genuinely-pending calls.
+ *
+ * `collectToolApprovals` only scans the LAST message
+ * (`collect-tool-approvals.ts:30-37`), so the synthesized tool message
+ * must end up at the tail of the partial. The agent's run-loop
+ * splices the partial onto the end of the accumulator, which keeps
+ * this invariant.
+ */
+function reshapeForHandoverResume(responseMessages) {
+    // First pass: gather the set of tool-call IDs that already have a
+    // matching tool-result. Those are "complete" — leave them alone.
+    const completedToolCallIds = new Set();
+    for (const message of responseMessages) {
+        if (message.role !== "tool" || typeof message.content === "string")
+            continue;
+        for (const part of message.content) {
+            if (part.type === "tool-result" && part.toolCallId) {
+                completedToolCallIds.add(part.toolCallId);
+            }
+        }
+    }
+    // Second pass: clone the messages, appending a tool-approval-request
+    // alongside each pending tool-call. Collect the matching responses.
+    const approvalResponses = [];
+    let approvalCounter = 0;
+    const reshaped = responseMessages.map((message) => {
+        if (message.role !== "assistant" || typeof message.content === "string") {
+            return message;
+        }
+        const newContent = [...message.content];
+        for (const part of message.content) {
+            if (part.type === "tool-call" &&
+                part.toolCallId &&
+                !completedToolCallIds.has(part.toolCallId)) {
+                const approvalId = `handover-approval-${++approvalCounter}`;
+                newContent.push({
+                    type: "tool-approval-request",
+                    approvalId,
+                    toolCallId: part.toolCallId,
+                });
+                approvalResponses.push({
+                    type: "tool-approval-response",
+                    approvalId,
+                    approved: true,
+                });
+            }
+        }
+        return { ...message, content: newContent };
+    });
+    if (approvalResponses.length > 0) {
+        reshaped.push({
+            role: "tool",
+            content: approvalResponses,
+        });
+    }
+    return reshaped;
+}
+//# sourceMappingURL=chat-server.js.map