@tanstack/ai 0.41.0 → 0.43.0

package/dist/esm/stream-to-response.d.ts

@@ -1,4 +1,10 @@
+import { LockStore } from './activities/chat/middleware/locks.js';
+import { RunStore } from './activities/chat/middleware/run-store.js';
+import { InternalLogger } from './logger/internal-logger.js';
+import { DebugOption } from './logger/types.js';
+import { StreamDurability } from './stream-durability.js';
 import { StreamChunk } from './types.js';
+export { resolveResumeRunId } from './stream-durability.js';
 /**
  * Collect all text content from a StreamChunk async iterable and return as a string.
  *
@@ -11,8 +17,7 @@ import { StreamChunk } from './types.js';
  * @example
  * ```typescript
  * const stream = chat({
- *   adapter: openaiText(),
- *   model: 'gpt-4o',
+ *   adapter: openaiText('gpt-5.5'),
  *   messages: [{ role: 'user', content: 'Hello!' }]
  * });
  * const text = await streamToText(stream);
@@ -30,9 +35,30 @@ export declare function streamToText(stream: AsyncIterable<StreamChunk>): Promis
  *
  * @param stream - AsyncIterable of StreamChunks from chat()
  * @param abortController - Optional AbortController to abort when stream is cancelled
+ * @param getId - Optional per-chunk durability offset; when present, each event gets an `id:` line
  * @returns ReadableStream in Server-Sent Events format
  */
-export declare function toServerSentEventsStream(stream: AsyncIterable<StreamChunk>, abortController?: AbortController): ReadableStream<Uint8Array>;
+export declare function toServerSentEventsStream(stream: AsyncIterable<StreamChunk>, abortController?: AbortController, getId?: (chunk: StreamChunk, index: number) => string | undefined): ReadableStream<Uint8Array>;
+/**
+ * Name of the synthetic `CUSTOM` chunk a fresh durable producer appends to its
+ * log before pulling the first real chunk.
+ *
+ * Flushing `RUN_STARTED` (above) makes a run joinable from the instant the
+ * stream EMITS something — but a `chat()` whose middleware boots a sandbox
+ * (create a container, install a CLI) legitimately emits nothing for minutes,
+ * and during that window the log is empty. Every joiner's empty-log fail-fast
+ * (`memoryStream`'s first-chunk deadline, the client's rejoin connect deadline)
+ * then reads the run as gone — and the client clears its resume pointer, so a
+ * reload during the boot window permanently orphans a run that is still going.
+ *
+ * This marker closes the window: it is appended (and flushed) before the
+ * producer stream is first pulled, so a join always finds a first chunk within
+ * milliseconds of the run being accepted. Takeover alignment is unaffected — a
+ * journal replay cannot reproduce the marker, and alignment already skips
+ * stored `CUSTOM` chunks as out-of-band for exactly that reason (see
+ * `isBridgeCustomChunk` in `@tanstack/ai-sandbox`).
+ */
+export declare const RUN_ACCEPTED_EVENT = "run.accepted";
 /**
  * Convert a StreamChunk async iterable to a Response in Server-Sent Events format
  *
@@ -41,19 +67,113 @@ export declare function toServerSentEventsStream(stream: AsyncIterable<StreamChu
  * - Each chunk is followed by "\n\n"
  * - Stream ends when the underlying iterable is exhausted (RUN_FINISHED is the terminal event)
  *
+ * Pass a `durability` sink (`memoryStream(request)` / `durableStream(request)`)
+ * to make the stream resumable: fresh runs are appended to the log and each SSE
+ * event is tagged with an `id:` offset; a reconnect (native `Last-Event-ID`) or
+ * a `?offset` join replays from the log without re-running the producer. `batch`
+ * controls how many chunks are buffered per `append` (default 32).
+ *
  * @param stream - AsyncIterable of StreamChunks from chat()
- * @param init - Optional Response initialization options (including `abortController`)
+ * @param init - Optional Response initialization options (including `abortController`, `durability` with its optional `batch`, and `debug`)
  * @returns Response in Server-Sent Events format
  *
  * @example
  * ```typescript
- * const stream = chat({ adapter: openaiText(), model: "gpt-4o", messages: [...] });
- * return toServerSentEventsResponse(stream, { abortController });
+ * export async function POST(request: Request) {
+ *   const stream = chat({ adapter: openaiText('gpt-5.5'), messages: [...] });
+ *   return toServerSentEventsResponse(stream, { durability: { adapter: memoryStream(request) } });
+ * }
  * ```
  */
-export declare function toServerSentEventsResponse(stream: AsyncIterable<StreamChunk>, init?: ResponseInit & {
+export declare function toServerSentEventsResponse<TOffset extends string = string>(stream: AsyncIterable<StreamChunk>, init?: ResponseInit & {
     abortController?: AbortController;
+    durability?: {
+        adapter: StreamDurability<TOffset>;
+        batch?: number;
+    };
+    /**
+     * Customize logging for durability failure paths (terminal-append and
+     * close). These failures are always logged server-side by default (the
+     * `errors` category is on even without `debug`, via a `ConsoleLogger`);
+     * pass `debug` to route them to a custom `Logger` or raise verbosity. A
+     * joiner replaying the log only ever sees a generic incomplete error, so
+     * server-side logging is where the real cause is recoverable.
+     */
+    debug?: DebugOption;
 }): Response;
+/**
+ * Everything the resume helpers need to take a run over as a side effect of
+ * serving its log.
+ *
+ * `claim` and `pipe` are **injected**, not imported. The two mechanisms a
+ * takeover needs (`withRunClaim` and `pipeToRunLog`) live in
+ * `@tanstack/ai-sandbox`, and `@tanstack/ai` must not depend on that package —
+ * that layering inversion is exactly what moving `LockStore` into core was meant
+ * to prevent, and it would make core depend on the sandbox package to serve a
+ * plain chat run. Injecting them keeps only the *shape* of a takeover in core
+ * (parse the run id, read the record, skip if terminal, claim, drive) and lets a
+ * background-worker-driven run supply its own pair.
+ * `@tanstack/ai-sandbox`'s `sandboxRunDriver` fills both in.
+ */
+export interface RunDriverOptions {
+    /** The attach request; its run id is read with {@link resolveResumeRunId}. */
+    request: Request;
+    runs: RunStore;
+    locks: LockStore;
+    /** Produce the run's remaining events. Called only once the claim is held. */
+    drive: (input: {
+        runId: string;
+        threadId: string;
+        signal: AbortSignal;
+    }) => AsyncIterable<StreamChunk>;
+    /** Run `fn` under exclusive ownership of the run, or reject if refused. */
+    claim: <T>(input: {
+        runs: RunStore;
+        locks: LockStore;
+        runId: string;
+    }, fn: (claim: {
+        runId: string;
+        epoch: number;
+        signal: AbortSignal;
+    }) => Promise<T>) => Promise<T>;
+    /** Persist the driven stream to the run's producer-side durability log. */
+    pipe: (stream: AsyncIterable<StreamChunk>, input: {
+        runId: string;
+        threadId: string;
+        signal: AbortSignal;
+    }) => Promise<unknown>;
+    /** Platform keep-alive (e.g. `ctx.waitUntil`) for the background drive. */
+    waitUntil?: (promise: Promise<unknown>) => void;
+    logger?: InternalLogger;
+}
+/** Shared options for the resume-only response helpers. */
+type ResumeResponseOptions<TOffset extends string> = ResponseInit & {
+    adapter: StreamDurability<TOffset>;
+    batch?: number;
+    debug?: DebugOption;
+    /**
+     * Take the run over while serving its log. Omit to serve the log only —
+     * the response is byte-identical either way.
+     */
+    driver?: RunDriverOptions;
+};
+/**
+ * Serve a resumable run from its durability log over Server-Sent Events, without
+ * re-running the model. Use this in a `GET` handler so a reload or a second tab
+ * can re-attach to an in-flight or finished run.
+ *
+ * The adapter (`memoryStream(request)` / `durableStream(request)`) captures the
+ * resume offset from the request. If there is none (no `Last-Event-ID` header
+ * and no `?offset`), there is nothing to replay and this returns a 400.
+ *
+ * @example
+ * ```typescript
+ * export async function GET(request: Request) {
+ *   return resumeServerSentEventsResponse({ adapter: memoryStream(request) });
+ * }
+ * ```
+ */
+export declare function resumeServerSentEventsResponse<TOffset extends string = string>(options: ResumeResponseOptions<TOffset>): Response;
 /**
  * Convert a StreamChunk async iterable to a ReadableStream in HTTP stream format (newline-delimited JSON)
  *
@@ -63,13 +183,20 @@ export declare function toServerSentEventsResponse(stream: AsyncIterable<StreamC
  *
  * This format is compatible with `fetchHttpStream` connection adapter.
  *
+ * When `getId` is supplied (delivery durability), each chunk is emitted as an
+ * envelope `{"id":"<offset>","chunk":{…}}` instead of a bare chunk. NDJSON has
+ * no native event-id field like SSE's `id:` line, so the resumable offset rides
+ * inside the payload. Untagged chunks (no id) stay bare, so a non-durable
+ * stream is byte-identical to before and the client auto-detects either form.
+ *
  * @param stream - AsyncIterable of StreamChunks from chat()
  * @param abortController - Optional AbortController to abort when stream is cancelled
+ * @param getId - Optional per-chunk durability offset; when present, chunks are envelope-encoded
  * @returns ReadableStream in HTTP stream format (newline-delimited JSON)
  *
  * @example
  * ```typescript
- * const stream = chat({ adapter: openaiText(), model: "gpt-4o", messages: [...] });
+ * const stream = chat({ adapter: openaiText('gpt-5.5'), messages: [...] });
  * const readableStream = toHttpStream(stream);
  * // Use with Response for HTTP streaming (not SSE)
  * return new Response(readableStream, {
@@ -77,7 +204,7 @@ export declare function toServerSentEventsResponse(stream: AsyncIterable<StreamC
  * });
  * ```
  */
-export declare function toHttpStream(stream: AsyncIterable<StreamChunk>, abortController?: AbortController): ReadableStream<Uint8Array>;
+export declare function toHttpStream(stream: AsyncIterable<StreamChunk>, abortController?: AbortController, getId?: (chunk: StreamChunk, index: number) => string | undefined): ReadableStream<Uint8Array>;
 /**
  * Convert a StreamChunk async iterable to a Response in HTTP stream format (newline-delimited JSON)
  *
@@ -87,16 +214,54 @@ export declare function toHttpStream(stream: AsyncIterable<StreamChunk>, abortCo
  *
  * This format is compatible with `fetchHttpStream` connection adapter.
  *
+ * Pass a `durability` sink (`memoryStream(request)` / `durableStream(request)`)
+ * to make the stream resumable: fresh runs are appended to the log and each
+ * NDJSON line is emitted as an `{ id, chunk }` envelope carrying an opaque
+ * offset; a reconnect (native `Last-Event-ID` header) or a `?offset` join
+ * replays from the log without re-running the producer. `batch` controls how
+ * many chunks are buffered per `append` (default 32). This shares the exact
+ * `durableStreamSource` used by `toServerSentEventsResponse` — only the wire
+ * encoding differs.
+ *
  * @param stream - AsyncIterable of StreamChunks from chat()
- * @param init - Optional Response initialization options (including `abortController`)
+ * @param init - Optional Response initialization options (including `abortController`, `durability` with its optional `batch`, and `debug`)
  * @returns Response in HTTP stream format (newline-delimited JSON)
  *
  * @example
  * ```typescript
- * const stream = chat({ adapter: openaiText(), model: "gpt-4o", messages: [...] });
- * return toHttpResponse(stream, { abortController });
+ * export async function POST(request: Request) {
+ *   const stream = chat({ adapter: openaiText('gpt-5.5'), messages: [...] });
+ *   return toHttpResponse(stream, { durability: { adapter: memoryStream(request) } });
+ * }
  * ```
  */
-export declare function toHttpResponse(stream: AsyncIterable<StreamChunk>, init?: ResponseInit & {
+export declare function toHttpResponse<TOffset extends string = string>(stream: AsyncIterable<StreamChunk>, init?: ResponseInit & {
     abortController?: AbortController;
+    durability?: {
+        adapter: StreamDurability<TOffset>;
+        batch?: number;
+    };
+    /**
+     * Customize logging for durability failure paths (terminal-append and
+     * close). These failures are always logged server-side by default (the
+     * `errors` category is on even without `debug`, via a `ConsoleLogger`);
+     * pass `debug` to route them to a custom `Logger` or raise verbosity. A
+     * joiner replaying the log only ever sees a generic incomplete error, so
+     * server-side logging is where the real cause is recoverable.
+     */
+    debug?: DebugOption;
 }): Response;
+/**
+ * Serve a resumable run from its durability log over NDJSON, without re-running
+ * the model. The NDJSON counterpart of {@link resumeServerSentEventsResponse};
+ * pair it with a `toHttpResponse` producer. Returns a 400 when the request
+ * carries no resume offset (no `Last-Event-ID` header and no `?offset`).
+ *
+ * @example
+ * ```typescript
+ * export async function GET(request: Request) {
+ *   return resumeHttpResponse({ adapter: memoryStream(request) });
+ * }
+ * ```
+ */
+export declare function resumeHttpResponse<TOffset extends string = string>(options: ResumeResponseOptions<TOffset>): Response;