ai.ts

import {
  accessoryAttributes,
  AnyTask,
  apiClientManager,
  controlSubtype,
  getSchemaParseFn,
  headerValue,
  InputStreamOncePromise,
  type InputStreamOnceOptions,
  type InputStreamWaitOptions,
  type InputStreamWaitWithIdleTimeoutOptions,
  isSchemaZodEsque,
  logger,
  type MachinePresetName,
  ManualWaitpointPromise,
  OutOfMemoryError,
  sessionStreams,
  type PipeStreamResult,
  type RealtimeDefinedInputStream,
  type RealtimeDefinedStream,
  type ReadStreamOptions,
  SemanticInternalAttributes,
  type SendInputStreamOptions,
  Task,
  taskContext,
  type AppendStreamOptions,
  type InputStreamOnceResult,
  type inferSchemaIn,
  type inferSchemaOut,
  type PipeStreamOptions,
  type TaskIdentifier,
  type TaskOptions,
  type TaskSchema,
  type TaskRunContext,
  type TaskWithSchema,
  SESSION_IN_EVENT_ID_HEADER,
  TRIGGER_CONTROL_SUBTYPE,
  generateJWT,
  type WriterStreamOptions,
} from "@trigger.dev/core/v3";
import type {
  FinishReason,
  LanguageModelUsage,
  ModelMessage,
  Tool,
  ToolSet,
  UIMessage,
  UIMessageChunk,
  UIMessageStreamOptions,
} from "ai";
import type { ChatSnapshotV1, StreamWriteResult } from "@trigger.dev/core/v3";
// Runtime VALUES go through the ESM/CJS shim so the CJS build can `require`
// ESM-only `ai@7` (see ../imports/ai-runtime.ts).
import {
  convertToModelMessages,
  dynamicTool,
  generateId as generateMessageId,
  getToolName,
  isToolUIPart,
  jsonSchema,
  readUIMessageStream,
  tool as aiTool,
  zodSchema,
} from "../imports/ai-runtime.js";
import type { JSONSchema7, Schema } from "ai";

// `ToolCallOptions` is defined locally rather than imported from `ai`: v7
// renamed/removed that export (it's `ToolExecutionOptions<CONTEXT>` now), so a
// direct import breaks on v7. This structural shape is wider than both majors'
// and reads the user-context field under both names (`experimental_context` on
// v6, `context` on v7).
type ToolCallOptions = {
  toolCallId: string;
  messages?: ModelMessage[];
  abortSignal?: AbortSignal;
  experimental_context?: unknown;
  context?: unknown;
};
import { type Attributes, trace } from "@opentelemetry/api";
import { auth } from "./auth.js";
import { locals } from "./locals.js";
import { metadata } from "./metadata.js";
import type { ResolvedPrompt } from "./prompt.js";
import type { ResolvedSkill } from "./skill.js";
// Bash-skill runtime lives in `./agentSkillsRuntime.ts` (exposed as
// the `@trigger.dev/sdk/ai/skills-runtime` subpath). It's a normal
// static import — `ai.ts` is server-only by reachability now that
// browser-side primitives (PENDING_MESSAGE_INJECTED_TYPE and the
// chat-task wire types) live in `./ai-shared.ts`. Any browser bundle
// that wants those primitives imports `./ai-shared.js` directly and
// never touches `ai.ts`'s module graph, so the `node:*` builtins
// pulled in transitively here never reach a client chunk.
import { runBashInSkill, readFileInSkill } from "./agentSkillsRuntime.js";
import { streams, markChatAgentRunForStreamsWarning } from "./streams.js";
import {
  sessions,
  type SessionHandle,
  type SessionInputChannel,
  type SessionOutputChannel,
  type SessionPipeStreamOptions,
  type SessionSubscribeOptions,
} from "./sessions.js";
import { createTask } from "./shared.js";
import { ensureAiSdkTelemetry } from "./aiAutoTelemetry.js";
import { resourceCatalog, type SessionTriggerConfig } from "@trigger.dev/core/v3";
import { tracer } from "./tracer.js";

/** Re-export for typing `ctx` in `chat.agent` hooks without importing `@trigger.dev/core`. */
export type { TaskRunContext } from "@trigger.dev/core/v3";

const METADATA_KEY = "tool.execute.options";

/**
 * Wrapper around `convertToModelMessages` that always passes
 * `ignoreIncompleteToolCalls: true` to prevent failures from
 * stopped/aborted conversations with partial tool parts.
 */
function toModelMessages(messages: UIMessage[]): Promise<ModelMessage[]> {
  // Pass the resolved per-turn `tools` (if any) so the AI SDK can look up each
  // tool's `toModelOutput` and re-apply it to prior-turn tool results. Without
  // `tools` it falls back to JSON-stringifying the raw output (TRI-10149). The
  // conditional spread keeps the options object byte-identical to the no-tools
  // path when nothing was declared.
  const tools = locals.get(chatResolvedToolsKey);
  return convertToModelMessages(messages, {
    ignoreIncompleteToolCalls: true,
    ...(tools ? { tools } : {}),
  });
}

export type ToolCallExecutionOptions = {
  toolCallId: string;
  experimental_context?: unknown;
  /** v7 name for the user context (`experimental_context` on v6). */
  context?: unknown;
  /** Chat context — only present when the tool runs inside a chat.agent turn. */
  chatId?: string;
  turn?: number;
  continuation?: boolean;
  clientData?: unknown;
  /** Serialized chat.local values from the parent run. @internal */
  chatLocals?: Record<string, unknown>;
};

/** Chat context stored in locals during each chat.agent turn for auto-detection. */
type ChatTurnContext<TClientData = unknown> = {
  chatId: string;
  turn: number;
  continuation: boolean;
  clientData?: TClientData;
};
const chatTurnContextKey = locals.create<ChatTurnContext>("chat.turnContext");

/**
 * Per-run slot holding the Session handle that backs this chat's `.in` /
 * `.out` channels. Populated at the top of `chatAgent`'s run function from
 * `payload.sessionId`; read by every module-level helper (`chatStream`,
 * `messagesInput`, `stopInput`) so the chat.agent internals can remain
 * the same module-level shape they were when the I/O was run-scoped.
 * @internal
 */
const chatSessionHandleKey = locals.create<SessionHandle>("chat.sessionHandle");

/**
 * S2 seq_num of the most recent `turn-complete` control record written by
 * this worker. Read by `writeTurnCompleteChunk` to know what to trim back
 * to when the next turn finishes, keeping `session.out` bounded to ~one
 * turn at steady state.
 *
 * Seeded at boot from `ChatSnapshotV1.lastOutEventId` (which is exactly
 * the previous turn-complete's seq_num). Wrapped in a mutable holder so
 * `writeTurnCompleteChunk` can advance it without going through a setter.
 * @internal
 */
const lastTurnCompleteSeqNumKey = locals.create<{ value: number | undefined }>(
  "chat.lastTurnCompleteSeqNum"
);

/**
 * Scan `session.out` for the latest `turn-complete` control record and
 * return its `session-in-event-id` header value — the committed-consume
 * cursor on `.in` as of that turn-complete. Used at worker boot to seed
 * the `.in` subscription so already-processed user messages don't get
 * replayed from S2.
 *
 * Implementation streams the SSE endpoint and listens for `turn-complete`
 * via the transport's `onControl` callback; the data-chunk for-await is
 * just there to drive the stream. The scan is O(1 turn) because
 * `session.out` is bounded to roughly one turn at steady state — every
 * successful turn-complete is followed by an S2 trim back to the
 * previous one (see `writeTurnCompleteChunk`).
 *
 * Returns `undefined` if no `turn-complete` carrying the header has been
 * written yet — first-turn-ever, first turn post-OOM-with-no-prior-runs,
 * or a `turn-complete` written before this header existed (cross-version
 * boot). Callers fall back to subscribing `.in` from seq 0 in that case;
 * the slim-wire merge handles any dedup against snapshot-restored
 * messages.
 * @internal
 */
async function findLatestSessionInCursor(
  chatId: string
): Promise<number | undefined> {
  const apiClient = apiClientManager.clientOrThrow();
  let latestCursor: number | undefined;
  const stream = await apiClient.subscribeToSessionStream<unknown>(chatId, "out", {
    // 5s rather than 1s: S2 trim is eventually-consistent (10-60s
    // window), so a worker booting just after a trim could still see
    // pre-trim records and need a bit longer to drain them all before
    // the SSE long-poll closes. Without enough headroom the scan would
    // fall back to `undefined`, the `.in` cursor wouldn't be seeded,
    // and the next subscribe would replay messages already processed.
    timeoutInSeconds: 5,
    onControl: (event) => {
      if (event.subtype !== TRIGGER_CONTROL_SUBTYPE.TURN_COMPLETE) return;
      const raw = headerValue(event.headers, SESSION_IN_EVENT_ID_HEADER);
      if (!raw) return;
      const parsed = Number.parseInt(raw, 10);
      if (Number.isFinite(parsed)) latestCursor = parsed;
    },
  });
  // Drain the stream so the underlying SSE reader runs to completion. We
  // don't accumulate chunks; `onControl` fires inline as turn-complete
  // records arrive.
  for await (const _ of stream) {
    // intentionally empty
  }
  return latestCursor;
}

/**
 * Versioned blob written to S3 after every turn completes (when no
 * `hydrateMessages` hook is registered). Read at run boot to seed the
 * accumulator with prior conversation state, replacing the old wire-borne
 * full-history seed.
 *
 * The shape is shared with the Sessions dashboard (which reads the same
 * blob to render the full conversation transcript) via
 * `@trigger.dev/core/v3`. Customer code shouldn't reach in here — the
 * SDK transports surface the messages through the standard `messages`
 * accumulator.
 *
 * @internal
 */
export type { ChatSnapshotV1 } from "@trigger.dev/core/v3";

/**
 * Test-only override hook — `mockChatAgent` installs a fake to return
 * synthetic snapshots without hitting S3. Mirrors the `__set*ImplForTests`
 * pattern in `sessions.ts`. Not part of the public API.
 * @internal
 */
type ReadChatSnapshotImpl = <TUIMessage extends UIMessage>(
  sessionId: string
) => Promise<ChatSnapshotV1<TUIMessage> | undefined> | ChatSnapshotV1<TUIMessage> | undefined;
let readChatSnapshotImpl: ReadChatSnapshotImpl | undefined;

export function __setReadChatSnapshotImplForTests(impl: ReadChatSnapshotImpl | undefined): void {
  readChatSnapshotImpl = impl;
}

/**
 * Test-only override hook — see `__setReadChatSnapshotImplForTests`. The
 * mock harness records writes for assertion via this setter. Not public.
 * @internal
 */
type WriteChatSnapshotImpl = <TUIMessage extends UIMessage>(
  sessionId: string,
  snapshot: ChatSnapshotV1<TUIMessage>
) => Promise<void> | void;
let writeChatSnapshotImpl: WriteChatSnapshotImpl | undefined;

export function __setWriteChatSnapshotImplForTests(impl: WriteChatSnapshotImpl | undefined): void {
  writeChatSnapshotImpl = impl;
}

/**
 * Read the persisted snapshot for a session. Returns `undefined` on:
 *   - missing object (404 from the presigned GET — fresh session, never
 *     persisted)
 *   - presign failure (network/auth issue)
 *   - malformed JSON
 *   - version mismatch (forward-compat — older runtimes ignore newer blobs)
 *
 * Always swallows errors via `logger.warn`. The agent boot loop must stay
 * available even if S3 hiccups; the worst case is replaying more of
 * `session.out` than strictly necessary.
 * @internal
 */
async function readChatSnapshot<TUIMessage extends UIMessage>(
  sessionId: string
): Promise<ChatSnapshotV1<TUIMessage> | undefined> {
  if (readChatSnapshotImpl) {
    return (await readChatSnapshotImpl<TUIMessage>(sessionId)) ?? undefined;
  }
  const apiClient = apiClientManager.clientOrThrow();
  let presignedUrl: string;
  try {
    const resp = await apiClient.getChatSnapshotUrl(sessionId);
    presignedUrl = resp.presignedUrl;
  } catch (error) {
    logger.warn("chat.agent: snapshot presign (read) failed; continuing without snapshot", {
      error: error instanceof Error ? error.message : String(error),
      sessionId,
    });
    return undefined;
  }
  let response: Response;
  try {
    response = await fetch(presignedUrl, { method: "GET" });
  } catch (error) {
    logger.warn("chat.agent: snapshot fetch failed; continuing without snapshot", {
      error: error instanceof Error ? error.message : String(error),
      sessionId,
    });
    return undefined;
  }
  if (response.status === 404) {
    // First-ever boot for this session — no snapshot yet. Caller falls
    // through to replay-only.
    return undefined;
  }
  if (!response.ok) {
    logger.warn("chat.agent: snapshot fetch returned non-OK; continuing without snapshot", {
      status: response.status,
      sessionId,
    });
    return undefined;
  }
  let parsed: unknown;
  try {
    parsed = await response.json();
  } catch (error) {
    logger.warn("chat.agent: snapshot JSON parse failed; continuing without snapshot", {
      error: error instanceof Error ? error.message : String(error),
      sessionId,
    });
    return undefined;
  }
  if (!parsed || typeof parsed !== "object") return undefined;
  const candidate = parsed as Partial<ChatSnapshotV1<TUIMessage>>;
  if (candidate.version !== 1 || !Array.isArray(candidate.messages)) {
    logger.warn("chat.agent: snapshot version/shape mismatch; ignoring", {
      version: candidate.version,
      sessionId,
    });
    return undefined;
  }
  return candidate as ChatSnapshotV1<TUIMessage>;
}

/**
 * Persist the snapshot for a session. Awaited by callers immediately after
 * `onTurnComplete` — the agent may suspend right after this point, and
 * fire-and-forget promises don't reliably complete on suspend.
 *
 * Errors are swallowed via `logger.warn`. A failed write means the next
 * boot replays slightly more of `session.out` (back to the previous
 * snapshot's cursor) instead of failing — the conversation stays
 * coherent, only the boot path does marginally more work.
 * @internal
 */
async function writeChatSnapshot<TUIMessage extends UIMessage>(
  sessionId: string,
  snapshot: ChatSnapshotV1<TUIMessage>
): Promise<void> {
  if (writeChatSnapshotImpl) {
    await writeChatSnapshotImpl<TUIMessage>(sessionId, snapshot);
    return;
  }
  const apiClient = apiClientManager.clientOrThrow();
  let presignedUrl: string;
  try {
    const resp = await apiClient.createChatSnapshotUploadUrl(sessionId);
    presignedUrl = resp.presignedUrl;
  } catch (error) {
    logger.warn("chat.agent: snapshot presign (write) failed; next run will replay further", {
      error: error instanceof Error ? error.message : String(error),
      sessionId,
    });
    return;
  }
  let response: Response;
  try {
    response = await fetch(presignedUrl, {
      method: "PUT",
      headers: { "content-type": "application/json" },
      body: JSON.stringify(snapshot),
    });
  } catch (error) {
    logger.warn("chat.agent: snapshot upload failed; next run will replay further", {
      error: error instanceof Error ? error.message : String(error),
      sessionId,
    });
    return;
  }
  if (!response.ok) {
    logger.warn("chat.agent: snapshot upload returned non-OK; next run will replay further", {
      status: response.status,
      sessionId,
    });
  }
}

/**
 * Test-only entry point that bypasses `__setReadChatSnapshotImplForTests`
 * and reaches the real `apiClient.getPayloadUrl` + `fetch` + JSON-parse path.
 * Used by `chat-snapshot.test.ts` to verify 404 / 500 / malformed JSON /
 * version-mismatch / network-error behavior end-to-end. Tests mock global
 * `fetch` and the api-client config; this wrapper lets them drive the
 * production code without the override hook short-circuiting.
 *
 * Not part of the public API. The `__` prefix and `ForTests` suffix mirror
 * the override-hook setters above.
 * @internal
 */
export async function __readChatSnapshotProductionPathForTests<TUIMessage extends UIMessage>(
  sessionId: string
): Promise<ChatSnapshotV1<TUIMessage> | undefined> {
  const saved = readChatSnapshotImpl;
  readChatSnapshotImpl = undefined;
  try {
    return await readChatSnapshot<TUIMessage>(sessionId);
  } finally {
    readChatSnapshotImpl = saved;
  }
}

/**
 * Test-only entry point that bypasses `__setWriteChatSnapshotImplForTests`
 * and reaches the real `apiClient.createUploadPayloadUrl` + `fetch` PUT
 * path. Pairs with `__readChatSnapshotProductionPathForTests` — see that
 * function's note for the rationale.
 *
 * Not part of the public API.
 * @internal
 */
export async function __writeChatSnapshotProductionPathForTests<TUIMessage extends UIMessage>(
  sessionId: string,
  snapshot: ChatSnapshotV1<TUIMessage>
): Promise<void> {
  const saved = writeChatSnapshotImpl;
  writeChatSnapshotImpl = undefined;
  try {
    await writeChatSnapshot<TUIMessage>(sessionId, snapshot);
  } finally {
    writeChatSnapshotImpl = saved;
  }
}

/**
 * Merge two `UIMessage[]` lists by `id`, with the second list winning on
 * collision. Used at run boot to combine the snapshot's persisted history
 * with the replayed `session.out` tail — replay produces the freshest
 * representation of any assistant message that landed after the snapshot's
 * cursor, so it should overwrite the older copy from the snapshot.
 *
 * Order: items unique to `a` keep their original positions; items unique to
 * `b` are appended at the end in their `b` order; collisions take `b`'s
 * value but keep the position they had in `a`.
 *
 * @internal
 */
function mergeByIdReplaceWins<TUIMessage extends UIMessage>(
  a: TUIMessage[],
  b: TUIMessage[]
): TUIMessage[] {
  if (b.length === 0) return [...a];
  if (a.length === 0) return [...b];
  const indexById = new Map<string, number>();
  for (let i = 0; i < a.length; i++) {
    const id = a[i]!.id;
    if (typeof id === "string" && id.length > 0) indexById.set(id, i);
  }
  const result = [...a];
  for (const next of b) {
    const id = next.id;
    if (typeof id === "string" && id.length > 0 && indexById.has(id)) {
      result[indexById.get(id)!] = next;
    } else {
      const newIdx = result.length;
      result.push(next);
      if (typeof id === "string" && id.length > 0) indexById.set(id, newIdx);
    }
  }
  return result;
}

/**
 * Test-only entry point for `mergeByIdReplaceWins`. The merge helper is the
 * one piece of slim-wire boot logic that's purely functional, so it earns a
 * direct unit test that exercises empty inputs, id collisions, no-id append,
 * order preservation, and the replay-wins-on-collision invariant. Mirrors
 * the `__*ProductionPathForTests` pattern used for the snapshot/replay
 * helpers above.
 *
 * Not part of the public API.
 * @internal
 */
export function __mergeByIdReplaceWinsForTests<TUIMessage extends UIMessage>(
  a: TUIMessage[],
  b: TUIMessage[]
): TUIMessage[] {
  return mergeByIdReplaceWins<TUIMessage>(a, b);
}

/**
 * Test-only override hook — `mockChatAgent` installs a fake replay that
 * returns a synthetic `UIMessage[]` so unit tests can drive the boot loop
 * without an SSE subscription. Mirrors the snapshot setters above. Not
 * part of the public API.
 * @internal
 */
type ReplaySessionOutTailResult<TUIMessage extends UIMessage> = {
  /** Messages whose `finish` chunk landed before the run died. Safe to seed the chain. */
  settled: TUIMessage[];
  /**
   * The trailing assistant message whose `finish` chunk never arrived —
   * an orphan from a cancel / crash / OOM. `cleanupAbortedParts` has
   * already stripped streaming-in-progress fragments. `undefined` if
   * the tail ended cleanly (every segment closed).
   */
  partial: TUIMessage | undefined;
  /**
   * The trailing assistant message BEFORE `cleanupAbortedParts` ran. Same
   * `undefined` semantics as `partial`. Use this when you need to inspect
   * tool parts the cleanup would strip (e.g. `input-available` /
   * `input-streaming` orphans surfaced via `pendingToolCalls`).
   */
  partialRaw: TUIMessage | undefined;
};

type ReplaySessionOutTailImpl = <TUIMessage extends UIMessage>(
  sessionId: string,
  options?: { lastEventId?: string }
) => Promise<ReplaySessionOutTailResult<TUIMessage>>;
let replaySessionOutTailImpl: ReplaySessionOutTailImpl | undefined;

export function __setReplaySessionOutTailImplForTests(
  impl: ReplaySessionOutTailImpl | undefined
): void {
  replaySessionOutTailImpl = impl;
}

/**
 * Drain `session.out` from `lastEventId` (or the start) and reduce the
 * remaining `UIMessageChunk`s back into `UIMessage[]`. Used at run boot to
 * catch any chunks that landed AFTER the last persisted snapshot — typically
 * the chunks from the turn whose `onTurnComplete` ran but whose snapshot
 * write didn't make it to S3 before the run crashed / suspended.
 *
 * Implementation:
 *   1. `apiClient.readSessionStreamRecords` — non-SSE, `wait=0` drain.
 *      Returns immediately with whatever records exist after the cursor.
 *      The previous SSE-subscribe path paid a fixed ~1s long-poll tax on
 *      every fresh chat (timeout duration on empty streams) — unacceptable
 *      for the first-message TTFC budget.
 *   2. Filter out the agent's control chunks (`type: "trigger:*"`) — they
 *      ride on the same stream as the user-visible UIMessageChunks.
 *   3. Split chunks at `start`/`finish` boundaries so each segment is a
 *      single message, then feed each segment through the AI SDK's
 *      `readUIMessageStream` reducer (the same one `useChat` uses on the
 *      browser side) and grab the final emitted snapshot.
 *   4. The trailing message — if it never received a `finish` chunk —
 *      goes through `cleanupAbortedParts` so partial in-flight parts
 *      don't leak into the next turn's accumulator. Drop it entirely
 *      if cleanup empties it.
 *
 * Errors are propagated to the caller (the boot loop wraps in try/catch and
 * `logger.warn`s); we don't swallow here so test code can observe failures
 * directly.
 * @internal
 */
async function replaySessionOutTail<TUIMessage extends UIMessage>(
  sessionId: string,
  options?: { lastEventId?: string }
): Promise<ReplaySessionOutTailResult<TUIMessage>> {
  if (replaySessionOutTailImpl) {
    return await replaySessionOutTailImpl<TUIMessage>(sessionId, options);
  }
  const apiClient = apiClientManager.clientOrThrow();
  const response = await apiClient.readSessionStreamRecords(sessionId, "out", {
    afterEventId: options?.lastEventId,
  });
  const collected: UIMessageChunk[] = [];
  for (const record of response.records) {
    // `data` is the chunk object as written by the SDK's session-out
    // writer (an AI SDK `UIMessageChunk` or a Trigger control object).
    // The route forwards it as-is — no JSON envelope to unwrap here.
    // Defensive shape checks below tolerate malformed records by
    // skipping them instead of throwing.
    const chunk: unknown = record.data;
    if (!chunk || typeof chunk !== "object") continue;
    const type = (chunk as { type?: unknown }).type;
    if (typeof type !== "string") continue;
    // Drop agent control chunks (`trigger:turn-complete`, `trigger:upgrade-required`,
    // session-state telemetry, etc.). They ride the same stream but aren't part
    // of the UIMessageChunk discriminated union and would confuse the reducer.
    if (type.startsWith("trigger:")) continue;
    collected.push(chunk as UIMessageChunk);
  }
  if (collected.length === 0) return { settled: [], partial: undefined, partialRaw: undefined };

  // Split chunks into per-message segments. A `start` chunk demarcates the
  // beginning of an assistant message; chunks before any `start` (rare —
  // but possible if the stream begins mid-message after a resume) get
  // bundled into a leading "implicit" segment so we don't drop them silently.
  type Segment = { chunks: UIMessageChunk[]; closed: boolean };
  const segments: Segment[] = [];
  let current: Segment | undefined;
  for (const chunk of collected) {
    if (chunk.type === "start") {
      current = { chunks: [chunk], closed: false };
      segments.push(current);
      continue;
    }
    if (!current) {
      // Chunk arrived before any `start`. Synthesize a segment so the reducer
      // has something to work with — `readUIMessageStream` tolerates a missing
      // `start` because we pass `message: undefined`.
      current = { chunks: [], closed: false };
      segments.push(current);
    }
    current.chunks.push(chunk);
    if (chunk.type === "finish") {
      current.closed = true;
      current = undefined;
    }
  }

  const settled: TUIMessage[] = [];
  let partial: TUIMessage | undefined;
  let partialRaw: TUIMessage | undefined;
  for (let i = 0; i < segments.length; i++) {
    const seg = segments[i]!;
    const isTrailing = i === segments.length - 1 && !seg.closed;
    const segmentStream = new ReadableStream<UIMessageChunk>({
      start(controller) {
        for (const c of seg.chunks) controller.enqueue(c);
        controller.close();
      },
    });
    let last: UIMessage | undefined;
    try {
      for await (const snapshot of readUIMessageStream({ stream: segmentStream })) {
        last = snapshot;
      }
    } catch (error) {
      // Reducer error — the segment is malformed. Skip it and keep going so a
      // single corrupt chunk doesn't sink the entire replay.
      logger.warn("chat.agent: replay reducer failed for segment; skipping", {
        sessionId,
        segmentIndex: i,
        error: error instanceof Error ? error.message : String(error),
      });
      continue;
    }
    if (!last) continue;
    if (isTrailing) {
      const cleaned = cleanupAbortedParts(last as TUIMessage);
      if (cleaned.parts.length === 0) continue;
      partial = cleaned;
      // Keep the raw pre-cleanup message too — recovery boot extracts
      // `pendingToolCalls` from it, since `cleanupAbortedParts` strips
      // exactly the input-streaming / input-available tool parts that
      // we want to surface.
      partialRaw = last as TUIMessage;
    } else {
      settled.push(last as TUIMessage);
    }
  }
  return { settled, partial, partialRaw };
}

/**
 * Test-only entry point that bypasses `__setReplaySessionOutTailImplForTests`
 * and reaches the real `apiClient.subscribeToSessionStream` + chunk-segment
 * splitter + `readUIMessageStream` reducer. Pairs with the snapshot
 * production-path wrappers above. Lets `replay-session-out.test.ts` drive
 * synthetic chunk sequences through the real reducer to lock down chunk-
 * stream → `UIMessage[]` correctness — if the AI SDK's chunk semantics
 * shift in a future version, the test catches it before customers do.
 *
 * Tests should mock `apiClient.subscribeToSessionStream` (e.g. via
 * `vi.spyOn(apiClient, ...)`) to feed a `ReadableStream<UIMessageChunk>`.
 *
 * Not part of the public API.
 * @internal
 */
export async function __replaySessionOutTailProductionPathForTests<
  TUIMessage extends UIMessage,
>(
  sessionId: string,
  options?: { lastEventId?: string }
): Promise<TUIMessage[]> {
  const saved = replaySessionOutTailImpl;
  replaySessionOutTailImpl = undefined;
  try {
    const { settled, partial } = await replaySessionOutTail<TUIMessage>(sessionId, options);
    return partial !== undefined ? [...settled, partial] : settled;
  } finally {
    replaySessionOutTailImpl = saved;
  }
}

/**
 * Test-only override hook for `replaySessionInTail`. Mirrors
 * `__setReplaySessionOutTailImplForTests` so unit tests can drive the boot
 * loop's chain-reconstruction logic without an HTTP round-trip.
 * @internal
 */
type ReplaySessionInTailImpl = <TUIMessage extends UIMessage>(
  sessionId: string,
  options?: { lastEventId?: string }
) => Promise<{ message: TUIMessage; metadata: unknown; seqNum: number }[]>;
let replaySessionInTailImpl: ReplaySessionInTailImpl | undefined;

export function __setReplaySessionInTailImplForTests(
  impl: ReplaySessionInTailImpl | undefined
): void {
  replaySessionInTailImpl = impl;
}

/**
 * Drain `session.in` from `lastEventId` (or the start) and surface the user
 * messages that landed past the cursor. Mirror of `replaySessionOutTail` —
 * both reads run at continuation boot so the SDK can reconstruct
 * conversational order across a dead run that never wrote `onTurnComplete`.
 *
 * `session.in` carries the {@link ChatInputChunk} tagged union:
 *   - `kind: "message"` — a `ChatTaskWirePayload` envelope for a new user
 *     message (`trigger: "submit-message"`) or a regeneration. Only the
 *     submit-message records carry a `payload.message`; regenerations,
 *     preload / close / action / handover-prepare have no message.
 *   - `kind: "stop"` — mid-turn cancellation signal. Not a message.
 *   - `kind: "handover"` / `kind: "handover-skip"` — head-start signals.
 *     Not user messages.
 *
 * This function filters to the first variant and returns the embedded
 * `UIMessage`s in seq_num order, paired with their seq_num so the caller
 * can advance the session.in cursor past them.
 *
 * Errors are propagated to the caller (the boot loop wraps in try/catch
 * and `logger.warn`s); we don't swallow here so test code can observe
 * failures directly.
 * @internal
 */
async function replaySessionInTail<TUIMessage extends UIMessage>(
  sessionId: string,
  options?: { lastEventId?: string }
): Promise<{ message: TUIMessage; metadata: unknown; seqNum: number }[]> {
  if (replaySessionInTailImpl) {
    return await replaySessionInTailImpl<TUIMessage>(sessionId, options);
  }
  const apiClient = apiClientManager.clientOrThrow();
  const response = await apiClient.readSessionStreamRecords(sessionId, "in", {
    afterEventId: options?.lastEventId,
  });
  const out: { message: TUIMessage; metadata: unknown; seqNum: number }[] = [];
  for (const record of response.records) {
    // session.in writers POST `JSON.stringify(chunk)` directly; the
    // webapp wraps that in `{ data: <string>, id }` and stores it on
    // S2. The records endpoint hands `data` back as the original
    // string — unlike session.out (where the writer puts a chunk
    // OBJECT into the envelope and the route forwards it as an
    // object). Defensive: handle both shapes so future writer changes
    // on either side don't silently lose records.
    let chunk: unknown = record.data;
    if (typeof chunk === "string") {
      try {
        chunk = JSON.parse(chunk);
      } catch {
        continue;
      }
    }
    if (!chunk || typeof chunk !== "object") continue;
    const kind = (chunk as { kind?: unknown }).kind;
    if (kind !== "message") continue;
    const payload = (
      chunk as {
        payload?: { trigger?: unknown; message?: unknown; metadata?: unknown };
      }
    ).payload;
    if (!payload || payload.trigger !== "submit-message") continue;
    const message = payload.message;
    if (!message || typeof message !== "object") continue;
    out.push({
      message: message as TUIMessage,
      metadata: payload.metadata,
      seqNum: record.seqNum,
    });
  }
  return out;
}

/**
 * Test-only entry point that bypasses
 * `__setReplaySessionInTailImplForTests` and reaches the real
 * `apiClient.readSessionStreamRecords` + filter pipeline. Mirrors
 * `__replaySessionOutTailProductionPathForTests`.
 * @internal
 */
export async function __replaySessionInTailProductionPathForTests<
  TUIMessage extends UIMessage,
>(
  sessionId: string,
  options?: { lastEventId?: string }
): Promise<{ message: TUIMessage; metadata: unknown; seqNum: number }[]> {
  const saved = replaySessionInTailImpl;
  replaySessionInTailImpl = undefined;
  try {
    return await replaySessionInTail<TUIMessage>(sessionId, options);
  } finally {
    replaySessionInTailImpl = saved;
  }
}

/**
 * Resolve the Session handle for the current chat.agent run.
 *
 * Two contexts populate this:
 *   1. Inside a `chat.agent` run — `locals.chatSessionHandleKey` is set
 *      at boot (lines 4665 / 4760).
 *   2. Inside an `ai.toolExecute` subtask — the tool wrapper threads the
 *      parent's `chatId` through tool metadata under `METADATA_KEY`. We
 *      lazily open the session from that chatId here so subtasks can use
 *      `chat.stream.writer({ target: "root" })` to write back to the
 *      parent's chat session without any wiring.
 *
 * Throws if neither is available.
 * @internal
 */
function getChatSession(): SessionHandle {
  let handle = locals.get(chatSessionHandleKey);
  if (handle) return handle;

  // Fallback: subtask context. The parent threaded chatId via tool metadata.
  const toolMeta = metadata.get(METADATA_KEY) as ToolCallExecutionOptions | undefined;
  if (toolMeta?.chatId) {
    handle = sessions.open(toolMeta.chatId);
    locals.set(chatSessionHandleKey, handle);
    return handle;
  }

  throw new Error(
    "chat.agent session handle is not initialized. This indicates a chat.agent helper was used outside of a chat.agent run, or the transport did not send a sessionId."
  );
}

/**
 * Stamp `gen_ai.conversation.id` on the active span at chat-run boot.
 * The run-level span is already alive when the run callback fires, so
 * `TaskContextSpanProcessor.onStart` (which stamps subsequent spans
 * automatically) won't catch it — set explicitly here.
 */
function stampConversationIdOnActiveSpan(
  conversationId: string | undefined,
  span = trace.getActiveSpan()
): void {
  if (!span || !conversationId) return;
  span.setAttribute(SemanticInternalAttributes.GEN_AI_CONVERSATION_ID, conversationId);
}

type ToolResultContent = Array<
  | {
    type: "text";
    text: string;
  }
  | {
    type: "image";
    data: string;
    mimeType?: string;
  }
>;

export type ToolOptions<TResult> = {
  experimental_toToolResultContent?: (result: TResult) => ToolResultContent;
};

/** Satisfies AI SDK `ToolSet` index signature alongside concrete `Tool` input/output types. */
type ToolSetCompatible<T extends Tool<any, any>> = T & NonNullable<ToolSet[string]>;

function assertTaskUsableAsTool(task: AnyTask): void {
  if (("schema" in task && !task.schema) || ("jsonSchema" in task && !task.jsonSchema)) {
    throw new Error(
      "Cannot convert this task to to a tool because the task has no schema. Make sure to either use schemaTask or a task with an input jsonSchema."
    );
  }
}

/**
 * Shared implementation: run a task as a tool invocation (`triggerAndSubscribe` + tool metadata).
 * Used by {@link toolExecute} and the deprecated `ai.tool()` wrapper.
 */
function createTaskToolExecuteHandler<
  TIdentifier extends string,
  TTaskSchema extends TaskSchema | undefined = undefined,
  TInput = void,
  TOutput = unknown,
>(
  task: TaskWithSchema<TIdentifier, TTaskSchema, TOutput> | Task<TIdentifier, TInput, TOutput>
): (input: unknown, toolOpts: ToolCallOptions | undefined) => Promise<TOutput> {
  assertTaskUsableAsTool(task);

  return async function taskToolExecuteHandler(
    input: unknown,
    toolOpts: ToolCallOptions | undefined
  ): Promise<TOutput> {
    const toolMeta: ToolCallExecutionOptions = {
      toolCallId: toolOpts?.toolCallId ?? "",
    };
    // v6 passes user context as `experimental_context`, v7 as `context`. Read
    // whichever is set and stamp both so subtasks reading either name work.
    const toolContext = toolOpts?.context ?? toolOpts?.experimental_context;
    if (toolContext !== undefined) {
      try {
        const serialized = JSON.parse(JSON.stringify(toolContext));
        toolMeta.experimental_context = serialized;
        toolMeta.context = serialized;
      } catch {
        /* non-serializable */
      }
    }

    const chatCtx = locals.get(chatTurnContextKey);
    if (chatCtx) {
      toolMeta.chatId = chatCtx.chatId;
      toolMeta.turn = chatCtx.turn;
      toolMeta.continuation = chatCtx.continuation;
      toolMeta.clientData = chatCtx.clientData;
    }

    const chatLocals: Record<string, unknown> = {};
    for (const entry of chatLocalRegistry) {
      const value = locals.get(entry.key);
      if (value !== undefined) {
        chatLocals[entry.id] = value;
      }
    }
    if (Object.keys(chatLocals).length > 0) {
      toolMeta.chatLocals = chatLocals;
    }

    return await task
      .triggerAndSubscribe(input as inferSchemaIn<TTaskSchema>, {
        metadata: {
          [METADATA_KEY]: toolMeta as any,
        },
        tags: toolOpts?.toolCallId ? [`toolCallId:${toolOpts.toolCallId}`] : undefined,
        signal: toolOpts?.abortSignal,
      })
      .unwrap();
  };
}

/**
 * Returns an `execute` function for the AI SDK `tool()` helper (or any compatible tool definition).
 * Preferred API for task-backed tools: the same Trigger wiring as the deprecated `ai.tool()`
 * (`triggerAndSubscribe`, tool-call metadata, chat context, `chat.local` serialization) without
 * building the tool object. You supply `description`, `inputSchema`, and any AI-SDK-only options
 * (e.g. `experimental_toToolResultContent`) on `tool()` yourself.
 *
 * @example
 * ```ts
 * import { tool } from "ai";
 * import { z } from "zod";
 * import { ai } from "@trigger.dev/sdk/ai";
 * import { myTask } from "./trigger/myTask";
 *
 * export const myTool = tool({
 *   description: myTask.description ?? "",
 *   inputSchema: z.object({ id: z.string() }),
 *   execute: ai.toolExecute(myTask),
 * });
 * ```
 */
function toolExecute<TIdentifier extends string, TInput = void, TOutput = unknown>(
  task: Task<TIdentifier, TInput, TOutput>
): (input: TInput, toolOpts: ToolCallOptions) => Promise<TOutput>;
function toolExecute<
  TIdentifier extends string,
  TTaskSchema extends TaskSchema | undefined = undefined,
  TOutput = unknown,
>(
  task: TaskWithSchema<TIdentifier, TTaskSchema, TOutput>
): (input: inferSchemaIn<TTaskSchema>, toolOpts: ToolCallOptions) => Promise<TOutput>;
function toolExecute<
  TIdentifier extends string,
  TTaskSchema extends TaskSchema | undefined = undefined,
  TInput = void,
  TOutput = unknown,
>(
  task: TaskWithSchema<TIdentifier, TTaskSchema, TOutput> | Task<TIdentifier, TInput, TOutput>
): (
  input: TTaskSchema extends TaskSchema ? inferSchemaIn<TTaskSchema> : TInput,
  toolOpts: ToolCallOptions
) => Promise<TOutput> {
  return createTaskToolExecuteHandler(task) as (
    input: TTaskSchema extends TaskSchema ? inferSchemaIn<TTaskSchema> : TInput,
    toolOpts: ToolCallOptions
  ) => Promise<TOutput>;
}

/**
 * @deprecated Use `tool()` from the `ai` package with `execute: ai.toolExecute(task)` instead.
 * This helper may be removed in a future major release.
 */
function toolFromTask<TIdentifier extends string, TInput = void, TOutput = unknown>(
  task: Task<TIdentifier, TInput, TOutput>,
  options?: ToolOptions<TOutput>
): ToolSetCompatible<Tool<TInput, TOutput>>;
/** @deprecated Use `tool()` from `ai` with `execute: ai.toolExecute(task)`. */
function toolFromTask<
  TIdentifier extends string,
  TTaskSchema extends TaskSchema | undefined = undefined,
  TOutput = unknown,
>(
  task: TaskWithSchema<TIdentifier, TTaskSchema, TOutput>,
  options?: ToolOptions<TOutput>
): ToolSetCompatible<Tool<inferSchemaIn<TTaskSchema>, TOutput>>;
/** @deprecated Use `tool()` from `ai` with `execute: ai.toolExecute(task)`. */
function toolFromTask<
  TIdentifier extends string,
  TTaskSchema extends TaskSchema | undefined = undefined,
  TInput = void,
  TOutput = unknown,
>(
  task: TaskWithSchema<TIdentifier, TTaskSchema, TOutput> | Task<TIdentifier, TInput, TOutput>,
  options?: ToolOptions<TOutput>
): ToolSetCompatible<
  TTaskSchema extends TaskSchema ? Tool<inferSchemaIn<TTaskSchema>, TOutput> : Tool<TInput, TOutput>
> {
  const executeFromTaskInput = createTaskToolExecuteHandler(task);

  // Zod-backed tasks: use static `tool()` so runtime shape matches `ToolSet`. Generic task context
  // prevents `tool()` overloads from inferring input; `as any` is localized to this call only.
  if ("schema" in task && task.schema && isSchemaZodEsque(task.schema)) {
    const staticTool = aiTool({
      description: task.description ?? "",
      inputSchema: zodSchema(task.schema as any),
      execute: async (input: unknown, toolOpts: ToolCallOptions) =>
        executeFromTaskInput(input, toolOpts),
      ...(options?.experimental_toToolResultContent !== undefined
        ? { experimental_toToolResultContent: options.experimental_toToolResultContent }
        : {}),
    } as any);
    return staticTool as unknown as ToolSetCompatible<
      TTaskSchema extends TaskSchema ? Tool<inferSchemaIn<TTaskSchema>, TOutput> : Tool<TInput, TOutput>
    >;
  }

  const toolDefinition = dynamicTool({
    description: task.description,
    inputSchema: convertTaskSchemaToToolParameters(task),
    ...(options?.experimental_toToolResultContent !== undefined
      ? { experimental_toToolResultContent: options.experimental_toToolResultContent }
      : {}),
    execute: async (input: unknown, toolOpts: ToolCallOptions) =>
      executeFromTaskInput(input, toolOpts),
  });

  return toolDefinition as unknown as ToolSetCompatible<
    TTaskSchema extends TaskSchema ? Tool<inferSchemaIn<TTaskSchema>, TOutput> : Tool<TInput, TOutput>
  >;
}

function getToolOptionsFromMetadata(): ToolCallExecutionOptions | undefined {
  const tool = metadata.get(METADATA_KEY);
  if (!tool) {
    return undefined;
  }
  return tool as ToolCallExecutionOptions;
}

/**
 * Get the current tool call ID from inside a subtask invoked via `ai.toolExecute()` (or legacy `ai.tool()`).
 * Returns `undefined` if not running as a tool subtask.
 */
function getToolCallId(): string | undefined {
  return getToolOptionsFromMetadata()?.toolCallId;
}

/**
 * Get the chat context from inside a subtask invoked via `ai.toolExecute()` (or legacy `ai.tool()`) within a `chat.agent`.
 * Pass `typeof yourChatTask` as the type parameter to get typed `clientData`.
 * Returns `undefined` if the parent is not a chat task.
 *
 * @example
 * ```ts
 * const ctx = ai.chatContext<typeof myChat>();
 * // ctx?.clientData is typed based on myChat's clientDataSchema
 * ```
 */
function getToolChatContext<TChatTask extends AnyTask = AnyTask>():
  | ChatTurnContext<InferChatClientData<TChatTask>>
  | undefined {
  const opts = getToolOptionsFromMetadata();
  if (!opts?.chatId) return undefined;
  return {
    chatId: opts.chatId,
    turn: opts.turn ?? 0,
    continuation: opts.continuation ?? false,
    clientData: opts.clientData as InferChatClientData<TChatTask>,
  };
}

/**
 * Get the chat context from inside a subtask, throwing if not in a chat context.
 * Pass `typeof yourChatTask` as the type parameter to get typed `clientData`.
 *
 * @example
 * ```ts
 * const ctx = ai.chatContextOrThrow<typeof myChat>();
 * // ctx.chatId, ctx.clientData are guaranteed non-null
 * ```
 */
function getToolChatContextOrThrow<TChatTask extends AnyTask = AnyTask>(): ChatTurnContext<
  InferChatClientData<TChatTask>
> {
  const ctx = getToolChatContext<TChatTask>();
  if (!ctx) {
    throw new Error(
      "ai.chatContextOrThrow() called outside of a chat.agent context. " +
      "This helper can only be used inside a subtask invoked via ai.toolExecute() (or legacy ai.tool()) from a chat.agent."
    );
  }
  return ctx;
}

function convertTaskSchemaToToolParameters(
  task: AnyTask | TaskWithSchema<any, any, any>
): Schema<unknown> {
  if ("schema" in task) {
    // If TaskSchema is ArkTypeEsque, use ai.jsonSchema to convert it to a Schema
    if ("toJsonSchema" in task.schema && typeof task.schema.toJsonSchema === "function") {
      return jsonSchema((task.schema as any).toJsonSchema());
    }

    // If TaskSchema is ZodEsque, use ai.zodSchema to convert it to a Schema
    if (isSchemaZodEsque(task.schema)) {
      return zodSchema(task.schema as any);
    }
  }

  if ("jsonSchema" in task) {
    return jsonSchema(task.jsonSchema as JSONSchema7);
  }

  throw new Error(
    "Cannot convert task to a tool. Make sure to use a task with a schema or jsonSchema."
  );
}

export const ai = {
  /**
   * @deprecated Use `tool()` from the `ai` package with `execute: ai.toolExecute(task)` instead.
   */
  tool: toolFromTask,
  /**
   * Preferred: return value for the `execute` field of AI SDK `tool()`. Keeps Trigger subtask and
   * metadata behavior without coupling to a specific `ai` version’s `Tool` / `ToolSet` types.
   */
  toolExecute,
  currentToolOptions: getToolOptionsFromMetadata,
  /** Get the tool call ID from inside a subtask invoked via `ai.toolExecute()` (or legacy `ai.tool()`). */
  toolCallId: getToolCallId,
  /** Get chat context (chatId, turn, clientData, etc.) from inside a subtask of a `chat.agent`. Returns undefined if not in a chat context. */
  chatContext: getToolChatContext,
  /** Get chat context or throw if not in a chat context. Pass `typeof yourChatTask` for typed clientData. */
  chatContextOrThrow: getToolChatContextOrThrow,
};

/**
 * Creates a public access token for a chat task.
 *
 * This is a convenience helper that creates a multi-use trigger public token
 * scoped to the given task. Use it in a server action to provide the frontend
 * `TriggerChatTransport` with an `accessToken`.
 *
 * @example
 * ```ts
 * // actions.ts
 * "use server";
 * import { chat } from "@trigger.dev/sdk/ai";
 * import type { myChat } from "@/trigger/chat";
 *
 * export const getChatToken = () => chat.createAccessToken<typeof myChat>("my-chat");
 * ```
 */
function createChatAccessToken<TTask extends AnyTask>(
  taskId: TaskIdentifier<TTask>
): Promise<string> {
  return auth.createTriggerPublicToken(taskId as string, { expirationTime: "24h" });
}

// ---------------------------------------------------------------------------
// Chat transport helpers — backend side
// ---------------------------------------------------------------------------

/**
 * Typed chat output stream — `.writer()`, `.pipe()`, `.append()`, and
 * `.read()` methods pre-bound to this run's Session `.out` channel and
 * typed to `UIMessageChunk`.
 *
 * Use from within a `chat.agent` run to write custom chunks:
 * ```ts
 * const { waitUntilComplete } = chat.stream.writer({
 *   execute: ({ write }) => {
 *     write({ type: "text-start", id: "status-1" });
 *     write({ type: "text-delta", id: "status-1", delta: "Processing..." });
 *     write({ type: "text-end", id: "status-1" });
 *   },
 * });
 * await waitUntilComplete();
 * ```
 *
 * Backed by the Session primitive so a chat's output outlives any single
 * run — subscribers (browser transport, server-side `ChatStream`) read
 * the session's `.out`, not a per-run stream. Run-scoped `target`
 * options on `.pipe()` are honoured as no-ops; the session is the target.
 */
const chatStream: RealtimeDefinedStream<UIMessageChunk> = {
  // Stable opaque label for the run-scoped `RealtimeDefinedStream` shape.
  // `chatStream` is backed by the Session's `.out` channel — this id is
  // not the real addressing key (the session is). Kept as a literal so
  // the facade type stays satisfied without re-introducing a top-level
  // constant; dashboards/telemetry that already read "chat" keep working.
  id: "chat",
  pipe(value, options) {
    const { target: _target, ...sessionOptions } = (options ?? {}) as PipeStreamOptions;
    return getChatSession().out.pipe<UIMessageChunk>(
      value,
      sessionOptions as SessionPipeStreamOptions
    );
  },
  async read(_runId, options) {
    // Session channels don't need a runId — the session is the address.
    // Keep the signature for backward compatibility with the run-scoped
    // RealtimeDefinedStream shape, but ignore the argument.
    return getChatSession().out.read<UIMessageChunk>(
      options as SessionSubscribeOptions<UIMessageChunk> | undefined
    );
  },
  async append(value, options) {
    const { target: _target, ...sessionOptions } = (options ?? {}) as AppendStreamOptions;
    return getChatSession().out.append(value, sessionOptions as SessionPipeStreamOptions);
  },
  writer(options) {
    return getChatSession().out.writer<UIMessageChunk>(options);
  },
};

// ---------------------------------------------------------------------------
// chat.response — write data parts that persist to the response message
// ---------------------------------------------------------------------------

/**
 * Write data parts that both stream to the frontend AND persist in
 * `onTurnComplete`'s `responseMessage` and `uiMessages`.
 *
 * Non-transient data chunks (`type` starts with `data-`, no `transient: true`)
 * are queued for accumulation into the assistant response message.
 * Transient or non-data chunks are streamed only (same as `chat.stream`).
 *
 * @example
 * ```ts
 * // Persists to responseMessage.parts
 * chat.response.write({ type: "data-handover", data: { context: summary } });
 *
 * // Transient — streams only, not in responseMessage
 * chat.response.write({ type: "data-progress", data: { percent: 50 }, transient: true });
 * ```
 */
const chatResponse = {
  /**
   * Write a single chunk. Non-transient data parts are accumulated into the
   * response message; everything else is stream-only.
   */
  write(part: UIMessageChunk): void {
    queueResponsePart(part);
    const { waitUntilComplete } = chatStream.writer({
      spanName: "chat.response.write",
      collapsed: true,
      execute: ({ write }) => {
        write(part);
      },
    });
    waitUntilComplete().catch(() => {});
  },
};

// ---------------------------------------------------------------------------
// ChatWriter — stream writer for callbacks
// ---------------------------------------------------------------------------

/**
 * A stream writer passed to chat lifecycle callbacks (`onPreload`, `onChatStart`,
 * `onTurnStart`, `onTurnComplete`, `onCompacted`).
 *
 * Write custom `UIMessageChunk` parts (e.g. `data-*` parts) directly to the chat
 * stream without the ceremony of `chat.stream.writer({ execute })`.
 *
 * The writer is lazy — no stream overhead if you don't call `write()` or `merge()`.
 *
 * @example
 * ```ts
 * onTurnStart: async ({ writer }) => {
 *   writer.write({ type: "data-status", data: { loading: true } });
 * },
 * onTurnComplete: async ({ writer, uiMessages }) => {
 *   writer.write({ type: "data-analytics", data: { messageCount: uiMessages.length } });
 * },
 * ```
 */
export type ChatWriter = {
  /** Write a single UIMessageChunk to the chat stream. */
  write(part: UIMessageChunk): void;
  /** Merge another stream's chunks into the chat stream. */
  merge(stream: ReadableStream<UIMessageChunk>): void;
};

/**
 * Creates a lazy ChatWriter that only opens a realtime stream on first use.
 * Call `flush()` after the callback returns to await stream completion.
 * @internal
 */
function createLazyChatWriter(): { writer: ChatWriter; flush: () => Promise<void> } {
  let writeImpl: ((part: UIMessageChunk) => void) | null = null;
  let mergeImpl: ((stream: ReadableStream<UIMessageChunk>) => void) | null = null;
  let waitPromise: (() => Promise<unknown>) | null = null;
  let resolveExecute: (() => void) | null = null;

  function ensureInitialized() {
    if (writeImpl) return;

    const executePromise = new Promise<void>((resolve) => {
      resolveExecute = resolve;
    });

    const { waitUntilComplete } = chatStream.writer({
      collapsed: true,
      spanName: "callback writer",
      execute: ({ write, merge }) => {
        writeImpl = write;
        mergeImpl = merge;
        return executePromise; // Keep execute alive until flush()
      },
    });
    waitPromise = waitUntilComplete;
  }

  return {
    writer: {
      write(part: UIMessageChunk) {
        ensureInitialized();
        queueResponsePart(part);
        writeImpl!(part);
      },
      merge(stream: ReadableStream<UIMessageChunk>) {
        ensureInitialized();
        mergeImpl!(stream);
      },
    },
    async flush() {
      if (resolveExecute) {
        resolveExecute(); // Signal execute to complete
        await waitPromise!(); // Wait for stream to finish piping
      }
    },
  };
}

/**
 * Runs a callback with a lazy ChatWriter, flushing the stream after completion.
 * @internal
 */
async function withChatWriter<T>(fn: (writer: ChatWriter) => Promise<T> | T): Promise<T> {
  const { writer, flush } = createLazyChatWriter();
  const result = await fn(writer);
  await flush();
  return result;
}

// `ChatTaskWirePayload` and `ChatInputChunk` live in `./ai-shared.ts` so
// browser bundles (which import them via `chat-client.ts` / `chat.ts`)
// can pull the types without dragging `ai.ts` into the client graph.
// Re-exported here so `@trigger.dev/sdk/ai` consumers see them.
import type { ChatTaskWirePayload, ChatInputChunk } from "./ai-shared.js";
export type { ChatTaskWirePayload, ChatInputChunk } from "./ai-shared.js";

/**
 * The payload shape passed to the `chatAgent` run function.
 *
 * - `messages` contains model-ready messages (converted via `convertToModelMessages`) —
 *   pass these directly to `streamText`.
 * - `clientData` contains custom data from the frontend (the `metadata` field from `sendMessage()`).
 *
 * The backend accumulates the full conversation history across turns, so the frontend
 * only needs to send new messages after the first turn.
 */
export type ChatTaskPayload<TClientData = unknown> = {
  /** Model-ready messages — pass directly to `streamText({ messages })`. */
  messages: ModelMessage[];

  /** The unique identifier for the chat session */
  chatId: string;

  /**
   * The trigger type:
   * - `"submit-message"`: A new user message
   * - `"regenerate-message"`: Regenerate the last assistant response
   * - `"preload"`: Run was preloaded before the first message (only on turn 0)
   * - `"action"`: A typed action from the frontend (see `actionSchema` + `onAction`).
   *   The action has already been applied before `run()` fires — check `trigger === "action"`
   *   to short-circuit the LLM call when an action doesn't need a response.
   * - `"close"`: The chat session is being closed (internal; `run()` is not called).
   */
  trigger: "submit-message" | "regenerate-message" | "preload" | "action" | "close";

  /** The ID of the message to regenerate (only for `"regenerate-message"`) */
  messageId?: string;

  /** Custom data from the frontend (passed via `metadata` on `sendMessage()` or the transport). */
  clientData?: TClientData;

  /** Whether this run is continuing an existing chat (previous run timed out or was cancelled). False for brand new chats. */
  continuation: boolean;
  /** The run ID of the previous run (only set when `continuation` is true). */
  previousRunId?: string;
  /** Whether this run was preloaded before the first message. */
  preloaded: boolean;
  /**
   * The friendlyId of the Session primitive backing this chat. Use with
   * `sessions.open(sessionId)` when you need direct access to the session's
   * `.in` / `.out` channels outside the hooks the agent already wires for
   * you. Undefined only for legacy transports that predate the sessions
   * migration.
   */
  sessionId?: string;
};

/**
 * Abort signals provided to the `chatAgent` run function.
 */
export type ChatTaskSignals = {
  /** Combined signal — fires on run cancel OR stop generation. Pass to `streamText`. */
  signal: AbortSignal;
  /** Fires only when the run is cancelled, expired, or exceeds maxDuration. */
  cancelSignal: AbortSignal;
  /** Fires only when the frontend stops generation for this turn (per-turn, reset each turn). */
  stopSignal: AbortSignal;
};

/**
 * The full payload passed to a `chatAgent` run function.
 * Extends `ChatTaskPayload` (the wire payload) with abort signals.
 */
export type ChatTaskRunPayload<
  TClientData = unknown,
  TTools extends ToolSet = ToolSet,
> = ChatTaskPayload<TClientData> &
  ChatTaskSignals & {
    /**
     * Task run context — same object as the `ctx` passed to a standard `task({ run })` handler’s second argument.
     * Use for tags, metadata, parent run links, or any API that needs the full run record.
     */
    ctx: TaskRunContext;
    /** Token usage from the previous turn. Undefined on turn 0. */
    previousTurnUsage?: LanguageModelUsage;
    /** Cumulative token usage across all completed turns so far. */
    totalUsage: LanguageModelUsage;
    /**
     * The resolved tool set for this turn, the same `tools` you declared on
     * `chat.agent({ tools })` (or the result of the per-turn `tools` function).
     * Pass straight to `streamText({ tools })` so you don't redeclare them:
     *
     * ```ts
     * run: ({ messages, tools, signal }) =>
     *   streamText({ model, messages, tools, abortSignal: signal })
     * ```
     *
     * Declaring `tools` on the config is also what lets the SDK re-run each
     * tool's `toModelOutput` when it re-converts prior-turn history (see the
     * `tools` option on `chat.agent`). Empty object when no `tools` were declared.
     */
    tools: TTools;
  };

// Input streams for bidirectional chat communication
//
// Both `messagesInput` and `stopInput` are thin facades over the current
// run's Session `.in` channel. The Session carries a single tagged stream
// (`ChatInputChunk`); these facades filter by `kind` so existing call
// sites (both internal and exposed via `chat.messages` / `chat.createStopSignal`)
// keep their original shape. Each accessor resolves the session handle
// lazily via `getChatSession()` so the module-level references stay
// compatible with the pre-migration wiring.
const messagesInput: RealtimeDefinedInputStream<ChatTaskWirePayload> = {
  id: "chat-messages",
  on(handler) {
    return getChatSession().in.on<ChatInputChunk>((chunk) => {
      if (chunk.kind === "message") {
        return handler(chunk.payload);
      }
    });
  },
  once(options) {
    const ctx = taskContext.ctx;
    const runId = ctx?.run.id;

    return new InputStreamOncePromise<ChatTaskWirePayload>((resolve, reject) => {
      tracer
        .startActiveSpan(
          options?.spanName ?? `chat.messages.once()`,
          async () => {
            while (true) {
              const result = await getChatSession().in.once<ChatInputChunk>(options);
              if (!result.ok) {
                resolve(result as InputStreamOnceResult<ChatTaskWirePayload>);
                return;
              }
              if (result.output.kind === "message") {
                resolve({ ok: true, output: result.output.payload });
                return;
              }
              // Non-message chunks (stops) are handled by the stopInput
              // facade's persistent listener; loop and wait for the next.
            }
          },
          {
            attributes: {
              [SemanticInternalAttributes.STYLE_ICON]: "streams",
              [SemanticInternalAttributes.ENTITY_TYPE]: "input-stream",
              ...(runId
                ? {
                    [SemanticInternalAttributes.ENTITY_ID]: `${runId}:chat-messages`,
                  }
                : {}),
              streamId: "chat-messages",
              ...accessoryAttributes({
                items: [{ text: "chat-messages", variant: "normal" }],
                style: "codepath",
              }),
            },
          }
        )
        .catch(reject);
    });
  },
  peek() {
    const chunk = getChatSession().in.peek<ChatInputChunk>();
    if (chunk && chunk.kind === "message") return chunk.payload;
    return undefined;
  },
  wait(options) {
    return new ManualWaitpointPromise<ChatTaskWirePayload>(async (resolve, reject) => {
      try {
        while (true) {
          const result = await getChatSession().in.wait<ChatInputChunk>(options);
          if (!result.ok) {
            resolve(result);
            return;
          }
          if (result.output.kind === "message") {
            resolve({ ok: true, output: result.output.payload });
            return;
          }
          // Stop chunks are handled by the stopInput facade's persistent
          // listener; loop back into the suspending wait.
        }
      } catch (error) {
        reject(error);
      }
    });
  },
  async waitWithIdleTimeout(options) {
    while (true) {
      const result = await getChatSession().in.waitWithIdleTimeout<ChatInputChunk>(options);
      if (!result.ok) return result;
      if (result.output.kind === "message") {
        return { ok: true, output: result.output.payload };
      }
      // Swallow stop-kind chunks — persistent stop listener already handled
      // the abort; we just loop for the next message.
    }
  },
  async send(_runId, data, options) {
    // The `runId` argument is kept for signature parity with
    // `RealtimeDefinedInputStream` but ignored — sessions are addressed
    // by sessionId, not runId. Callers producing messages from outside
    // the run should prefer the transport's `session.in.send(...)` path.
    await getChatSession().in.send(
      { kind: "message", payload: data } satisfies ChatInputChunk,
      options?.requestOptions
    );
  },
};

const stopInput: RealtimeDefinedInputStream<{ stop: true; message?: string }> = {
  id: "chat-stop",
  on(handler) {
    return getChatSession().in.on<ChatInputChunk>((chunk) => {
      if (chunk.kind === "stop") {
        return handler({ stop: true, message: chunk.message });
      }
    });
  },
  once(options) {
    const ctx = taskContext.ctx;
    const runId = ctx?.run.id;

    return new InputStreamOncePromise<{ stop: true; message?: string }>((resolve, reject) => {
      tracer
        .startActiveSpan(
          options?.spanName ?? `chat.stop.once()`,
          async () => {
            while (true) {
              const result = await getChatSession().in.once<ChatInputChunk>(options);
              if (!result.ok) {
                resolve(result as InputStreamOnceResult<{ stop: true; message?: string }>);
                return;
              }
              if (result.output.kind === "stop") {
                resolve({
                  ok: true,
                  output: { stop: true, message: result.output.message },
                });
                return;
              }
            }
          },
          {
            attributes: {
              [SemanticInternalAttributes.STYLE_ICON]: "streams",
              [SemanticInternalAttributes.ENTITY_TYPE]: "input-stream",
              ...(runId
                ? {
                    [SemanticInternalAttributes.ENTITY_ID]: `${runId}:chat-stop`,
                  }
                : {}),
              streamId: "chat-stop",
              ...accessoryAttributes({
                items: [{ text: "chat-stop", variant: "normal" }],
                style: "codepath",
              }),
            },
          }
        )
        .catch(reject);
    });
  },
  peek() {
    const chunk = getChatSession().in.peek<ChatInputChunk>();
    if (chunk && chunk.kind === "stop") {
      return { stop: true, message: chunk.message };
    }
    return undefined;
  },
  wait(options) {
    return new ManualWaitpointPromise<{ stop: true; message?: string }>(async (resolve, reject) => {
      try {
        while (true) {
          const result = await getChatSession().in.wait<ChatInputChunk>(options);
          if (!result.ok) {
            resolve(result);
            return;
          }
          if (result.output.kind === "stop") {
            resolve({
              ok: true,
              output: { stop: true, message: result.output.message },
            });
            return;
          }
        }
      } catch (error) {
        reject(error);
      }
    });
  },
  async waitWithIdleTimeout(options) {
    while (true) {
      const result = await getChatSession().in.waitWithIdleTimeout<ChatInputChunk>(options);
      if (!result.ok) return result;
      if (result.output.kind === "stop") {
        return { ok: true, output: { stop: true, message: result.output.message } };
      }
    }
  },
  async send(_runId, data, options) {
    await getChatSession().in.send(
      { kind: "stop", message: data?.message } satisfies ChatInputChunk,
      options?.requestOptions
    );
  },
};

/**
 * Signal received by a `handover-prepare` agent run waiting on
 * `session.in`. Either the customer's first-turn `streamText` finished
 * with pending tool calls (`"handover"` — agent picks up from tool
 * execution), or it finished pure-text (`"handover-skip"` — agent
 * exits cleanly without making an LLM call).
 * @internal
 */
type HandoverSignal =
  | {
      kind: "handover";
      partialAssistantMessage: ModelMessage[];
      messageId?: string;
      /**
       * Whether the customer's step 1 is the final response. When
       * true, the agent's turn loop runs hooks but skips the LLM
       * call (the partial IS the response). When false, the agent
       * runs `streamText` which executes pending tool-calls via the
       * approval round and continues from step 2.
       */
      isFinal: boolean;
    }
  | { kind: "handover-skip" };

/**
 * Internal facade for waiting on the handover signal. Mirrors
 * `messagesInput` / `stopInput` so the wait paths and tracing
 * attributes stay consistent across all input-stream branches.
 * @internal
 */
const handoverInput = {
  async waitWithIdleTimeout(options: {
    idleTimeoutInSeconds: number;
    timeout?: string;
    spanName?: string;
    skipSuspend?: boolean;
  }) {
    while (true) {
      const result = await getChatSession().in.waitWithIdleTimeout<ChatInputChunk>(options);
      if (!result.ok) return result;
      if (
        result.output.kind === "handover" ||
        result.output.kind === "handover-skip"
      ) {
        return { ok: true as const, output: result.output as HandoverSignal };
      }
      // Other kinds (message, stop) are not expected during handover-prepare.
      // Loop back; the message and stop facades have their own listeners
      // running so signals on those kinds aren't lost.
    }
  },
};

/**
 * Per-turn deferred promises. Registered via `chat.defer()`, awaited
 * before `onTurnComplete` fires. Reset each turn.
 * @internal
 */
const chatDeferKey = locals.create<Set<Promise<unknown>>>("chat.defer");

/**
 * Run-scoped slot holding the partial assistant message handed over by
 * `chat.handover` from a customer's first-turn `streamText`. Appended
 * to `accumulatedMessages` during turn 0 setup so `streamText` resumes
 * at tool execution. Cleared (read once) after consumption.
 * @internal
 */
const chatHandoverPartialKey = locals.create<ModelMessage[]>("chat.handoverPartial");

/**
 * Run-scoped slot holding the assistant `messageId` the customer's
 * `chat.handover` handler used for its step-1 stream. The agent reuses
 * it on the agent-side `toUIMessageStream` (and the synthesized
 * partial UIMessage in `originalMessages`) so all chunks merge into a
 * single assistant message on the browser side.
 * @internal
 */
const chatHandoverMessageIdKey = locals.create<string>("chat.handoverMessageId");

/**
 * Run-scoped slot indicating that the customer's step-1 head-start
 * response is the FINAL turn response. When true, turn 0 runs through
 * the full turn-loop hooks but SKIPS the `userRun` / `streamText`
 * call — the customer's partial already IS the response. The agent's
 * `onTurnComplete` fires with that partial so persistence + any
 * post-turn work happens normally. Cleared after consumption.
 * @internal
 */
const chatHandoverIsFinalKey = locals.create<boolean>("chat.handoverIsFinal");

/**
 * Build a UIMessage representation of a `chat.handover` partial so AI
 * SDK's `processUIMessageStream` can transition `tool-output-available`
 * chunks (emitted by the initial-tool-execution branch when the
 * approval round runs) onto the existing tool-call. Without this,
 * `state.message.parts` is empty when the agent's `streamText`
 * finishes, and AI SDK throws
 * `UIMessageStreamError: No tool invocation found`.
 *
 * Only the assistant message matters — the synthesized
 * `tool-approval-response` rows are AI-SDK-internal and don't need a
 * UIMessage representation. We map:
 *   - `text` parts → `{ type: "text", text }`
 *   - `tool-call` parts → `{ type: "tool-${name}", toolCallId,
 *      state: "input-available", input }`
 *   - `tool-approval-request` parts → skipped (AI SDK derives the
 *      approval state from chunks during processing)
 *
 * @internal
 */
function synthesizeHandoverUIMessage(
  partial: ModelMessage[],
  messageId?: string
): UIMessage | undefined {
  const assistant = partial.find((m) => m.role === "assistant");
  if (!assistant || typeof assistant.content === "string") return undefined;

  const parts: UIMessage["parts"] = [];
  for (const part of assistant.content as Array<{
    type: string;
    text?: string;
    toolCallId?: string;
    toolName?: string;
    input?: unknown;
  }>) {
    if (part.type === "text" && typeof part.text === "string") {
      parts.push({ type: "text", text: part.text } as UIMessage["parts"][number]);
    } else if (part.type === "tool-call" && part.toolCallId && part.toolName) {
      parts.push({
        type: `tool-${part.toolName}`,
        toolCallId: part.toolCallId,
        state: "input-available",
        input: part.input,
      } as unknown as UIMessage["parts"][number]);
    }
    // tool-approval-request parts intentionally skipped — they're an
    // AI-SDK protocol detail, not a UI surface.
  }

  if (parts.length === 0) return undefined;

  // Use the customer's step-1 messageId if provided (so the agent's
  // post-handover chunks merge into the same assistant message on the
  // browser). Fall back to a fresh id only if the handover signal
  // didn't carry one.
  return {
    id: messageId ?? generateMessageId(),
    role: "assistant",
    parts,
  } as UIMessage;
}

/**
 * Per-turn background context queue. Messages added via `chat.backgroundWork.inject()`
 * are drained at the next `prepareStep` boundary and appended to the model messages.
 * @internal
 */
const chatBackgroundQueueKey = locals.create<ModelMessage[]>("chat.backgroundQueue");

/**
 * Run-scoped pipe counter. Stored in locals so concurrent runs in the
 * same worker don't share state.
 * @internal
 */
const chatPipeCountKey = locals.create<number>("chat.pipeCount");
const chatStopControllerKey = locals.create<AbortController>("chat.stopController");
/** Static (task-level) UIMessageStream options, set once during chatAgent setup. @internal */
const chatUIStreamStaticKey = locals.create<ChatUIMessageStreamOptions<UIMessage>>(
  "chat.uiMessageStreamOptions.static"
);
/** Per-turn UIMessageStream options, set via chat.setUIMessageStreamOptions(). @internal */
const chatUIStreamPerTurnKey = locals.create<ChatUIMessageStreamOptions<UIMessage>>(
  "chat.uiMessageStreamOptions.perTurn"
);

/**
 * Run-scoped `toolCallId → assistant messageId` map. Records the head
 * assistant id whenever the accumulator absorbs an assistant message
 * containing tool parts. Used as a fallback in the id-merge for
 * incoming tool-answer messages — if the AI SDK regenerates the
 * assistant id on a HITL `addToolOutput` resume, we look up the
 * original head id by `toolCallId` and rewrite it before the merge.
 *
 * Customer-side workaround for the same case is documented in Arena
 * AI's chat-agent task; lifting it into the SDK so customers don't
 * have to. See TRI-9137.
 * @internal
 */
const chatToolCallToMessageIdKey = locals.create<Map<string, string>>(
  "chat.toolCallToMessageId"
);

function recordToolCallIdsFromMessage(message: { id?: string; role?: string; parts?: unknown[] } | undefined) {
  if (!message || message.role !== "assistant" || !message.id) return;
  let map = locals.get(chatToolCallToMessageIdKey);
  if (!map) {
    map = new Map();
    locals.set(chatToolCallToMessageIdKey, map);
  }
  for (const part of message.parts ?? []) {
    if (typeof part !== "object" || part == null) continue;
    const toolCallId = (part as { toolCallId?: unknown }).toolCallId;
    if (typeof toolCallId === "string" && toolCallId.length > 0) {
      map.set(toolCallId, message.id);
    }
  }
}

function rewriteIncomingIdViaToolCallMap<T extends { id?: string; parts?: unknown[] }>(
  incoming: T
): T {
  const map = locals.get(chatToolCallToMessageIdKey);
  if (!map || map.size === 0) return incoming;
  for (const part of incoming.parts ?? []) {
    if (typeof part !== "object" || part == null) continue;
    const toolCallId = (part as { toolCallId?: unknown }).toolCallId;
    if (typeof toolCallId !== "string" || toolCallId.length === 0) continue;
    const headId = map.get(toolCallId);
    if (headId && headId !== incoming.id) {
      return { ...incoming, id: headId };
    }
  }
  return incoming;
}

// ---------------------------------------------------------------------------
// Token usage helpers (internal)
// ---------------------------------------------------------------------------

/** Convenience re-export of the AI SDK's `LanguageModelUsage` type. */
export type ChatTurnUsage = LanguageModelUsage;

function emptyUsage(): LanguageModelUsage {
  return {
    inputTokens: undefined,
    outputTokens: undefined,
    totalTokens: undefined,
    inputTokenDetails: {
      noCacheTokens: undefined,
      cacheReadTokens: undefined,
      cacheWriteTokens: undefined,
    },
    outputTokenDetails: { textTokens: undefined, reasoningTokens: undefined },
  };
}

function addUsage(a: LanguageModelUsage, b: LanguageModelUsage): LanguageModelUsage {
  const add = (x: number | undefined, y: number | undefined) =>
    x != null || y != null ? (x ?? 0) + (y ?? 0) : undefined;
  return {
    inputTokens: add(a.inputTokens, b.inputTokens),
    outputTokens: add(a.outputTokens, b.outputTokens),
    totalTokens: add(a.totalTokens, b.totalTokens),
    inputTokenDetails: {
      noCacheTokens: add(a.inputTokenDetails?.noCacheTokens, b.inputTokenDetails?.noCacheTokens),
      cacheReadTokens: add(
        a.inputTokenDetails?.cacheReadTokens,
        b.inputTokenDetails?.cacheReadTokens
      ),
      cacheWriteTokens: add(
        a.inputTokenDetails?.cacheWriteTokens,
        b.inputTokenDetails?.cacheWriteTokens
      ),
    },
    outputTokenDetails: {
      textTokens: add(a.outputTokenDetails?.textTokens, b.outputTokenDetails?.textTokens),
      reasoningTokens: add(
        a.outputTokenDetails?.reasoningTokens,
        b.outputTokenDetails?.reasoningTokens
      ),
    },
  };
}

// ---------------------------------------------------------------------------
// chat.setMessages — replace accumulated messages for compaction
// ---------------------------------------------------------------------------

/** @internal */
const chatOverrideMessagesKey = locals.create<UIMessage[]>("chat.overrideMessages");

/**
 * Tracks the current accumulated UI messages so chat.history.all() can
 * read them from outside the chatAgent closure.
 * @internal
 */
const chatCurrentUIMessagesKey = locals.create<UIMessage[]>("chat.currentUIMessages");

/**
 * Replace the accumulated conversation messages for the current run.
 *
 * Call from `onTurnStart` to compact before `run()` executes, or from
 * `onTurnComplete` to compact before the next turn. Takes `UIMessage[]`
 * and converts to `ModelMessage[]` internally.
 */
function setChatMessages<TUIM extends UIMessage = UIMessage>(uiMessages: TUIM[]): void {
  locals.set(chatOverrideMessagesKey, uiMessages);
}

// ---------------------------------------------------------------------------
// chat.history — imperative message history mutations
// ---------------------------------------------------------------------------

/**
 * Read the current message history state, accounting for pending overrides.
 * @internal
 */
function getChatHistoryState(): UIMessage[] {
  const pending = locals.get(chatOverrideMessagesKey);
  if (pending) return pending;
  return locals.get(chatCurrentUIMessagesKey) ?? [];
}

/**
 * A tool call surfaced by `chat.history.getPendingToolCalls()` /
 * `getResolvedToolCalls()`. Identifies the call by its `toolCallId` plus
 * the `messageId` of the assistant message that hosts it, so callers can
 * locate the part precisely without re-walking the chain.
 */
export type ChatToolCallRef = {
  toolCallId: string;
  toolName: string;
  messageId: string;
};

/**
 * A new tool result surfaced by `chat.history.extractNewToolResults()`.
 * `errorText` is set iff the part is in `output-error` state; otherwise
 * `output` carries the resolved value.
 */
export type ChatNewToolResult = {
  toolCallId: string;
  toolName: string;
  output: unknown;
  errorText?: string;
};

/**
 * Tool parts that are "done" — either succeeded with a value or failed
 * with an error. Excludes pending (`input-streaming`/`input-available`)
 * and approval (`approval-requested`/`approval-responded`) states.
 * @internal
 */
function isResolvedToolState(state: unknown): state is "output-available" | "output-error" {
  return state === "output-available" || state === "output-error";
}

/** @internal */
function isPendingToolState(state: unknown): state is "input-available" {
  return state === "input-available";
}

/**
 * Walk an assistant message and yield each tool part with its callId,
 * name, and state. Skips non-assistant messages and non-tool parts.
 * @internal
 */
function* iterateToolParts(
  message: UIMessage
): Generator<{ part: any; toolCallId: string; toolName: string; state: unknown }> {
  if (message.role !== "assistant") return;
  for (const part of (message.parts ?? []) as any[]) {
    if (!isToolUIPart(part)) continue;
    const toolCallId = part.toolCallId;
    if (typeof toolCallId !== "string" || toolCallId.length === 0) continue;
    yield {
      part,
      toolCallId,
      toolName: getToolName(part),
      state: part.state,
    };
  }
}

/**
 * Walk a partial assistant message and surface the tool calls the model
 * had started but never received a result for. Used at recovery boot to
 * populate `RecoveryBootEvent.pendingToolCalls`.
 *
 * The partial assistant in question is the orphan from a dead run — its
 * `turn-complete` never fired, so any `input-available` tool part is
 * truly orphan (NOT a stable HITL pause; HITL parts live on settled
 * messages, not on the partial).
 *
 * @internal
 */
function extractPendingToolCallsFromPartial(
  partial: UIMessage | undefined
): RecoveryPendingToolCall[] {
  if (!partial) return [];
  const out: RecoveryPendingToolCall[] = [];
  const parts = (partial.parts ?? []) as any[];
  for (let i = 0; i < parts.length; i++) {
    const part = parts[i];
    if (!isToolUIPart(part)) continue;
    if (!isPendingToolState(part.state)) continue;
    const toolCallId = part.toolCallId;
    if (typeof toolCallId !== "string" || toolCallId.length === 0) continue;
    out.push({
      toolCallId,
      toolName: getToolName(part),
      input: part.input,
      partIndex: i,
    });
  }
  return out;
}

/**
 * Tool parts on the *leaf* assistant message that are still waiting on
 * an answer (`input-available` state). Used to gate fresh user turns
 * during HITL flows.
 * @internal
 */
function getPendingToolCallsFromHistory(messages: UIMessage[]): ChatToolCallRef[] {
  for (let i = messages.length - 1; i >= 0; i--) {
    const msg = messages[i]!;
    if (msg.role !== "assistant") continue;
    const pending: ChatToolCallRef[] = [];
    for (const { toolCallId, toolName, state } of iterateToolParts(msg)) {
      if (isPendingToolState(state)) {
        pending.push({ toolCallId, toolName, messageId: msg.id });
      }
    }
    return pending;
  }
  return [];
}

/**
 * All tool parts across the chain that have already produced an output
 * (`output-available` or `output-error`). Used to dedup re-saves when
 * the AI SDK resends an assistant with progressively more answered
 * parts.
 * @internal
 */
function getResolvedToolCallsFromHistory(messages: UIMessage[]): ChatToolCallRef[] {
  const out: ChatToolCallRef[] = [];
  for (const msg of messages) {
    for (const { toolCallId, toolName, state } of iterateToolParts(msg)) {
      if (isResolvedToolState(state)) {
        out.push({ toolCallId, toolName, messageId: msg.id });
      }
    }
  }
  return out;
}

/**
 * Pure helper: tool parts in `message` that have a fresh result not
 * already represented by the resolved toolCallIds in `messages`. The
 * `errorText` field is present only for `output-error` parts.
 *
 * Within a single `message`, duplicate `toolCallId`s emit only once
 * (first occurrence wins). This guards against malformed assistants
 * with repeated tool parts.
 * @internal
 */
function extractNewToolResultsFromHistory(
  message: UIMessage,
  messages: UIMessage[]
): ChatNewToolResult[] {
  const resolved = new Set(
    getResolvedToolCallsFromHistory(messages).map((r) => r.toolCallId)
  );
  const seen = new Set<string>();
  const out: ChatNewToolResult[] = [];
  for (const { part, toolCallId, toolName, state } of iterateToolParts(message)) {
    if (!isResolvedToolState(state)) continue;
    if (resolved.has(toolCallId)) continue;
    if (seen.has(toolCallId)) continue;
    seen.add(toolCallId);
    if (state === "output-error") {
      out.push({ toolCallId, toolName, output: undefined, errorText: part.errorText });
    } else {
      out.push({ toolCallId, toolName, output: part.output });
    }
  }
  return out;
}

/**
 * Per-turn merge of an incoming wire `UIMessage` onto the matching entry
 * a `hydrateMessages` hook (or the default accumulator) provides. Used
 * to fold tool-state advances from the client into the agent's
 * authoritative chain without trusting the wire copy for fields the
 * LLM consumes.
 *
 * `hydrated` is treated as the source of truth for everything outside
 * tool-state advancement: text, reasoning blobs, provider metadata,
 * and tool `input` all stay as hydrated had them. We only overlay
 * tool parts whose incoming state is wire-advanced — `output-available`
 * / `output-error` (HITL `addToolOutput`) or `approval-responded` /
 * `output-denied` (approval flow) — and only the corresponding
 * resolution fields (`output` / `errorText` / `approval`). Hydrated
 * `input` and everything else stay put.
 *
 * Without this, a slim wire copy (which `TriggerChatTransport` /
 * `AgentChat.sendRaw` ship by default on HITL continuations) would
 * clobber the hydrated assistant — the next LLM call would receive a
 * tool call with no `input` and 4xx.
 *
 * @internal
 */
function mergeIncomingIntoHydrated<TMsg extends UIMessage>(
  hydrated: TMsg,
  incoming: UIMessage
): TMsg {
  const incomingAdvancedByCallId = new Map<string, any>();
  for (const part of (incoming.parts ?? []) as any[]) {
    if (!isToolUIPart(part)) continue;
    const toolCallId = part.toolCallId;
    if (typeof toolCallId !== "string" || toolCallId.length === 0) continue;
    if (!isWireAdvanceableToolState(part.state)) continue;
    incomingAdvancedByCallId.set(toolCallId, part);
  }

  if (incomingAdvancedByCallId.size === 0) return hydrated;

  let mutated = false;
  const hydratedParts = (hydrated.parts ?? []) as any[];
  const mergedParts = hydratedParts.map((part) => {
    if (!isToolUIPart(part)) return part;
    const toolCallId = part.toolCallId;
    if (typeof toolCallId !== "string" || toolCallId.length === 0) return part;
    const incomingPart = incomingAdvancedByCallId.get(toolCallId);
    if (!incomingPart) return part;
    // Terminal hydrated states (`output-available`, `output-error`,
    // `output-denied`) are authoritative — never regressed by a stale
    // wire arrival (replay, retry, out-of-order). `output-denied`
    // matters here because the wire's `approval-responded` could
    // otherwise overwrite a hydrated denial back to a non-terminal
    // state.
    if (isResolvedToolState(part.state) || part.state === "output-denied") {
      return part;
    }
    // Same state on both sides — no progression to apply.
    if (part.state === incomingPart.state) return part;
    mutated = true;
    if (incomingPart.state === "output-available") {
      return {
        ...part,
        state: incomingPart.state,
        output: incomingPart.output,
        ...(incomingPart.approval !== undefined ? { approval: incomingPart.approval } : {}),
      };
    }
    if (incomingPart.state === "output-error") {
      return {
        ...part,
        state: incomingPart.state,
        errorText: incomingPart.errorText,
        ...(incomingPart.approval !== undefined ? { approval: incomingPart.approval } : {}),
      };
    }
    // approval-responded / output-denied — overlay state + approval.
    return {
      ...part,
      state: incomingPart.state,
      ...(incomingPart.approval !== undefined ? { approval: incomingPart.approval } : {}),
    };
  });

  if (!mutated) return hydrated;
  return { ...hydrated, parts: mergedParts };
}

/**
 * Mirror of `slimSubmitMessageForWire`'s predicate. Kept here so the
 * agent runtime doesn't have to import from `ai-shared.ts` for a
 * one-liner. See that file for the full state-machine docs.
 *
 * @internal
 */
function isWireAdvanceableToolState(
  state: unknown
): state is "output-available" | "output-error" | "approval-responded" | "output-denied" {
  return (
    state === "output-available" ||
    state === "output-error" ||
    state === "approval-responded" ||
    state === "output-denied"
  );
}

/**
 * Imperative API for reading and modifying the accumulated message history.
 *
 * Mutations use the same deferred override mechanism as `chat.setMessages()`:
 * they are applied at lifecycle checkpoints (after hooks return). Reads are
 * synchronous against the current accumulator state.
 *
 * Can be called from `onTurnStart`, `onBeforeTurnComplete`, `onTurnComplete`,
 * `run()`, `onAction`, or AI SDK tools.
 */
const chatHistory = {
  /** Read the current accumulated UI messages (copy). */
  all(): UIMessage[] {
    return [...getChatHistoryState()];
  },

  /**
   * Read the current chain as an ordered `UIMessage[]`. Identical to
   * `all()`; use whichever name reads better in context.
   */
  getChain(): UIMessage[] {
    return chatHistory.all();
  },

  /**
   * Find a message by id. Returns `undefined` if no message with that id
   * is present in the current chain.
   */
  findMessage(messageId: string): UIMessage | undefined {
    return getChatHistoryState().find((m) => m.id === messageId);
  },

  /**
   * Tool calls on the *most recent* assistant message that are still in
   * `input-available` state (waiting on an `addToolOutput` answer). The
   * scan walks back from the tail and stops at the first assistant
   * message it finds, so a trailing user message does not change the
   * result — pending tool calls remain pending until they're resolved
   * on that assistant or the assistant is removed.
   *
   * Use this to gate fresh user turns or actions during HITL flows: if
   * `getPendingToolCalls().length > 0`, an `addToolOutput` is expected.
   *
   * Returns `[]` if there is no assistant message yet, or if the most
   * recent assistant has no pending tool calls.
   *
   * Approval flows (`approval-requested` / `approval-responded` states)
   * are not surfaced here. Those are about the user authorizing a tool
   * to run; "pending" is about the user *answering* a tool call.
   */
  getPendingToolCalls(): ChatToolCallRef[] {
    return getPendingToolCallsFromHistory(getChatHistoryState());
  },

  /**
   * Tool calls across the chain with a final result (`output-available`
   * or `output-error`). Use this to dedup re-saves when the AI SDK
   * resends an assistant message with progressively more answered parts.
   */
  getResolvedToolCalls(): ChatToolCallRef[] {
    return getResolvedToolCallsFromHistory(getChatHistoryState());
  },

  /**
   * Pure helper: returns the tool parts in `message` whose results are
   * not already represented in the current chain. Use this when
   * persisting tool results to your own store: each call surfaces only
   * the *new* answers, so writes stay idempotent across re-streams.
   * Duplicate `toolCallId`s within `message` itself are also collapsed
   * to a single entry.
   */
  extractNewToolResults(message: UIMessage): ChatNewToolResult[] {
    return extractNewToolResultsFromHistory(message, getChatHistoryState());
  },

  /** Replace all accumulated messages. Same as `chat.setMessages()`. */
  set(messages: UIMessage[]): void {
    locals.set(chatOverrideMessagesKey, messages);
  },

  /** Remove a specific message by ID. */
  remove(messageId: string): void {
    chatHistory.set(getChatHistoryState().filter((m) => m.id !== messageId));
  },

  /** Keep messages up to and including the given ID (undo/rollback). */
  rollbackTo(messageId: string): void {
    const current = getChatHistoryState();
    const idx = current.findIndex((m) => m.id === messageId);
    if (idx !== -1) {
      chatHistory.set(current.slice(0, idx + 1));
    }
  },

  /** Replace a specific message by ID (edit). */
  replace(messageId: string, message: UIMessage): void {
    chatHistory.set(getChatHistoryState().map((m) => (m.id === messageId ? message : m)));
  },

  /** Keep only messages in the given range. */
  slice(start: number, end?: number): void {
    chatHistory.set(getChatHistoryState().slice(start, end));
  },
};

/**
 * Model-only message override. Set by compaction to replace only the model
 * messages (what goes to the LLM) without affecting UI messages (what gets
 * persisted and displayed). This preserves full conversation history for the
 * user while keeping LLM context compact.
 * @internal
 */
const chatOverrideModelMessagesKey = locals.create<ModelMessage[]>("chat.overrideModelMessages");

// ---------------------------------------------------------------------------
// chat.compaction — prepareStep compaction API
// ---------------------------------------------------------------------------

/** State stored in locals during prepareStep compaction. */
interface CompactionState {
  summary: string;
  baseResponseMessageCount: number;
}

/** @internal */
const chatCompactionStateKey = locals.create<CompactionState>("chat.compaction");
const chatOnCompactedKey =
  locals.create<(event: CompactedEvent) => Promise<void> | void>("chat.onCompacted");
/** @internal Full task `ctx` for the active `chat.agent` run (for hooks invoked from nested compaction). */
const chatAgentRunContextKey = locals.create<TaskRunContext>("chat.agentRunContext");
const chatPrepareMessagesKey =
  locals.create<(event: PrepareMessagesEvent<unknown>) => ModelMessage[] | Promise<ModelMessage[]>>(
    "chat.prepareMessages"
  );
/**
 * @internal The raw `tools` option from `chat.agent({ tools })`, either a
 * static `ToolSet` or a per-turn function. Set once at boot.
 */
const chatToolsOptionKey = locals.create<
  ToolSet | ((event: ResolveToolsEvent<unknown>) => ToolSet | Promise<ToolSet>)
>("chat.toolsOption");
/**
 * @internal The concrete `ToolSet` resolved for the current turn. Read by
 * `toModelMessages` so `convertToModelMessages` can re-run `toModelOutput` on
 * prior-turn tool results. Unset when no `tools` were declared (preserves the
 * exact pre-feature conversion behavior).
 */
const chatResolvedToolsKey = locals.create<ToolSet>("chat.resolvedTools");

/** @internal Flag set by `chat.requestUpgrade()` to exit the loop after the current turn. */
const chatUpgradeRequestedKey = locals.create<boolean>("chat.upgradeRequested");

/**
 * @internal Flag set by `chat.endRun()` to exit the loop after the current
 * turn completes, without any upgrade semantics. Checked at the same
 * post-turn / pre-wait sites as `chatUpgradeRequestedKey`.
 */
const chatEndRunRequestedKey = locals.create<boolean>("chat.endRunRequested");

/**
 * Event passed to `summarize` callbacks.
 */
export type SummarizeEvent = {
  /** The current model messages to summarize. */
  messages: ModelMessage[];
  /** Full usage object from the triggering step/turn. */
  usage?: LanguageModelUsage;
  /** Cumulative token usage across all completed turns. Present in chat.agent contexts. */
  totalUsage?: LanguageModelUsage;
  /** The chat session ID (if running inside a chat.agent). */
  chatId?: string;
  /** The current turn number (0-indexed, if inside a chat.agent). */
  turn?: number;
  /** Custom data from the frontend (if inside a chat.agent). */
  clientData?: unknown;
  /**
   * Where compaction is running:
   * - `"inner"` — between tool-call steps (prepareStep)
   * - `"outer"` — between turns
   */
  source?: "inner" | "outer";
  /** The step number (0-indexed). Only present when `source` is `"inner"`. */
  stepNumber?: number;
};

/**
 * Event passed to `compactUIMessages` and `compactModelMessages` callbacks.
 */
export type CompactMessagesEvent<TUIM extends UIMessage = UIMessage> = {
  /** The generated summary text. */
  summary: string;
  /** The current UI messages (full conversation). */
  uiMessages: TUIM[];
  /** The current model messages (full conversation). */
  modelMessages: ModelMessage[];
  /** The chat session ID. */
  chatId: string;
  /** The current turn number (0-indexed). */
  turn: number;
  /** Custom data from the frontend. */
  clientData?: unknown;
  /**
   * Where compaction is running:
   * - `"inner"` — between tool-call steps (prepareStep)
   * - `"outer"` — between turns
   */
  source: "inner" | "outer";
};

/**
 * Options for the `compaction` field on `chat.agent()`.
 *
 * Handles compaction automatically in both the inner loop (prepareStep, between
 * tool-call steps) and the outer loop (between turns, for single-step responses
 * where prepareStep never fires).
 */
export type ChatAgentCompactionOptions<TUIM extends UIMessage = UIMessage> = {
  /** Decide whether to compact. Return true to trigger compaction. */
  shouldCompact: (event: ShouldCompactEvent) => boolean | Promise<boolean>;
  /** Generate a summary from the current messages. Return the summary text. */
  summarize: (event: SummarizeEvent) => Promise<string>;
  /**
   * Transform UI messages after compaction (what gets persisted and displayed).
   * Default: preserve all UI messages unchanged.
   *
   * @example
   * ```ts
   * // Flatten to summary
   * compactUIMessages: ({ summary }) => [{
   *   id: generateId(), role: "assistant",
   *   parts: [{ type: "text", text: `[Summary]\n\n${summary}` }],
   * }],
   *
   * // Summary + keep last 4 messages
   * compactUIMessages: ({ uiMessages, summary }) => [
   *   { id: generateId(), role: "assistant",
   *     parts: [{ type: "text", text: `[Summary]\n\n${summary}` }] },
   *   ...uiMessages.slice(-4),
   * ],
   * ```
   */
  compactUIMessages?: (event: CompactMessagesEvent<TUIM>) => TUIM[] | Promise<TUIM[]>;
  /**
   * Transform model messages after compaction (what gets sent to the LLM).
   * Default: replace all with a single summary message.
   *
   * @example
   * ```ts
   * // Summary + keep last 2 model messages
   * compactModelMessages: ({ modelMessages, summary }) => [
   *   { role: "user", content: summary },
   *   ...modelMessages.slice(-2),
   * ],
   * ```
   */
  compactModelMessages?: (
    event: CompactMessagesEvent<TUIM>
  ) => ModelMessage[] | Promise<ModelMessage[]>;
};

/** @internal */
const chatAgentCompactionKey =
  locals.create<ChatAgentCompactionOptions<UIMessage>>("chat.agentCompaction");

// ---------------------------------------------------------------------------
// Pending messages — mid-execution message injection via prepareStep
// ---------------------------------------------------------------------------

/**
 * Event passed to `shouldInject` and `prepareMessages` callbacks.
 */
export type PendingMessagesBatchEvent<TUIM extends UIMessage = UIMessage> = {
  /** All pending UI messages that arrived during streaming (batch). */
  messages: TUIM[];
  /** Current model messages in the conversation. */
  modelMessages: ModelMessage[];
  /** Completed steps so far. */
  steps: CompactionStep[];
  /** Current step number (0-indexed). */
  stepNumber: number;
  /** Chat session ID. */
  chatId: string;
  /** Current turn number (0-indexed). */
  turn: number;
  /** Custom data from the frontend. */
  clientData?: unknown;
};

/**
 * Event passed to `onReceived` callback (per-message, as they arrive).
 */
export type PendingMessageReceivedEvent<TUIM extends UIMessage = UIMessage> = {
  /** The UI message that arrived during streaming. */
  message: TUIM;
  /** Chat session ID. */
  chatId: string;
  /** Current turn number (0-indexed). */
  turn: number;
};

/**
 * Event passed to `onInjected` callback (batch, after injection).
 */
export type PendingMessagesInjectedEvent<TUIM extends UIMessage = UIMessage> = {
  /** All UI messages that were injected. */
  messages: TUIM[];
  /** The model messages that were injected. */
  injectedModelMessages: ModelMessage[];
  /** Chat session ID. */
  chatId: string;
  /** Current turn number (0-indexed). */
  turn: number;
  /** Step number where injection occurred. */
  stepNumber: number;
};

/**
 * Options for the `pendingMessages` field on `chat.agent()`, `chat.createSession()`,
 * or `ChatMessageAccumulator`.
 *
 * Configures how messages that arrive during streaming are handled. When
 * `shouldInject` is provided and returns `true`, the full batch of pending
 * messages is injected between tool-call steps via `prepareStep`.
 * Otherwise, messages queue for the next turn.
 */
export type PendingMessagesOptions<TUIM extends UIMessage = UIMessage> = {
  /**
   * Decide whether to inject pending messages between tool-call steps.
   * Called once per step boundary with the full batch of pending messages.
   * If absent, no injection happens — messages only queue for the next turn.
   */
  shouldInject?: (event: PendingMessagesBatchEvent<TUIM>) => boolean | Promise<boolean>;
  /**
   * Transform the batch of pending messages before injection.
   * Return the model messages to inject.
   * Default: convert each UI message via `convertToModelMessages`.
   */
  prepare?: (event: PendingMessagesBatchEvent<TUIM>) => ModelMessage[] | Promise<ModelMessage[]>;
  /** Called when a message arrives during streaming (per-message). */
  onReceived?: (event: PendingMessageReceivedEvent<TUIM>) => void | Promise<void>;
  /** Called after a batch of messages is injected via `prepareStep`. */
  onInjected?: (event: PendingMessagesInjectedEvent<TUIM>) => void | Promise<void>;
};

/**
 * The data part type used to signal that pending messages were injected
 * between tool-call steps. The frontend can match on this to render
 * injection points inline in the assistant response.
 */
// `PENDING_MESSAGE_INJECTED_TYPE` lives in `./ai-shared.ts` so the chat
// React hooks (`@trigger.dev/sdk/chat/react`) can import it without
// dragging `ai.ts` into the browser graph. Re-exported here so
// `@trigger.dev/sdk/ai` consumers still see it.
export { PENDING_MESSAGE_INJECTED_TYPE, upsertIncomingMessage } from "./ai-shared.js";
import { PENDING_MESSAGE_INJECTED_TYPE } from "./ai-shared.js";

/** @internal */
type SteeringQueueEntry = { uiMessage: UIMessage; modelMessages: ModelMessage[] };
/** @internal */
const chatPendingMessagesKey = locals.create<PendingMessagesOptions>("chat.pendingMessages");
/** @internal */
const chatSteeringQueueKey = locals.create<SteeringQueueEntry[]>("chat.steeringQueue");
/** @internal — IDs of messages that were successfully injected via prepareStep */
const chatInjectedMessageIdsKey = locals.create<Set<string>>("chat.injectedMessageIds");
/** @internal — non-transient data parts queued via chat.response or writer.write() for accumulation into the response message */
const chatResponsePartsKey = locals.create<unknown[]>("chat.responseParts");

/**
 * Check if a chunk is a non-transient data part that should persist to the response message.
 * @internal
 */
function isNonTransientDataPart(part: unknown): boolean {
  if (typeof part !== "object" || part === null) return false;
  const p = part as Record<string, unknown>;
  return typeof p.type === "string" && p.type.startsWith("data-") && p.transient !== true;
}

/**
 * Queue a chunk for accumulation into the response message (if it's a non-transient data part).
 * Called by `chat.response.write()` and `ChatWriter.write()`.
 * @internal
 */
function queueResponsePart(part: unknown): void {
  if (!isNonTransientDataPart(part)) return;
  const parts = locals.get(chatResponsePartsKey) ?? [];
  parts.push(part);
  locals.set(chatResponsePartsKey, parts);
}

/**
 * Event passed to the `prepareMessages` hook.
 */
export type PrepareMessagesEvent<TClientData = unknown> = {
  /** The messages to transform. Return the transformed array. */
  messages: ModelMessage[];
  /** Why messages are being prepared. */
  reason:
  | "run" // Messages being passed to run() for streamText
  | "compaction-rebuild" // Rebuilding from a previous compaction summary
  | "compaction-result"; // Fresh compaction just produced these messages
  /** The chat session ID. */
  chatId: string;
  /** The current turn number (0-indexed). */
  turn: number;
  /** Custom data from the frontend. */
  clientData?: TClientData;
};

/**
 * Event passed to the per-turn `tools` function form on `chat.agent`.
 *
 * Use this when the active tool set depends on per-turn context (the user, a
 * feature flag, etc.). Return the `ToolSet` to use for converting this turn's
 * history. Only `inputSchema` and `toModelOutput` are read during conversion,
 * so a lightweight map (no `execute`) is fine.
 */
export type ResolveToolsEvent<TClientData = unknown> = {
  /** The chat session ID. */
  chatId: string;
  /** The current turn number (0-indexed). */
  turn: number;
  /** Whether this run is continuing an existing chat. */
  continuation: boolean;
  /** Custom data from the frontend. */
  clientData?: TClientData;
};

/**
 * Data shape for `data-compaction` stream chunks emitted during compaction.
 * Use to type the `data` field when rendering compaction parts in the frontend.
 */
export type CompactionChunkData = {
  status: "compacting" | "complete";
  totalTokens: number | undefined;
};

/**
 * Event passed to the `onCompacted` callback.
 */
export type CompactedEvent = {
  /** Task run context — same as `task` lifecycle hooks and `chat.agent` `run({ ctx })`. */
  ctx: TaskRunContext;
  /** The generated summary text. */
  summary: string;
  /** The messages that were compacted (pre-compaction). */
  messages: ModelMessage[];
  /** Number of messages before compaction. */
  messageCount: number;
  /** Token usage from the step that triggered compaction. */
  usage: LanguageModelUsage;
  /** Total token count that triggered compaction. */
  totalTokens: number | undefined;
  /** Input token count from the triggering step. */
  inputTokens: number | undefined;
  /** Output token count from the triggering step. */
  outputTokens: number | undefined;
  /** The step number where compaction occurred (0-indexed). */
  stepNumber: number;
  /** The chat session ID (if running inside a chat.agent). */
  chatId?: string;
  /** The current turn number (if running inside a chat.agent). */
  turn?: number;
  /** Stream writer — write custom `UIMessageChunk` parts to the chat stream. Lazy: no overhead if unused. */
  writer: ChatWriter;
};

/**
 * Event passed to `shouldCompact` callbacks.
 */
export type ShouldCompactEvent = {
  /** The current model messages (full conversation). */
  messages: ModelMessage[];
  /** Total token count from the triggering step/turn. */
  totalTokens: number | undefined;
  /** Input token count from the triggering step/turn. */
  inputTokens: number | undefined;
  /** Output token count from the triggering step/turn. */
  outputTokens: number | undefined;
  /** Full usage object from the triggering step/turn. */
  usage?: LanguageModelUsage;
  /** Cumulative token usage across all completed turns. Present in chat.agent contexts. */
  totalUsage?: LanguageModelUsage;
  /** The chat session ID (if running inside a chat.agent). */
  chatId?: string;
  /** The current turn number (0-indexed, if inside a chat.agent). */
  turn?: number;
  /** Custom data from the frontend (if inside a chat.agent). */
  clientData?: unknown;
  /**
   * Where this check is running:
   * - `"inner"` — between tool-call steps (prepareStep)
   * - `"outer"` — between turns (after response, before onBeforeTurnComplete)
   */
  source?: "inner" | "outer";
  /** The step number (0-indexed). Only present when `source` is `"inner"`. */
  stepNumber?: number;
  /** The steps array from prepareStep. Only present when `source` is `"inner"`. */
  steps?: CompactionStep[];
};

/**
 * Options for `chat.compaction()` — the high-level prepareStep factory.
 */
export type CompactionOptions = {
  /** Generate a summary from the current messages. Return the summary text. */
  summarize: (messages: ModelMessage[]) => Promise<string>;
  /** Token threshold — compact when totalTokens exceeds this. Ignored if `shouldCompact` is provided. */
  threshold?: number;
  /** Custom compaction trigger. When provided, used instead of `threshold`. */
  shouldCompact?: (event: ShouldCompactEvent) => boolean | Promise<boolean>;
};

/** A step object as received in prepareStep's `steps` array. */
export type CompactionStep = {
  usage: LanguageModelUsage;
  finishReason: string;
  content: Array<{ type: string; toolCallId?: string }>;
  response: { messages: Array<any> };
};

/**
 * Result of `chat.compact()`. Discriminated union so you can inspect
 * what happened, but also directly compatible with prepareStep's return type.
 *
 * - `"skipped"` — no compaction needed (first step, boundary unsafe, or under threshold). Return `undefined` to prepareStep.
 * - `"rebuilt"` — previous compaction exists, messages rebuilt from summary + new response messages.
 * - `"compacted"` — compaction just happened, includes the generated summary.
 */
export type CompactResult =
  | { type: "skipped" }
  | { type: "rebuilt"; messages: ModelMessage[] }
  | { type: "compacted"; messages: ModelMessage[]; summary: string };

/**
 * Options for `chat.compact()` — the low-level compaction function.
 */
export type CompactOptions = {
  /** Generate a summary from the current messages. Return the summary text. */
  summarize: (messages: ModelMessage[]) => Promise<string>;
  /** Token threshold — compact when totalTokens exceeds this. Ignored if `shouldCompact` is provided. */
  threshold?: number;
  /** Custom compaction trigger. When provided, used instead of `threshold`. */
  shouldCompact?: (event: ShouldCompactEvent) => boolean | Promise<boolean>;
};

/**
 * Check that no tool calls are in-flight in a step's content.
 * Used before compaction to avoid losing tool state mid-execution.
 * @internal
 */
function isStepBoundarySafe(step: {
  finishReason: string;
  content: Array<{ type: string; toolCallId?: string }>;
}): boolean {
  if (step.finishReason === "error") return false;
  const callIds = new Set(
    step.content.filter((p) => p.type === "tool-call").map((p) => p.toolCallId)
  );
  const settledIds = new Set(
    step.content
      .filter((p) => p.type === "tool-result" || p.type === "tool-error")
      .map((p) => p.toolCallId)
  );
  return ![...callIds].some((id) => !settledIds.has(id));
}

/**
 * Apply the prepareMessages hook if one is set in locals.
 * @internal
 */
async function applyPrepareMessages(
  messages: ModelMessage[],
  reason: PrepareMessagesEvent["reason"]
): Promise<ModelMessage[]> {
  const hook = locals.get(chatPrepareMessagesKey);
  if (!hook) return messages;

  const turnCtx = locals.get(chatTurnContextKey);

  return tracer.startActiveSpan(
    "prepareMessages()",
    async () => {
      return hook({
        messages,
        reason,
        chatId: turnCtx?.chatId ?? "",
        turn: turnCtx?.turn ?? 0,
        clientData: turnCtx?.clientData,
      });
    },
    {
      attributes: {
        [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
        [SemanticInternalAttributes.COLLAPSED]: true,
        "chat.prepareMessages.reason": reason,
        "chat.prepareMessages.messageCount": messages.length,
      },
    }
  );
}

/**
 * Resolve the `tools` option into a concrete `ToolSet` and cache it in locals so
 * `toModelMessages` can pass it to `convertToModelMessages`. For the function
 * form, invokes the user function with the given context (or the current turn
 * context when no override is passed). Pass an `override` for the boot-time
 * history conversion, which runs before the per-turn context exists and uses
 * the run/continuation payload's `clientData`.
 *
 * Fails closed: a throwing resolver propagates rather than carrying a prior
 * turn's set forward. The function form can gate capabilities by user or flag,
 * so reusing stale tools would leak capabilities. No-op when no `tools` were
 * declared.
 * @internal
 */
async function resolveTurnTools(
  override?: { chatId: string; turn: number; continuation: boolean; clientData: unknown }
): Promise<void> {
  const option = locals.get(chatToolsOptionKey);
  if (!option) return;

  if (typeof option !== "function") {
    locals.set(chatResolvedToolsKey, option);
    return;
  }

  const ctx = override ?? locals.get(chatTurnContextKey);
  const resolved = await option({
    chatId: ctx?.chatId ?? "",
    turn: ctx?.turn ?? 0,
    continuation: ctx?.continuation ?? false,
    clientData: ctx?.clientData,
  });
  locals.set(chatResolvedToolsKey, resolved);
}

/**
 * Read the current compaction state. Returns the summary and base message count
 * if compaction has occurred in this turn, or `undefined` if not.
 *
 * Use in a custom `prepareStep` to rebuild from a previous compaction:
 * ```ts
 * const state = chat.getCompactionState();
 * if (state) {
 *   return { messages: [{ role: "user", content: state.summary }, ...newMsgs] };
 * }
 * ```
 */
function getCompactionState(): CompactionState | undefined {
  return locals.get(chatCompactionStateKey);
}

/**
 * Low-level compaction for use inside a custom `prepareStep`.
 *
 * Handles the full decision tree: first step, already-compacted rebuild,
 * boundary safety, threshold check, summarization, stream chunks, state
 * storage, and accumulator update.
 *
 * Returns a `CompactResult` — inspect `result.type` to see what happened,
 * or convert to a prepareStep return with `result.type === "skipped" ? undefined : result`.
 *
 * @example
 * ```ts
 * prepareStep: async ({ messages, steps }) => {
 *   // your custom logic here...
 *   const result = await chat.compact(messages, steps, {
 *     threshold: 80_000,
 *     summarize: async (msgs) => generateText({ model, messages: msgs }).then(r => r.text),
 *   });
 *   if (result.type === "compacted") {
 *     logger.info("Compacted!", { summary: result.summary });
 *   }
 *   return result.type === "skipped" ? undefined : result;
 * },
 * ```
 */
async function chatCompact(
  messages: ModelMessage[],
  steps: CompactionStep[],
  options: CompactOptions
): Promise<CompactResult> {
  const currentStep = steps.at(-1);

  // First step — nothing to check
  if (!currentStep) {
    return { type: "skipped" };
  }

  // Already compacted — rebuild from summary + new response messages
  const state = locals.get(chatCompactionStateKey);
  if (state && isStepBoundarySafe(currentStep)) {
    return {
      type: "rebuilt",
      messages: await applyPrepareMessages(
        [
          { role: "user" as const, content: state.summary },
          ...currentStep.response.messages.slice(state.baseResponseMessageCount),
        ],
        "compaction-rebuild"
      ),
    };
  }

  // Boundary unsafe — skip
  if (!isStepBoundarySafe(currentStep)) {
    return { type: "skipped" };
  }

  const totalTokens = currentStep.usage.totalTokens;
  const inputTokens = currentStep.usage.inputTokens;
  const outputTokens = currentStep.usage.outputTokens;

  const turnCtx = locals.get(chatTurnContextKey);
  const stepNumber = steps.length - 1;

  const shouldTrigger = options.shouldCompact
    ? await options.shouldCompact({
      messages,
      totalTokens,
      inputTokens,
      outputTokens,
      usage: currentStep.usage,
      source: "inner",
      stepNumber,
      steps,
      chatId: turnCtx?.chatId,
      turn: turnCtx?.turn,
      clientData: turnCtx?.clientData,
    })
    : totalTokens != null && options.threshold != null && totalTokens > options.threshold;

  if (!shouldTrigger) {
    return { type: "skipped" };
  }

  const result = await tracer.startActiveSpan(
    "context compaction",
    async (span) => {
      const compactionId = generateMessageId();
      let summary!: string;

      const { waitUntilComplete } = chatStream.writer({
        spanName: "stream compaction chunks",
        collapsed: true,
        execute: async ({ write, merge }) => {
          // Control chunks aren't part of UIMessageChunk's discriminated
          // union but flow on the same session.out so subscribers can
          // intercept them — cast on the way out.
          write({ type: "step-start" } as unknown as UIMessageChunk);
          write({
            type: "data-compaction",
            id: compactionId,
            data: { status: "compacting", totalTokens },
            transient: true,
          });

          // Generate summary
          summary = await options.summarize(messages);

          // Store state in locals for subsequent steps
          locals.set(chatCompactionStateKey, {
            summary,
            baseResponseMessageCount: currentStep.response.messages.length,
          });

          // Set model-only override — UI messages stay intact for persistence.
          // The summary becomes the model message history for the next turn,
          // while accumulatedUIMessages keeps the full conversation for display.
          locals.set(chatOverrideModelMessagesKey, [
            {
              role: "assistant" as const,
              content: [{ type: "text" as const, text: `[Conversation summary]\n\n${summary}` }],
            },
          ]);

          // Fire onCompacted hook — pass the existing writer so the callback
          // can write custom chunks without creating a separate stream.
          const onCompactedHook = locals.get(chatOnCompactedKey);
          if (onCompactedHook) {
            await onCompactedHook({
              ctx: locals.get(chatAgentRunContextKey)!,
              summary,
              messages,
              messageCount: messages.length,
              usage: currentStep.usage,
              totalTokens,
              inputTokens,
              outputTokens,
              stepNumber,
              chatId: turnCtx?.chatId,
              turn: turnCtx?.turn,
              writer: { write, merge },
            });
          }

          write({
            type: "data-compaction",
            id: compactionId,
            data: { status: "complete", totalTokens },
            transient: true,
          });
          write({ type: "finish-step" });
        },
      });
      await waitUntilComplete();

      // Set attributes after we have the summary
      span.setAttribute("compaction.summary_length", summary.length);

      return {
        type: "compacted" as const,
        messages: await applyPrepareMessages(
          [{ role: "user" as const, content: summary }],
          "compaction-result"
        ),
        summary,
      };
    },
    {
      attributes: {
        [SemanticInternalAttributes.STYLE_ICON]: "tabler-scissors",
        "compaction.threshold": options.threshold,
        "compaction.total_tokens": totalTokens ?? 0,
        "compaction.input_tokens": inputTokens ?? 0,
        "compaction.message_count": messages.length,
        "compaction.step_number": stepNumber,
        ...(turnCtx?.chatId ? { "compaction.chat_id": turnCtx.chatId } : {}),
        ...(turnCtx?.turn != null ? { "compaction.turn": turnCtx.turn } : {}),
        ...accessoryAttributes({
          items: [
            { text: `${totalTokens ?? 0} tokens`, variant: "normal" },
            { text: `${messages.length} msgs`, variant: "normal" },
          ],
          style: "codepath",
        }),
      },
    }
  );

  return result;
}

/**
 * Returns a `prepareStep` function that handles context compaction automatically.
 *
 * Monitors token usage between tool-call steps. When `totalTokens` exceeds
 * the threshold, generates a summary via `summarize()`, replaces the message
 * history, and emits `data-compaction` stream chunks for the frontend.
 *
 * @example
 * ```ts
 * return streamText({
 *   ...chat.toStreamTextOptions({ registry }),
 *   messages: chat.addCacheBreaks(messages),
 *   prepareStep: chat.compactionStep({
 *     threshold: 80_000,
 *     summarize: async (messages) => {
 *       return generateText({ model, messages: [...messages, { role: "user", content: "Summarize." }] })
 *         .then((r) => r.text);
 *     },
 *   }),
 *   tools: { ... },
 * });
 * ```
 */
function chatCompactionStep(
  options: CompactionOptions
): (args: {
  messages: ModelMessage[];
  steps: CompactionStep[];
}) => Promise<{ messages: ModelMessage[] } | undefined> {
  return async ({ messages, steps }) => {
    const result = await chatCompact(messages, steps, options);
    return result.type === "skipped" ? undefined : result;
  };
}

// ---------------------------------------------------------------------------
// Steering queue drain — shared by toStreamTextOptions, session, accumulator
// ---------------------------------------------------------------------------

/**
 * Drain the steering queue as a batch. Calls `shouldInject` once with all
 * pending messages. If it returns true, calls `prepareMessages` once to
 * transform the batch, then clears the queue.
 * Returns the model messages to inject (empty if none).
 * @internal
 */
async function drainSteeringQueue(
  config: PendingMessagesOptions,
  messages: ModelMessage[],
  steps: CompactionStep[],
  queueOverride?: SteeringQueueEntry[]
): Promise<ModelMessage[]> {
  const queue = queueOverride ?? locals.get(chatSteeringQueueKey);
  if (!queue || queue.length === 0) return [];

  const ctx = locals.get(chatTurnContextKey);
  const stepNumber = steps.length - 1;
  const uiMessages = queue.map((e) => e.uiMessage);

  const batchEvent: PendingMessagesBatchEvent = {
    messages: uiMessages,
    modelMessages: messages,
    steps,
    stepNumber,
    chatId: ctx?.chatId ?? "",
    turn: ctx?.turn ?? 0,
    clientData: ctx?.clientData,
  };

  // Call shouldInject once for the whole batch
  const shouldInject = config.shouldInject ? await config.shouldInject(batchEvent) : false;

  if (!shouldInject) return [];

  // Extract message texts for span attributes
  const messageTexts = uiMessages.map(
    (m) =>
      (m.parts ?? [])
        .filter((p: any) => p.type === "text")
        .map((p: any) => p.text)
        .join("") || ""
  );
  const previewText =
    messageTexts.length === 1 ? messageTexts[0]!.slice(0, 80) : `${queue.length} messages`;

  return tracer.startActiveSpan(
    "pending message injected",
    async () => {
      // Transform the batch — default: concatenate all pre-converted model messages
      const injected = config.prepare
        ? await config.prepare(batchEvent)
        : queue.flatMap((e) => e.modelMessages);

      // Clear the queue and record injected IDs
      queue.length = 0;
      const injectedIds = locals.get(chatInjectedMessageIdsKey);
      if (injectedIds) {
        for (const m of uiMessages) injectedIds.add(m.id);
      }

      // Write injection confirmation chunk to the stream so the frontend
      // knows which messages were injected and where in the response.
      if (injected.length > 0) {
        try {
          const { waitUntilComplete } = chatStream.writer({
            collapsed: true,
            execute: ({ write }) => {
              write({
                type: PENDING_MESSAGE_INJECTED_TYPE,
                id: generateMessageId(),
                data: {
                  messageIds: uiMessages.map((m) => m.id),
                  messages: uiMessages.map((m, idx) => ({
                    id: m.id,
                    text: messageTexts[idx] ?? "",
                  })),
                },
              });
            },
          });
          await waitUntilComplete();
        } catch {
          /* non-fatal — stream write failed */
        }
      }

      // Fire onInjected callback
      if (config.onInjected && injected.length > 0) {
        try {
          await config.onInjected({
            messages: uiMessages,
            injectedModelMessages: injected,
            chatId: ctx?.chatId ?? "",
            turn: ctx?.turn ?? 0,
            stepNumber,
          });
        } catch {
          /* non-fatal */
        }
      }

      return injected;
    },
    {
      attributes: {
        [SemanticInternalAttributes.STYLE_ICON]: "tabler-message-forward",
        "pending.message_count": uiMessages.length,
        "pending.step_number": stepNumber,
        "pending.messages": messageTexts,
        ...(ctx?.chatId ? { "pending.chat_id": ctx.chatId } : {}),
        ...(ctx?.turn != null ? { "pending.turn": ctx.turn } : {}),
        ...accessoryAttributes({
          items: [
            {
              text: `${uiMessages.length} message${uiMessages.length === 1 ? "" : "s"}`,
              variant: "normal",
            },
            { text: `between steps ${stepNumber} and ${stepNumber + 1}`, variant: "normal" },
          ],
          style: "codepath",
        }),
      },
    }
  );
}

// ---------------------------------------------------------------------------
// chat.isCompactionSafe — check if it's safe to compact messages
// ---------------------------------------------------------------------------

/**
 * Checks whether it's safe to compact the message history. Returns `false`
 * if any tool calls are in-flight (incomplete tool invocations without results).
 *
 * Call before `chat.setMessages()` to avoid corrupting tool-call state.
 */
function isCompactionSafe(messages: UIMessage[]): boolean {
  for (const msg of messages) {
    if (msg.role !== "assistant") continue;
    for (const part of msg.parts as any[]) {
      if (part.type === "tool-invocation") {
        const state = part.toolInvocation?.state ?? part.state;
        if (state !== "result" && state !== "error") {
          return false;
        }
      }
    }
  }
  return true;
}

// ---------------------------------------------------------------------------
// chat.prompt — store and retrieve a resolved prompt for the current run
// ---------------------------------------------------------------------------

/**
 * A resolved prompt stored via `chat.prompt.set()`. Either a full `ResolvedPrompt`
 * from `prompts.define().resolve()`, or a lightweight wrapper around a plain string.
 */
export type ChatPromptValue =
  | ResolvedPrompt
  | {
    text: string;
    model: undefined;
    config: undefined;
    promptId: string;
    version: number;
    labels: string[];
    toAISDKTelemetry: (additionalMetadata?: Record<string, string>) => {
      experimental_telemetry: { isEnabled: true; metadata: Record<string, string> };
    };
  };

/** @internal */
const chatPromptKey = locals.create<ChatPromptValue>("chat.prompt");

/**
 * Store a resolved prompt (or plain string) for the current run.
 * Call from any hook (`onPreload`, `onChatStart`, `onTurnStart`) or `run()`.
 */
function setChatPrompt(resolved: ResolvedPrompt | string): void {
  if (typeof resolved === "string") {
    locals.set(chatPromptKey, {
      text: resolved,
      model: undefined,
      config: undefined,
      promptId: "",
      version: 0,
      labels: [],
      toAISDKTelemetry: () => ({
        experimental_telemetry: { isEnabled: true, metadata: {} },
      }),
    });
  } else {
    locals.set(chatPromptKey, resolved);
  }
}

/**
 * Read the stored prompt. Throws if `chat.prompt.set()` has not been called.
 */
function getChatPrompt(): ChatPromptValue {
  const prompt = locals.get(chatPromptKey);
  if (!prompt) {
    throw new Error(
      "chat.prompt() called before chat.prompt.set(). Set a prompt in onPreload, onChatStart, onTurnStart, or run() first."
    );
  }
  return prompt;
}

// ---------------------------------------------------------------------------
// chat.skills — store resolved agent skills and inject them into streamText
// ---------------------------------------------------------------------------

/** @internal */
const chatSkillsKey = locals.create<ResolvedSkill[]>("chat.skills");

/**
 * Store resolved skills for the current run. Call from any hook
 * (`onPreload`, `onChatStart`, `onTurnStart`) or `run()`.
 */
function setChatSkills(skills: ResolvedSkill[]): void {
  locals.set(chatSkillsKey, skills);
}

/** Read the stored skills. Returns `undefined` if none set. */
function getChatSkills(): ResolvedSkill[] | undefined {
  return locals.get(chatSkillsKey);
}

/**
 * Build the system-prompt preamble advertising available skills. Only the
 * frontmatter description surfaces here — full SKILL.md body is loaded
 * on-demand via the `loadSkill` tool.
 */
function buildSkillsSystemPrompt(skills: ResolvedSkill[]): string {
  if (skills.length === 0) return "";
  const lines = skills.map(
    (s) => `- ${s.frontmatter.name}: ${s.frontmatter.description}`
  );
  return [
    "Available skills (call `loadSkill` to read the full instructions before using one):",
    ...lines,
  ].join("\n");
}

/** Resolve a skill by its frontmatter `name`. */
function findSkillByName(skills: ResolvedSkill[], name: string): ResolvedSkill | undefined {
  return skills.find((s) => s.frontmatter.name === name);
}

/**
 * Build the three tools we auto-inject into `streamText` when skills are
 * set: `loadSkill`, `readFile`, `bash`. Scoped per-skill by name.
 *
 * Exported so callers can use the same tools outside the auto-wired path
 * (e.g. in a `chat.createSession` loop with custom streamText).
 */
export function buildSkillTools(skills: ResolvedSkill[]): Record<string, Tool> {
  const loadSkill = aiTool({
    description:
      "Load the full instructions for a skill by its name. Call this first before using a skill.",
    inputSchema: jsonSchema<{ name: string }>({
      type: "object",
      properties: {
        name: {
          type: "string",
          description: "The `name` field from the skill's frontmatter.",
        },
      },
      required: ["name"],
      additionalProperties: false,
    } as JSONSchema7),
    execute: async ({ name }: { name: string }) => {
      const skill = findSkillByName(skills, name);
      if (!skill) {
        return {
          error: `Skill "${name}" not found. Available: ${skills
            .map((s) => s.frontmatter.name)
            .join(", ")}`,
        };
      }
      return {
        name: skill.frontmatter.name,
        description: skill.frontmatter.description,
        body: skill.body,
        path: skill.path,
      };
    },
  });

  const readFile = aiTool({
    description:
      "Read a file from a skill's bundled folder. Paths must be relative to the skill's root.",
    inputSchema: jsonSchema<{ skill: string; path: string }>({
      type: "object",
      properties: {
        skill: { type: "string", description: "The skill's name (from frontmatter)." },
        path: {
          type: "string",
          description: "Relative path inside the skill folder (e.g. `references/citation-style.md`).",
        },
      },
      required: ["skill", "path"],
      additionalProperties: false,
    } as JSONSchema7),
    execute: async ({ skill: skillName, path: relPath }: { skill: string; path: string }) => {
      const skill = findSkillByName(skills, skillName);
      if (!skill) {
        return { error: `Skill "${skillName}" not found.` };
      }
      try {
        return await readFileInSkill({
          skillPath: skill.path,
          relativePath: relPath,
        });
      } catch (err) {
        return { error: (err as Error).message };
      }
    },
  });

  const bash = aiTool({
    description:
      "Run a bash command inside a skill's bundled folder. Use this to invoke the skill's scripts. The working directory is the skill's root.",
    inputSchema: jsonSchema<{ skill: string; command: string }>({
      type: "object",
      properties: {
        skill: { type: "string", description: "The skill's name (from frontmatter)." },
        command: {
          type: "string",
          description: "Bash command to run. Relative script paths resolve against the skill's root.",
        },
      },
      required: ["skill", "command"],
      additionalProperties: false,
    } as JSONSchema7),
    execute: async (
      { skill: skillName, command }: { skill: string; command: string },
      { abortSignal }: { abortSignal?: AbortSignal } = {}
    ) => {
      const skill = findSkillByName(skills, skillName);
      if (!skill) {
        return { error: `Skill "${skillName}" not found.` };
      }
      try {
        return await runBashInSkill({
          skillPath: skill.path,
          command,
          abortSignal,
        });
      } catch (err) {
        return { error: (err as Error).message };
      }
    },
  });

  return { loadSkill, readFile, bash };
}

/**
 * Options for {@link toStreamTextOptions}.
 */
export type ToStreamTextOptionsOptions = {
  /** Additional telemetry metadata merged into `experimental_telemetry.metadata`. */
  telemetry?: Record<string, string>;
  /**
   * An AI SDK provider registry (from `createProviderRegistry`) or any object
   * with a `languageModel(id)` method. When provided and the stored prompt has
   * a `model` string, the resolved `LanguageModel` is included in the returned
   * options so `streamText` uses it directly.
   *
   * The model string should use the `"provider:model-id"` format
   * (e.g. `"openai:gpt-4o"`, `"anthropic:claude-sonnet-4-6"`).
   */
  registry?: { languageModel(modelId: string): unknown };
  /**
   * User-defined tools to merge alongside the auto-injected skill tools
   * (`loadSkill`, `readFile`, `bash`). User tools win on name conflicts.
   *
   * If you don't pass `tools` here and skills are set, the returned options
   * will include just the skill tools — spread after any `tools` you pass
   * directly to `streamText` and they'll be replaced. Easiest: pass all
   * your tools here.
   */
  tools?: Record<string, Tool>;
};

/**
 * Returns an options object ready to spread into `streamText()`.
 *
 * Includes `system`, `experimental_telemetry`, and any config fields
 * (temperature, maxTokens, etc.) from the stored prompt.
 *
 * When a `registry` is provided and the prompt has a `model` string,
 * the resolved `LanguageModel` is included as `model`.
 *
 * If no prompt has been set, returns `{}` (no-op spread).
 */
function toStreamTextOptions(options?: ToStreamTextOptionsOptions): Record<string, unknown> {
  const prompt = locals.get(chatPromptKey);
  const skills = locals.get(chatSkillsKey);
  const result: Record<string, unknown> = {};

  // Build the combined system prompt: stored prompt + skills preamble.
  const promptText = prompt?.text ?? "";
  const skillsText = skills && skills.length > 0 ? buildSkillsSystemPrompt(skills) : "";
  if (promptText || skillsText) {
    result.system = [promptText, skillsText].filter(Boolean).join("\n\n");
  }

  // Prompt-related options (only if chat.prompt.set() was called)
  if (prompt) {
    // Resolve model via registry if both are present
    if (options?.registry && prompt.model) {
      result.model = options.registry.languageModel(prompt.model);
    }

    // Spread config (temperature, maxTokens, etc.)
    if (prompt.config) {
      Object.assign(result, prompt.config);
    }

    // Add telemetry (forward additional metadata from caller)
    const telemetry = prompt.toAISDKTelemetry(options?.telemetry);
    Object.assign(result, telemetry);
  }

  // Skills: merge auto-injected tools with any user-provided tools.
  // User tools override on name conflict (though we namespace ours).
  if (skills && skills.length > 0) {
    const skillTools = buildSkillTools(skills);
    result.tools = { ...skillTools, ...(options?.tools ?? {}) };
  } else if (options?.tools) {
    result.tools = options.tools;
  }

  // Auto-inject prepareStep for compaction, pending messages, and background context injection.
  // This runs regardless of whether a prompt is set — these features are independent.
  const taskCompaction = locals.get(chatAgentCompactionKey);
  const taskPendingMessages = locals.get(chatPendingMessagesKey);

  {
    result.prepareStep = async ({
      messages,
      steps,
    }: {
      messages: ModelMessage[];
      steps: CompactionStep[];
    }) => {
      let resultMessages: ModelMessage[] | undefined;

      // 1. Compaction
      if (taskCompaction) {
        const compactResult = await chatCompact(messages, steps, {
          shouldCompact: taskCompaction.shouldCompact,
          summarize: (msgs) => {
            const ctx = locals.get(chatTurnContextKey);
            const lastStep = steps.at(-1);
            return taskCompaction.summarize({
              messages: msgs,
              usage: lastStep?.usage,
              source: "inner",
              stepNumber: steps.length - 1,
              chatId: ctx?.chatId,
              turn: ctx?.turn,
              clientData: ctx?.clientData,
            });
          },
        });
        if (compactResult.type !== "skipped") {
          resultMessages = compactResult.messages;
        }
      }

      // 2. Pending message injection (steering)
      if (taskPendingMessages) {
        const injected = await drainSteeringQueue(
          taskPendingMessages,
          resultMessages ?? messages,
          steps
        );
        if (injected.length > 0) {
          resultMessages = [...(resultMessages ?? messages), ...injected];
        }
      }

      // 3. Background context injection
      const bgQueue = locals.get(chatBackgroundQueueKey);
      if (bgQueue && bgQueue.length > 0) {
        const injected = bgQueue.splice(0); // drain
        resultMessages = [...(resultMessages ?? messages), ...injected];
      }

      return resultMessages ? { messages: resultMessages } : undefined;
    };
  }

  return result;
}

/**
 * Options for `pipeChat`.
 */
export type PipeChatOptions = {
  /**
   * Override the stream key. Must match the `streamKey` on `TriggerChatTransport`.
   * @default "chat"
   */
  streamKey?: string;

  /** An AbortSignal to cancel the stream. */
  signal?: AbortSignal;

  /**
   * The target run ID to pipe to.
   * @default "self" (current run)
   */
  target?: string;

  /** Override the default span name for this operation. */
  spanName?: string;
};

/**
 * Options for customizing the `toUIMessageStream()` call used when piping
 * `streamText` results to the frontend.
 *
 * Set static defaults via `uiMessageStreamOptions` on `chat.agent()`, or
 * override per-turn via `chat.setUIMessageStreamOptions()`.
 *
 * `onFinish` is omitted because it is managed internally for response capture.
 * Use `streamText`'s `onFinish` for custom finish handling, or drop down to
 * raw task mode with `chat.pipe()` for full control.
 *
 * `originalMessages` is omitted because it is automatically set from the
 * accumulated conversation history, ensuring message IDs are reused across
 * turns (e.g. for tool approval continuations).
 *
 * `generateMessageId` can be set to control ID generation for response
 * messages (e.g. UUID-v7). If not set, the AI SDK's default `generateId` is used.
 */
export type ChatUIMessageStreamOptions<TUIM extends UIMessage = UIMessage> = Omit<
  UIMessageStreamOptions<TUIM>,
  "onFinish" | "originalMessages"
>;

/**
 * An object with a `toUIMessageStream()` method (e.g. `StreamTextResult` from `streamText()`).
 */
type UIMessageStreamable = {
  toUIMessageStream: (...args: any[]) => AsyncIterable<unknown> | ReadableStream<unknown>;
};

function isUIMessageStreamable(value: unknown): value is UIMessageStreamable {
  return (
    typeof value === "object" &&
    value !== null &&
    "toUIMessageStream" in value &&
    typeof (value as any).toUIMessageStream === "function"
  );
}

let warnedMissingOnAction = false;
function warnMissingOnActionOnce() {
  if (warnedMissingOnAction) return;
  warnedMissingOnAction = true;
  console.warn(
    "[chat.agent] Received an action but no `onAction` handler is configured. " +
      "The action is being ignored. Define `onAction` (and optionally `actionSchema`) on " +
      "your agent to handle it."
  );
}

function isAsyncIterable(value: unknown): value is AsyncIterable<unknown> {
  return typeof value === "object" && value !== null && Symbol.asyncIterator in value;
}

function isReadableStream(value: unknown): value is ReadableStream<unknown> {
  return (
    typeof value === "object" && value !== null && typeof (value as any).getReader === "function"
  );
}

/**
 * Pipes a chat stream to the realtime stream, making it available to the
 * `TriggerChatTransport` on the frontend.
 *
 * Accepts:
 * - A `StreamTextResult` from `streamText()` (has `.toUIMessageStream()`)
 * - An `AsyncIterable` of `UIMessageChunk`s
 * - A `ReadableStream` of `UIMessageChunk`s
 *
 * Must be called from inside a Trigger.dev task's `run` function.
 *
 * @example
 * ```ts
 * import { task } from "@trigger.dev/sdk";
 * import { chat, type ChatTaskPayload } from "@trigger.dev/sdk/ai";
 * import { streamText, convertToModelMessages } from "ai";
 *
 * export const myChatTask = task({
 *   id: "my-chat-task",
 *   run: async (payload: ChatTaskPayload) => {
 *     const result = streamText({
 *       model: openai("gpt-4o"),
 *       messages: payload.messages,
 *     });
 *
 *     await chat.pipe(result);
 *   },
 * });
 * ```
 *
 * @example
 * ```ts
 * // Works from anywhere inside a task — even deep in your agent code
 * async function runAgentLoop(messages: CoreMessage[]) {
 *   const result = streamText({ model, messages });
 *   await chat.pipe(result);
 * }
 * ```
 */
async function pipeChat(
  source: UIMessageStreamable | AsyncIterable<unknown> | ReadableStream<unknown>,
  options?: PipeChatOptions
): Promise<void> {
  locals.set(chatPipeCountKey, (locals.get(chatPipeCountKey) ?? 0) + 1);

  let stream: AsyncIterable<unknown> | ReadableStream<unknown>;

  if (isUIMessageStreamable(source)) {
    stream = source.toUIMessageStream();
  } else if (isAsyncIterable(source) || isReadableStream(source)) {
    stream = source;
  } else {
    throw new Error(
      "pipeChat: source must be a StreamTextResult (with .toUIMessageStream()), " +
      "an AsyncIterable, or a ReadableStream"
    );
  }

  const pipeOptions: SessionPipeStreamOptions = {};
  if (options?.signal) {
    pipeOptions.signal = options.signal;
  }
  if (options?.spanName) {
    pipeOptions.spanName = options.spanName;
  }
  // `options.target` / `options.streamKey` are accepted for API parity
  // with the pre-migration run-scoped pipe but no longer have meaning —
  // sessions are the address (single stream per session, no sub-run
  // targeting). Sub-agents that need to write into a parent's chat now
  // open that session explicitly via `sessions.open(parentSessionId).out.pipe`.

  // The generic is typed for `UIMessageChunk`, but `pipeChat` also
  // accepts opaque UIMessageStreamable / raw iterables whose element
  // type we don't know at compile time. Cast — runtime behaviour is
  // identical (bytes go to session.out either way).
  const { waitUntilComplete } = chatStream.pipe(
    stream as ReadableStream<UIMessageChunk> | AsyncIterable<UIMessageChunk>,
    pipeOptions
  );
  await waitUntilComplete();
}

/**
 * Options for defining a chat task.
 *
 * Extends the standard `TaskOptions` but pre-types the payload as `ChatTaskPayload`
 * and overrides `run` to accept `ChatTaskRunPayload` (with abort signals).
 *
 * **Auto-piping:** If the `run` function returns a value with `.toUIMessageStream()`
 * (like a `StreamTextResult`), the stream is automatically piped to the frontend.
 *
 * **Single-run mode:** By default, the task uses input streams so that the
 * entire conversation lives inside one run. After each AI response, the task
 * emits a control chunk and suspends via `messagesInput.wait()`. The frontend
 * transport resumes the same run by sending the next message via input streams.
 */
/**
 * Event passed to the `onPreload` callback.
 */
export type PreloadEvent<TClientData = unknown> = {
  /** Task run context — same as `task({ run })` second-argument `ctx`. */
  ctx: TaskRunContext;
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The Trigger.dev run ID for this conversation. */
  runId: string;
  /** A scoped access token for this chat run. */
  chatAccessToken: string;
  /** Custom data from the frontend. */
  clientData?: TClientData;
  /** Stream writer — write custom `UIMessageChunk` parts to the chat stream. Lazy: no overhead if unused. */
  writer: ChatWriter;
};

/**
 * Event passed to the `onBoot` callback.
 *
 * Fires once at the start of every run boot — for the initial first-message
 * run, for preloaded runs, AND for reactive continuation runs (post-cancel,
 * post-crash, post-`endRun`, `chat.requestUpgrade`, OOM-retry attempts).
 * Does NOT fire when the SAME run resumes from snapshot via the
 * idle-window suspend/resume path — use `onChatResume` for that.
 *
 * Use this for per-process setup that needs to run every time a fresh
 * worker picks up the chat: initialize `chat.local` state, open
 * per-process resources (DB connections, sandboxes, etc.), or
 * re-hydrate customer state from your DB on continuation.
 *
 * Ordering:
 *   - First message of a chat: `onBoot` → `onChatStart` → `onTurnStart` → `run()`
 *   - Preloaded run:           `onBoot` → `onPreload` → (wait) → `onChatStart` → ...
 *   - Continuation run:        `onBoot` → (wait for first message) → `onTurnStart` → `run()`
 *                              (`onChatStart` does NOT fire on continuation runs)
 */
export type BootEvent<TClientData = unknown> = {
  /** Task run context — same as `task({ run })` second-argument `ctx`. */
  ctx: TaskRunContext;
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The Trigger.dev run ID for this run boot. */
  runId: string;
  /** A scoped access token for this chat run. Persist this for frontend reconnection. */
  chatAccessToken: string;
  /** Custom data from the frontend (passed via `metadata` on `sendMessage()` or the transport). */
  clientData: TClientData;
  /**
   * True when this run is a reactive continuation — the prior run died
   * (cancel / crash / `endRun` / `requestUpgrade` / OOM retry) and this
   * fresh worker is taking over the chat. Branch on this to re-load
   * customer-owned state from your DB.
   */
  continuation: boolean;
  /** Public id of the prior run when `continuation` is true. */
  previousRunId?: string;
  /** Whether this run was triggered as a preload. */
  preloaded: boolean;
};

/**
 * A tool call extracted from the partial assistant message of a dead run.
 * Surfaced on `RecoveryBootEvent.pendingToolCalls` so the customer can
 * decide how to repair the chain (synthesize a result, drop the partial,
 * etc.).
 */
export type RecoveryPendingToolCall = {
  /** The AI SDK tool call id. */
  toolCallId: string;
  /** The tool name (the `tool-${name}` suffix). */
  toolName: string;
  /** The input the model produced for the tool call. */
  input: unknown;
  /** The part index inside `partialAssistant.parts` for in-place edits. */
  partIndex: number;
};

/**
 * Event passed to the `onRecoveryBoot` callback.
 *
 * Fires once at boot when a continuation run inherits in-flight state from
 * a dead predecessor (cancel / crash / OOM / deploy eviction / graceful
 * `chat.requestUpgrade`). The runtime reads both `session.in` and
 * `session.out` past the last `turn-complete` cursor and surfaces the
 * recovered pieces here so the customer can shape the conversational
 * chain before the first turn fires.
 *
 * Does NOT fire when there's nothing to recover (clean continuation after
 * `chat.endRun()` with no buffered user messages, fresh chat, OOM retry
 * after a successful turn-complete with no in-flight tail).
 *
 * Does NOT fire when `hydrateMessages` is registered (the customer owns
 * persistence; recovery decisions live in their own DB query).
 */
export type RecoveryBootEvent<TUIM extends UIMessage = UIMessage> = {
  /** Task run context — same as `task({ run })` second-argument `ctx`. */
  ctx: TaskRunContext;
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The Trigger.dev run ID for this run boot. */
  runId: string;
  /** Public id of the prior run that died. */
  previousRunId: string;
  /**
   * Best-effort cause of the predecessor's death. Currently always
   * `"unknown"` — the run engine doesn't yet plumb the real reason
   * into the continuation payload. Future SDK versions will narrow
   * this. Don't branch behavior on it yet.
   */
  cause: "cancelled" | "crashed" | "unknown";
  /**
   * The conversation chain that was successfully persisted by the
   * predecessor's last `onTurnComplete`. Empty if the predecessor died
   * before turn 1 ever completed.
   */
  settledMessages: TUIM[];
  /**
   * User messages that arrived on `session.in` past the cursor — i.e.
   * the message(s) the predecessor was processing or had queued when
   * it died. The runtime's default is to re-dispatch each as a fresh
   * turn after the chain is restored. Return a different list via
   * `recoveredTurns` to skip / reorder / collapse them.
   */
  inFlightUsers: TUIM[];
  /**
   * The trailing assistant message the predecessor was streaming when
   * it died — the orphan whose `turn-complete` never fired. Undefined
   * if the predecessor died before any assistant output reached
   * `session.out` (cancel-before-first-token, snapshot-only path).
   */
  partialAssistant: TUIM | undefined;
  /**
   * Tool calls extracted from `partialAssistant.parts` that the model
   * had started but the tool runtime never resolved. Empty when
   * `partialAssistant` is undefined or carries no `input-available`
   * tool parts.
   */
  pendingToolCalls: RecoveryPendingToolCall[];
  /**
   * Lazy session.out writer — identical to the `writer` passed to
   * `onTurnStart` / `onTurnComplete` / `onChatStart`. Use this to emit
   * a recovery signal (e.g. a `data-chat-recovery` UIMessage chunk)
   * BEFORE the first recovered turn fires so the bridge can render a
   * "recovering..." banner. Lazy: no overhead if unused.
   */
  writer: ChatWriter;
};

/**
 * Return shape for the `onRecoveryBoot` callback. Every field is optional —
 * omit one to accept the default.
 */
export type RecoveryBootResult<TUIM extends UIMessage = UIMessage> = {
  /**
   * The chain the new run boots with. Replaces the default
   * (`settledMessages`). Use this to keep the partial assistant in
   * context, mutate its tool parts to inject synthesized results,
   * collapse history, etc.
   *
   * Ignored when `hydrateMessages` is registered (the hydrate hook
   * runs per-turn and overwrites the chain).
   */
  chain?: TUIM[];
  /**
   * The user messages to re-dispatch as fresh turns after the chain is
   * restored. Default: `inFlightUsers` (re-process every in-flight
   * user). Return `[]` to suppress all of them; return a filtered /
   * reordered subset to skip specific ones.
   */
  recoveredTurns?: TUIM[];
  /**
   * Awaitable run AFTER the writer flushes and BEFORE the first
   * recovered turn fires. Use for blocking persistence (e.g. write the
   * partial assistant to your DB so a follow-up turn can reference
   * it). Errors bubble — wrap your own try/catch if you want to soft-
   * fail.
   */
  beforeBoot?: () => Promise<void>;
};

/**
 * Event passed to the `onChatStart` callback.
 *
 * Fires exactly once per chat, on the very first user message of the chat's
 * lifetime. Does NOT fire on continuation runs (post-`endRun`,
 * post-waitpoint-timeout, `chat.requestUpgrade`) or on OOM-retry attempts —
 * those are runs of an already-started chat.
 */
export type ChatStartEvent<TClientData = unknown> = {
  /** Task run context — same as `task({ run })` second-argument `ctx`. */
  ctx: TaskRunContext;
  /** The unique identifier for the chat session. */
  chatId: string;
  /**
   * The initial model-ready messages for this conversation. Typically just
   * the first user message (or empty if `chat.headStart` is in play and the
   * seed message is supplied elsewhere). Since this hook only fires for the
   * chat's very first message, there's no prior history to load here.
   */
  messages: ModelMessage[];
  /** Custom data from the frontend (passed via `metadata` on `sendMessage()` or the transport). */
  clientData: TClientData;
  /** The Trigger.dev run ID for this conversation. */
  runId: string;
  /** A scoped access token for this chat run. Persist this for frontend reconnection. */
  chatAccessToken: string;
  /**
   * @deprecated Always `false` — `onChatStart` no longer fires on continuation
   * runs. Kept for backward compatibility; remove your `continuation` checks
   * from `onChatStart` and rely on the contract (this hook fires exactly once
   * per chat, on the very first message).
   */
  continuation: boolean;
  /**
   * @deprecated Always `undefined` — `onChatStart` no longer fires on
   * continuation runs.
   */
  previousRunId?: string;
  /** Whether this run was preloaded before the first message. */
  preloaded: boolean;
  /** Stream writer — write custom `UIMessageChunk` parts to the chat stream. Lazy: no overhead if unused. */
  writer: ChatWriter;
};

/**
 * Event passed to the `hydrateMessages` callback.
 */
export type HydrateMessagesEvent<TClientData = unknown, TUIM extends UIMessage = UIMessage> = {
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The turn number (0-indexed). */
  turn: number;
  /** The trigger type for this turn. */
  trigger: "submit-message" | "regenerate-message" | "action";
  /** Validated incoming UI messages from the wire payload (what the frontend sent). Empty for actions. */
  incomingMessages: TUIM[];
  /** The accumulated UI messages before this turn (empty on turn 0). */
  previousMessages: TUIM[];
  /** Parsed client data from the transport metadata. */
  clientData?: TClientData;
  /** Whether this run is continuing from a previous run. */
  continuation: boolean;
  /** The ID of the previous run (if continuation). */
  previousRunId?: string;
};

/**
 * Event passed to the `onValidateMessages` callback.
 */
export type ValidateMessagesEvent<TUIM extends UIMessage = UIMessage> = {
  /**
   * The incoming UI messages for this turn (after cleanup of aborted tool parts).
   *
   * For HITL continuations the assistant entry is slim — `state` + `output` /
   * `errorText` / `approval` only, no `input` or other parts. Don't pass the
   * full `messages` array to `validateUIMessages` from `ai`; filter to user
   * messages (or your own subset) first.
   */
  messages: TUIM[];
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The turn number (0-indexed). */
  turn: number;
  /** The trigger type for this turn. */
  trigger: "submit-message" | "regenerate-message" | "preload" | "close";
};

/**
 * Event passed to the `onAction` callback.
 */
export type ActionEvent<
  TAction = unknown,
  TClientData = unknown,
  TUIM extends UIMessage = UIMessage,
> = {
  /** The parsed and validated action payload. */
  action: TAction;
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The turn number (0-indexed). */
  turn: number;
  /** Parsed client data from the transport metadata. */
  clientData?: TClientData;
  /** The accumulated UI messages (after hydration, if set). */
  uiMessages: TUIM[];
  /** The accumulated model messages (after hydration, if set). */
  messages: ModelMessage[];
};

/**
 * Event passed to the `onTurnStart` callback.
 */
export type TurnStartEvent<TClientData = unknown, TUIM extends UIMessage = UIMessage> = {
  /** Task run context — same as `task({ run })` second-argument `ctx`. */
  ctx: TaskRunContext;
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The accumulated model-ready messages (all turns so far, including new user message). */
  messages: ModelMessage[];
  /** The accumulated UI messages (all turns so far, including new user message). */
  uiMessages: TUIM[];
  /** The turn number (0-indexed). */
  turn: number;
  /** The Trigger.dev run ID for this conversation. */
  runId: string;
  /** A scoped access token for this chat run. */
  chatAccessToken: string;
  /** Custom data from the frontend. */
  clientData?: TClientData;
  /** Whether this run is continuing an existing chat (previous run timed out or was cancelled). False for brand new chats. */
  continuation: boolean;
  /** The run ID of the previous run (only set when `continuation` is true). */
  previousRunId?: string;
  /** Whether this run was preloaded before the first message. */
  preloaded: boolean;
  /** Token usage from the previous turn. Undefined on turn 0. */
  previousTurnUsage?: LanguageModelUsage;
  /** Cumulative token usage across all completed turns so far. */
  totalUsage: LanguageModelUsage;
  /** Stream writer — write custom `UIMessageChunk` parts to the chat stream. Lazy: no overhead if unused. */
  writer: ChatWriter;
};

/**
 * Event passed to the `onTurnComplete` callback.
 */
export type TurnCompleteEvent<TClientData = unknown, TUIM extends UIMessage = UIMessage> = {
  /** Task run context — same as `task({ run })` second-argument `ctx`. */
  ctx: TaskRunContext;
  /** The unique identifier for the chat session. */
  chatId: string;
  /** The full accumulated conversation in model format (all turns so far). */
  messages: ModelMessage[];
  /**
   * The full accumulated conversation in UI format (all turns so far).
   * This is the format expected by `useChat` — store this for persistence.
   */
  uiMessages: TUIM[];
  /**
   * Only the new model messages from this turn (user message(s) + assistant response).
   * Useful for appending to an existing conversation record.
   */
  newMessages: ModelMessage[];
  /**
   * Only the new UI messages from this turn (user message(s) + assistant response).
   * Useful for inserting individual message records instead of overwriting the full history.
   */
  newUIMessages: TUIM[];
  /** The assistant's response for this turn, with aborted parts cleaned up when `stopped` is true. Undefined if `pipeChat` was used manually. */
  responseMessage: TUIM | undefined;
  /**
   * The raw assistant response before abort cleanup. Includes incomplete tool parts
   * (`input-available`, `partial-call`) and streaming reasoning/text parts.
   * Use this if you need custom cleanup logic. Same as `responseMessage` when not stopped.
   */
  rawResponseMessage: TUIM | undefined;
  /** The turn number (0-indexed). */
  turn: number;
  /** The Trigger.dev run ID for this conversation. */
  runId: string;
  /** A fresh scoped access token for this chat run (renewed each turn). Persist this for frontend reconnection. */
  chatAccessToken: string;
  /** The last event ID from the stream writer. Use this with `resume: true` to avoid replaying events after refresh. */
  lastEventId?: string;
  /** Custom data from the frontend. */
  clientData?: TClientData;
  /** Whether the user stopped generation during this turn. */
  stopped: boolean;
  /** Whether this run is continuing an existing chat (previous run timed out or was cancelled). False for brand new chats. */
  continuation: boolean;
  /** The run ID of the previous run (only set when `continuation` is true). */
  previousRunId?: string;
  /** Whether this run was preloaded before the first message. */
  preloaded: boolean;
  /** Token usage for this turn. Undefined if usage couldn't be captured (e.g. manual pipeChat). */
  usage?: LanguageModelUsage;
  /** Cumulative token usage across all turns in this run (including this turn). */
  totalUsage: LanguageModelUsage;
  /**
   * Why the LLM stopped generating this turn:
   * - `"stop"` — model generated a stop sequence (normal completion)
   * - `"tool-calls"` — model stopped on one or more tool calls. If any tool
   *   has no `execute` function (e.g. an `ask_user` HITL tool), the turn is
   *   paused awaiting user input; inspect `responseMessage.parts` for tool
   *   parts in `input-available` state to distinguish.
   * - `"length"` — max tokens reached
   * - `"content-filter"` — content filter stopped the model
   * - `"error"` — model errored
   * - `"other"` — provider-specific reason
   *
   * Undefined if the underlying stream didn't provide a finish reason (e.g.
   * manual `pipeChat()` or an aborted stream).
   */
  finishReason?: FinishReason;
};

/**
 * Event passed to the `onBeforeTurnComplete` callback.
 * Same as `TurnCompleteEvent` but includes a `writer` since the stream is still open.
 */
export type BeforeTurnCompleteEvent<
  TClientData = unknown,
  TUIM extends UIMessage = UIMessage,
> = TurnCompleteEvent<TClientData, TUIM> & {
  /** Stream writer — write custom `UIMessageChunk` parts to the chat stream. Lazy: no overhead if unused. */
  writer: ChatWriter;
};

/**
 * Discriminated event passed to the `onChatSuspend` callback.
 * Use `phase` to distinguish preload vs turn suspension.
 */
export type ChatSuspendEvent<TClientData = unknown, TUIM extends UIMessage = UIMessage> =
  | {
    /** Suspend is happening after onPreload, before the first message. */
    phase: "preload";
    /** Task run context. */
    ctx: TaskRunContext;
    /** The chat session ID. */
    chatId: string;
    /** The Trigger.dev run ID. */
    runId: string;
    /** Custom data from the frontend. */
    clientData?: TClientData;
  }
  | {
    /**
     * Suspend is happening on a continuation run that booted with no incoming
     * message (post-`endRun`, post-waitpoint-timeout, etc.) and is waiting
     * for the next session.in record before running any turn. Distinct from
     * `phase: "preload"` — the chat already started; `onPreload` has not
     * fired and will not fire on this run.
     */
    phase: "continuation";
    /** Task run context. */
    ctx: TaskRunContext;
    /** The chat session ID. */
    chatId: string;
    /** The Trigger.dev run ID. */
    runId: string;
    /** Custom data from the frontend. */
    clientData?: TClientData;
  }
  | {
    /** Suspend is happening after a completed turn, waiting for the next message. */
    phase: "turn";
    /** Task run context. */
    ctx: TaskRunContext;
    /** The chat session ID. */
    chatId: string;
    /** The Trigger.dev run ID. */
    runId: string;
    /** The turn number (0-indexed) that just completed. */
    turn: number;
    /** The accumulated model messages after the completed turn. */
    messages: ModelMessage[];
    /** The accumulated UI messages after the completed turn. */
    uiMessages: TUIM[];
    /** Custom data from the frontend. */
    clientData?: TClientData;
  };

/**
 * Discriminated event passed to the `onChatResume` callback.
 * Use `phase` to distinguish preload vs turn resumption.
 */
export type ChatResumeEvent<TClientData = unknown, TUIM extends UIMessage = UIMessage> =
  | {
    /** First message arrived after preload suspension. */
    phase: "preload";
    /** Task run context. */
    ctx: TaskRunContext;
    /** The chat session ID. */
    chatId: string;
    /** The Trigger.dev run ID. */
    runId: string;
    /** Custom data from the frontend. */
    clientData?: TClientData;
  }
  | {
    /**
     * First message arrived after continuation-wait suspension. Distinct
     * from `phase: "preload"` — the chat already started; this is a new
     * run picking up after a prior run ended (`endRun`, waitpoint timeout,
     * etc.).
     */
    phase: "continuation";
    /** Task run context. */
    ctx: TaskRunContext;
    /** The chat session ID. */
    chatId: string;
    /** The Trigger.dev run ID. */
    runId: string;
    /** Custom data from the frontend. */
    clientData?: TClientData;
  }
  | {
    /** Next message arrived after turn suspension. */
    phase: "turn";
    /** Task run context. */
    ctx: TaskRunContext;
    /** The chat session ID. */
    chatId: string;
    /** The Trigger.dev run ID. */
    runId: string;
    /** The turn number that was completed before suspension. */
    turn: number;
    /** The accumulated model messages (from before suspension). */
    messages: ModelMessage[];
    /** The accumulated UI messages (from before suspension). */
    uiMessages: TUIM[];
    /** Custom data from the frontend. */
    clientData?: TClientData;
  };

export type ChatAgentOptions<
  TIdentifier extends string,
  TClientDataSchema extends TaskSchema | undefined = undefined,
  TUIMessage extends UIMessage = UIMessage,
  TActionSchema extends TaskSchema | undefined = undefined,
  TTools extends ToolSet = ToolSet,
> = Omit<
  TaskOptions<
    TIdentifier,
    ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>,
    unknown
  >,
  "run" | "retry"
> & {
  /**
   * Fallback machine preset to use when an attempt fails with an
   * out-of-memory (OOM) error. Setting this enables a single OOM retry:
   * the next attempt boots on the larger machine, and the chat picks
   * up via the standard continuation path (same `chatId` / Session,
   * accumulator rebuilds via `hydrateMessages` or post-`onTurnStart`
   * persisted state).
   *
   * Set `machine` (top-level `TaskOptions`) to control the *default*
   * machine the agent runs on. `oomMachine` is the *retry-only* swap.
   *
   * Note: an OOM retry restarts the entire turn from the top — the
   * model call and any in-flight tool executes re-run on the larger
   * machine. Make tool executes idempotent or persist results before
   * returning if you can't tolerate re-execution.
   *
   * Generic `retry` options are not exposed on `chat.agent` because
   * arbitrary retries against an LLM-driven loop tend to be expensive
   * and side-effecting. If you need richer retry semantics, drop down
   * to `chat.task` (the raw primitive).
   *
   * @example
   * ```ts
   * chat.agent({
   *   id: "my-chat",
   *   machine: "small-1x",
   *   oomMachine: "medium-2x",
   *   run: async ({ messages, signal }) =>
   *     streamText({ model, messages, abortSignal: signal }),
   * });
   * ```
   */
  oomMachine?: MachinePresetName;

  /**
   * Schema for validating `clientData` from the frontend.
   * Accepts Zod, ArkType, Valibot, or any supported schema library.
   * When provided, `clientData` is parsed and typed in all hooks and `run`.
   *
   * @example
   * ```ts
   * import { z } from "zod";
   *
   * chat.agent({
   *   id: "my-chat",
   *   clientDataSchema: z.object({ model: z.string().optional(), userId: z.string() }),
   *   run: async ({ messages, clientData, ctx, signal }) => {
   *     // clientData is typed as { model?: string; userId: string }
   *     // ctx is the same TaskRunContext as in task({ run: (payload, { ctx }) => ... })
   *   },
   * });
   * ```
   */
  clientDataSchema?: TClientDataSchema;

  /**
   * Schema for validating custom actions sent via `transport.sendAction()`.
   *
   * When the frontend sends `trigger: "action"`, the `action` payload is
   * parsed against this schema before reaching `onAction`. Invalid actions
   * throw and abort the turn.
   *
   * @example
   * ```ts
   * import { z } from "zod";
   *
   * chat.agent({
   *   id: "my-chat",
   *   actionSchema: z.discriminatedUnion("type", [
   *     z.object({ type: z.literal("undo") }),
   *     z.object({ type: z.literal("rollback"), targetMessageId: z.string() }),
   *   ]),
   *   onAction: async ({ action }) => {
   *     if (action.type === "undo") chat.history.slice(0, -2);
   *     if (action.type === "rollback") chat.history.rollbackTo(action.targetMessageId);
   *   },
   *   run: async ({ messages, signal }) => { ... },
   * });
   * ```
   */
  actionSchema?: TActionSchema;

  /**
   * Called when the frontend sends a custom action via `transport.sendAction()`.
   *
   * Actions are not turns. They fire `hydrateMessages` (if configured) and
   * `onAction` only — no `onTurnStart` / `prepareMessages` /
   * `onBeforeTurnComplete` / `onTurnComplete`, no `run()`. Use
   * `chat.history.*` inside `onAction` to mutate state.
   *
   * To produce a model response from an action, return a
   * `StreamTextResult` (auto-piped), `string`, or `UIMessage`. Returning
   * `void` or nothing is the side-effect-only default.
   */
  onAction?: (
    event: ActionEvent<
      [TActionSchema] extends [TaskSchema] ? inferSchemaOut<TActionSchema> : unknown,
      inferSchemaOut<TClientDataSchema>,
      TUIMessage
    >
  ) => Promise<unknown> | unknown;

  /**
   * The tools available to this agent.
   *
   * `chat.agent` doesn't call the model for you. Your tools still go to
   * `streamText({ tools })` inside `run()`. Declaring them here additionally
   * lets the SDK re-run each tool's
   * [`toModelOutput`](https://ai-sdk.dev/docs/ai-sdk-core/tools-and-tool-calling#tomodeloutput)
   * when it re-converts persisted history on later turns. Without this, the
   * AI SDK has no `tools` to look up `toModelOutput` against, so a tool's
   * transformed result (e.g. raw image bytes → an image content part, or a
   * sub-agent summary) silently degrades to its raw JSON output from turn 2
   * onward.
   *
   * Only `inputSchema` and `toModelOutput` are read during conversion (never
   * `execute`), so you may pass a lightweight map if you keep heavy execute
   * deps out of this module.
   *
   * Pass either a static `ToolSet` or a function of per-turn context (for
   * tools that depend on the user, a feature flag, etc.). The resolved set is
   * available on the `run()` payload as `tools`.
   *
   * @example
   * ```ts
   * const tools = { read_file, search };
   * chat.agent({
   *   tools,
   *   run: async ({ messages, tools, signal }) =>
   *     streamText({ model, messages, tools, abortSignal: signal }),
   * });
   * ```
   */
  tools?:
    | TTools
    | ((event: ResolveToolsEvent<inferSchemaOut<TClientDataSchema>>) => TTools | Promise<TTools>);

  /**
   * The run function for the chat task.
   *
   * Receives a `ChatTaskRunPayload` with the conversation messages, chat session ID,
   * trigger type, task `ctx` (same as `task({ run })`’s second argument), and abort signals
   * (`signal`, `cancelSignal`, `stopSignal`).
   *
   * **Auto-piping:** If this function returns a value with `.toUIMessageStream()`,
   * the stream is automatically piped to the frontend.
   */
  run: (
    payload: ChatTaskRunPayload<inferSchemaOut<TClientDataSchema>, TTools>
  ) => Promise<unknown>;

  /**
   * Called once at the start of every run boot — for the initial run, for
   * preloaded runs, AND for reactive continuation runs (post-cancel /
   * crash / `endRun` / `requestUpgrade` / OOM retry).
   *
   * Use this for per-process setup that needs to run every time a fresh
   * worker picks up the chat: initialize `chat.local` state, open
   * per-process resources, or re-hydrate customer state from your DB
   * on continuation. Branch on `continuation` to decide whether to load
   * existing state vs. start fresh.
   *
   * Fires BEFORE `onPreload` (preloaded path) and `onChatStart`
   * (first-message path), so `chat.local` is safe to read in those hooks.
   *
   * Does NOT fire when the same run resumes from snapshot via the
   * idle-window suspend/resume path — use `onChatResume` for that.
   *
   * @example
   * ```ts
   * onBoot: async ({ chatId, clientData, continuation }) => {
   *   const user = await db.user.findFirst({ where: { id: clientData.userId } });
   *   userContext.init({ name: user.name, plan: user.plan });
   *   if (continuation) {
   *     // re-hydrate any per-chat in-memory state from your DB
   *   }
   * }
   * ```
   */
  onBoot?: (event: BootEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void;

  /**
   * Recovery boot hook — fires once on a continuation run that inherited
   * in-flight state from a dead predecessor (cancel / crash / OOM /
   * deploy eviction / `chat.requestUpgrade()`). The runtime reads both
   * stream tails past the last `turn-complete` cursor and hands the
   * customer the recovered pieces (settled chain, in-flight users,
   * partial assistant, pending tool calls) so the chain can be shaped
   * before the first recovered turn fires.
   *
   * Does NOT fire when there's nothing to recover — e.g. a clean
   * continuation after `chat.endRun()` with no buffered user, a fresh
   * chat, or an OOM retry on top of a complete snapshot.
   *
   * Does NOT fire when `hydrateMessages` is registered — that hook owns
   * the per-turn chain and overlapping recovery decisions belong in the
   * customer's DB.
   *
   * Defaults (returned when the hook is omitted or returns no field):
   *   - `chain` = `settledMessages` (drop the orphan partial)
   *   - `recoveredTurns` = `inFlightUsers` (re-dispatch every user)
   *
   * @example
   * ```ts
   * onRecoveryBoot: async ({ partialAssistant, inFlightUsers, writer, cause }) => {
   *   writer.write({
   *     type: "data-chat-recovery",
   *     id: generateId(),
   *     data: { cause, partial: partialAssistant?.id },
   *   });
   *   return {}; // accept defaults: drop partial, re-dispatch users
   * }
   * ```
   */
  onRecoveryBoot?: (
    event: RecoveryBootEvent<TUIMessage>
  ) => Promise<RecoveryBootResult<TUIMessage> | void> | RecoveryBootResult<TUIMessage> | void;

  /**
   * Called when a preloaded run starts, before the first message arrives.
   *
   * Use this to initialize state, create DB records, and load context early —
   * so everything is ready when the user's first message comes through.
   *
   * @example
   * ```ts
   * onPreload: async ({ ctx, chatId, clientData }) => {
   *   await db.chat.create({ data: { id: chatId } });
   *   userContext.init(await loadUser(clientData.userId));
   * }
   * ```
   */
  onPreload?: (event: PreloadEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void;

  /**
   * Called exactly once per chat, on the very first user message of the
   * chat's lifetime. Does NOT fire on continuation runs (post-`endRun`,
   * post-waitpoint-timeout, `chat.requestUpgrade`) or on OOM-retry attempts —
   * those are runs of an already-started chat.
   *
   * Use this for one-time chat-setup work — creating the Chat DB row,
   * initializing per-chat in-memory state, minting resources tied to the
   * chat's lifetime. Safe to assume no prior history exists here.
   *
   * For per-turn work, use `onTurnStart`.
   *
   * @example
   * ```ts
   * onChatStart: async ({ ctx, chatId, messages, clientData }) => {
   *   await db.chat.create({ data: { id: chatId, userId: clientData.userId } });
   * }
   * ```
   */
  onChatStart?: (event: ChatStartEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void;

  /**
   * Validate or transform incoming UI messages before they are converted to model
   * messages and accumulated. Fires once per turn with the raw `UIMessage[]` from
   * the wire payload (after cleanup of aborted tool parts).
   *
   * Return the validated messages array. Throw to abort the turn with an error.
   *
   * This is the right place to call the AI SDK's `validateUIMessages` on fresh
   * user input. For HITL continuations (`addToolOutput` /
   * `addToolApproveResponse`), the wire carries a slim assistant message — only
   * the resolved tool parts, with `state` + `output` / `errorText` / `approval`
   * and no `input`. `validateUIMessages` against the AI SDK schema rejects
   * that shape, so filter to user messages (or skip validation entirely) on
   * those turns.
   *
   * @example
   * ```ts
   * import { validateUIMessages } from "ai";
   *
   * chat.agent({
   *   id: "my-chat",
   *   onValidateMessages: async ({ messages }) => {
   *     const userMessages = messages.filter((m) => m.role === "user");
   *     if (userMessages.length > 0) {
   *       await validateUIMessages({ messages: userMessages, tools: chatTools });
   *     }
   *     return messages;
   *   },
   *   run: async ({ messages }) => {
   *     return streamText({ model, messages, tools: chatTools });
   *   },
   * });
   * ```
   */
  onValidateMessages?: (
    event: ValidateMessagesEvent<TUIMessage>
  ) => TUIMessage[] | Promise<TUIMessage[]>;

  /**
   * Load the full message history from your backend on every turn,
   * replacing the built-in linear accumulator.
   *
   * When set, the returned messages become the accumulated state for this turn.
   * The normal accumulation logic (append for submit, replace for regenerate)
   * is skipped entirely — the hook is the source of truth.
   *
   * After the hook returns, any incoming wire messages with matching IDs are
   * auto-merged (handles tool approval responses transparently).
   *
   * Use cases:
   * - Backend trust: prevent clients from injecting fabricated history
   * - Branching conversations (DAGs): load only the active branch
   * - Rollback/undo: exclude undone messages from history
   *
   * @example
   * ```ts
   * chat.agent({
   *   id: "my-chat",
   *   hydrateMessages: async ({ chatId, trigger, incomingMessages }) => {
   *     // Persist the new message
   *     const newMsg = incomingMessages[incomingMessages.length - 1];
   *     if (newMsg && trigger === "submit-message") {
   *       await db.chatMessages.create({ chatId, message: newMsg });
   *     }
   *     // Return the full authoritative history
   *     return db.chatMessages.findMany({ where: { chatId } });
   *   },
   *   run: async ({ messages, signal }) => {
   *     return streamText({ model: openai("gpt-4o"), messages, abortSignal: signal });
   *   },
   * });
   * ```
   */
  hydrateMessages?: (
    event: HydrateMessagesEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
  ) => TUIMessage[] | Promise<TUIMessage[]>;

  /**
   * Called at the start of every turn, after message accumulation and `onChatStart` (turn 0),
   * but before the `run` function executes.
   *
   * Use this to persist messages before streaming begins, so a mid-stream page refresh
   * still shows the user's message.
   *
   * @example
   * ```ts
   * onTurnStart: async ({ ctx, chatId, uiMessages }) => {
   *   await db.chat.update({ where: { id: chatId }, data: { messages: uiMessages } });
   * }
   * ```
   */
  onTurnStart?: (
    event: TurnStartEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
  ) => Promise<void> | void;

  /**
   * Called after the response is captured but before the stream closes.
   * The stream is still open, so you can write custom chunks to the frontend
   * (e.g. compaction progress). Use this for compaction, post-processing,
   * or any work where the user should see real-time status updates.
   *
   * @example
   * ```ts
   * onBeforeTurnComplete: async ({ ctx, writer, usage }) => {
   *   if (usage?.inputTokens && usage.inputTokens > 5000) {
   *     writer.write({ type: "data-compaction", id: generateId(), data: { status: "compacting" } });
   *     // ... compact messages ...
   *     chat.setMessages(compactedMessages);
   *     writer.write({ type: "data-compaction", id: generateId(), data: { status: "complete" } });
   *   }
   * }
   * ```
   */
  onBeforeTurnComplete?: (
    event: BeforeTurnCompleteEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
  ) => Promise<void> | void;

  /**
   * Called when conversation compaction occurs (via `chat.compact()` or
   * `chat.compactionStep()`). Use for logging, billing, or persisting the summary.
   *
   * @example
   * ```ts
   * onCompacted: async ({ ctx, summary, totalTokens, chatId }) => {
   *   logger.info("Compacted", { totalTokens, chatId });
   *   await db.compactionLog.create({ data: { chatId, summary } });
   * }
   * ```
   */
  onCompacted?: (event: CompactedEvent) => Promise<void> | void;

  /**
   * Automatic context compaction. When provided, compaction runs automatically
   * in both the inner loop (prepareStep, between tool-call steps) and the
   * outer loop (between turns, for single-step responses where prepareStep
   * never fires).
   *
   * The `shouldCompact` callback decides when to compact, and `summarize`
   * generates the summary. The prepareStep is auto-injected into
   * `chat.toStreamTextOptions()` — if you provide your own `prepareStep`
   * after spreading, it overrides the auto-injected one.
   *
   * @example
   * ```ts
   * chat.agent({
   *   id: "my-chat",
   *   compaction: {
   *     shouldCompact: ({ totalTokens }) => (totalTokens ?? 0) > 80_000,
   *     summarize: async (messages) =>
   *       generateText({ model, messages: [...messages, { role: "user", content: "Summarize." }] })
   *         .then((r) => r.text),
   *   },
   *   run: async ({ messages, signal }) => {
   *     return streamText({ ...chat.toStreamTextOptions({ registry }), messages });
   *   },
   * });
   * ```
   */
  compaction?: ChatAgentCompactionOptions<TUIMessage>;

  /**
   * Configure how messages that arrive during streaming are handled.
   *
   * By default, messages queue for the next turn. When `shouldInject` is provided
   * and returns `true`, messages are injected between tool-call steps via
   * `prepareStep` — allowing users to steer the agent mid-execution.
   *
   * @example
   * ```ts
   * pendingMessages: {
   *   shouldInject: ({ steps }) => steps.length > 0,
   *   onReceived: ({ message }) => logger.info("Steering message received"),
   * },
   * ```
   */
  pendingMessages?: PendingMessagesOptions<TUIMessage>;

  /**
   * Called after each assistant response completes. Use to persist the
   * conversation to your database after each assistant response.
   *
   * @example
   * ```ts
   * onTurnComplete: async ({ ctx, chatId, messages }) => {
   *   await db.chat.update({ where: { id: chatId }, data: { messages } });
   * }
   * ```
   */
  onTurnComplete?: (
    event: TurnCompleteEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
  ) => Promise<void> | void;

  /**
   * Maximum number of conversational turns (message round-trips) a single run
   * will handle before ending. After this many turns the run completes
   * normally and the next message will start a fresh run.
   *
   * @default 100
   */
  maxTurns?: number;

  /**
   * How long to wait for the next message before timing out and ending the run.
   * Accepts any duration string (e.g. `"1h"`, `"30m"`).
   *
   * @default "1h"
   */
  turnTimeout?: string;

  /**
   * How long (in seconds) the run stays idle (active, using compute) after each
   * turn, waiting for the next message. During this window responses are instant.
   * After this timeout the run suspends (frees compute) and waits via
   * `inputStream.wait()`.
   *
   * Set to `0` to suspend immediately after each turn.
   *
   * @default 30
   */
  idleTimeoutInSeconds?: number;

  /**
   * How long the `chatAccessToken` (scoped to this run) remains valid.
   * A fresh token is minted after each turn, so this only needs to cover
   * the gap between turns.
   *
   * Accepts a duration string (e.g. `"1h"`, `"30m"`, `"2h"`).
   *
   * @default "1h"
   */
  chatAccessTokenTTL?: string;

  /**
   * How long (in seconds) the run stays idle after `onPreload` fires,
   * waiting for the first message before suspending.
   *
   * Only applies to preloaded runs (triggered via `transport.preload()`).
   * Takes precedence over `transport.preload(..., { idleTimeoutInSeconds })`
   * and over {@link ChatAgentOptions.idleTimeoutInSeconds}.
   *
   * @default Same as `idleTimeoutInSeconds`
   */
  preloadIdleTimeoutInSeconds?: number;

  /**
   * How long to wait (suspended) for the first message after a preloaded run starts.
   * If no message arrives within this time, the run ends.
   *
   * Only applies to preloaded runs.
   *
   * @default Same as `turnTimeout`
   */
  preloadTimeout?: string;

  /**
   * Transform model messages before they're used anywhere — in `run()`,
   * in compaction rebuilds, and in compaction results.
   *
   * Define once, applied everywhere. Use for Anthropic cache breaks,
   * injecting system context, stripping PII, etc.
   *
   * @example
   * ```ts
   * prepareMessages: async ({ messages, reason }) => {
   *   // Add Anthropic cache breaks to the last message
   *   if (messages.length === 0) return messages;
   *   const last = messages[messages.length - 1];
   *   return [...messages.slice(0, -1), {
   *     ...last,
   *     providerOptions: { ...last.providerOptions, anthropic: { cacheControl: { type: "ephemeral" } } },
   *   }];
   * }
   * ```
   */
  prepareMessages?: (
    event: PrepareMessagesEvent<inferSchemaOut<TClientDataSchema>>
  ) => ModelMessage[] | Promise<ModelMessage[]>;

  /**
   * Default options for `toUIMessageStream()` when auto-piping or using
   * `turn.complete()` / `chat.pipeAndCapture()`.
   *
   * Controls how the `StreamTextResult` is converted to a `UIMessageChunk`
   * stream — error handling, reasoning/source visibility, metadata, etc.
   *
   * Can be overridden per-turn by calling `chat.setUIMessageStreamOptions()`
   * inside `run()` or lifecycle hooks. Per-turn values are merged on top
   * of these defaults (per-turn wins on conflicts).
   *
   * `onFinish` and `originalMessages` are managed internally and cannot be
   * overridden here. Use `streamText`'s `onFinish` for custom finish
   * handling. `generateMessageId` can be set to control response message
   * ID generation (e.g. UUID-v7).
   *
   * @example
   * ```ts
   * chat.agent({
   *   id: "my-chat",
   *   uiMessageStreamOptions: {
   *     sendReasoning: true,
   *     onError: (error) => error instanceof Error ? error.message : "An error occurred.",
   *   },
   *   run: async ({ messages, signal }) => { ... },
   * });
   * ```
   */
  uiMessageStreamOptions?: ChatUIMessageStreamOptions<TUIMessage>;

  /**
   * Called right before the run suspends to wait for a message.
   *
   * The `phase` discriminator tells you when the suspend happened:
   * - `"preload"`: after `onPreload`, waiting for the first message
   * - `"turn"`: after `onTurnComplete`, waiting for the next message
   *
   * Use this for cleanup before suspension (e.g. disposing sandboxes, closing connections).
   *
   * @example
   * ```ts
   * onChatSuspend: async (event) => {
   *   await disposeExpensiveResources(event.ctx.run.id);
   *   if (event.phase === "turn") {
   *     logger.info("Suspending after turn", { turn: event.turn });
   *   }
   * }
   * ```
   */
  onChatSuspend?: (
    event: ChatSuspendEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
  ) => Promise<void> | void;

  /**
   * Called right after the run resumes from suspension with a new message.
   *
   * The `phase` discriminator tells you when the resume happened:
   * - `"preload"`: first message arrived after preload suspension
   * - `"turn"`: next message arrived after turn suspension
   *
   * Use this for re-initialization after wake (e.g. warming caches, reconnecting).
   *
   * @example
   * ```ts
   * onChatResume: async (event) => {
   *   warmCache(event.ctx.run.id);
   *   if (event.phase === "turn") {
   *     logger.info("Resumed after turn", { turn: event.turn });
   *   }
   * }
   * ```
   */
  onChatResume?: (
    event: ChatResumeEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
  ) => Promise<void> | void;

  /**
   * When `true`, the run exits successfully after the preload idle timeout
   * instead of suspending and waiting. The run completes with no turn executed.
   *
   * Use this for "fire and forget" preloads where you only want to do eager
   * initialization. If the user doesn't send a message during the idle window,
   * the run ends cleanly.
   *
   * Only applies to preloaded runs (triggered via `transport.preload()`).
   *
   * @default false
   */
  exitAfterPreloadIdle?: boolean;
};

/**
 * Creates a Trigger.dev task pre-configured for AI SDK chat.
 *
 * - **Pre-types the payload** as `ChatTaskRunPayload` — includes abort signals
 * - **Auto-pipes the stream** if `run` returns a `StreamTextResult`
 * - **Multi-turn**: keeps the conversation in a single run using input streams
 * - **Stop support**: frontend can stop generation mid-stream via the stop input stream
 * - For complex flows, use `pipeChat()` from anywhere inside your task code
 *
 * @example
 * ```ts
 * import { chat } from "@trigger.dev/sdk/ai";
 * import { streamText, convertToModelMessages } from "ai";
 * import { openai } from "@ai-sdk/openai";
 *
 * export const myChat = chat.agent({
 *   id: "my-chat",
 *   run: async ({ messages, signal }) => {
 *     return streamText({
 *       model: openai("gpt-4o"),
 *       messages, // already converted via convertToModelMessages
 *       abortSignal: signal,
 *     });
 *   },
 * });
 * ```
 */
// ─── chat.customAgent ──────────────────────────────────────────────
// A thin wrapper around createTask that marks the task as an agent
// (triggerSource: "agent") so it appears in the playground, but does
// NOT implement the managed lifecycle (no turn loop, no preload, etc.).
// The user's run function receives the raw ChatTaskWirePayload and
// uses composable primitives (chat.messages, chat.MessageAccumulator, etc.).
// ────────────────────────────────────────────────────────────────────

type ChatCustomAgentOptions<
  TIdentifier extends string,
  TClientDataSchema extends TaskSchema | undefined = undefined,
  TUIMessage extends UIMessage = UIMessage,
> = Omit<
  TaskOptions<
    TIdentifier,
    ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>,
    unknown
  >,
  "triggerSource" | "agentConfig"
> & {
  clientDataSchema?: TClientDataSchema;
};

function chatCustomAgent<
  TIdentifier extends string,
  TClientDataSchema extends TaskSchema | undefined = undefined,
  TUIMessage extends UIMessage = UIMessage,
>(
  options: ChatCustomAgentOptions<TIdentifier, TClientDataSchema, TUIMessage>
): Task<TIdentifier, ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>, unknown> {
  const { clientDataSchema, run: userRun, ...restOptions } = options;

  const task = createTask<
    TIdentifier,
    ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>,
    unknown
  >({
    ...restOptions,
    triggerSource: "agent",
    agentConfig: { type: "ai-sdk-chat" },
    run: async (
      payload: ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>,
      runOptions
    ) => {
      // Bind the run to its backing Session so module-level helpers
      // (chat.messages, chat.stream, chat.createStopSignal, chat.createSession)
      // resolve to this chat's `.in` / `.out` channels. Address
      // everywhere by `payload.chatId` (the session externalId) so the
      // agent's writes and the transport's reads converge on the same
      // S2 stream key + waitpoint key.
      //
      // The Session row is created server-side by `POST /sessions` (or
      // `chat.createStartSessionAction`) before this run is triggered.
      // No client-side upsert needed.
      locals.set(chatSessionHandleKey, sessions.open(payload.chatId));
      locals.set(chatAgentRunContextKey, runOptions.ctx);
      markChatAgentRunForStreamsWarning();
      taskContext.setConversationId(payload.chatId);
      stampConversationIdOnActiveSpan(payload.chatId);
      return userRun(payload, runOptions);
    },
  });

  // Register clientDataSchema so the CLI converts it to JSONSchema
  // and stores it as payloadSchema — used by the Playground UI
  if (clientDataSchema) {
    resourceCatalog.updateTaskMetadata(options.id, {
      schema: clientDataSchema as any,
    });
  }

  return task;
}

function chatAgent<
  TIdentifier extends string,
  TClientDataSchema extends TaskSchema | undefined = undefined,
  TUIMessage extends UIMessage = UIMessage,
  TActionSchema extends TaskSchema | undefined = undefined,
  TTools extends ToolSet = ToolSet,
>(
  options: ChatAgentOptions<TIdentifier, TClientDataSchema, TUIMessage, TActionSchema, TTools>
): Task<TIdentifier, ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>, unknown> {
  const {
    run: userRun,
    clientDataSchema,
    onBoot,
    onRecoveryBoot,
    onPreload,
    onChatStart,
    onValidateMessages,
    hydrateMessages,
    actionSchema,
    onAction,
    onTurnStart,
    onBeforeTurnComplete,
    onCompacted,
    compaction,
    pendingMessages: pendingMessagesConfig,
    prepareMessages,
    tools: toolsOption,
    onTurnComplete,
    maxTurns = 100,
    turnTimeout = "1h",
    idleTimeoutInSeconds = 30,
    chatAccessTokenTTL = "1h",
    preloadIdleTimeoutInSeconds,
    preloadTimeout,
    uiMessageStreamOptions,
    onChatSuspend,
    onChatResume,
    exitAfterPreloadIdle = false,
    oomMachine,
    ...restOptions
  } = options;

  const parseClientData = clientDataSchema ? getSchemaParseFn(clientDataSchema) : undefined;
  const parseAction = actionSchema ? getSchemaParseFn(actionSchema) : undefined;

  // chat.agent does not expose generic retry options (see docstring on
  // `oomMachine`). The only opt-in is an OOM-triggered machine swap. If
  // `oomMachine` is set we allow one retry on a larger machine; otherwise
  // we keep the historical no-retry default.
  const retry = oomMachine
    ? { maxAttempts: 2, outOfMemory: { machine: oomMachine } }
    : { maxAttempts: 1 };

  const task = createTask<
    TIdentifier,
    ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>,
    unknown
  >({
    ...restOptions,
    retry,
    triggerSource: "agent",
    agentConfig: { type: "ai-sdk-chat" },
    run: async (
      payload: ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>,
      { signal: runSignal, ctx }
    ) => {
      locals.set(chatAgentRunContextKey, ctx);

      // On AI SDK 7, register the `@ai-sdk/otel` integration (once per process)
      // so `experimental_telemetry` spans flow into the run trace. Awaited here
      // at run boot — before any `streamText` — and a no-op on v5/v6 or when the
      // optional `@ai-sdk/otel` peer isn't installed. See ./aiAutoTelemetry.ts.
      await ensureAiSdkTelemetry();

      // Bind the run to its backing Session so every module-level helper
      // (chat.stream, chat.messages, chat.stopSignal) resolves to this
      // chat's `.in` / `.out` channels.
      //
      // Address everywhere by `payload.chatId` (the session externalId):
      // matches what the transport puts in URL paths and waitpoint keys,
      // and what the server-side trigger flow uses as the session
      // identity. The Session row is created by `POST /sessions` (via
      // `chat.createStartSessionAction` or browser-direct) before this
      // run is triggered — no client-side upsert needed here.
      locals.set(chatSessionHandleKey, sessions.open(payload.chatId));
      // Mutable holder; advances in `writeTurnCompleteChunk` after each turn
      // and is the trim target for the NEXT turn's trim record.
      locals.set(lastTurnCompleteSeqNumKey, { value: undefined });
      markChatAgentRunForStreamsWarning();
      taskContext.setConversationId(payload.chatId);

      // Stamp `gen_ai.conversation.id` on the run-level span. Every
      // nested span inherits the same attribute via
      // `TaskContextSpanProcessor.onStart`.
      const activeSpan = trace.getActiveSpan();
      stampConversationIdOnActiveSpan(payload.chatId, activeSpan);

      // Store static UIMessageStream options in locals so resolveUIMessageStreamOptions() can read them
      if (uiMessageStreamOptions) {
        locals.set(chatUIStreamStaticKey, uiMessageStreamOptions);
      }

      // Store onCompacted hook in locals so chat.compact() can call it
      if (onCompacted) {
        locals.set(chatOnCompactedKey, onCompacted);
      }

      if (prepareMessages) {
        locals.set(chatPrepareMessagesKey, prepareMessages);
      }

      if (toolsOption) {
        // Cast: the option's function form is typed against the parsed
        // `clientData` (`ResolveToolsEvent<inferSchemaOut<...>>`), but the
        // locals key uses the erased `ResolveToolsEvent<unknown>`. The runtime
        // value is identical; this mirrors how `prepareMessages` is stored.
        locals.set(
          chatToolsOptionKey,
          toolsOption as
            | ToolSet
            | ((event: ResolveToolsEvent<unknown>) => ToolSet | Promise<ToolSet>)
        );
        // Static tools are usable immediately. The function form is resolved
        // just before the boot history conversion (with the payload's
        // clientData) and again per-turn (see resolveTurnTools).
        if (typeof toolsOption !== "function") {
          locals.set(chatResolvedToolsKey, toolsOption);
        }
      }

      if (compaction) {
        locals.set(
          chatAgentCompactionKey,
          compaction as unknown as ChatAgentCompactionOptions<UIMessage>
        );
      }

      if (pendingMessagesConfig) {
        locals.set(chatPendingMessagesKey, pendingMessagesConfig);
      }

      let currentWirePayload = payload;
      const continuation = payload.continuation ?? false;
      const previousRunId = payload.previousRunId;
      const preloaded = payload.trigger === "preload";

      // Accumulated model messages across turns. Seeded at boot from a
      // durable snapshot + `session.out` replay (or `hydrateMessages` if
      // registered) — the wire is delta-only now, no longer a seed.
      let accumulatedMessages: ModelMessage[] = [];

      // Accumulated UI messages for persistence. Mirrors the model accumulator
      // but in frontend-friendly UIMessage format (with parts, id, etc.).
      let accumulatedUIMessages: TUIMessage[] = [];

      // ── Snapshot + replay (gated on prior-state signals) ─────────────
      //
      // With `hydrateMessages` registered the customer owns history — the
      // hook fires per-turn and produces the canonical chain from their DB.
      // Skip both reads entirely: no need to load a blob the customer's
      // hook will overwrite.
      //
      // Without it, both reads are gated on `couldHavePriorState`. A fresh
      // chat (no continuation, attempt 1) can't have a snapshot OR replay
      // records by definition — `readChatSnapshot` would 404 and
      // `replaySessionOutTail` would return [], and the round-trips
      // collectively cost ~600ms on every first-message TTFC. Both reads
      // swallow errors internally; the agent stays available either way.
      const sessionIdForSnapshot = payload.sessionId ?? payload.chatId;
      let bootSnapshot: ChatSnapshotV1<TUIMessage> | undefined;
      let replayedSettled: TUIMessage[] = [];
      let replayedPartial: TUIMessage | undefined;
      let replayedPartialRaw: TUIMessage | undefined;
      let replayedInTail: { message: TUIMessage; metadata: unknown; seqNum: number }[] = [];
      // Wire payloads to dispatch as turns before the regular session.in
      // pump kicks in. Populated by `onRecoveryBoot.recoveredTurns` (or its
      // default, `inFlightUsers`). The turn-loop checks this queue ahead of
      // `messagesInput.waitWithIdleTimeout` so recovered turns fire first.
      const bootInjectedQueue: ChatTaskWirePayload<
        TUIMessage,
        inferSchemaIn<TClientDataSchema>
      >[] = [];
      const couldHavePriorState =
        payload.continuation === true || ctx.attempt.number > 1;

      if (!hydrateMessages && couldHavePriorState) {
        // Single parent span for the whole boot read phase — snapshot
        // read, session.out replay, session.in replay. Per-phase timing
        // + result counts are attributes on the span.
        await tracer.startActiveSpan(
          "chat.boot",
          async (bootSpan) => {
            // snapshot read
            const snapStart = Date.now();
            try {
              bootSnapshot = await readChatSnapshot<TUIMessage>(sessionIdForSnapshot);
            } catch (error) {
              // `readChatSnapshot` already swallows + warns internally; this catch
              // is just belt-and-suspenders against tracer/span errors.
              logger.warn(
                "chat.agent: snapshot read failed; continuing without snapshot",
                {
                  error: error instanceof Error ? error.message : String(error),
                  sessionId: sessionIdForSnapshot,
                }
              );
            }
            bootSpan.setAttribute("chat.boot.snapshot.durationMs", Date.now() - snapStart);
            bootSpan.setAttribute("chat.boot.snapshot.present", !!bootSnapshot);
            bootSpan.setAttribute(
              "chat.boot.snapshot.messageCount",
              bootSnapshot?.messages?.length ?? 0
            );

            // Seed the trim chain from the snapshot's `lastOutEventId` (the SSE
            // id of the previous turn's `turn-complete` control record). The
            // first turn-complete this worker writes will then trim back to it.
            // Without seeding, the new worker would emit no trim on its first
            // turn (chain self-bootstraps from turn 2), so this is purely an
            // optimization to keep continuation runs bounded from the first turn.
            if (bootSnapshot?.lastOutEventId !== undefined) {
              const seeded = Number.parseInt(bootSnapshot.lastOutEventId, 10);
              if (Number.isFinite(seeded)) {
                const slot = locals.get(lastTurnCompleteSeqNumKey);
                if (slot) slot.value = seeded;
              }
            }

            // session.out replay
            const replayOutStart = Date.now();
            try {
              const replayResult = await replaySessionOutTail<TUIMessage>(
                sessionIdForSnapshot,
                { lastEventId: bootSnapshot?.lastOutEventId }
              );
              replayedSettled = replayResult.settled;
              replayedPartial = replayResult.partial;
              replayedPartialRaw = replayResult.partialRaw;
            } catch (error) {
              logger.warn(
                "chat.agent: session.out replay failed; using snapshot only",
                {
                  error: error instanceof Error ? error.message : String(error),
                  sessionId: sessionIdForSnapshot,
                }
              );
            }
            bootSpan.setAttribute(
              "chat.boot.replay.out.durationMs",
              Date.now() - replayOutStart
            );
            bootSpan.setAttribute("chat.boot.replay.out.settledCount", replayedSettled.length);
            bootSpan.setAttribute(
              "chat.boot.replay.out.partialPresent",
              replayedPartial !== undefined
            );

            // session.in tail read
            //
            // session.in carries the user-side of the conversation
            // (ChatInputChunk records). On a continuation boot we read past
            // the last turn-complete's `session-in-event-id` header so any
            // user message the dead predecessor hadn't acknowledged surfaces
            // here. Without this read, in-flight user messages would only be
            // visible via the live SSE subscription — by which point they
            // would arrive AFTER the partial-assistant orphan and look like
            // brand-new turns to the model, producing inverted chains.
            const replayInStart = Date.now();
            const lastInEventId = await findLatestSessionInCursor(payload.chatId)
              .then((cursor) => (cursor !== undefined ? String(cursor) : undefined))
              .catch(() => undefined);
            try {
              replayedInTail = await replaySessionInTail<TUIMessage>(payload.chatId, {
                lastEventId: lastInEventId,
              });
            } catch (error) {
              logger.warn(
                "chat.agent: session.in replay failed; in-flight users may not be recovered",
                { error: error instanceof Error ? error.message : String(error) }
              );
            }
            bootSpan.setAttribute(
              "chat.boot.replay.in.durationMs",
              Date.now() - replayInStart
            );
            bootSpan.setAttribute(
              "chat.boot.replay.in.userCount",
              replayedInTail.length
            );
          },
          {
            attributes: {
              [SemanticInternalAttributes.STYLE_ICON]: "tabler-rotate-clockwise",
              [SemanticInternalAttributes.COLLAPSED]: true,
              "chat.id": payload.chatId,
              "chat.continuation": payload.continuation ?? false,
              "chat.attempt": ctx.attempt.number,
            },
          }
        );
      }

      // ── session.in resume cursor ───────────────────────────────────
      //
      // A fresh worker subscribes to `session.in` from seq 0 and would
      // re-deliver every record ever appended — including user messages
      // from turns already completed on a prior run. Without a cursor,
      // the loop would re-process them as fresh turns and the slim-wire
      // merge would replace-by-id against snapshot-restored copies,
      // yielding no-op replaces while the customer's actual new message
      // waits in the queue.
      //
      // The cursor is the seq_num of the last `.in` record the prior
      // worker committed to processing, persisted on each `turn-complete`
      // control record as a `session-in-event-id` sibling header. The
      // boot scan reads the header off `.out`'s latest turn-complete and
      // seeds the manager so the upcoming `.in` SSE subscribe opens with
      // `Last-Event-ID: <cursor>` — S2 starts after that seq and old
      // messages never reach this worker.
      //
      // Applies in three cases (any of which means `.in` has records
      // belonging to completed turns the new run should skip):
      //  - OOM retry (`ctx.attempt.number > 1`)
      //  - Continuation run (`payload.continuation === true`) — prior run
      //    crashed / was canceled / requested upgrade
      //  - Snapshot exists at all (catches edge cases where the wire
      //    didn't set `continuation` but a snapshot indicates prior turns)
      const needsResumeCursor =
        ctx.attempt.number > 1 ||
        payload.continuation === true ||
        bootSnapshot !== undefined;

      if (needsResumeCursor) {
        try {
          const cursor = await findLatestSessionInCursor(payload.chatId);
          if (cursor !== undefined) {
            sessionStreams.setLastSeqNum(payload.chatId, "in", cursor);
            sessionStreams.setLastDispatchedSeqNum(payload.chatId, "in", cursor);
          }
        } catch (error) {
          logger.warn(
            "chat.agent: session.in resume cursor lookup failed; old messages may replay",
            { error: error instanceof Error ? error.message : String(error) }
          );
        }
      }

      // ── Recovery boot + chain reconstruction ────────────────────────
      if (!hydrateMessages) {
        const settledMessages = mergeByIdReplaceWins<TUIMessage>(
          (bootSnapshot?.messages as TUIMessage[]) ?? [],
          replayedSettled
        );
        const inFlightUsers = replayedInTail.map((r) => r.message);
        const partialAssistant = replayedPartial;
        // Fire the hook only when there's a partial assistant — the
        // mid-stream-died signal. In-flight users alone (no partial)
        // cover graceful exits like `chat.requestUpgrade()` where the
        // predecessor chose to end before processing the message;
        // those route through the normal continuation-wait path.
        const hasRecoveredState = partialAssistant !== undefined;

        let hookChain: TUIMessage[] | undefined;
        let hookRecoveredTurns: TUIMessage[] | undefined;
        let hookBeforeBoot: (() => Promise<void>) | undefined;
        if (couldHavePriorState && hasRecoveredState && onRecoveryBoot) {
          // Extract from the RAW partial (pre-cleanup). `cleanupAbortedParts`
          // strips exactly the input-streaming / input-available tool parts
          // we want to surface here, so the cleaned `partialAssistant` would
          // always report zero pending tool calls.
          const pendingToolCalls = extractPendingToolCallsFromPartial(replayedPartialRaw);
          const previousRunIdForHook = previousRunId ?? "";
          let hookResult: RecoveryBootResult<TUIMessage> | void = undefined;
          const { writer: hookWriter, flush: hookFlush } = createLazyChatWriter();
          try {
            hookResult = await tracer.startActiveSpan(
              "onRecoveryBoot()",
              async () =>
                onRecoveryBoot({
                  ctx,
                  chatId: payload.chatId,
                  runId: ctx.run.id,
                  previousRunId: previousRunIdForHook,
                  cause: "unknown",
                  settledMessages,
                  inFlightUsers,
                  partialAssistant,
                  pendingToolCalls,
                  writer: hookWriter,
                }),
              {
                attributes: {
                  [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                  [SemanticInternalAttributes.COLLAPSED]: true,
                  "chat.id": payload.chatId,
                },
              }
            );
          } catch (error) {
            logger.warn("chat.agent: onRecoveryBoot threw; using defaults", {
              error: error instanceof Error ? error.message : String(error),
              chatId: payload.chatId,
            });
          }
          // Flush any chunks the hook wrote so they land on session.out
          // BEFORE the first recovered turn fires.
          try {
            await hookFlush();
          } catch (error) {
            logger.warn("chat.agent: onRecoveryBoot writer flush failed", {
              error: error instanceof Error ? error.message : String(error),
            });
          }
          if (hookResult && typeof hookResult === "object") {
            if (Array.isArray(hookResult.chain)) hookChain = hookResult.chain;
            if (Array.isArray(hookResult.recoveredTurns))
              hookRecoveredTurns = hookResult.recoveredTurns;
            if (typeof hookResult.beforeBoot === "function")
              hookBeforeBoot = hookResult.beforeBoot;
          }
        }

        // Default: splice partial + the user it was answering into
        // the chain so follow-ups like "keep going" still have context.
        let seedChain: TUIMessage[];
        let recoveredTurns: TUIMessage[];
        if (hookChain !== undefined) {
          seedChain = hookChain;
        } else if (partialAssistant !== undefined && inFlightUsers.length > 0) {
          seedChain = [...settledMessages, inFlightUsers[0]!, partialAssistant];
        } else {
          seedChain = settledMessages;
        }
        if (hookRecoveredTurns !== undefined) {
          recoveredTurns = hookRecoveredTurns;
        } else if (partialAssistant !== undefined && inFlightUsers.length > 0) {
          recoveredTurns = inFlightUsers.slice(1);
        } else {
          recoveredTurns = inFlightUsers;
        }
        // `beforeBoot` errors bubble — the customer opted into blocking
        // persistence and a failure there should fail the run rather than
        // dispatch recovered turns against half-persisted state.
        if (hookBeforeBoot) {
          await hookBeforeBoot();
        }

        // Advance the session.in cursor past every recovered user so
        // the live subscription doesn't re-deliver them.
        if (replayedInTail.length > 0) {
          const lastRecoveredSeq = replayedInTail[replayedInTail.length - 1]!.seqNum;
          const currentCursor = sessionStreams.lastSeqNum(payload.chatId, "in");
          if (currentCursor === undefined || lastRecoveredSeq > currentCursor) {
            sessionStreams.setLastSeqNum(payload.chatId, "in", lastRecoveredSeq);
            sessionStreams.setLastDispatchedSeqNum(payload.chatId, "in", lastRecoveredSeq);
          }
        }

        // Synthesize wire payloads for each recoveredTurn. The turn-loop
        // pops these ahead of `messagesInput.waitWithIdleTimeout` so they
        // dispatch as normal turns with the existing hook stack.
        //
        // Per-record metadata preservation: each session.in record
        // carries its own `payload.metadata` (the transport sets it at
        // send time). Look up the original by message id so a recovered
        // turn dispatches with the metadata its writer actually sent.
        // Fall back to the boot payload's metadata for hook-synthesized
        // messages (customer returned a recoveredTurn with no matching
        // session.in record).
        //
        // OOM-retry dedup: if `payload.message` is the same user message
        // the queue is about to redispatch (the wire payload survives
        // across attempts, but session.in records it once), the wire
        // payload already runs turn 0 — drop the duplicate from the queue
        // so we don't fire the same turn twice.
        const wireMessageId =
          (payload.message as { id?: string } | undefined)?.id;
        const metadataById = new Map<string, unknown>();
        for (const entry of replayedInTail) {
          metadataById.set(entry.message.id, entry.metadata);
        }
        for (const msg of recoveredTurns) {
          if (wireMessageId && msg.id === wireMessageId) continue;
          const recoveredMetadata = metadataById.has(msg.id)
            ? metadataById.get(msg.id)
            : payload.metadata;
          bootInjectedQueue.push({
            chatId: payload.chatId,
            sessionId: payload.sessionId,
            metadata: recoveredMetadata,
            trigger: "submit-message",
            message: msg,
            messageId: msg.id,
            continuation: payload.continuation,
            previousRunId: payload.previousRunId,
          } as ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>);
        }

        accumulatedUIMessages = seedChain;

        // ── Head-start bootstrap ─────────────────────────────────────
        //
        // The very-first turn of a head-start handover has no snapshot
        // (it doesn't exist yet) and no `session.out` history (the run
        // just woke up). The customer's HTTP route handler ships full
        // UIMessage history via `headStartMessages` — that's the only
        // path where wire-borne UIMessage[] still seeds the accumulator,
        // and it's safe because the route handler isn't subject to the
        // `/in/append` 512 KiB cap.
        if (
          accumulatedUIMessages.length === 0 &&
          payload.trigger === "handover-prepare" &&
          Array.isArray(payload.headStartMessages) &&
          payload.headStartMessages.length > 0
        ) {
          accumulatedUIMessages = [...(payload.headStartMessages as TUIMessage[])];
        }

        if (accumulatedUIMessages.length > 0) {
          // Resolve a function-form `tools` with the run/continuation payload's
          // clientData so this conversion of the restored history applies each
          // tool's toModelOutput (static tools were already seeded above). This
          // only re-renders saved history, so it fails open: a resolver hiccup
          // logs and converts without tools rather than blocking the resume.
          // Per-turn resolveTurnTools still fails closed for live turns.
          if (typeof toolsOption === "function") {
            try {
              await resolveTurnTools({
                chatId: payload.chatId,
                turn: 0,
                continuation: payload.continuation ?? false,
                clientData: parseClientData
                  ? await parseClientData(payload.metadata)
                  : payload.metadata,
              });
            } catch (error) {
              logger.warn(
                "chat.agent: tools() resolver threw at boot; restored history converted without toModelOutput",
                { error: error instanceof Error ? error.message : String(error) }
              );
            }
          }
          try {
            accumulatedMessages = await toModelMessages(accumulatedUIMessages);
          } catch (error) {
            logger.warn("chat.agent: toModelMessages failed at boot; starting empty", {
              error: error instanceof Error ? error.message : String(error),
              sessionId: sessionIdForSnapshot,
            });
            accumulatedMessages = [];
          }
        }

        // Make the seeded UI accumulator visible to `chat.history.*`
        // before any hook (`onChatStart`, `onTurnStart`, etc.) fires.
        locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);

      }

      // Token usage tracking across turns
      let previousTurnUsage: LanguageModelUsage | undefined;
      let cumulativeUsage: LanguageModelUsage = emptyUsage();

      // Mutable reference to the current turn's stop controller so the
      // stop input stream listener (registered once) can abort the right turn.
      let currentStopController: AbortController | undefined;

      // Stop-input subscription is registered AFTER preload's wait resolves
      // (see the post-preload block below). Registering it earlier would
      // cause it to drain any session.in records buffered before the runtime
      // started — including the customer's first user message arriving on
      // `kind: "message"`. The persistent-listener semantics of session.in
      // pop the buffer when a handler attaches; the stop listener filters
      // out anything that isn't `kind: "stop"` and silently swallows the
      // user message instead of leaving it for `messagesInput.waitWithIdleTimeout`
      // to pick up.
      let stopSub: { off: () => void } | undefined;

      try {
        // ── onBoot — fires once per fresh worker process picking up the chat.
        //
        // Runs BEFORE `onPreload`, `onChatStart`, the continuation-wait
        // branch, and any turn. This is the hook customers use to do
        // per-process setup that has to repeat for every boot (initial
        // run, preloaded run, AND reactive continuation): `chat.local`
        // initialization, opening per-process resources, re-hydrating
        // customer state from their DB on continuation.
        //
        // `onChatStart` is once-per-chat (`!couldHavePriorState` gate);
        // when the prior run dies and a fresh worker boots, `onChatStart`
        // does NOT fire — which is why customers initializing
        // `chat.local` only in `onChatStart` previously crashed in
        // `run()` on continuation. `onBoot` fills that gap.
        if (onBoot) {
          const bootClientData = (
            parseClientData ? await parseClientData(payload.metadata) : payload.metadata
          ) as inferSchemaOut<TClientDataSchema>;

          let bootAccessToken = "";
          const bootRunId = ctx.run.id;
          if (bootRunId) {
            try {
              bootAccessToken = await auth.createPublicToken({
                scopes: {
                  read: {
                    runs: bootRunId,
                    sessions: payload.chatId,
                  },
                  write: {
                    inputStreams: bootRunId,
                    sessions: payload.chatId,
                  },
                },
                expirationTime: chatAccessTokenTTL,
              });
            } catch {
              // Token mint is best-effort — customers can re-mint from
              // their own backend if they need one.
            }
          }

          await tracer.startActiveSpan(
            "onBoot()",
            async () => {
              await onBoot({
                ctx,
                chatId: payload.chatId,
                runId: bootRunId,
                chatAccessToken: bootAccessToken,
                clientData: bootClientData,
                continuation,
                previousRunId,
                preloaded,
              });
            },
            {
              attributes: {
                [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                [SemanticInternalAttributes.COLLAPSED]: true,
                "chat.id": payload.chatId,
                "chat.continuation": continuation,
                "chat.preloaded": preloaded,
              },
            }
          );
        }

        // Handle preloaded runs — fire onPreload, then wait for the first real message
        if (preloaded) {
          if (activeSpan) {
            activeSpan.setAttribute("chat.preloaded", true);
          }

          const currentRunId = ctx.run.id;
          let preloadAccessToken = "";
          if (currentRunId) {
            try {
              preloadAccessToken = await auth.createPublicToken({
                scopes: {
                  read: {
                    runs: currentRunId,
                    sessions: payload.chatId,
                  },
                  write: {
                    inputStreams: currentRunId,
                    sessions: payload.chatId,
                  },
                },
                expirationTime: chatAccessTokenTTL,
              });
            } catch {
              // Token creation failed
            }
          }

          // Parse client data for the preload hook
          const preloadClientData = (
            parseClientData ? await parseClientData(payload.metadata) : payload.metadata
          ) as inferSchemaOut<TClientDataSchema>;

          // Fire onPreload hook
          if (onPreload) {
            await tracer.startActiveSpan(
              "onPreload()",
              async () => {
                await withChatWriter(async (writer) => {
                  await onPreload({
                    ctx,
                    chatId: payload.chatId,
                    runId: currentRunId,
                    chatAccessToken: preloadAccessToken,
                    clientData: preloadClientData,
                    writer,
                  });
                });
              },
              {
                attributes: {
                  [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                  [SemanticInternalAttributes.COLLAPSED]: true,
                  "chat.id": payload.chatId,
                  "chat.preloaded": true,
                },
              }
            );
          }

          // Wait for the first real message — task-level idle settings win over
          // `transport.preload(..., { idleTimeoutInSeconds })` / wire payload so
          // `chat.agent({ idleTimeoutInSeconds, preloadIdleTimeoutInSeconds })` is authoritative.
          const effectivePreloadIdleTimeout =
            preloadIdleTimeoutInSeconds ??
            idleTimeoutInSeconds ??
            payload.idleTimeoutInSeconds;

          const effectivePreloadTimeout =
            (metadata.get(TURN_TIMEOUT_METADATA_KEY) as string | undefined) ??
            preloadTimeout ??
            turnTimeout;

          const preloadResult = await messagesInput.waitWithIdleTimeout({
            idleTimeoutInSeconds: effectivePreloadIdleTimeout,
            timeout: effectivePreloadTimeout,
            spanName: "waiting for first message",
            skipSuspend: exitAfterPreloadIdle,
            onSuspend: onChatSuspend
              ? async () => {
                await tracer.startActiveSpan(
                  "onChatSuspend()",
                  async () => {
                    await onChatSuspend({
                      phase: "preload",
                      ctx,
                      chatId: payload.chatId,
                      runId: currentRunId,
                      clientData: preloadClientData,
                    });
                  },
                  {
                    attributes: {
                      [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onComplete",
                      [SemanticInternalAttributes.COLLAPSED]: true,
                      "chat.id": payload.chatId,
                      "chat.suspend.phase": "preload",
                    },
                  }
                );
              }
              : undefined,
            onResume: onChatResume
              ? async () => {
                await tracer.startActiveSpan(
                  "onChatResume()",
                  async () => {
                    await onChatResume({
                      phase: "preload",
                      ctx,
                      chatId: payload.chatId,
                      runId: currentRunId,
                      clientData: preloadClientData,
                    });
                  },
                  {
                    attributes: {
                      [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                      [SemanticInternalAttributes.COLLAPSED]: true,
                      "chat.id": payload.chatId,
                      "chat.resume.phase": "preload",
                    },
                  }
                );
              }
              : undefined,
          });

          if (!preloadResult.ok) {
            return; // Timed out waiting for first message — end run
          }

          let firstMessage = preloadResult.output;

          currentWirePayload = firstMessage as ChatTaskWirePayload<
            TUIMessage,
            inferSchemaIn<TClientDataSchema>
          >;

          // Close signal during preload — exit before first turn
          if (currentWirePayload.trigger === "close") {
            return;
          }
        }

        // Listen for stop signals for the rest of the run. Registered AFTER
        // the preload wait resolves (or skipped immediately for non-preload
        // triggers) so the persistent-listener semantics on session.in
        // don't drain the buffered user message before
        // `messagesInput.waitWithIdleTimeout` can pick it up. A stop signal
        // that lands during the preload window is dropped — acceptable, the
        // customer can't reasonably stop a chat that hasn't started.
        stopSub = stopInput.on((data) => {
          currentStopController?.abort(data?.message || "stopped");
        });

        // Handle handover-prepare runs — wait on session.in for the
        // customer's `chat.handover` route handler to either hand off
        // mid-turn (tool calls) or signal pure-text completion.
        if (payload.trigger === "handover-prepare") {
          if (activeSpan) {
            activeSpan.setAttribute("chat.handoverPreparing", true);
          }

          const handoverResult = await handoverInput.waitWithIdleTimeout({
            idleTimeoutInSeconds: idleTimeoutInSeconds ?? payload.idleTimeoutInSeconds ?? 60,
            spanName: "waiting for handover signal",
          });

          if (!handoverResult.ok) {
            // Handler crashed before signaling — exit cleanly.
            return;
          }

          if (handoverResult.output.kind === "handover-skip") {
            // Sent only when the customer's handler aborts before
            // producing a finishReason. Normal pure-text and
            // tool-call finishes go through `kind: "handover"` with
            // `isFinal: true | false`. Exit without firing any turn
            // hooks.
            return;
          }

          // kind === "handover": stash the partial assistant message
          // so turn-0 setup can append it after loading user
          // messages. Two branches downstream, switched by `isFinal`:
          //   - `false`: customer's step 1 ended with `tool-calls`.
          //     The agent's `streamText` sees pending tool-calls (via
          //     the approval round in the partial) and executes them,
          //     then runs step 2's LLM call.
          //   - `true`: customer's step 1 ended pure-text. The agent
          //     runs the turn-loop hooks but SKIPS the `streamText`
          //     call entirely (the response is already complete).
          //     `onTurnComplete` fires with the partial as
          //     `responseMessage` so persistence works normally.
          locals.set(
            chatHandoverPartialKey,
            handoverResult.output.partialAssistantMessage
          );
          // Stash the customer-side step-1 messageId. Turn-0 setup
          // uses it to seed the synthesized partial UIMessage with the
          // SAME id, so the agent's post-handover chunks merge into
          // the same assistant message on the browser side.
          if (handoverResult.output.messageId) {
            locals.set(chatHandoverMessageIdKey, handoverResult.output.messageId);
          }
          locals.set(chatHandoverIsFinalKey, handoverResult.output.isFinal);

          // Synthesize a wire payload that the turn loop treats as a
          // normal first-turn message. The accumulator was already seeded
          // at boot from `payload.headStartMessages` (see B.3 head-start
          // bootstrap), so the rewritten `submit-message` carries no
          // delta — the loop runs streamText against the seeded state.
          currentWirePayload = {
            ...payload,
            trigger: "submit-message",
            message: undefined,
            headStartMessages: undefined,
          } as ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>;
        }

        // Continuation-wait: a continuation run (post-`endRun`,
        // post-waitpoint-timeout, etc.) booted with no incoming message.
        // The server strips the Session's basePayload `message` /
        // `messages` / `trigger` on continuation so a stale one-shot boot
        // payload doesn't re-fire on every resume. Without a real first
        // message, we have nothing to process at turn 0 — wait silently
        // for the next session.in record before entering the turn loop.
        // Unlike preload, `onPreload` does NOT fire (the chat already
        // started); `onChatStart` will fire on the first turn with
        // `continuation: true, preloaded: false`.
        if (
          !preloaded &&
          payload.continuation === true &&
          !payload.message &&
          payload.trigger !== "handover-prepare" &&
          payload.trigger !== "close"
        ) {
          if (activeSpan) {
            activeSpan.setAttribute("chat.continuationWaiting", true);
          }

          const continuationClientData = (
            parseClientData ? await parseClientData(payload.metadata) : payload.metadata
          ) as inferSchemaOut<TClientDataSchema>;

          // Recovery-boot injection: if `onRecoveryBoot` (or its default
          // `inFlightUsers`) populated `bootInjectedQueue`, dispatch the
          // first synthesized payload as the very first turn instead of
          // waiting on the live session.in. Subsequent recovered turns
          // get drained by the end-of-turn picker below.
          if (bootInjectedQueue.length > 0) {
            currentWirePayload = bootInjectedQueue.shift()!;
          } else {

          const effectiveIdleTimeout =
            idleTimeoutInSeconds ?? payload.idleTimeoutInSeconds;
          const effectiveTurnTimeout =
            (metadata.get(TURN_TIMEOUT_METADATA_KEY) as string | undefined) ?? turnTimeout;

          const continuationResult = await messagesInput.waitWithIdleTimeout({
            idleTimeoutInSeconds: effectiveIdleTimeout,
            timeout: effectiveTurnTimeout,
            spanName: "waiting for first message (continuation)",
            onSuspend: onChatSuspend
              ? async () => {
                  await tracer.startActiveSpan(
                    "onChatSuspend()",
                    async () => {
                      await onChatSuspend({
                        phase: "continuation",
                        ctx,
                        chatId: payload.chatId,
                        runId: ctx.run.id,
                        clientData: continuationClientData,
                      });
                    },
                    {
                      attributes: {
                        [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onComplete",
                        [SemanticInternalAttributes.COLLAPSED]: true,
                        "chat.id": payload.chatId,
                        "chat.suspend.phase": "continuation",
                      },
                    }
                  );
                }
              : undefined,
            onResume: onChatResume
              ? async () => {
                  await tracer.startActiveSpan(
                    "onChatResume()",
                    async () => {
                      await onChatResume({
                        phase: "continuation",
                        ctx,
                        chatId: payload.chatId,
                        runId: ctx.run.id,
                        clientData: continuationClientData,
                      });
                    },
                    {
                      attributes: {
                        [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                        [SemanticInternalAttributes.COLLAPSED]: true,
                        "chat.id": payload.chatId,
                        "chat.resume.phase": "continuation",
                      },
                    }
                  );
                }
              : undefined,
          });

          if (!continuationResult.ok) {
            // Timed out waiting for the customer's next message — exit.
            return;
          }

          currentWirePayload = continuationResult.output as ChatTaskWirePayload<
            TUIMessage,
            inferSchemaIn<TClientDataSchema>
          >;

          if (currentWirePayload.trigger === "close") {
            return;
          }
          } // end else (no boot-injected first turn)
        }

        for (let turn = 0; turn < maxTurns; turn++) {
          try {
              // Extract turn-level context before entering the span. Slim
              // wire: at most one delta message per record. `headStartMessages`
              // is consumed at boot only (via `payload.headStartMessages`)
              // and intentionally discarded here.
              const {
                metadata: wireMetadata,
                message: incomingMessage,
                headStartMessages: _hsm,
                ...restWire
              } = currentWirePayload;
              void _hsm;
              const incomingMessages: TUIMessage[] = incomingMessage
                ? [incomingMessage as TUIMessage]
                : [];
              // Cleaning happens once here so `extractLastUserMessageText` and
              // every downstream consumer see the same message shape — and
              // `cleanupAbortedParts` no longer has to be re-applied below.
              const cleanedIncomingMessages: TUIMessage[] = incomingMessages.map((msg) =>
                msg.role === "assistant" ? cleanupAbortedParts(msg) : msg
              );
              const clientData = (
                parseClientData ? await parseClientData(wireMetadata) : wireMetadata
              ) as inferSchemaOut<TClientDataSchema>;
              const lastUserMessage = extractLastUserMessageText(cleanedIncomingMessages);

              // Actions are not turns. They use a different span name
              // and don't carry a turn.number. Branched on at `isAction`.
              const isAction = currentWirePayload.trigger === "action";
              const spanName = isAction ? "chat action" : `chat turn ${turn + 1}`;

              const turnAttributes: Attributes = {
                ...(isAction ? {} : { "turn.number": turn + 1 }),
                "gen_ai.conversation.id": currentWirePayload.chatId,
                "gen_ai.operation.name": "chat",
                "chat.trigger": currentWirePayload.trigger,
                [SemanticInternalAttributes.STYLE_ICON]: isAction
                  ? "tabler-bolt"
                  : "tabler-message-chatbot",
                [SemanticInternalAttributes.ENTITY_TYPE]: isAction ? "chat-action" : "chat-turn",
              };

              if (lastUserMessage) {
                turnAttributes["chat.user_message"] = lastUserMessage;

                // Show a truncated preview of the user message as an accessory
                const preview =
                  lastUserMessage.length > 80 ? lastUserMessage.slice(0, 80) + "..." : lastUserMessage;
                Object.assign(
                  turnAttributes,
                  accessoryAttributes({
                    items: [{ text: preview, variant: "normal" }],
                    style: "codepath",
                  })
                );
              }

              if (wireMetadata !== undefined) {
                turnAttributes["chat.client_data"] =
                  typeof wireMetadata === "string" ? wireMetadata : JSON.stringify(wireMetadata);
              }

              const turnResult = await tracer.startActiveSpan(
                spanName,
                async (turnSpan) => {
                  // (errors are caught by the outer try/catch which writes an error chunk)
                  locals.set(chatPipeCountKey, 0);
                  locals.set(chatDeferKey, new Set());
                  locals.set(chatCompactionStateKey, undefined);
                  locals.set(chatSteeringQueueKey, []);
                  locals.set(chatResponsePartsKey, []);
                  // NOTE: chatBackgroundQueueKey is NOT reset here — messages injected
                  // by deferred work from the previous turn's onTurnComplete need to
                  // survive into the next turn. The queue is drained before run().
                  locals.set(chatInjectedMessageIdsKey, new Set());

                  // Store chat context for auto-detection by task-tool subtasks (ai.toolExecute / legacy ai.tool)
                  locals.set(chatTurnContextKey, {
                    chatId: currentWirePayload.chatId,
                    turn,
                    continuation,
                    clientData,
                  });

                  // Resolve the per-turn `tools` set now that turn context
                  // (incl. parsed clientData) exists, so every toModelMessages
                  // call this turn can re-apply tool `toModelOutput`.
                  await resolveTurnTools();

                  // Per-turn stop controller (reset each turn)
                  const stopController = new AbortController();
                  currentStopController = stopController;
                  locals.set(chatStopControllerKey, stopController);

                  // Three signals for the user's run function
                  const stopSignal = stopController.signal;
                  const cancelSignal = runSignal;
                  const combinedSignal = AbortSignal.any([runSignal, stopController.signal]);

                  // Buffer messages that arrive during streaming
                  const pendingMessages: ChatTaskWirePayload<
                    TUIMessage,
                    inferSchemaIn<TClientDataSchema>
                  >[] = [];
                  const pmConfig = locals.get(chatPendingMessagesKey);
                  const msgSub = messagesInput.on(async (msg) => {
                    // If pendingMessages is configured, route to the steering queue
                    // instead of the wire buffer. The frontend handles re-sending
                    // non-injected messages via sendMessage on turn complete.
                    if (pmConfig) {
                      // Slim wire: at most one delta message per record. The
                      // pendingMessages handler reads `msg.message` directly
                      // instead of slicing an array — a wire record arrives
                      // with the new user message in `.message`, or no message
                      // at all (regenerate / preload / close / handover-prepare).
                      const lastUIMessage = msg.message as TUIMessage | undefined;
                      if (lastUIMessage) {
                        if (pmConfig.onReceived) {
                          try {
                            await pmConfig.onReceived({
                              message: lastUIMessage as TUIMessage,
                              chatId: currentWirePayload.chatId,
                              turn,
                            });
                          } catch {
                            /* non-fatal */
                          }
                        }

                        try {
                          const queue = locals.get(chatSteeringQueueKey) ?? [];
                          // Deduplicate by message ID — guards against double-sends
                          if (
                            lastUIMessage.id &&
                            queue.some((e) => e.uiMessage.id === lastUIMessage.id)
                          ) {
                            return;
                          }
                          const modelMsgs = await toModelMessages([lastUIMessage]);
                          queue.push({
                            uiMessage: lastUIMessage as UIMessage,
                            modelMessages: modelMsgs,
                          });
                          locals.set(chatSteeringQueueKey, queue);
                        } catch {
                          /* conversion failed — skip steering queue */
                        }
                      }
                      return; // Don't add to wire buffer — frontend handles non-injected case
                    }

                    // No pendingMessages config — standard wire buffer for next turn
                    pendingMessages.push(
                      msg as ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>
                    );
                  });

                  // Track new messages for this turn (user input + assistant response).
                  const turnNewModelMessages: ModelMessage[] = [];
                  const turnNewUIMessages: TUIMessage[] = [];

                  // ── Action handling ──────────────────────────────────────
                  // Actions arrive on the same input stream but with
                  // trigger === "action". They are NOT turns — only
                  // `hydrateMessages` and `onAction` fire. No turn lifecycle
                  // hooks (`onTurnStart` / `prepareMessages` /
                  // `onBeforeTurnComplete` / `onTurnComplete`) and no
                  // `run()` invocation. To produce a model response from
                  // an action, return a `StreamTextResult` (auto-piped),
                  // string, or UIMessage from `onAction`. Turn counter
                  // does not advance.
                  let actionStreamResult: unknown = undefined;
                  if (isAction) {
                    // Parse and validate the action payload
                    const parsedAction = parseAction
                      ? await parseAction(currentWirePayload.action)
                      : currentWirePayload.action;

                    // Hydrate messages from backend if configured
                    if (hydrateMessages) {
                      const hydrated = await tracer.startActiveSpan(
                        "hydrateMessages()",
                        async () => {
                          return hydrateMessages({
                            chatId: currentWirePayload.chatId,
                            turn,
                            trigger: "action",
                            incomingMessages: [] as TUIMessage[],
                            previousMessages: [...accumulatedUIMessages],
                            clientData,
                            continuation,
                            previousRunId,
                          });
                        },
                        {
                          attributes: {
                            [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                            [SemanticInternalAttributes.COLLAPSED]: true,
                            "chat.id": currentWirePayload.chatId,
                            "chat.trigger": "action",
                          },
                        }
                      );
                      accumulatedUIMessages = [...hydrated] as TUIMessage[];
                      accumulatedMessages = await toModelMessages(hydrated);
                      locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                    }

                    // Fire onAction — handler may mutate state via
                    // `chat.history.*` and / or return a model response.
                    if (onAction) {
                      actionStreamResult = await tracer.startActiveSpan(
                        "onAction()",
                        async () => {
                          return await onAction({
                            action: parsedAction as any,
                            chatId: currentWirePayload.chatId,
                            turn,
                            clientData,
                            uiMessages: accumulatedUIMessages,
                            messages: accumulatedMessages,
                          });
                        },
                        {
                          attributes: {
                            [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                            [SemanticInternalAttributes.COLLAPSED]: true,
                            "chat.id": currentWirePayload.chatId,
                            "chat.action":
                              typeof parsedAction === "object" && parsedAction !== null
                                ? JSON.stringify(parsedAction)
                                : String(parsedAction),
                          },
                        }
                      );

                      // Apply chat.history mutations from onAction
                      const actionOverride = locals.get(chatOverrideMessagesKey);
                      if (actionOverride) {
                        locals.set(chatOverrideMessagesKey, undefined);
                        accumulatedUIMessages = [...actionOverride] as TUIMessage[];
                        accumulatedMessages = await toModelMessages(actionOverride);
                        locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                      }
                    } else {
                      warnMissingOnActionOnce();
                    }
                  }

                  // ── Message handling (non-action turns) ───────────────────
                  //
                  // Slim wire: at most one delta message arrives per record.
                  // The accumulator was already seeded at boot from a durable
                  // snapshot + `session.out` replay (or `hydrateMessages`,
                  // which also fires per-turn below). Per-turn handling is
                  // therefore a delta merge, not a full-history reset.
                  if (currentWirePayload.trigger !== "action") {

                  let cleanedUIMessages: TUIMessage[] = cleanedIncomingMessages;

                  // Validate/transform UIMessages before conversion — catches malformed
                  // messages from storage or untrusted input before they reach the model.
                  // Slim wire: triggers like `regenerate-message` carry no incoming
                  // message; nothing to validate, so skip the hook to avoid feeding
                  // `[]` to validators (AI SDK's `validateUIMessages` rejects empty).
                  if (onValidateMessages && cleanedUIMessages.length > 0) {
                    cleanedUIMessages = (await tracer.startActiveSpan(
                      "onValidateMessages()",
                      async () => {
                        return onValidateMessages({
                          messages: cleanedUIMessages as TUIMessage[],
                          chatId: currentWirePayload.chatId,
                          turn,
                          trigger: currentWirePayload.trigger as ValidateMessagesEvent["trigger"],
                        });
                      },
                      {
                        attributes: {
                          [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                          [SemanticInternalAttributes.COLLAPSED]: true,
                          "chat.id": currentWirePayload.chatId,
                          "chat.turn": turn + 1,
                          "chat.messages.count": cleanedUIMessages.length,
                        },
                      }
                    )) as TUIMessage[];
                  }

                  if (hydrateMessages) {
                    // Snapshot the ids the accumulator knew BEFORE this
                    // turn ran — used below to decide whether an
                    // incoming wire message is genuinely new or just a
                    // state advance on an existing entry. We can't use
                    // the post-`hydrateMessages` array for this because
                    // the canonical hook pattern pushes the incoming
                    // user message into the persisted chain and
                    // returns it.
                    const previouslyKnownMessageIds = new Set(
                      accumulatedUIMessages
                        .map((m) => m.id)
                        .filter((id): id is string => typeof id === "string")
                    );

                    // Backend hydration: load the full message history from the user's
                    // backend, replacing the built-in accumulator entirely. With slim
                    // wire, `incomingMessages` is consistently 0-or-1-length — what
                    // was always true for `submit-message` is now true for every
                    // trigger.
                    const hydrated = await tracer.startActiveSpan(
                      "hydrateMessages()",
                      async () => {
                        return hydrateMessages({
                          chatId: currentWirePayload.chatId,
                          turn,
                          trigger: currentWirePayload.trigger as
                            | "submit-message"
                            | "regenerate-message",
                          incomingMessages: cleanedUIMessages as TUIMessage[],
                          previousMessages: [...accumulatedUIMessages],
                          clientData,
                          continuation,
                          previousRunId,
                        });
                      },
                      {
                        attributes: {
                          [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                          [SemanticInternalAttributes.COLLAPSED]: true,
                          "chat.id": currentWirePayload.chatId,
                          "chat.turn": turn + 1,
                          "chat.trigger": currentWirePayload.trigger,
                          "chat.incoming_messages.count": cleanedUIMessages.length,
                        },
                      }
                    );

                    // Per-turn merge of incoming wire messages onto the hydrated
                    // chain. Hydrated stays authoritative for text, reasoning
                    // blobs, provider metadata, and tool `input`; we only
                    // overlay tool-part state/output/errorText for tool calls
                    // the wire copy has just resolved. Apps that slim the wire
                    // copy to fit the .in/append cap (or drop fields they
                    // re-source from their own DB) get the hydrated copy
                    // through unchanged.
                    const merged = [...hydrated] as TUIMessage[];
                    for (const incoming of cleanedUIMessages) {
                      if (!incoming.id) continue;
                      const idx = merged.findIndex((m) => m.id === incoming.id);
                      if (idx !== -1) {
                        merged[idx] = mergeIncomingIntoHydrated(
                          merged[idx]!,
                          incoming
                        ) as TUIMessage;
                      }
                    }

                    accumulatedUIMessages = merged;
                    accumulatedMessages = await toModelMessages(merged);
                    locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);

                    // Track new messages for onTurnComplete.newUIMessages.
                    // Only push for genuinely new ids — HITL continuations
                    // whose incoming wire id matches an existing entry are
                    // state advances on an old message, not new messages.
                    // We compare against `previouslyKnownMessageIds`
                    // captured BEFORE hydration, not against `hydrated`:
                    // the canonical hydrate pattern pushes the incoming
                    // user message into the persisted chain and returns
                    // it, so the new id IS in `hydrated`, which would
                    // wrongly drop every fresh user turn from
                    // `newUIMessages`. The non-hydrate branch below has
                    // the same "push only on append" semantic via its
                    // own append-vs-replace path.
                    if (
                      currentWirePayload.trigger === "submit-message" &&
                      cleanedUIMessages.length > 0
                    ) {
                      const lastUI = cleanedUIMessages[cleanedUIMessages.length - 1]!;
                      const matchedExisting =
                        lastUI.id !== undefined &&
                        previouslyKnownMessageIds.has(lastUI.id);
                      if (!matchedExisting) {
                        turnNewUIMessages.push(lastUI);
                        const lastModel = (await toModelMessages([lastUI]))[0];
                        if (lastModel) turnNewModelMessages.push(lastModel);
                      }
                    }
                  } else {
                    // Default delta-merge accumulation.
                    //
                    // The accumulator was seeded at boot from snapshot+replay,
                    // so it already reflects prior history. Per-turn handling
                    // appends/replaces the single incoming delta message and
                    // (for regenerate) trims the tail.
                    if (currentWirePayload.trigger === "regenerate-message") {
                      // Regenerate: trim trailing assistant messages from the
                      // accumulator until the tail is a user message. AI SDK's
                      // frontend `regenerate()` already removed the trailing
                      // assistant from its local store; the wire signals "do
                      // the same here", and the agent re-runs from the new
                      // tail. No incoming message accompanies this trigger.
                      while (
                        accumulatedUIMessages.length > 0 &&
                        accumulatedUIMessages[accumulatedUIMessages.length - 1]!.role !== "user"
                      ) {
                        accumulatedUIMessages.pop();
                      }
                      accumulatedMessages = await toModelMessages(accumulatedUIMessages);
                    } else if (cleanedUIMessages.length > 0) {
                      // Submit-message (and the special-cased
                      // handover-prepare → submit-message rewrite earlier in
                      // this scope): merge-or-append for the single delta
                      // message.
                      //
                      // Tool approval responses arrive as a single assistant
                      // message whose id collides with the existing assistant
                      // in the accumulator — we merge the resolved tool-part
                      // resolutions onto the existing entry, keeping text,
                      // reasoning, and tool `input` from the prior snapshot.
                      // The fallback for HITL `addToolOutput` continuations
                      // where AI SDK regenerates the id (TRI-9137) still
                      // applies via `rewriteIncomingIdViaToolCallMap`.
                      let replaced = false;
                      for (const raw of cleanedUIMessages) {
                        let incoming = raw;
                        let idx = accumulatedUIMessages.findIndex(
                          (m) => m.id === incoming.id
                        );
                        if (idx === -1) {
                          const rewritten = rewriteIncomingIdViaToolCallMap(incoming);
                          if (rewritten.id !== incoming.id) {
                            incoming = rewritten as typeof raw;
                            idx = accumulatedUIMessages.findIndex(
                              (m) => m.id === incoming.id
                            );
                          }
                        }
                        if (idx !== -1) {
                          accumulatedUIMessages[idx] = mergeIncomingIntoHydrated(
                            accumulatedUIMessages[idx]!,
                            incoming
                          ) as TUIMessage;
                          replaced = true;
                        } else {
                          accumulatedUIMessages.push(incoming as TUIMessage);
                          turnNewUIMessages.push(incoming as TUIMessage);
                        }
                        recordToolCallIdsFromMessage(incoming);
                      }
                      if (replaced) {
                        // Replacement changes structure — reconvert all model
                        // messages instead of appending.
                        accumulatedMessages = await toModelMessages(accumulatedUIMessages);
                      } else {
                        const incomingModelMessages = await toModelMessages(cleanedUIMessages);
                        accumulatedMessages.push(...incomingModelMessages);
                      }
                      if (turnNewUIMessages.length > 0) {
                        turnNewModelMessages.push(
                          ...(await toModelMessages(turnNewUIMessages))
                        );
                      }
                    }
                    // `preload` / `close` / `handover-prepare` and submits
                    // with no incoming message fall through with the boot-
                    // seeded accumulator unchanged.

                    if (turn === 0) {
                      // Head-start handover splice (turn 0 only): the
                      // `chat.handover` route handler signalled a mid-turn
                      // handover, so splice its partial assistant response
                      // (text + pending tool-calls + the synthesized
                      // tool-approval round) onto the accumulator.
                      // `streamText` then hits AI SDK's initial-tool-
                      // execution branch, runs the agent-side tool executes,
                      // and resumes from step 2 — skipping the first model
                      // call (already done by the handler).
                      //
                      // We also synthesize a UIMessage form of the partial
                      // assistant and push it to `accumulatedUIMessages` so
                      // AI SDK's `processUIMessageStream` (invoked when the
                      // run loop calls `runResult.toUIMessageStream({
                      // onFinish })`) can initialize `state.message` from
                      // the trailing assistant in `originalMessages`. Without
                      // that, the `tool-output-available` chunks emitted by
                      // the initial-tool-execution branch can't find their
                      // matching tool-call in state and AI SDK throws
                      // `UIMessageStreamError: No tool invocation found`.
                      const pendingHandoverPartial = locals.get(chatHandoverPartialKey);
                      if (pendingHandoverPartial && pendingHandoverPartial.length > 0) {
                        accumulatedMessages.push(...pendingHandoverPartial);
                        const handoverMessageId = locals.get(chatHandoverMessageIdKey);
                        const partialUI = synthesizeHandoverUIMessage(
                          pendingHandoverPartial,
                          handoverMessageId
                        );
                        if (partialUI) {
                          accumulatedUIMessages.push(partialUI as TUIMessage);
                        }
                        locals.set(chatHandoverPartialKey, []); // consume once
                      }
                    }

                    locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                  }

                  } // end if (trigger !== "action")

                  // ── Action result handling ──────────────────────────────
                  // For action turns, skip the turn machinery entirely.
                  // If `onAction` returned a stream / string / UIMessage,
                  // pipe it as the response. Either way, emit
                  // `trigger:turn-complete` and then fall through to the
                  // wait-for-next-message logic (shared with message turns).
                  // The turn counter is decremented so the next iteration
                  // sees the same `turn` value — actions don't count.
                  if (isAction) {
                    msgSub.off();

                    if (
                      (locals.get(chatPipeCountKey) ?? 0) === 0 &&
                      isUIMessageStreamable(actionStreamResult)
                    ) {
                      try {
                        const resolvedOptions = resolveUIMessageStreamOptions();
                        const uiStream = (
                          actionStreamResult as UIMessageStreamable
                        ).toUIMessageStream({
                          ...resolvedOptions,
                          generateMessageId:
                            resolvedOptions.generateMessageId ?? generateMessageId,
                        });
                        await pipeChat(uiStream, {
                          signal: combinedSignal,
                          spanName: "stream response",
                        });
                      } catch (error) {
                        if (
                          error instanceof Error &&
                          error.name === "AbortError" &&
                          runSignal.aborted
                        ) {
                          return "exit";
                        }
                        throw error;
                      }
                    }

                    await writeTurnCompleteChunk(currentWirePayload.chatId);

                    // Don't consume a turn iteration — actions aren't turns.
                    turn--;
                  }

                  if (!isAction) {

                  // Mint a scoped public access token once per turn, reused for
                  // onChatStart, onTurnStart, onTurnComplete, and the turn-complete chunk.
                  const currentRunId = ctx.run.id;
                  let turnAccessToken = "";
                  if (currentRunId) {
                    try {
                      turnAccessToken = await auth.createPublicToken({
                        scopes: {
                          read: {
                            runs: currentRunId,
                            sessions: payload.chatId,
                          },
                          write: {
                            inputStreams: currentRunId,
                            sessions: payload.chatId,
                          },
                        },
                        expirationTime: chatAccessTokenTTL,
                      });
                    } catch {
                      // Token creation failed
                    }
                  }

                  // Fire onChatStart on the very first message of a chat
                  // (across the chat's entire lifetime). Gated on
                  // `!couldHavePriorState` so it does NOT re-fire on
                  // continuation runs (post-`endRun`, post-waitpoint-timeout)
                  // or on OOM-retry attempts. Customers put one-time
                  // chat-setup work in `onChatStart` (e.g. create the Chat
                  // DB row, init user context) and that contract relies on
                  // it firing exactly once per chat.
                  if (turn === 0 && onChatStart && !couldHavePriorState) {
                    await tracer.startActiveSpan(
                      "onChatStart()",
                      async () => {
                        await withChatWriter(async (writer) => {
                          await onChatStart({
                            ctx,
                            chatId: currentWirePayload.chatId,
                            messages: accumulatedMessages,
                            clientData,
                            runId: currentRunId,
                            chatAccessToken: turnAccessToken,
                            continuation,
                            previousRunId,
                            preloaded,
                            writer,
                          });
                        });
                      },
                      {
                        attributes: {
                          [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                          [SemanticInternalAttributes.COLLAPSED]: true,
                          "chat.id": currentWirePayload.chatId,
                          "chat.messages.count": accumulatedMessages.length,
                          "chat.continuation": continuation,
                          "chat.preloaded": preloaded,
                          ...(previousRunId ? { "chat.previous_run_id": previousRunId } : {}),
                        },
                      }
                    );
                  }

                  // Fire onTurnStart before running user code — persist messages
                  // so a mid-stream page refresh still shows the user's message.
                  if (onTurnStart) {
                    await tracer.startActiveSpan(
                      "onTurnStart()",
                      async () => {
                        await withChatWriter(async (writer) => {
                          await onTurnStart({
                            ctx,
                            chatId: currentWirePayload.chatId,
                            messages: accumulatedMessages,
                            uiMessages: accumulatedUIMessages,
                            turn,
                            runId: currentRunId,
                            chatAccessToken: turnAccessToken,
                            clientData,
                            continuation,
                            previousRunId,
                            preloaded,
                            previousTurnUsage,
                            totalUsage: cumulativeUsage,
                            writer,
                          });
                        });

                        // Check if onTurnStart replaced messages (compaction or chat.history)
                        const turnStartOverride = locals.get(chatOverrideMessagesKey);
                        if (turnStartOverride) {
                          locals.set(chatOverrideMessagesKey, undefined);
                          accumulatedUIMessages = [...turnStartOverride] as TUIMessage[];
                          accumulatedMessages = await toModelMessages(turnStartOverride);
                          locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                        }
                      },
                      {
                        attributes: {
                          [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                          [SemanticInternalAttributes.COLLAPSED]: true,
                          "chat.id": currentWirePayload.chatId,
                          "chat.turn": turn + 1,
                          "chat.messages.count": accumulatedMessages.length,
                          "chat.trigger": currentWirePayload.trigger,
                          "chat.continuation": continuation,
                          "chat.preloaded": preloaded,
                          ...(previousRunId ? { "chat.previous_run_id": previousRunId } : {}),
                        },
                      }
                    );
                  }

                  // chat.requestUpgrade() called in onTurnStart (or onValidateMessages) —
                  // skip run() and signal the transport to re-trigger the same message
                  // on the new version.
                  if (locals.get(chatUpgradeRequestedKey)) {
                    await writeUpgradeRequiredChunk();
                    return "exit";
                  }

                  // Captured by the onFinish callback below — works even on abort/stop.
                  let capturedResponseMessage: TUIMessage | undefined;
                  let capturedFinishReason: FinishReason | undefined;

                  // Promise that resolves when the AI SDK's onFinish fires.
                  // On abort, the stream's cancel() handler calls onFinish
                  // asynchronously AFTER pipeChat resolves, so we must await
                  // this to avoid a race where we check capturedResponseMessage
                  // before it's been set.
                  let resolveOnFinish: () => void;
                  const onFinishPromise = new Promise<void>((r) => {
                    resolveOnFinish = r;
                  });
                  let onFinishAttached = false;
                  let runResult: unknown;

                  // Pure-text head-start: customer's step 1 IS the
                  // final response. Skip the user's `run` callback
                  // (no LLM call) and use the synthesized partial
                  // UIMessage as `capturedResponseMessage`. The post-
                  // turn flow (`onBeforeTurnComplete` →
                  // `onTurnComplete` → trigger:turn-complete) fires
                  // normally so persistence works.
                  const headStartIsFinal = locals.get(chatHandoverIsFinalKey);
                  const isHeadStartFinalTurn = turn === 0 && headStartIsFinal === true;
                  if (isHeadStartFinalTurn) {
                    locals.set(chatHandoverIsFinalKey, undefined); // consume once
                  }

                  try {
                    // Drain any messages injected by background work (e.g. self-review from previous turn).
                    // Skip if the last message is a tool message — appending after it would
                    // prevent streamText from finding pending tool approvals (it checks
                    // the last message). The queued messages will be picked up by prepareStep
                    // at the next step boundary instead.
                    const lastAccumulated = accumulatedMessages[accumulatedMessages.length - 1];
                    const bgQueue = locals.get(chatBackgroundQueueKey);
                    if (bgQueue && bgQueue.length > 0 && lastAccumulated?.role !== "tool") {
                      accumulatedMessages.push(...bgQueue.splice(0));
                    }

                    if (isHeadStartFinalTurn) {
                      // The synthesized partial UIMessage IS the response.
                      // It was pushed to `accumulatedUIMessages` during the
                      // submit-message branch's splice; recover it as the
                      // last assistant.
                      const lastUI = accumulatedUIMessages[accumulatedUIMessages.length - 1];
                      if (lastUI && lastUI.role === "assistant") {
                        capturedResponseMessage = lastUI;
                        capturedFinishReason = "stop";
                      }
                      // Don't call userRun. Don't pipe. Skip directly
                      // to the post-turn flow below.
                    } else {
                      const preparedMessages = await applyPrepareMessages(accumulatedMessages, "run");
                      runResult = await userRun({
                        ...restWire,
                        messages: preparedMessages,
                        clientData,
                        continuation,
                        previousRunId,
                        preloaded,
                        previousTurnUsage,
                        totalUsage: cumulativeUsage,
                        ctx,
                        tools: locals.get(chatResolvedToolsKey) ?? {},
                        signal: combinedSignal,
                        cancelSignal,
                        stopSignal,
                      } as any);
                    }

                    // Auto-pipe if the run function returned a StreamTextResult or similar,
                    // but only if pipeChat() wasn't already called manually during this turn.
                    // We call toUIMessageStream ourselves to attach onFinish for response capture.
                    // Pass originalMessages so the AI SDK reuses message IDs across turns
                    // (e.g. for tool approval continuations / HITL flows).
                    if ((locals.get(chatPipeCountKey) ?? 0) === 0 && isUIMessageStreamable(runResult)) {
                      onFinishAttached = true;
                      const resolvedOptions = resolveUIMessageStreamOptions();
                      // For action turns, don't pass originalMessages: the response
                      // should always be a fresh assistant message, not a continuation
                      // of whatever trailing assistant was left after chat.history
                      // mutations.
                      const isActionTurn = currentWirePayload.trigger === "action";
                      const uiStream = runResult.toUIMessageStream({
                        ...resolvedOptions,
                        // Pass originalMessages so the AI SDK reuses message IDs across
                        // turns (e.g. for tool approval continuations / HITL flows).
                        // Omit for action turns to force a fresh response ID.
                        ...(isActionTurn
                          ? {}
                          : { originalMessages: accumulatedUIMessages }),
                        // Always provide generateMessageId so the start chunk carries a
                        // messageId. Without this, the frontend and backend generate IDs
                        // independently and they won't match for ID-based dedup.
                        generateMessageId: resolvedOptions.generateMessageId ?? generateMessageId,
                        onFinish: ({
                          responseMessage,
                          finishReason,
                        }: {
                          responseMessage: UIMessage;
                          finishReason?: FinishReason;
                        }) => {
                          capturedResponseMessage = responseMessage as TUIMessage;
                          capturedFinishReason = finishReason;
                          resolveOnFinish!();
                        },
                      });
                      await pipeChat(uiStream, { signal: combinedSignal, spanName: "stream response" });
                    }
                  } catch (error) {
                    // Handle AbortError from streamText gracefully
                    if (error instanceof Error && error.name === "AbortError") {
                      if (runSignal.aborted) {
                        return "exit"; // Full run cancellation — exit
                      }
                      // Stop generation — fall through to continue the loop
                    } else {
                      throw error;
                    }
                  } finally {
                    msgSub.off();
                  }

                  // Wait for onFinish to fire — on abort this may resolve slightly
                  // after pipeChat, since the stream's cancel() handler is async.
                  // Race with a timeout so a stop-abort that prevents onFinish from
                  // firing doesn't hang the turn loop indefinitely.
                  if (onFinishAttached) {
                    await Promise.race([
                      onFinishPromise,
                      new Promise<void>((r) => setTimeout(r, 2_000)),
                    ]);
                  }

                  // Capture token usage from the streamText result (if available).
                  // totalUsage is a PromiseLike that resolves after the stream is consumed.
                  // Race with a 2s timeout — on stop-abort the AI SDK's totalUsage
                  // promise can hang indefinitely (the underlying provider stream
                  // never reports final usage), which would block the turn loop
                  // from ever firing onTurnComplete / writeTurnComplete.
                  let turnUsage: LanguageModelUsage | undefined;
                  if (runResult != null && typeof (runResult as any).totalUsage?.then === "function") {
                    try {
                      turnUsage = (await Promise.race([
                        (runResult as any).totalUsage,
                        new Promise<undefined>((r) => setTimeout(() => r(undefined), 2_000)),
                      ])) as LanguageModelUsage | undefined;
                    } catch {
                      /* non-fatal — usage capture failed */
                    }
                  }
                  if (turnUsage) {
                    cumulativeUsage = addUsage(cumulativeUsage, turnUsage);
                    previousTurnUsage = turnUsage;

                    // Add usage attributes to the turn span
                    if (turnUsage.inputTokens != null) {
                      turnSpan.setAttribute("gen_ai.usage.input_tokens", turnUsage.inputTokens);
                    }
                    if (turnUsage.outputTokens != null) {
                      turnSpan.setAttribute("gen_ai.usage.output_tokens", turnUsage.outputTokens);
                    }
                    if (turnUsage.totalTokens != null) {
                      turnSpan.setAttribute("gen_ai.usage.total_tokens", turnUsage.totalTokens);
                    }
                    if (cumulativeUsage.totalTokens != null) {
                      turnSpan.setAttribute(
                        "gen_ai.usage.cumulative_total_tokens",
                        cumulativeUsage.totalTokens
                      );
                    }
                    if (cumulativeUsage.inputTokens != null) {
                      turnSpan.setAttribute(
                        "gen_ai.usage.cumulative_input_tokens",
                        cumulativeUsage.inputTokens
                      );
                    }
                    if (cumulativeUsage.outputTokens != null) {
                      turnSpan.setAttribute(
                        "gen_ai.usage.cumulative_output_tokens",
                        cumulativeUsage.outputTokens
                      );
                    }
                  }

                  // Check if run() (e.g. via prepareStep or chat.history) replaced messages
                  // during this turn. The updated messages become the new base, and the
                  // response gets appended on top.
                  const runOverride = locals.get(chatOverrideMessagesKey);
                  if (runOverride) {
                    locals.set(chatOverrideMessagesKey, undefined);
                    accumulatedUIMessages = [...runOverride] as TUIMessage[];
                    accumulatedMessages = await toModelMessages(runOverride);
                    locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                  }

                  // Check if compaction set a model-only override (preserves UI messages).
                  // Apply compactUIMessages/compactModelMessages callbacks if configured.
                  const modelOnlyOverride = locals.get(chatOverrideModelMessagesKey);
                  if (modelOnlyOverride) {
                    const compactionSummary = locals.get(chatCompactionStateKey)?.summary ?? "";
                    const taskCompactionConfig = locals.get(chatAgentCompactionKey);
                    locals.set(chatOverrideModelMessagesKey, undefined);

                    const compactEvent: CompactMessagesEvent<TUIMessage> = {
                      summary: compactionSummary,
                      uiMessages: accumulatedUIMessages,
                      modelMessages: accumulatedMessages,
                      chatId: currentWirePayload.chatId,
                      turn,
                      clientData,
                      source: "inner",
                    };

                    // Apply model messages: callback or default (use override)
                    accumulatedMessages = taskCompactionConfig?.compactModelMessages
                      ? await taskCompactionConfig.compactModelMessages(compactEvent)
                      : modelOnlyOverride;

                    // Apply UI messages: callback or default (preserve all)
                    if (taskCompactionConfig?.compactUIMessages) {
                      accumulatedUIMessages = (await taskCompactionConfig.compactUIMessages(
                        compactEvent
                      )) as TUIMessage[];
                    }
                  }

                  // Determine if the user stopped generation this turn (not a full run cancel).
                  const wasStopped = stopController.signal.aborted && !runSignal.aborted;

                  // Append the assistant's response (partial or complete) to the accumulator.
                  // The onFinish callback fires even on abort/stop, so partial responses
                  // from stopped generation are captured correctly.
                  let rawResponseMessage: TUIMessage | undefined;
                  if (capturedResponseMessage) {
                    // Keep the raw message before cleanup for users who want custom handling
                    rawResponseMessage = capturedResponseMessage;
                    // Clean up aborted parts (streaming tool calls, reasoning) when stopped
                    if (wasStopped) {
                      capturedResponseMessage = cleanupAbortedParts(capturedResponseMessage);
                    }
                    // Ensure the response message has an ID (the stream's onFinish
                    // may produce a message with an empty ID since IDs are normally
                    // assigned by the frontend's useChat).
                    if (!capturedResponseMessage.id) {
                      capturedResponseMessage = { ...capturedResponseMessage, id: generateMessageId() };
                    }
                    // Append any non-transient data parts queued via chat.response or writer.write()
                    const queuedParts = locals.get(chatResponsePartsKey);
                    if (queuedParts && queuedParts.length > 0) {
                      capturedResponseMessage = {
                        ...capturedResponseMessage,
                        parts: [...capturedResponseMessage.parts, ...queuedParts],
                      } as TUIMessage;
                      locals.set(chatResponsePartsKey, []);
                    }
                    // Tool-approval continuations: the AI SDK reuses the trailing
                    // assistant's ID (via originalMessages) so the captured response
                    // carries the same ID as an existing message. Replace in place
                    // instead of pushing a duplicate. For action turns this never
                    // matches because originalMessages is omitted (fresh ID).
                    const existingIdx = capturedResponseMessage.id
                      ? accumulatedUIMessages.findIndex(
                          (m) => m.id === capturedResponseMessage!.id
                        )
                      : -1;
                    if (existingIdx !== -1) {
                      accumulatedUIMessages[existingIdx] = capturedResponseMessage;
                    } else {
                      accumulatedUIMessages.push(capturedResponseMessage);
                    }
                    turnNewUIMessages.push(capturedResponseMessage);
                    locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                    // Record toolCallId → head messageId so a HITL
                    // continuation next turn can recover the head id
                    // even if the AI SDK regenerates it. See
                    // `chatToolCallToMessageIdKey` for the full
                    // rationale (TRI-9137).
                    recordToolCallIdsFromMessage(capturedResponseMessage);
                    try {
                      const responseModelMessages = await toModelMessages([
                        stripProviderMetadata(capturedResponseMessage),
                      ]);
                      if (existingIdx !== -1) {
                        // Reconvert all model messages since we replaced rather than appended
                        accumulatedMessages = await toModelMessages(accumulatedUIMessages);
                      } else {
                        accumulatedMessages.push(...responseModelMessages);
                      }
                      turnNewModelMessages.push(...responseModelMessages);
                    } catch {
                      // Conversion failed — skip accumulation for this turn
                    }
                  }
                  // If there's no captured response (manual pipe mode) but there are
                  // queued data parts, create a minimal response message to hold them.
                  if (!capturedResponseMessage) {
                    const remainingParts = locals.get(chatResponsePartsKey);
                    if (remainingParts && remainingParts.length > 0) {
                      capturedResponseMessage = {
                        id: generateMessageId(),
                        role: "assistant" as const,
                        parts: [...remainingParts],
                      } as TUIMessage;
                      locals.set(chatResponsePartsKey, []);
                      accumulatedUIMessages.push(capturedResponseMessage);
                      turnNewUIMessages.push(capturedResponseMessage);
                      locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                    }
                  }

                  if (runSignal.aborted) return "exit";

                  // Await deferred background work (e.g. DB writes from onTurnStart)
                  // before firing hooks so they can rely on the work being done.
                  const deferredWork = locals.get(chatDeferKey);
                  if (deferredWork && deferredWork.size > 0) {
                    await Promise.race([
                      Promise.allSettled(deferredWork),
                      new Promise<void>((r) => setTimeout(r, 5_000)),
                    ]);
                  }

                  // Outer-loop compaction: runs between turns for single-step responses
                  // where prepareStep never fires (no tool calls = no step boundaries).
                  // Only triggers when: task has compaction configured, prepareStep didn't
                  // already compact this turn, and shouldCompact returns true.
                  const outerCompaction = locals.get(chatAgentCompactionKey);
                  const innerCompactionState = locals.get(chatCompactionStateKey);

                  if (outerCompaction && !innerCompactionState && turnUsage && !wasStopped) {
                    const shouldTrigger = await outerCompaction.shouldCompact({
                      messages: accumulatedMessages,
                      totalTokens: turnUsage.totalTokens,
                      inputTokens: turnUsage.inputTokens,
                      outputTokens: turnUsage.outputTokens,
                      usage: turnUsage,
                      totalUsage: cumulativeUsage,
                      chatId: currentWirePayload.chatId,
                      turn,
                      clientData,
                      source: "outer",
                    });

                    if (shouldTrigger) {
                      await tracer.startActiveSpan(
                        "context compaction (outer loop)",
                        async (compactionSpan) => {
                          const compactionId = generateMessageId();

                          const { waitUntilComplete } = chatStream.writer({
                            spanName: "stream compaction chunks",
                            collapsed: true,
                            execute: async ({ write, merge }) => {
                              write({
                                type: "data-compaction",
                                id: compactionId,
                                data: { status: "compacting", totalTokens: turnUsage.totalTokens },
                                transient: true,
                              });

                              const summary = await outerCompaction.summarize({
                                messages: accumulatedMessages,
                                usage: turnUsage,
                                totalUsage: cumulativeUsage,
                                chatId: currentWirePayload.chatId,
                                turn,
                                clientData,
                                source: "outer",
                              });

                              // Apply compactModelMessages/compactUIMessages callbacks, or defaults.

                              const outerCompactEvent: CompactMessagesEvent<TUIMessage> = {
                                summary,
                                uiMessages: accumulatedUIMessages,
                                modelMessages: accumulatedMessages,
                                chatId: currentWirePayload.chatId,
                                turn,
                                clientData,
                                source: "outer",
                              };

                              // Model messages: callback or default (replace with summary)
                              accumulatedMessages = outerCompaction.compactModelMessages
                                ? await outerCompaction.compactModelMessages(outerCompactEvent)
                                : [
                                  {
                                    role: "assistant" as const,
                                    content: [
                                      {
                                        type: "text" as const,
                                        text: `[Conversation summary]\n\n${summary}`,
                                      },
                                    ],
                                  },
                                ];

                              // UI messages: callback or default (preserve all)
                              if (outerCompaction.compactUIMessages) {
                                accumulatedUIMessages = (await outerCompaction.compactUIMessages(
                                  outerCompactEvent
                                )) as TUIMessage[];
                              }

                              // Fire onCompacted hook
                              const onCompactedHook = locals.get(chatOnCompactedKey);
                              if (onCompactedHook) {
                                await onCompactedHook({
                                  ctx,
                                  summary,
                                  messages: accumulatedMessages,
                                  messageCount: accumulatedMessages.length,
                                  usage: turnUsage,
                                  totalTokens: turnUsage.totalTokens,
                                  inputTokens: turnUsage.inputTokens,
                                  outputTokens: turnUsage.outputTokens,
                                  stepNumber: -1, // outer loop, not a step
                                  chatId: currentWirePayload.chatId,
                                  turn,
                                  writer: { write, merge },
                                });
                              }

                              compactionSpan.setAttribute("compaction.summary_length", summary.length);

                              write({
                                type: "data-compaction",
                                id: compactionId,
                                data: { status: "complete", totalTokens: turnUsage.totalTokens },
                                transient: true,
                              });
                            },
                          });
                          await waitUntilComplete();
                        },
                        {
                          attributes: {
                            [SemanticInternalAttributes.STYLE_ICON]: "tabler-scissors",
                            "compaction.total_tokens": turnUsage.totalTokens ?? 0,
                            "compaction.input_tokens": turnUsage.inputTokens ?? 0,
                            "compaction.message_count": accumulatedMessages.length,
                            "compaction.outer_loop": true,
                            "compaction.turn": turn,
                            ...(currentWirePayload.chatId
                              ? { "compaction.chat_id": currentWirePayload.chatId }
                              : {}),
                            ...accessoryAttributes({
                              items: [
                                { text: `${turnUsage.totalTokens ?? 0} tokens`, variant: "normal" },
                                { text: `${accumulatedMessages.length} msgs`, variant: "normal" },
                                { text: "outer loop", variant: "normal" },
                              ],
                              style: "codepath",
                            }),
                          },
                        }
                      );
                    }
                  }

                  const turnCompleteEvent = {
                    ctx,
                    chatId: currentWirePayload.chatId,
                    messages: accumulatedMessages,
                    uiMessages: accumulatedUIMessages,
                    newMessages: turnNewModelMessages,
                    newUIMessages: turnNewUIMessages,
                    responseMessage: capturedResponseMessage,
                    rawResponseMessage,
                    turn,
                    runId: currentRunId,
                    chatAccessToken: turnAccessToken,
                    clientData,
                    stopped: wasStopped,
                    continuation,
                    previousRunId,
                    preloaded,
                    usage: turnUsage,
                    totalUsage: cumulativeUsage,
                    finishReason: capturedFinishReason,
                  };

                  // Fire onBeforeTurnComplete — stream is still open so the hook
                  // can write custom chunks to the frontend (e.g. compaction progress).
                  if (onBeforeTurnComplete) {
                    await tracer.startActiveSpan(
                      "onBeforeTurnComplete()",
                      async () => {
                        await withChatWriter(async (writer) => {
                          await onBeforeTurnComplete({ ...turnCompleteEvent, writer });
                        });

                        // Check if the hook replaced messages (compaction or chat.history)
                        const override = locals.get(chatOverrideMessagesKey);
                        if (override) {
                          locals.set(chatOverrideMessagesKey, undefined);
                          accumulatedUIMessages = [...override] as TUIMessage[];
                          accumulatedMessages = await toModelMessages(override);
                          locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                          // Update event so onTurnComplete sees compacted messages
                          turnCompleteEvent.messages = accumulatedMessages;
                          turnCompleteEvent.uiMessages = accumulatedUIMessages;
                        }
                      },
                      {
                        attributes: {
                          [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onComplete",
                          [SemanticInternalAttributes.COLLAPSED]: true,
                          "chat.id": currentWirePayload.chatId,
                          "chat.turn": turn + 1,
                        },
                      }
                    );
                  }

                  // Drain any late response parts added during onBeforeTurnComplete
                  const lateParts = locals.get(chatResponsePartsKey);
                  if (lateParts && lateParts.length > 0 && capturedResponseMessage) {
                    const idx = accumulatedUIMessages.findIndex((m) => m.id === capturedResponseMessage!.id);
                    if (idx !== -1) {
                      const msg = accumulatedUIMessages[idx]!;
                      accumulatedUIMessages[idx] = {
                        ...msg,
                        parts: [...(msg.parts ?? []), ...lateParts],
                      } as TUIMessage;
                      capturedResponseMessage = accumulatedUIMessages[idx] as TUIMessage;
                      turnCompleteEvent.responseMessage = capturedResponseMessage;
                      turnCompleteEvent.uiMessages = accumulatedUIMessages;
                    }
                    locals.set(chatResponsePartsKey, []);
                  }

                  // Write turn-complete control chunk — closes the frontend stream.
                  const turnCompleteResult = await writeTurnCompleteChunk(
                    currentWirePayload.chatId,
                    turnAccessToken
                  );

                  // Fire onTurnComplete — stream is closed, use for persistence.
                  if (onTurnComplete) {
                    await tracer.startActiveSpan(
                      "onTurnComplete()",
                      async () => {
                        await onTurnComplete({
                          ...turnCompleteEvent,
                          lastEventId: turnCompleteResult?.lastEventId,
                        });

                        // Check if onTurnComplete replaced messages (compaction or chat.history)
                        const turnCompleteOverride = locals.get(chatOverrideMessagesKey);
                        if (turnCompleteOverride) {
                          locals.set(chatOverrideMessagesKey, undefined);
                          accumulatedUIMessages = [...turnCompleteOverride] as TUIMessage[];
                          accumulatedMessages = await toModelMessages(turnCompleteOverride);
                          locals.set(chatCurrentUIMessagesKey, accumulatedUIMessages);
                        }
                      },
                      {
                        attributes: {
                          [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onComplete",
                          [SemanticInternalAttributes.COLLAPSED]: true,
                          "chat.id": currentWirePayload.chatId,
                          "chat.turn": turn + 1,
                          "chat.stopped": wasStopped,
                          "chat.continuation": continuation,
                          "chat.preloaded": preloaded,
                          ...(previousRunId ? { "chat.previous_run_id": previousRunId } : {}),
                          "chat.messages.count": accumulatedMessages.length,
                          "chat.response.parts.count": capturedResponseMessage?.parts?.length ?? 0,
                          "chat.new_messages.count": turnNewUIMessages.length,
                          ...(turnUsage?.inputTokens != null
                            ? { "gen_ai.usage.input_tokens": turnUsage.inputTokens }
                            : {}),
                          ...(turnUsage?.outputTokens != null
                            ? { "gen_ai.usage.output_tokens": turnUsage.outputTokens }
                            : {}),
                          ...(turnUsage?.totalTokens != null
                            ? { "gen_ai.usage.total_tokens": turnUsage.totalTokens }
                            : {}),
                          ...(cumulativeUsage.totalTokens != null
                            ? { "gen_ai.usage.cumulative_total_tokens": cumulativeUsage.totalTokens }
                            : {}),
                        },
                      }
                    );
                  }

                  // ── Snapshot write ─────────────────────────────────────
                  //
                  // Persist the post-turn accumulator so the next run boot
                  // can replay history without the wire shipping it. Skipped
                  // when `hydrateMessages` is registered — those customers
                  // own persistence and would never read this blob.
                  //
                  // AWAITED, not fire-and-forget: the agent may suspend (idle
                  // timeout) immediately after this point, and in-flight
                  // promises don't reliably complete on suspend. The user-
                  // visible turn already finished (the turn-complete chunk
                  // closed the response stream above), so the await delay
                  // only affects "when can the NEXT turn start," gated by
                  // user typing — not TTFC.
                  //
                  // `writeChatSnapshot` swallows errors internally; this
                  // outer try/catch is just belt-and-suspenders against
                  // tracer/span failures.
                  if (!hydrateMessages) {
                    try {
                      await tracer.startActiveSpan(
                        "snapshot.write",
                        async () => {
                          await writeChatSnapshot<TUIMessage>(sessionIdForSnapshot, {
                            version: 1,
                            savedAt: Date.now(),
                            messages: accumulatedUIMessages,
                            lastOutEventId: turnCompleteResult?.lastEventId,
                          });
                        },
                        {
                          attributes: {
                            [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onComplete",
                            [SemanticInternalAttributes.COLLAPSED]: true,
                            "chat.id": currentWirePayload.chatId,
                            "chat.turn": turn + 1,
                            "chat.messages.count": accumulatedUIMessages.length,
                          },
                        }
                      );
                    } catch (error) {
                      logger.warn(
                        "chat.agent: snapshot write failed; next run will replay further",
                        {
                          error: error instanceof Error ? error.message : String(error),
                          sessionId: sessionIdForSnapshot,
                        }
                      );
                    }
                  }

                  } // end if (!isAction)

                  // NOTE: We intentionally do NOT await deferred work from onTurnComplete here.
                  // Promises deferred in onTurnComplete (e.g. background self-review via
                  // chat.defer + chat.inject) run during the idle wait. If they complete
                  // before the next message, their injected context is picked up in prepareStep.
                  // The pre-onBeforeTurnComplete drain handles promises from onTurnStart/run().

                  // Recovery-boot injection: drain remaining recovered turns
                  // before any other source. `onRecoveryBoot` (or its default)
                  // produced these from in-flight user messages on session.in
                  // that the dead predecessor never acknowledged.
                  if (bootInjectedQueue.length > 0) {
                    currentWirePayload = bootInjectedQueue.shift()!;
                    return "continue";
                  }

                  // If messages arrived during streaming (without pendingMessages config),
                  // use the first one immediately as the next turn.
                  if (pendingMessages.length > 0) {
                    currentWirePayload = pendingMessages[0]!;
                    return "continue";
                  }

                  // chat.requestUpgrade() was called — exit the loop so the
                  // transport triggers a new run on the latest version.
                  // chat.endRun() — same exit, no upgrade semantics.
                  if (
                    locals.get(chatUpgradeRequestedKey) ||
                    locals.get(chatEndRunRequestedKey)
                  ) {
                    return "exit";
                  }

                  // Wait for the next message — stay idle briefly, then suspend
                  const effectiveIdleTimeout =
                    (metadata.get(IDLE_TIMEOUT_METADATA_KEY) as number | undefined) ??
                    idleTimeoutInSeconds;
                  const effectiveTurnTimeout =
                    (metadata.get(TURN_TIMEOUT_METADATA_KEY) as string | undefined) ?? turnTimeout;

                  const next = await messagesInput.waitWithIdleTimeout({
                    idleTimeoutInSeconds: effectiveIdleTimeout,
                    timeout: effectiveTurnTimeout,
                    spanName: "waiting for next message",
                    onSuspend: onChatSuspend
                      ? async () => {
                        await tracer.startActiveSpan(
                          "onChatSuspend()",
                          async () => {
                            await onChatSuspend({
                              phase: "turn",
                              ctx,
                              chatId: currentWirePayload.chatId,
                              runId: ctx.run.id,
                              turn,
                              messages: accumulatedMessages,
                              uiMessages: accumulatedUIMessages,
                              clientData,
                            });
                          },
                          {
                            attributes: {
                              [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onComplete",
                              [SemanticInternalAttributes.COLLAPSED]: true,
                              "chat.id": currentWirePayload.chatId,
                              "chat.suspend.phase": "turn",
                              "chat.turn": turn + 1,
                            },
                          }
                        );
                      }
                      : undefined,
                    onResume: onChatResume
                      ? async () => {
                        await tracer.startActiveSpan(
                          "onChatResume()",
                          async () => {
                            await onChatResume({
                              phase: "turn",
                              ctx,
                              chatId: currentWirePayload.chatId,
                              runId: ctx.run.id,
                              turn,
                              messages: accumulatedMessages,
                              uiMessages: accumulatedUIMessages,
                              clientData,
                            });
                          },
                          {
                            attributes: {
                              [SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
                              [SemanticInternalAttributes.COLLAPSED]: true,
                              "chat.id": currentWirePayload.chatId,
                              "chat.resume.phase": "turn",
                              "chat.turn": turn + 1,
                            },
                          }
                        );
                      }
                      : undefined,
                  });

                  if (!next.ok) {
                    return "exit";
                  }

                  currentWirePayload = next.output as ChatTaskWirePayload<
                    TUIMessage,
                    inferSchemaIn<TClientDataSchema>
                  >;

                  // Close signal — exit the loop gracefully
                  if (currentWirePayload.trigger === "close") {
                    return "exit";
                  }

                  return "continue";
                },
                {
                  attributes: turnAttributes,
                }
              );

              if (turnResult === "exit") return;
              // "continue" means proceed to next iteration
            } catch (turnError) {
              // Turn error handler: write an error chunk + turn-complete to the stream
              // so the client sees the error, then wait for the next message instead
              // of killing the entire run. This keeps the conversation alive.
              if (turnError instanceof Error && turnError.name === "AbortError" && runSignal.aborted) {
                // Full run cancellation — exit immediately
                throw turnError;
              }

              // OOM errors must escape the turn loop so the task runtime can
              // honor `retry.outOfMemory.machine` (set on chat.agent via
              // `oomMachine`). Catching them here would keep the dead worker
              // alive and defeat the machine swap. Re-throw and let the
              // runtime dispatch the retry on a larger machine; recovery on
              // attempt 2 picks up via the standard continuation path
              // (same chatId / Session, accumulator rehydrates).
              if (turnError instanceof OutOfMemoryError) {
                throw turnError;
              }

              try {
                await withChatWriter(async (writer) => {
                  const errorText =
                    turnError instanceof Error ? turnError.message : "An unexpected error occurred";
                  writer.write({ type: "error", errorText } as any);
                });
                // Signal turn complete so the client knows this turn is done
                await writeTurnCompleteChunk(currentWirePayload.chatId);
              } catch {
                // Best-effort — if stream write fails, let the run continue anyway
              }

              // chat.requestUpgrade() / chat.endRun() — exit after error turn too
              if (
                locals.get(chatUpgradeRequestedKey) ||
                locals.get(chatEndRunRequestedKey)
              ) {
                return;
              }

              // Drain remaining recovered turns before idling — a thrown
              // recovered turn shouldn't strand the rest of the boot queue
              // until an unrelated live message arrives.
              if (bootInjectedQueue.length > 0) {
                currentWirePayload = bootInjectedQueue.shift()!;
                continue;
              }

              // Wait for the next message — same as after a successful turn
              const effectiveIdleTimeout =
                (metadata.get(IDLE_TIMEOUT_METADATA_KEY) as number | undefined) ??
                idleTimeoutInSeconds;
              const effectiveTurnTimeout =
                (metadata.get(TURN_TIMEOUT_METADATA_KEY) as string | undefined) ?? turnTimeout;

              const next = await messagesInput.waitWithIdleTimeout({
                idleTimeoutInSeconds: effectiveIdleTimeout,
                timeout: effectiveTurnTimeout,
                spanName: "waiting for next message (after error)",
              });

              if (!next.ok) {
                return; // Timed out — end run gracefully
              }

              currentWirePayload = next.output as ChatTaskWirePayload<
                TUIMessage,
                inferSchemaIn<TClientDataSchema>
              >;
              // Continue to next iteration of the for loop
            }
          }
        } finally {
          // `stopSub` is registered post-preload so the close-during-preload
          // early-return path may exit before it ever attached. Guard the
          // cleanup so a missing subscription doesn't throw.
          stopSub?.off();
        }
    }
  });

  // Register clientDataSchema so the CLI converts it to JSONSchema
  // and stores it as payloadSchema — used by the Playground UI
  if (clientDataSchema) {
    resourceCatalog.updateTaskMetadata(options.id, {
      schema: clientDataSchema as any,
    });
  }

  return task;
}

/**
 * Optional config for {@link chat.withUIMessage}. `streamOptions` become default
 * static `toUIMessageStream()` settings; inner `chat.agent({ uiMessageStreamOptions })`
 * shallow-merges on top (task wins on conflicts).
 */
export type ChatWithUIMessageConfig<TUIM extends UIMessage = UIMessage> = {
  streamOptions?: ChatUIMessageStreamOptions<TUIM>;
};

// ---------------------------------------------------------------------------
// Chat builder
// ---------------------------------------------------------------------------

/**
 * A chainable builder for configuring chat tasks with fixed UI message types,
 * client data schemas, and builder-level hooks that compose with task-level hooks.
 *
 * Obtain a builder via {@link chat.withUIMessage} or {@link chat.withClientData}.
 *
 * @example
 * ```ts
 * export const myChat = chat
 *   .withUIMessage<AgentUiMessage>({ streamOptions: { sendReasoning: true } })
 *   .withClientData({ schema: z.object({ userId: z.string() }) })
 *   .onChatSuspend(async ({ ctx }) => { await disposeResources(ctx.run.id) })
 *   .task({
 *     id: "my-chat",
 *     run: async ({ messages, signal }) => streamText({ model, messages, abortSignal: signal }),
 *   });
 * ```
 */
export interface ChatBuilder<
  TUIMessage extends UIMessage = UIMessage,
  TClientDataSchema extends TaskSchema | undefined = undefined,
> {
  /** Fix the UI message type. Returns a new builder preserving all accumulated state. */
  withUIMessage<TUIM extends UIMessage = UIMessage>(
    config?: ChatWithUIMessageConfig<TUIM>
  ): ChatBuilder<TUIM, TClientDataSchema>;

  /** Fix the client data schema. Returns a new builder preserving all accumulated state. */
  withClientData<TSchema extends TaskSchema>(config: {
    schema: TSchema;
  }): ChatBuilder<TUIMessage, TSchema>;

  /** Register a builder-level `onBoot` hook. Runs before the task-level hook if both are set. */
  onBoot(
    fn: (event: BootEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onPreload` hook. Runs before the task-level hook if both are set. */
  onPreload(
    fn: (event: PreloadEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onChatStart` hook. Runs before the task-level hook if both are set. */
  onChatStart(
    fn: (event: ChatStartEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onTurnStart` hook. Runs before the task-level hook if both are set. */
  onTurnStart(
    fn: (
      event: TurnStartEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
    ) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onBeforeTurnComplete` hook. Runs before the task-level hook if both are set. */
  onBeforeTurnComplete(
    fn: (
      event: BeforeTurnCompleteEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
    ) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onTurnComplete` hook. Runs before the task-level hook if both are set. */
  onTurnComplete(
    fn: (
      event: TurnCompleteEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
    ) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onCompacted` hook. Runs before the task-level hook if both are set. */
  onCompacted(fn: (event: CompactedEvent) => Promise<void> | void): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onChatSuspend` hook. Runs before the task-level hook if both are set. */
  onChatSuspend(
    fn: (
      event: ChatSuspendEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
    ) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /** Register a builder-level `onChatResume` hook. Runs before the task-level hook if both are set. */
  onChatResume(
    fn: (
      event: ChatResumeEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
    ) => Promise<void> | void
  ): ChatBuilder<TUIMessage, TClientDataSchema>;

  /**
   * Create the chat agent with the accumulated builder configuration.
   *
   * When `withClientData` was called, `clientDataSchema` is injected automatically
   * and omitted from options. Otherwise, it can still be set directly in options
   * (backwards compatible).
   */
  agent: [TClientDataSchema] extends [undefined]
  ? <TId extends string, TInfer extends TaskSchema | undefined = undefined, TAction extends TaskSchema | undefined = undefined, TTools extends ToolSet = ToolSet>(
    options: ChatAgentOptions<TId, TInfer, TUIMessage, TAction, TTools>
  ) => Task<TId, ChatTaskWirePayload<TUIMessage, inferSchemaIn<TInfer>>, unknown>
  : <TId extends string, TAction extends TaskSchema | undefined = undefined, TTools extends ToolSet = ToolSet>(
    options: Omit<ChatAgentOptions<TId, TClientDataSchema, TUIMessage, TAction, TTools>, "clientDataSchema">
  ) => Task<TId, ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>, unknown>;

  /**
   * Create a custom agent with manual lifecycle control.
   *
   * The agent appears in the playground but you manage the turn loop,
   * message waiting, and streaming yourself using composable primitives
   * (`chat.messages`, `chat.MessageAccumulator`, `chat.pipeAndCapture`, etc.).
   *
   * Builder hooks (`onPreload`, `onChatStart`, etc.) are not applied —
   * those are managed-lifecycle concepts handled by `.agent()`.
   */
  customAgent: [TClientDataSchema] extends [undefined]
    ? <TId extends string>(
        options: ChatCustomAgentOptions<TId, undefined, TUIMessage>
      ) => Task<TId, ChatTaskWirePayload<TUIMessage, undefined>, unknown>
    : <TId extends string>(
        options: ChatCustomAgentOptions<TId, TClientDataSchema, TUIMessage>
      ) => Task<TId, ChatTaskWirePayload<TUIMessage, inferSchemaIn<TClientDataSchema>>, unknown>;
}

/** @internal */
type ChatBuilderHooks = {
  onBoot?: (event: any) => Promise<void> | void;
  onPreload?: (event: any) => Promise<void> | void;
  onChatStart?: (event: any) => Promise<void> | void;
  onTurnStart?: (event: any) => Promise<void> | void;
  onBeforeTurnComplete?: (event: any) => Promise<void> | void;
  onTurnComplete?: (event: any) => Promise<void> | void;
  onCompacted?: (event: any) => Promise<void> | void;
  onChatSuspend?: (event: any) => Promise<void> | void;
  onChatResume?: (event: any) => Promise<void> | void;
};

/** @internal */
type ChatBuilderConfig = {
  uiStreamOptions?: ChatUIMessageStreamOptions<any>;
  clientDataSchema?: TaskSchema;
  hooks: ChatBuilderHooks;
};

function composeHooks<T>(
  builderHook: ((event: T) => Promise<void> | void) | undefined,
  taskHook: ((event: T) => Promise<void> | void) | undefined
): ((event: T) => Promise<void>) | undefined {
  if (!builderHook) return taskHook as any;
  if (!taskHook) return builderHook as any;
  return async (event: T) => {
    await builderHook(event);
    await taskHook(event);
  };
}

function createChatBuilder<
  TUIMessage extends UIMessage = UIMessage,
  TClientDataSchema extends TaskSchema | undefined = undefined,
>(config: ChatBuilderConfig): ChatBuilder<TUIMessage, TClientDataSchema> {
  return {
    withUIMessage<TUIM extends UIMessage = UIMessage>(uimConfig?: ChatWithUIMessageConfig<TUIM>) {
      return createChatBuilder<TUIM, TClientDataSchema>({
        ...config,
        uiStreamOptions: uimConfig?.streamOptions ?? config.uiStreamOptions,
      });
    },

    withClientData<TSchema extends TaskSchema>(cdConfig: { schema: TSchema }) {
      return createChatBuilder<TUIMessage, TSchema>({
        ...config,
        clientDataSchema: cdConfig.schema,
      });
    },

    onBoot(
      fn: (event: BootEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onBoot: fn },
      });
    },
    onPreload(
      fn: (event: PreloadEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onPreload: fn },
      });
    },
    onChatStart(
      fn: (event: ChatStartEvent<inferSchemaOut<TClientDataSchema>>) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onChatStart: fn },
      });
    },
    onTurnStart(
      fn: (
        event: TurnStartEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
      ) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onTurnStart: fn },
      });
    },
    onBeforeTurnComplete(
      fn: (
        event: BeforeTurnCompleteEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
      ) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onBeforeTurnComplete: fn },
      });
    },
    onTurnComplete(
      fn: (
        event: TurnCompleteEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
      ) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onTurnComplete: fn },
      });
    },
    onCompacted(fn: (event: CompactedEvent) => Promise<void> | void) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onCompacted: fn },
      });
    },
    onChatSuspend(
      fn: (
        event: ChatSuspendEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
      ) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onChatSuspend: fn },
      });
    },
    onChatResume(
      fn: (
        event: ChatResumeEvent<inferSchemaOut<TClientDataSchema>, TUIMessage>
      ) => Promise<void> | void
    ) {
      return createChatBuilder<TUIMessage, TClientDataSchema>({
        ...config,
        hooks: { ...config.hooks, onChatResume: fn },
      });
    },

    agent(options: any) {
      const mergedUiStream =
        config.uiStreamOptions && options.uiMessageStreamOptions
          ? { ...config.uiStreamOptions, ...options.uiMessageStreamOptions }
          : options.uiMessageStreamOptions ?? config.uiStreamOptions;

      return chatAgent({
        ...options,
        ...(config.clientDataSchema ? { clientDataSchema: config.clientDataSchema } : {}),
        uiMessageStreamOptions: mergedUiStream,
        onBoot: composeHooks(config.hooks.onBoot, options.onBoot),
        onRecoveryBoot: options.onRecoveryBoot,
        onPreload: composeHooks(config.hooks.onPreload, options.onPreload),
        onChatStart: composeHooks(config.hooks.onChatStart, options.onChatStart),
        onTurnStart: composeHooks(config.hooks.onTurnStart, options.onTurnStart),
        onBeforeTurnComplete: composeHooks(
          config.hooks.onBeforeTurnComplete,
          options.onBeforeTurnComplete
        ),
        onTurnComplete: composeHooks(config.hooks.onTurnComplete, options.onTurnComplete),
        onCompacted: composeHooks(config.hooks.onCompacted, options.onCompacted),
        onChatSuspend: composeHooks(config.hooks.onChatSuspend, options.onChatSuspend),
        onChatResume: composeHooks(config.hooks.onChatResume, options.onChatResume),
      });
    },

    customAgent(options: any) {
      return chatCustomAgent({
        ...options,
        ...(config.clientDataSchema ? { clientDataSchema: config.clientDataSchema } : {}),
      });
    },
  } as unknown as ChatBuilder<TUIMessage, TClientDataSchema>;
}

/**
 * Fix the UI message type for a chat task (AI SDK `UIMessage` generics) while
 * keeping `id` and `clientDataSchema` inference on the inner {@link chat.agent} call.
 *
 * Returns a {@link ChatBuilder} that supports chaining `.withClientData()`,
 * hook methods (`.onPreload()`, `.onChatSuspend()`, etc.), and `.task()`.
 *
 * @example
 * ```ts
 * type AgentUiMessage = UIMessage<unknown, UIDataTypes, UITools>;
 *
 * export const myChat = chat.withUIMessage<AgentUiMessage>({
 *   streamOptions: { sendReasoning: true },
 * }).task({
 *   id: "my-chat",
 *   run: async ({ messages, signal }) => { ... },
 * });
 * ```
 */
function withUIMessage<TUIM extends UIMessage = UIMessage>(
  config?: ChatWithUIMessageConfig<TUIM>
): ChatBuilder<TUIM, undefined> {
  return createChatBuilder<TUIM, undefined>({
    uiStreamOptions: config?.streamOptions,
    hooks: {},
  });
}

/**
 * Fix the client data schema for a chat task, providing typed `clientData`
 * in all hooks and the `run` function.
 *
 * Returns a {@link ChatBuilder} that supports chaining `.withUIMessage()`,
 * hook methods (`.onPreload()`, `.onChatSuspend()`, etc.), and `.task()`.
 *
 * @example
 * ```ts
 * export const myChat = chat
 *   .withClientData({ schema: z.object({ userId: z.string() }) })
 *   .task({
 *     id: "my-chat",
 *     onPreload: async ({ clientData }) => {
 *       // clientData is typed as { userId: string }
 *     },
 *     run: async ({ messages, signal }) => { ... },
 *   });
 * ```
 */
function withClientData<TSchema extends TaskSchema>(config: {
  schema: TSchema;
}): ChatBuilder<UIMessage, TSchema> {
  return createChatBuilder<UIMessage, TSchema>({
    clientDataSchema: config.schema,
    hooks: {},
  });
}

/**
 * Namespace for AI SDK chat integration.
 *
 * @example
 * ```ts
 * import { chat } from "@trigger.dev/sdk/ai";
 *
 * // Define a chat task
 * export const myChat = chat.agent({
 *   id: "my-chat",
 *   run: async ({ messages, signal }) => {
 *     return streamText({ model, messages, abortSignal: signal });
 *   },
 * });
 *
 * // Pipe a stream manually (from inside a task)
 * await chat.pipe(streamTextResult);
 *
 * // Create an access token (from a server action)
 * const token = await chat.createAccessToken("my-chat");
 * ```
 */
// ---------------------------------------------------------------------------
// Runtime configuration helpers
// ---------------------------------------------------------------------------

const TURN_TIMEOUT_METADATA_KEY = "chat.turnTimeout";
const IDLE_TIMEOUT_METADATA_KEY = "chat.idleTimeout";

/**
 * Override the turn timeout for subsequent turns in the current run.
 *
 * The turn timeout controls how long the run stays suspended (freeing compute)
 * waiting for the next user message. When it expires, the run completes
 * gracefully and the next message starts a fresh run.
 *
 * Call from inside a `chatAgent` run function to adjust based on context.
 *
 * @param duration - A duration string (e.g. `"5m"`, `"1h"`, `"30s"`)
 *
 * @example
 * ```ts
 * run: async ({ messages, signal }) => {
 *   chat.setTurnTimeout("2h");
 *   return streamText({ model, messages, abortSignal: signal });
 * }
 * ```
 */
function setTurnTimeout(duration: string): void {
  metadata.set(TURN_TIMEOUT_METADATA_KEY, duration);
}

/**
 * Override the turn timeout in seconds for subsequent turns in the current run.
 *
 * @param seconds - Number of seconds to wait for the next message before ending the run
 *
 * @example
 * ```ts
 * run: async ({ messages, signal }) => {
 *   chat.setTurnTimeoutInSeconds(3600); // 1 hour
 *   return streamText({ model, messages, abortSignal: signal });
 * }
 * ```
 */
function setTurnTimeoutInSeconds(seconds: number): void {
  metadata.set(TURN_TIMEOUT_METADATA_KEY, `${seconds}s`);
}

/**
 * Override the idle timeout for subsequent turns in the current run.
 *
 * The idle timeout controls how long the run stays active (using compute)
 * after each turn, waiting for the next message. During this window,
 * responses are instant. After it expires, the run suspends.
 *
 * @param seconds - Number of seconds to stay idle (0 to suspend immediately)
 *
 * @example
 * ```ts
 * run: async ({ messages, signal }) => {
 *   chat.setIdleTimeoutInSeconds(60);
 *   return streamText({ model, messages, abortSignal: signal });
 * }
 * ```
 */
function setIdleTimeoutInSeconds(seconds: number): void {
  metadata.set(IDLE_TIMEOUT_METADATA_KEY, seconds);
}

/**
 * Override the `toUIMessageStream()` options for the current turn.
 *
 * These options control how the `StreamTextResult` is converted to a
 * `UIMessageChunk` stream — error handling, reasoning/source visibility,
 * message metadata, etc.
 *
 * Per-turn options are merged on top of the static `uiMessageStreamOptions`
 * set on `chat.agent()`. Per-turn values win on conflicts.
 *
 * @example
 * ```ts
 * run: async ({ messages, signal }) => {
 *   chat.setUIMessageStreamOptions({
 *     sendReasoning: true,
 *     onError: (error) => error instanceof Error ? error.message : "An error occurred.",
 *   });
 *   return streamText({ model, messages, abortSignal: signal });
 * }
 * ```
 */
function setUIMessageStreamOptions(options: ChatUIMessageStreamOptions<UIMessage>): void {
  locals.set(chatUIStreamPerTurnKey, options);
}

/**
 * Resolve the effective UIMessageStream options by merging:
 * 1. Static task-level options (from `chat.agent({ uiMessageStreamOptions })`)
 * 2. Per-turn overrides (from `chat.setUIMessageStreamOptions()`)
 *
 * Per-turn values win on conflicts. Clears the per-turn override after reading
 * so it doesn't leak into subsequent turns.
 * @internal
 */
function resolveUIMessageStreamOptions(): ChatUIMessageStreamOptions<UIMessage> {
  const staticOptions = locals.get(chatUIStreamStaticKey) ?? {};
  const perTurnOptions = locals.get(chatUIStreamPerTurnKey) ?? {};
  // Clear per-turn override so it doesn't leak into subsequent turns
  locals.set(chatUIStreamPerTurnKey, undefined);
  return { ...staticOptions, ...perTurnOptions };
}

// ---------------------------------------------------------------------------
// Stop detection
// ---------------------------------------------------------------------------

/**
 * Check whether the user stopped generation during the current turn.
 *
 * Works from **anywhere** inside a `chat.agent` run — including inside
 * `streamText`'s `onFinish` callback — without needing to thread the
 * `stopSignal` through closures.
 *
 * This is especially useful when the AI SDK's `isAborted` flag is unreliable
 * (e.g. when using `createUIMessageStream` + `writer.merge()`).
 *
 * @example
 * ```ts
 * onFinish: ({ isAborted }) => {
 *   const wasStopped = isAborted || chat.isStopped();
 *   if (wasStopped) {
 *     // handle stop
 *   }
 * }
 * ```
 */
function isStopped(): boolean {
  const controller = locals.get(chatStopControllerKey);
  return controller?.signal.aborted ?? false;
}

// ---------------------------------------------------------------------------
// Version upgrade
// ---------------------------------------------------------------------------

/**
 * Request that the current run exits so the next message starts on the latest
 * deployed version (via the standard continuation mechanism).
 *
 * When called from `onTurnStart` or `onValidateMessages`, `run()` is skipped
 * entirely — the run exits immediately and the transport re-triggers the
 * same message on the new version.
 *
 * When called from `run()` or `chat.defer()`, the current turn completes
 * normally and the run exits afterward instead of waiting for the next message.
 *
 * Call from `onTurnStart`, `onValidateMessages`, `onChatResume`, `run()`,
 * or inside `chat.defer()`.
 *
 * @example
 * ```ts
 * const SUPPORTED_VERSIONS = new Set(["v2", "v3"]);
 *
 * chat.agent({
 *   id: "my-chat",
 *   onTurnStart: async ({ clientData }) => {
 *     if (clientData?.protocolVersion && !SUPPORTED_VERSIONS.has(clientData.protocolVersion)) {
 *       chat.requestUpgrade();
 *     }
 *   },
 *   run: async ({ messages }) => { ... },
 * });
 * ```
 */
function requestUpgrade(): void {
  locals.set(chatUpgradeRequestedKey, true);
}

/**
 * Exit the run after the current turn completes, without waiting for the
 * next message. Unlike {@link requestUpgrade}, no upgrade-required signal
 * is sent to the client — the turn finishes normally, `onTurnComplete`
 * fires, and the loop exits instead of going idle.
 *
 * Call from `run()`, `chat.defer()`, `onBeforeTurnComplete`, or
 * `onTurnComplete` to end the run on your own terms (budget exhausted,
 * task complete, goal achieved, etc.).
 *
 * The next user message on the same `chatId` starts a fresh run via the
 * normal continuation mechanism.
 *
 * @example
 * ```ts
 * chat.agent({
 *   id: "one-shot-agent",
 *   run: async ({ messages, signal }) => {
 *     const result = streamText({ model: openai("gpt-4o"), messages, abortSignal: signal });
 *     // Single-response agent — exit after this turn.
 *     chat.endRun();
 *     return result;
 *   },
 * });
 * ```
 */
function endRun(): void {
  locals.set(chatEndRunRequestedKey, true);
}

// ---------------------------------------------------------------------------
// Per-turn deferred work
// ---------------------------------------------------------------------------

/**
 * Register a promise that runs in the background during the current turn.
 *
 * Use this to move non-blocking work (DB writes, analytics, etc.) out of
 * the critical path. The promise runs in parallel with streaming and is
 * awaited (with a 5 s timeout) before `onTurnComplete` fires.
 *
 * @example
 * ```ts
 * onTurnStart: async ({ chatId, uiMessages }) => {
 *   // Pass a promise directly
 *   chat.defer(db.chat.update({ where: { id: chatId }, data: { messages: uiMessages } }));
 *
 *   // Or pass an async function — cleaner for multi-step work
 *   chat.defer(async () => {
 *     const flags = await getFeatureFlags();
 *     if (flags.forceUpgrade) chat.requestUpgrade();
 *   });
 * },
 * ```
 */
function chatDefer(promiseOrFn: Promise<unknown> | (() => Promise<unknown>)): void {
  const promises = locals.get(chatDeferKey);
  if (promises) {
    promises.add(typeof promiseOrFn === "function" ? promiseOrFn() : promiseOrFn);
  }
}

// ---------------------------------------------------------------------------
// Background context injection
// ---------------------------------------------------------------------------

/**
 * Queue model messages for injection at the next `prepareStep` boundary.
 *
 * Use this to inject context from background work into the agent's conversation.
 * Messages are appended to the model messages before the next LLM inference call.
 *
 * Combine with `chat.defer()` to run background analysis and inject results:
 *
 * @example
 * ```ts
 * onTurnComplete: async ({ messages }) => {
 *   chat.defer((async () => {
 *     const review = await generateObject({
 *       model: openai("gpt-4o-mini"),
 *       messages: [...messages, { role: "user", content: "Review the last response." }],
 *       schema: z.object({ suggestions: z.array(z.string()) }),
 *     });
 *     if (review.object.suggestions.length > 0) {
 *       chat.inject([{
 *         role: "system",
 *         content: `Improvements for next response:\n${review.object.suggestions.join("\n")}`,
 *       }]);
 *     }
 *   })());
 * },
 * ```
 */
function injectBackgroundContext(messages: ModelMessage[]): void {
  const queue = locals.get(chatBackgroundQueueKey) ?? [];
  queue.push(...messages);
  locals.set(chatBackgroundQueueKey, queue);
}

// ---------------------------------------------------------------------------
// Aborted message cleanup
// ---------------------------------------------------------------------------

/**
 * Clean up a UIMessage that was captured during an aborted/stopped turn.
 *
 * When generation is stopped mid-stream, the captured message may contain:
 * - Tool parts stuck in incomplete states (`partial-call`, `input-available`,
 *   `input-streaming`) that cause permanent UI spinners
 * - Reasoning parts with `state: "streaming"` instead of `"done"`
 * - Text parts with `state: "streaming"` instead of `"done"`
 *
 * This function returns a cleaned copy with:
 * - Incomplete tool parts removed entirely
 * - Reasoning and text parts marked as `"done"`
 *
 * `chat.agent` calls this automatically when stop is detected before passing
 * the response to `onTurnComplete`. Use this manually when calling `pipeChat`
 * directly and capturing response messages yourself.
 *
 * @example
 * ```ts
 * onTurnComplete: async ({ responseMessage, stopped }) => {
 *   // Already cleaned automatically by chat.agent — but if you captured
 *   // your own message via pipeChat, clean it manually:
 *   const cleaned = chat.cleanupAbortedParts(myMessage);
 *   await db.messages.save(cleaned);
 * }
 * ```
 */
function cleanupAbortedParts<TUIM extends UIMessage>(message: TUIM): TUIM {
  if (!message.parts) return message;

  const isToolPart = (part: any) =>
    part.type === "tool-invocation" ||
    part.type?.startsWith("tool-") ||
    part.type === "dynamic-tool";

  return {
    ...message,
    parts: message.parts
      .filter((part: any) => {
        if (!isToolPart(part)) return true;
        // Remove tool parts that never completed execution.
        // partial-call: input was still streaming when aborted.
        // input-available: input was complete but tool never ran.
        // input-streaming: input was mid-stream.
        const state = part.toolInvocation?.state ?? part.state;
        return (
          state !== "partial-call" && state !== "input-available" && state !== "input-streaming"
        );
      })
      .map((part: any) => {
        // Mark streaming reasoning as done
        if (part.type === "reasoning" && part.state === "streaming") {
          return { ...part, state: "done" };
        }
        // Mark streaming text as done
        if (part.type === "text" && part.state === "streaming") {
          return { ...part, state: "done" };
        }
        return part;
      }),
  } as TUIM;
}

// ---------------------------------------------------------------------------
// Composable primitives for raw task chat
// ---------------------------------------------------------------------------

/**
 * Create a managed stop signal wired to the chat stop input stream.
 *
 * Call once at the start of your run. Use `signal` as the abort signal for
 * `streamText`. Call `reset()` at the start of each turn to get a fresh
 * per-turn signal. Call `cleanup()` when the run ends.
 *
 * @example
 * ```ts
 * const stop = chat.createStopSignal();
 * for (let turn = 0; turn < 100; turn++) {
 *   stop.reset();
 *   const result = streamText({ model, messages, abortSignal: stop.signal });
 *   await chat.pipe(result);
 *   // ...
 * }
 * stop.cleanup();
 * ```
 */
function createStopSignal(): {
  readonly signal: AbortSignal;
  reset: () => void;
  cleanup: () => void;
} {
  let controller = new AbortController();
  const sub = stopInput.on((data) => {
    controller.abort(data?.message || "stopped");
  });
  return {
    get signal() {
      return controller.signal;
    },
    reset() {
      controller = new AbortController();
    },
    cleanup() {
      sub.off();
    },
  };
}

/**
 * Signal the frontend that the current turn is complete.
 *
 * The `TriggerChatTransport` intercepts this to close the ReadableStream
 * for the current turn. Call after piping the response stream.
 *
 * @example
 * ```ts
 * await chat.pipe(result);
 * await chat.writeTurnComplete();
 * ```
 */
async function chatWriteTurnComplete(options?: { publicAccessToken?: string }): Promise<void> {
  await writeTurnCompleteChunk(undefined, options?.publicAccessToken);
}

/**
 * Pipe a `StreamTextResult` (or similar) to the chat stream and capture
 * the assistant's response message via `onFinish`.
 *
 * Combines `toUIMessageStream()` + `onFinish` callback + `chat.pipe()`.
 * Returns the captured `UIMessage`, or `undefined` if capture failed.
 *
 * @example
 * ```ts
 * const result = streamText({ model, messages, abortSignal: signal });
 * const response = await chat.pipeAndCapture(result, { signal });
 * if (response) conversation.addResponse(response);
 * ```
 */
async function pipeChatAndCapture(
  source: UIMessageStreamable,
  options?: { signal?: AbortSignal; spanName?: string }
): Promise<UIMessage | undefined> {
  let captured: UIMessage | undefined;
  let resolveOnFinish: () => void;
  const onFinishPromise = new Promise<void>((r) => {
    resolveOnFinish = r;
  });

  const uiStream = source.toUIMessageStream({
    ...resolveUIMessageStreamOptions(),
    onFinish: ({ responseMessage }: { responseMessage: UIMessage }) => {
      captured = responseMessage;
      resolveOnFinish!();
    },
  });

  await pipeChat(uiStream, {
    signal: options?.signal,
    spanName: options?.spanName ?? "stream response",
  });
  await onFinishPromise;

  return captured;
}

/**
 * Accumulates conversation messages across turns.
 *
 * Handles the transport protocol: turn 0 sends full history (replace),
 * subsequent turns send only new messages (append), regenerate sends
 * full history minus last assistant message (replace).
 *
 * @example
 * ```ts
 * const conversation = new chat.MessageAccumulator();
 * for (let turn = 0; turn < 100; turn++) {
 *   const messages = await conversation.addIncoming(payload.messages, payload.trigger, turn);
 *   const result = streamText({ model, messages });
 *   const response = await chat.pipeAndCapture(result);
 *   if (response) await conversation.addResponse(response);
 * }
 * ```
 */
class ChatMessageAccumulator {
  modelMessages: ModelMessage[] = [];
  uiMessages: UIMessage[] = [];
  private _compaction?: ChatAgentCompactionOptions;
  private _pendingMessages?: PendingMessagesOptions;
  private _steeringQueue: SteeringQueueEntry[] = [];

  constructor(options?: {
    compaction?: ChatAgentCompactionOptions;
    pendingMessages?: PendingMessagesOptions;
  }) {
    this._compaction = options?.compaction;
    this._pendingMessages = options?.pendingMessages;
  }

  /**
   * Add incoming messages from the transport payload.
   * Returns the full accumulated model messages for `streamText`.
   */
  async addIncoming(messages: UIMessage[], trigger: string, turn: number): Promise<ModelMessage[]> {
    const cleaned = messages.map((m) => (m.role === "assistant" ? cleanupAbortedParts(m) : m));
    const model = await toModelMessages(cleaned);

    if (turn === 0 || trigger === "regenerate-message") {
      this.modelMessages = model;
      this.uiMessages = [...cleaned];
    } else {
      this.modelMessages.push(...model);
      this.uiMessages.push(...cleaned);
    }
    return this.modelMessages;
  }

  /**
   * Add the assistant's response to the accumulator.
   * Call after `pipeAndCapture` with the captured response.
   */
  /**
   * Replace all accumulated messages (for compaction).
   * Converts UIMessages to ModelMessages internally.
   */
  async setMessages(uiMessages: UIMessage[]): Promise<void> {
    this.uiMessages = [...uiMessages];
    this.modelMessages = await toModelMessages(uiMessages);
  }

  async addResponse(response: UIMessage): Promise<void> {
    if (!response.id) {
      response = { ...response, id: generateMessageId() };
    }
    this.uiMessages.push(response);
    try {
      const msgs = await toModelMessages([stripProviderMetadata(response)]);
      this.modelMessages.push(...msgs);
    } catch {
      // Conversion failed — skip model message accumulation for this response
    }
  }

  /**
   * Queue a message for injection via `prepareStep`. Call from a
   * `messagesInput.on()` listener when a message arrives during streaming.
   */
  steer(message: UIMessage, modelMessages?: ModelMessage[]): void {
    if (modelMessages) {
      this._steeringQueue.push({ uiMessage: message, modelMessages });
    } else {
      // Defer conversion — will be done in prepareStep if needed
      this._steeringQueue.push({ uiMessage: message, modelMessages: [] });
    }
  }

  /**
   * Queue a message for injection, converting to model messages automatically.
   */
  async steerAsync(message: UIMessage): Promise<void> {
    const modelMsgs = await toModelMessages([message]);
    this._steeringQueue.push({ uiMessage: message, modelMessages: modelMsgs });
  }

  /**
   * Get and clear unconsumed steering messages.
   */
  drainSteering(): UIMessage[] {
    const result = this._steeringQueue.map((e) => e.uiMessage);
    this._steeringQueue = [];
    return result;
  }

  /**
   * Returns a `prepareStep` function that handles both compaction and
   * pending message injection. Pass to `streamText({ prepareStep: conversation.prepareStep() })`.
   */
  prepareStep():
    | ((args: {
      messages: ModelMessage[];
      steps: CompactionStep[];
    }) => Promise<{ messages: ModelMessage[] } | undefined>)
    | undefined {
    if (!this._compaction && !this._pendingMessages) return undefined;
    const comp = this._compaction;
    const pm = this._pendingMessages;
    const queue = this._steeringQueue;

    return async ({ messages, steps }) => {
      let resultMessages: ModelMessage[] | undefined;

      // 1. Compaction
      if (comp) {
        const result = await chatCompact(messages, steps, {
          shouldCompact: comp.shouldCompact,
          summarize: (msgs) => comp.summarize({ messages: msgs, source: "inner" }),
        });
        if (result.type !== "skipped") {
          resultMessages = result.messages;
        }
      }

      // 2. Pending message injection
      if (pm && queue.length > 0) {
        const injected = await drainSteeringQueue(pm, resultMessages ?? messages, steps, queue);
        if (injected.length > 0) {
          resultMessages = [...(resultMessages ?? messages), ...injected];
        }
      }

      return resultMessages ? { messages: resultMessages } : undefined;
    };
  }

  /**
   * Run outer-loop compaction if needed. Call after adding the response
   * and capturing usage. Applies `compactModelMessages` and `compactUIMessages`
   * callbacks if configured.
   *
   * @returns `true` if compaction was performed, `false` otherwise.
   */
  async compactIfNeeded(
    usage: LanguageModelUsage | undefined,
    context?: {
      chatId?: string;
      turn?: number;
      clientData?: unknown;
      totalUsage?: LanguageModelUsage;
    }
  ): Promise<boolean> {
    if (!this._compaction || !usage) return false;

    const shouldTrigger = await this._compaction.shouldCompact({
      messages: this.modelMessages,
      totalTokens: usage.totalTokens,
      inputTokens: usage.inputTokens,
      outputTokens: usage.outputTokens,
      usage,
      totalUsage: context?.totalUsage,
      chatId: context?.chatId,
      turn: context?.turn,
      clientData: context?.clientData,
      source: "outer",
    });

    if (!shouldTrigger) return false;

    const summary = await this._compaction.summarize({
      messages: this.modelMessages,
      usage,
      totalUsage: context?.totalUsage,
      chatId: context?.chatId,
      turn: context?.turn,
      clientData: context?.clientData,
      source: "outer",
    });

    const compactEvent: CompactMessagesEvent = {
      summary,
      uiMessages: this.uiMessages,
      modelMessages: this.modelMessages,
      chatId: context?.chatId ?? "",
      turn: context?.turn ?? 0,
      clientData: context?.clientData,
      source: "outer",
    };

    this.modelMessages = this._compaction.compactModelMessages
      ? await this._compaction.compactModelMessages(compactEvent)
      : [
        {
          role: "assistant" as const,
          content: [{ type: "text" as const, text: `[Conversation summary]\n\n${summary}` }],
        },
      ];

    if (this._compaction.compactUIMessages) {
      this.uiMessages = await this._compaction.compactUIMessages(compactEvent);
    }

    return true;
  }
}

// ---------------------------------------------------------------------------
// chat.createSession — async iterator for chat turns
// ---------------------------------------------------------------------------

export type ChatSessionOptions = {
  /** Run-level cancel signal (from task context). */
  signal: AbortSignal;
  /** Seconds to stay idle between turns before suspending. @default 30 */
  idleTimeoutInSeconds?: number;
  /** Duration string for suspend timeout. @default "1h" */
  timeout?: string;
  /** Max turns before ending. @default 100 */
  maxTurns?: number;
  /** Automatic context compaction — same options as `chat.agent({ compaction })`. */
  compaction?: ChatAgentCompactionOptions;
  /** Configure mid-execution message injection — same options as `chat.agent({ pendingMessages })`. */
  pendingMessages?: PendingMessagesOptions;
};

export type ChatTurn = {
  /** Turn number (0-indexed). */
  number: number;
  /** Chat session ID. */
  chatId: string;
  /** What triggered this turn. */
  trigger: string;
  /** Client data from the transport (`metadata` field on the wire payload). */
  clientData: unknown;
  /** Full accumulated model messages — pass directly to `streamText`. */
  readonly messages: ModelMessage[];
  /** Full accumulated UI messages — use for persistence. */
  readonly uiMessages: UIMessage[];
  /** Combined stop+cancel AbortSignal (fresh each turn). */
  signal: AbortSignal;
  /** Whether the user stopped generation this turn. */
  readonly stopped: boolean;
  /** Whether this is a continuation run. */
  continuation: boolean;
  /** Token usage from the previous turn. Undefined on turn 0. */
  previousTurnUsage?: LanguageModelUsage;
  /** Cumulative token usage across all completed turns so far. */
  totalUsage: LanguageModelUsage;

  /**
   * Replace accumulated messages (for compaction). Takes UIMessages and
   * converts to ModelMessages internally. After calling this, `turn.messages`
   * reflects the compacted history.
   */
  setMessages(uiMessages: UIMessage[]): Promise<void>;

  /**
   * Easy path: pipe stream, capture response, accumulate it,
   * clean up aborted parts if stopped, and write turn-complete chunk.
   */
  complete(source: UIMessageStreamable): Promise<UIMessage | undefined>;

  /**
   * Manual path: just write turn-complete chunk.
   * Use when you've already piped and accumulated manually.
   */
  done(): Promise<void>;

  /**
   * Add the response to the accumulator manually.
   * Use with `chat.pipeAndCapture` when you need control between pipe and done.
   */
  addResponse(response: UIMessage): Promise<void>;

  /**
   * Returns a `prepareStep` function that handles both compaction and
   * pending message injection. Pass to `streamText({ prepareStep: turn.prepareStep() })`.
   * Only needed when not using `chat.toStreamTextOptions()` (which auto-injects it).
   */
  prepareStep():
    | ((args: {
      messages: ModelMessage[];
      steps: CompactionStep[];
    }) => Promise<{ messages: ModelMessage[] } | undefined>)
    | undefined;
};

/**
 * Create a chat session that yields turns as an async iterator.
 *
 * Handles: preload wait, stop signals, message accumulation, turn-complete
 * signaling, and idle/suspend between turns. You control: initialization,
 * model/tool selection, persistence, and any custom per-turn logic.
 *
 * @example
 * ```ts
 * import { task } from "@trigger.dev/sdk";
 * import { chat, type ChatTaskWirePayload } from "@trigger.dev/sdk/ai";
 * import { streamText } from "ai";
 * import { openai } from "@ai-sdk/openai";
 *
 * export const myChat = task({
 *   id: "my-chat",
 *   run: async (payload: ChatTaskWirePayload, { signal }) => {
 *     const session = chat.createSession(payload, { signal });
 *
 *     for await (const turn of session) {
 *       const result = streamText({
 *         model: openai("gpt-4o"),
 *         messages: turn.messages,
 *         abortSignal: turn.signal,
 *       });
 *       await turn.complete(result);
 *     }
 *   },
 * });
 * ```
 */
function createChatSession(
  payload: ChatTaskWirePayload,
  options: ChatSessionOptions
): AsyncIterable<ChatTurn> {
  const {
    signal: runSignal,
    idleTimeoutInSeconds: sessionIdleTimeoutOpt,
    timeout = "1h",
    maxTurns = 100,
    compaction: sessionCompaction,
    pendingMessages: sessionPendingMessages,
  } = options;

  const idleTimeoutInSeconds = sessionIdleTimeoutOpt ?? 30;

  return {
    [Symbol.asyncIterator]() {
      let currentPayload = payload;
      let turn = -1;
      const stop = createStopSignal();
      const accumulator = new ChatMessageAccumulator();
      let previousTurnUsage: LanguageModelUsage | undefined;
      let cumulativeUsage: LanguageModelUsage = emptyUsage();

      return {
        async next(): Promise<IteratorResult<ChatTurn>> {
          turn++;

          // First turn: handle preload — wait for the first real message
          if (turn === 0 && currentPayload.trigger === "preload") {
            const result = await messagesInput.waitWithIdleTimeout({
              idleTimeoutInSeconds:
                sessionIdleTimeoutOpt ?? currentPayload.idleTimeoutInSeconds ?? 30,
              timeout,
              spanName: "waiting for first message",
            });
            if (!result.ok || runSignal.aborted) {
              stop.cleanup();
              return { done: true, value: undefined };
            }
            currentPayload = result.output;
          }

          // Subsequent turns: wait for the next message
          if (turn > 0) {
            // chat.requestUpgrade() / chat.endRun() — exit before waiting
            if (
              locals.get(chatUpgradeRequestedKey) ||
              locals.get(chatEndRunRequestedKey)
            ) {
              stop.cleanup();
              return { done: true, value: undefined };
            }

            const next = await messagesInput.waitWithIdleTimeout({
              idleTimeoutInSeconds,
              timeout,
              spanName: "waiting for next message",
            });
            if (!next.ok || runSignal.aborted) {
              stop.cleanup();
              return { done: true, value: undefined };
            }
            currentPayload = next.output;
          }

          // Check limits
          if (turn >= maxTurns || runSignal.aborted) {
            stop.cleanup();
            return { done: true, value: undefined };
          }

          // Reset stop signal for this turn
          stop.reset();

          // Reset per-turn state
          locals.set(chatResponsePartsKey, []);
          // Set up steering queue and pending messages config in locals
          // so toStreamTextOptions() auto-injects prepareStep for steering
          const turnSteeringQueue: SteeringQueueEntry[] = [];
          locals.set(chatSteeringQueueKey, turnSteeringQueue);
          if (sessionPendingMessages) {
            locals.set(chatPendingMessagesKey, sessionPendingMessages);
          }
          locals.set(chatTurnContextKey, {
            chatId: currentPayload.chatId,
            turn,
            continuation: currentPayload.continuation ?? false,
            clientData: currentPayload.metadata,
          });

          // Listen for messages during streaming (steering + next-turn buffer)
          const sessionPendingWire: ChatTaskWirePayload[] = [];
          const sessionMsgSub = messagesInput.on(async (msg) => {
            sessionPendingWire.push(msg);

            if (sessionPendingMessages) {
              // Slim wire: at most one delta message per record. Read
              // `msg.message` directly — no array slicing needed.
              const lastUIMessage = msg.message;
              if (lastUIMessage) {
                if (sessionPendingMessages.onReceived) {
                  try {
                    await sessionPendingMessages.onReceived({
                      message: lastUIMessage,
                      chatId: currentPayload.chatId,
                      turn,
                    });
                  } catch {
                    /* non-fatal */
                  }
                }
                try {
                  const modelMsgs = await toModelMessages([lastUIMessage]);
                  turnSteeringQueue.push({ uiMessage: lastUIMessage, modelMessages: modelMsgs });
                } catch {
                  /* non-fatal */
                }
              }
            }
          });

          // Accumulate messages. Slim wire: pass the single delta message as
          // a 0-or-1-length array. The accumulator's behavior is unchanged —
          // it still appends user messages and reconverts on regenerate.
          const incomingForAccumulator: UIMessage[] = currentPayload.message
            ? [currentPayload.message]
            : [];
          const messages = await accumulator.addIncoming(
            incomingForAccumulator,
            currentPayload.trigger,
            turn
          );

          // chat.requestUpgrade() called before this turn — signal transport and exit
          if (locals.get(chatUpgradeRequestedKey)) {
            await writeUpgradeRequiredChunk();
            sessionMsgSub.off();
            stop.cleanup();
            return { done: true, value: undefined };
          }

          const combinedSignal = AbortSignal.any([runSignal, stop.signal]);

          const turnObj: ChatTurn = {
            number: turn,
            chatId: currentPayload.chatId,
            trigger: currentPayload.trigger,
            clientData: currentPayload.metadata,
            get messages() {
              return accumulator.modelMessages;
            },
            get uiMessages() {
              return accumulator.uiMessages;
            },
            signal: combinedSignal,
            get stopped() {
              return stop.signal.aborted && !runSignal.aborted;
            },
            continuation: currentPayload.continuation ?? false,
            previousTurnUsage,
            totalUsage: cumulativeUsage,

            async setMessages(uiMessages: UIMessage[]) {
              await accumulator.setMessages(uiMessages);
            },

            async complete(source: UIMessageStreamable) {
              let response: UIMessage | undefined;
              try {
                response = await pipeChatAndCapture(source, { signal: combinedSignal });
              } catch (error) {
                if (error instanceof Error && error.name === "AbortError") {
                  if (runSignal.aborted) {
                    // Full cancel — don't accumulate
                    sessionMsgSub.off();
                    await chatWriteTurnComplete();
                    return undefined;
                  }
                  // Stop — fall through to accumulate partial response
                } else {
                  throw error;
                }
              }

              if (response) {
                const cleaned =
                  stop.signal.aborted && !runSignal.aborted
                    ? cleanupAbortedParts(response)
                    : response;
                // Append any non-transient data parts queued via chat.response or writer.write()
                const queuedParts = locals.get(chatResponsePartsKey);
                if (queuedParts && queuedParts.length > 0) {
                  (cleaned as any).parts = [...(cleaned.parts ?? []), ...queuedParts];
                  locals.set(chatResponsePartsKey, []);
                }
                await accumulator.addResponse(cleaned);
              } else {
                // No response (manual pipe mode) but there are queued data parts
                const queuedParts = locals.get(chatResponsePartsKey);
                if (queuedParts && queuedParts.length > 0) {
                  await accumulator.addResponse({
                    id: generateMessageId(),
                    role: "assistant" as const,
                    parts: queuedParts as UIMessage["parts"],
                  });
                  locals.set(chatResponsePartsKey, []);
                }
              }

              // Capture token usage from the streamText result
              let turnUsage: LanguageModelUsage | undefined;
              if (typeof (source as any).totalUsage?.then === "function") {
                try {
                  const usage: LanguageModelUsage = await (source as any).totalUsage;
                  turnUsage = usage;
                  previousTurnUsage = usage;
                  cumulativeUsage = addUsage(cumulativeUsage, usage);
                } catch {
                  /* non-fatal */
                }
              }

              // Outer-loop compaction (same logic as chat.agent)
              if (sessionCompaction && turnUsage && !turnObj.stopped) {
                const shouldTrigger = await sessionCompaction.shouldCompact({
                  messages: accumulator.modelMessages,
                  totalTokens: turnUsage.totalTokens,
                  inputTokens: turnUsage.inputTokens,
                  outputTokens: turnUsage.outputTokens,
                  usage: turnUsage,
                  totalUsage: cumulativeUsage,
                  chatId: currentPayload.chatId,
                  turn,
                  clientData: currentPayload.metadata,
                  source: "outer",
                });

                if (shouldTrigger) {
                  const summary = await sessionCompaction.summarize({
                    messages: accumulator.modelMessages,
                    usage: turnUsage,
                    totalUsage: cumulativeUsage,
                    chatId: currentPayload.chatId,
                    turn,
                    clientData: currentPayload.metadata,
                    source: "outer",
                  });

                  const compactEvent: CompactMessagesEvent = {
                    summary,
                    uiMessages: accumulator.uiMessages,
                    modelMessages: accumulator.modelMessages,
                    chatId: currentPayload.chatId,
                    turn,
                    clientData: currentPayload.metadata,
                    source: "outer",
                  };

                  accumulator.modelMessages = sessionCompaction.compactModelMessages
                    ? await sessionCompaction.compactModelMessages(compactEvent)
                    : [
                      {
                        role: "assistant" as const,
                        content: [
                          { type: "text" as const, text: `[Conversation summary]\n\n${summary}` },
                        ],
                      },
                    ];

                  if (sessionCompaction.compactUIMessages) {
                    accumulator.uiMessages = await sessionCompaction.compactUIMessages(
                      compactEvent
                    );
                  }
                }
              }

              sessionMsgSub.off();
              await chatWriteTurnComplete();
              return response;
            },

            async addResponse(response: UIMessage) {
              // Append any non-transient data parts queued via chat.response or writer.write()
              const queuedParts = locals.get(chatResponsePartsKey);
              if (queuedParts && queuedParts.length > 0) {
                response = { ...response, parts: [...(response.parts ?? []), ...(queuedParts as UIMessage["parts"])] };
                locals.set(chatResponsePartsKey, []);
              }
              await accumulator.addResponse(response);
            },

            async done() {
              sessionMsgSub.off();
              await chatWriteTurnComplete();
            },

            prepareStep() {
              const hasCompaction = !!sessionCompaction;
              const hasPending = !!sessionPendingMessages;
              if (!hasCompaction && !hasPending) return undefined;

              return async ({
                messages: stepMsgs,
                steps,
              }: {
                messages: ModelMessage[];
                steps: CompactionStep[];
              }) => {
                let resultMessages: ModelMessage[] | undefined;

                if (sessionCompaction) {
                  const compactResult = await chatCompact(stepMsgs, steps, {
                    shouldCompact: sessionCompaction.shouldCompact,
                    summarize: (msgs) =>
                      sessionCompaction.summarize({ messages: msgs, source: "inner" }),
                  });
                  if (compactResult.type !== "skipped") {
                    resultMessages = compactResult.messages;
                  }
                }

                if (sessionPendingMessages) {
                  const injected = await drainSteeringQueue(
                    sessionPendingMessages,
                    resultMessages ?? stepMsgs,
                    steps,
                    turnSteeringQueue
                  );
                  if (injected.length > 0) {
                    resultMessages = [...(resultMessages ?? stepMsgs), ...injected];
                  }
                }

                return resultMessages ? { messages: resultMessages } : undefined;
              };
            },
          };

          return { done: false, value: turnObj };
        },

        async return() {
          stop.cleanup();
          return { done: true, value: undefined };
        },
      };
    },
  };
}

// ---------------------------------------------------------------------------
// chat.local — per-run typed data with Proxy access
// ---------------------------------------------------------------------------

/** @internal Symbol for storing the locals key on the proxy target. */
const CHAT_LOCAL_KEY: unique symbol = Symbol("chatLocalKey");
/** @internal Symbol for storing the dirty-tracking locals key. */
const CHAT_LOCAL_DIRTY_KEY: unique symbol = Symbol("chatLocalDirtyKey");

// ---------------------------------------------------------------------------
// chat.local registry — tracks all declared locals for serialization
// ---------------------------------------------------------------------------

type ChatLocalEntry = { key: ReturnType<typeof locals.create>; id: string };
const chatLocalRegistry = new Set<ChatLocalEntry>();

/** @internal Run-scoped flag to ensure hydration happens at most once per run. */
const chatLocalsHydratedKey = locals.create<boolean>("chat.locals.hydrated");

/**
 * Hydrate chat.local values from subtask metadata (set by `ai.toolExecute()` or legacy `ai.tool()`).
 * Runs once per run — subsequent calls are no-ops.
 * @internal
 */
function hydrateLocalsFromMetadata(): void {
  if (locals.get(chatLocalsHydratedKey)) return;
  locals.set(chatLocalsHydratedKey, true);
  const opts = metadata.get(METADATA_KEY) as ToolCallExecutionOptions | undefined;
  if (!opts?.chatLocals) return;
  for (const [id, value] of Object.entries(opts.chatLocals)) {
    locals.set(locals.create(id), value);
  }
}

/**
 * A Proxy-backed, run-scoped data object that appears as `T` to users.
 * Includes helper methods for initialization, dirty tracking, and serialization.
 * Internal metadata is stored behind Symbols and invisible to
 * `Object.keys()`, `JSON.stringify()`, and spread.
 */
export type ChatLocal<T extends Record<string, unknown>> = T & {
  /** Initialize the local with a value. Call in `onChatStart` or `run()`. */
  init(value: T): void;
  /** Returns `true` if any property was set since the last check. Resets the dirty flag. */
  hasChanged(): boolean;
  /** Returns a plain object copy of the current value. Useful for persistence. */
  get(): T;
  readonly [CHAT_LOCAL_KEY]: ReturnType<typeof locals.create<T>>;
  readonly [CHAT_LOCAL_DIRTY_KEY]: ReturnType<typeof locals.create<boolean>>;
};

/**
 * Creates a per-run typed data object accessible from anywhere during task execution.
 *
 * Declare at module level, then initialize inside `onBoot` (recommended — fires
 * on every fresh worker including continuation runs). Do NOT initialize in
 * `onChatStart` alone: `onChatStart` only fires on the chat's very first
 * message, so `chat.local` would be uninitialized on continuation runs and
 * `run()` would throw.
 *
 * Multiple locals can coexist — each gets its own isolated run-scoped storage.
 *
 * The `id` is required and must be unique across all `chat.local()` calls in
 * your project. It's used to serialize values into subtask metadata so that
 * `ai.toolExecute()` (or legacy `ai.tool()`) subtasks can auto-hydrate parent locals (read-only).
 *
 * @example
 * ```ts
 * import { chat } from "@trigger.dev/sdk/ai";
 *
 * const userPrefs = chat.local<{ theme: string; language: string }>({ id: "userPrefs" });
 * const gameState = chat.local<{ score: number; streak: number }>({ id: "gameState" });
 *
 * export const myChat = chat.agent({
 *   id: "my-chat",
 *   onBoot: async ({ clientData }) => {
 *     const prefs = await db.prefs.findUnique({ where: { userId: clientData.userId } });
 *     userPrefs.init(prefs ?? { theme: "dark", language: "en" });
 *     gameState.init({ score: 0, streak: 0 });
 *   },
 *   onTurnComplete: async ({ chatId }) => {
 *     if (gameState.hasChanged()) {
 *       await db.save({ where: { chatId }, data: gameState.get() });
 *     }
 *   },
 *   run: async ({ messages }) => {
 *     gameState.score++;
 *     return streamText({
 *       system: `User prefers ${userPrefs.theme} theme. Score: ${gameState.score}`,
 *       messages,
 *     });
 *   },
 * });
 * ```
 */
function chatLocal<T extends Record<string, unknown>>(options: { id: string }): ChatLocal<T> {
  const id = `chat.local.${options.id}`;
  const localKey = locals.create<T>(id);
  const dirtyKey = locals.create<boolean>(`${id}.dirty`);

  chatLocalRegistry.add({ key: localKey, id });

  const target = {} as any;
  target[CHAT_LOCAL_KEY] = localKey;
  target[CHAT_LOCAL_DIRTY_KEY] = dirtyKey;

  return new Proxy(target, {
    get(_target, prop, _receiver) {
      // Internal Symbol properties
      if (prop === CHAT_LOCAL_KEY) return _target[CHAT_LOCAL_KEY];
      if (prop === CHAT_LOCAL_DIRTY_KEY) return _target[CHAT_LOCAL_DIRTY_KEY];

      // Instance methods
      if (prop === "init") {
        return (value: T) => {
          locals.set(localKey, value);
          locals.set(dirtyKey, false);
        };
      }
      if (prop === "hasChanged") {
        return () => {
          const dirty = locals.get(dirtyKey) ?? false;
          locals.set(dirtyKey, false);
          return dirty;
        };
      }
      if (prop === "get") {
        return () => {
          let current = locals.get(localKey);
          if (current === undefined) {
            hydrateLocalsFromMetadata();
            current = locals.get(localKey);
          }
          if (current === undefined) {
            throw new Error(
              "local.get() called before initialization. Call local.init() in onBoot (recommended — fires on every fresh worker including continuation runs) or run() first."
            );
          }
          return { ...current };
        };
      }
      // toJSON for serialization (JSON.stringify(local))
      if (prop === "toJSON") {
        return () => {
          let current = locals.get(localKey);
          if (current === undefined) {
            hydrateLocalsFromMetadata();
            current = locals.get(localKey);
          }
          return current ? { ...current } : undefined;
        };
      }

      let current = locals.get(localKey);
      if (current === undefined) {
        // Auto-hydrate from parent metadata in subtask context
        hydrateLocalsFromMetadata();
        current = locals.get(localKey);
      }
      if (current === undefined) return undefined;
      return (current as any)[prop];
    },

    set(_target, prop, value) {
      // Don't allow setting internal Symbols
      if (typeof prop === "symbol") return false;

      const current = locals.get(localKey);
      if (current === undefined) {
        throw new Error(
          "chat.local can only be modified after initialization. " +
          "Call local.init() in onBoot (recommended — fires on every fresh worker including continuation runs) or run() first. " +
          "If you previously initialized in onChatStart, move it to onBoot — onChatStart only fires on the chat's very first message and will not run on a continuation."
        );
      }
      locals.set(localKey, { ...current, [prop]: value });
      locals.set(dirtyKey, true);
      return true;
    },

    has(_target, prop) {
      if (typeof prop === "symbol") return prop in _target;
      let current = locals.get(localKey);
      if (current === undefined) {
        hydrateLocalsFromMetadata();
        current = locals.get(localKey);
      }
      return current !== undefined && prop in current;
    },

    ownKeys() {
      let current = locals.get(localKey);
      if (current === undefined) {
        hydrateLocalsFromMetadata();
        current = locals.get(localKey);
      }
      return current ? Reflect.ownKeys(current) : [];
    },

    getOwnPropertyDescriptor(_target, prop) {
      if (typeof prop === "symbol") return undefined;
      let current = locals.get(localKey);
      if (current === undefined) {
        hydrateLocalsFromMetadata();
        current = locals.get(localKey);
      }
      if (current === undefined || !(prop in current)) return undefined;
      return {
        configurable: true,
        enumerable: true,
        writable: true,
        value: (current as any)[prop],
      };
    },
  }) as ChatLocal<T>;
}

/**
 * Extracts the client data (metadata) type from a chat task.
 * Use this to type the `metadata` option on the transport.
 *
 * @example
 * ```ts
 * import type { InferChatClientData } from "@trigger.dev/sdk/ai";
 * import type { myChat } from "@/trigger/chat";
 *
 * type MyClientData = InferChatClientData<typeof myChat>;
 * // { model?: string; userId: string }
 * ```
 */
// `InferChatClientData` and `InferChatUIMessage` live in `./ai-shared.ts`
// so the chat React hooks can import them without dragging `ai.ts` into
// the browser graph. Re-exported here so `@trigger.dev/sdk/ai` consumers
// still see them.
import type { InferChatClientData, InferChatUIMessage } from "./ai-shared.js";
export type {
  InferChatClientData,
  InferChatUIMessage,
  InferChatUIMessageFromTools,
} from "./ai-shared.js";

/**
 * Options for {@link createChatStartSessionAction}.
 */
/**
 * Discriminator for per-endpoint `baseURL` / `fetch` callbacks on
 * `createChatStartSessionAction`.
 *
 * - `"sessions"` — `POST /api/v1/sessions` (session create + first run trigger).
 * - `"auth"` — `POST /api/v1/auth/jwt/claims` (only fired when
 *   `tokenTTL` is set; otherwise the publicAccessToken from session create
 *   is reused as-is).
 */
export type ChatStartSessionEndpoint = "sessions" | "auth";

export type ChatStartSessionEndpointContext = {
  endpoint: ChatStartSessionEndpoint;
  chatId: string;
};

export type ChatStartSessionBaseURLResolver = (
  ctx: ChatStartSessionEndpointContext
) => string;

export type ChatStartSessionFetchOverride = (
  url: string,
  init: RequestInit,
  ctx: ChatStartSessionEndpointContext
) => Promise<Response>;

export type CreateChatStartSessionActionOptions = {
  /** TTL for the session-scoped public access token. @default "1h" */
  tokenTTL?: string | number | Date;
  /**
   * Default trigger config used when starting a new session for a chat.
   * Per-call `params.triggerConfig` shallow-merges on top.
   */
  triggerConfig?: Partial<SessionTriggerConfig>;
  /**
   * Override the Trigger.dev API base URL. String applies to both
   * `/api/v1/sessions` and `/api/v1/auth/jwt/claims`; function picks per
   * endpoint. When unset, falls back to `apiClientManager.baseURL`
   * (typically the `TRIGGER_API_URL` env var). Set this to route session
   * create through a trusted edge proxy that injects server-side signal
   * into `basePayload.metadata` before forwarding upstream.
   */
  baseURL?: string | ChatStartSessionBaseURLResolver;
  /**
   * Per-request fetch override. Receives the resolved URL, RequestInit,
   * and endpoint context. Use for header injection, proxy routing, or
   * custom retry. Applies to both session-create and JWT-claims POSTs.
   */
  fetch?: ChatStartSessionFetchOverride;
};

/**
 * Params for the function returned by {@link createChatStartSessionAction}.
 */
export type ChatStartSessionParams<TChat extends AnyTask = AnyTask> = {
  /** Conversation id (mapped to the Session's `externalId`). */
  chatId: string;
  /**
   * Typed client data — folded into the first run's `payload.metadata` so
   * `onPreload`, `onChatStart`, etc. see the same `clientData` shape on the
   * first turn as subsequent turns get via the transport's `clientData`
   * option. Typed via the agent's `clientDataSchema` when the action is
   * parameterised with `createStartSessionAction<typeof myChat>(...)`.
   */
  clientData?: InferChatClientData<TChat>;
  /**
   * Per-call trigger config. Shallow-merged over the action's default
   * `triggerConfig`. `basePayload` is the customer's wire payload (for
   * `chat.agent`: anything beyond `chatId`/`messages`/`trigger`/`metadata`,
   * which the runtime injects automatically).
   */
  triggerConfig?: Partial<SessionTriggerConfig>;
  /**
   * Opaque session-level metadata stored on the Session row. Separate from
   * the per-turn `clientData` above. Use this when you want to attach
   * server-side metadata that doesn't go through the agent's `clientDataSchema`.
   */
  metadata?: Record<string, unknown>;
};

/**
 * Result from {@link createChatStartSessionAction}'s returned function.
 */
export type ChatStartSessionResult = {
  /**
   * Session-scoped public access token (`read:sessions:{chatId} +
   * write:sessions:{chatId}`). Pass this to the browser; the transport
   * uses it to call `.in/append`, `.out`, `end-and-continue`.
   */
  publicAccessToken: string;
  /** Friendly id of the run triggered alongside session create. */
  runId: string;
  /** Session friendlyId — informational. */
  sessionId: string;
};

/**
 * Creates a server-side helper that starts (or resumes) a Session for a
 * given chatId — atomically creating the row, triggering the first run,
 * and returning a session-scoped PAT for the browser to use.
 *
 * Wrap in a Next.js server action (or any server-side handler) so the
 * customer's secret key never crosses to the browser.
 *
 * Parameterise the action with `<typeof yourChatAgent>` to type the
 * `clientData` field against your agent's `clientDataSchema`.
 *
 * @example
 * ```ts
 * // actions.ts
 * "use server";
 * import { chat } from "@trigger.dev/sdk/ai";
 * import type { myChat } from "@/trigger/chat";
 *
 * export const startChatSession = chat.createStartSessionAction<typeof myChat>(
 *   "my-chat",
 *   { triggerConfig: { machine: "small-1x" } }
 * );
 * ```
 *
 * Then in the browser, threading the typed `clientData` from the transport:
 * ```tsx
 * const transport = useTriggerChatTransport<typeof myChat>({
 *   task: "my-chat",
 *   accessToken: ({ chatId }) => mintChatAccessToken(chatId),
 *   startSession: ({ chatId, clientData }) =>
 *     startChatSession({ chatId, clientData }),
 * });
 * ```
 */
function createChatStartSessionAction<TChat extends AnyTask = AnyTask>(
  taskId: string,
  options?: CreateChatStartSessionActionOptions
): (params: ChatStartSessionParams<TChat>) => Promise<ChatStartSessionResult> {
  return async (params: ChatStartSessionParams<TChat>): Promise<ChatStartSessionResult> => {
    if (!params.chatId) {
      throw new Error(
        "chat.createStartSessionAction: params.chatId is required — used as the session externalId."
      );
    }

    // The first run boots before the user's first message lands on
    // `.in/append`, so it sees an empty `messages` array and `trigger:
    // "preload"`. This matches the pre-Sessions preload semantics:
    // `onPreload` fires, the runtime opens its `.in` subscription, the
    // first user message arrives moments later via `.in/append`.
    //
    // `clientData` is folded into `basePayload.metadata` so the agent's
    // `clientDataSchema` validates on the very first turn against the same
    // shape per-turn `metadata` carries via the transport.
    // Auto-tag every chat.agent run with `chat:{chatId}` so the dashboard /
    // run-list filter by chat works without the customer having to wire it
    // up. Mirrors the browser-mediated `TriggerChatTransport.doStart` path.
    const userTags = params.triggerConfig?.tags ?? options?.triggerConfig?.tags ?? [];
    const tags = [`chat:${params.chatId}`, ...userTags].slice(0, 5);

    const clientDataMetadata =
      params.clientData !== undefined ? { metadata: params.clientData } : {};

    const triggerConfig: SessionTriggerConfig = {
      basePayload: {
        messages: [],
        trigger: "preload",
        ...(options?.triggerConfig?.basePayload ?? {}),
        ...(params.triggerConfig?.basePayload ?? {}),
        ...clientDataMetadata,
        chatId: params.chatId,
      },
      ...(options?.triggerConfig?.machine || params.triggerConfig?.machine
        ? { machine: params.triggerConfig?.machine ?? options?.triggerConfig?.machine }
        : {}),
      ...(options?.triggerConfig?.queue || params.triggerConfig?.queue
        ? { queue: params.triggerConfig?.queue ?? options?.triggerConfig?.queue }
        : {}),
      tags,
      ...(options?.triggerConfig?.maxAttempts !== undefined ||
      params.triggerConfig?.maxAttempts !== undefined
        ? {
            maxAttempts:
              params.triggerConfig?.maxAttempts ?? options?.triggerConfig?.maxAttempts!,
          }
        : {}),
      ...(options?.triggerConfig?.idleTimeoutInSeconds !== undefined ||
      params.triggerConfig?.idleTimeoutInSeconds !== undefined
        ? {
            idleTimeoutInSeconds:
              params.triggerConfig?.idleTimeoutInSeconds ??
              options?.triggerConfig?.idleTimeoutInSeconds!,
          }
        : {}),
    };

    const startBody = {
      type: "chat.agent" as const,
      externalId: params.chatId,
      taskIdentifier: taskId,
      triggerConfig,
      metadata: params.metadata,
    };

    const baseURLOption = options?.baseURL;
    const fetchOverride = options?.fetch;
    const hasOverride = baseURLOption !== undefined || fetchOverride !== undefined;

    const created: { id: string; runId: string; publicAccessToken: string } = hasOverride
      ? await callSessionsCreateWithOverride({
          chatId: params.chatId,
          body: startBody,
          baseURLOption,
          fetchOverride,
        })
      : await sessions.start(startBody);

    // Session create returns a session PAT directly when called with a
    // start token, but when the SDK call goes via the secret key we still
    // need to mint our own (the server returns a PAT regardless, but
    // re-minting here lets the customer override `tokenTTL`).
    const publicAccessToken =
      options?.tokenTTL !== undefined
        ? hasOverride
          ? await mintPublicTokenWithOverride({
              chatId: params.chatId,
              expirationTime: options.tokenTTL,
              baseURLOption,
              fetchOverride,
            })
          : await auth.createPublicToken({
              scopes: {
                read: { sessions: params.chatId },
                write: { sessions: params.chatId },
              },
              expirationTime: options.tokenTTL,
            })
        : created.publicAccessToken;

    return {
      publicAccessToken,
      runId: created.runId,
      sessionId: created.id,
    };
  };
}

function resolveChatStartBaseURL(
  endpoint: ChatStartSessionEndpoint,
  chatId: string,
  option: string | ChatStartSessionBaseURLResolver | undefined
): string {
  const fallback = apiClientManager.baseURL ?? "https://api.trigger.dev";
  const raw =
    typeof option === "function"
      ? option({ endpoint, chatId })
      : option ?? fallback;
  return raw.replace(/\/$/, "");
}

function overrideRequestHeaders(accessToken: string): Record<string, string> {
  const headers: Record<string, string> = {
    "Content-Type": "application/json",
    Authorization: `Bearer ${accessToken}`,
    "x-trigger-source": "sdk",
  };
  // Forward the preview-branch hint so override-mode requests land on the
  // same env the standard ApiClient path would have routed to. Mirrors
  // ApiClient.#getHeaders. Read from TRIGGER_PREVIEW_BRANCH /
  // VERCEL_GIT_COMMIT_REF via apiClientManager.branchName.
  if (apiClientManager.branchName) {
    headers["x-trigger-branch"] = apiClientManager.branchName;
  }
  return headers;
}

async function callSessionsCreateWithOverride(args: {
  chatId: string;
  body: { type: "chat.agent"; externalId: string; taskIdentifier: string; triggerConfig: SessionTriggerConfig; metadata?: Record<string, unknown> };
  baseURLOption: string | ChatStartSessionBaseURLResolver | undefined;
  fetchOverride: ChatStartSessionFetchOverride | undefined;
}): Promise<{ id: string; runId: string; publicAccessToken: string }> {
  const accessToken = apiClientManager.accessToken;
  if (!accessToken) {
    throw new Error(
      "chat.createStartSessionAction: no API access token configured. Set TRIGGER_SECRET_KEY or call apiClientManager.setGlobalAPIClientConfiguration before invoking the action."
    );
  }
  const ctx: ChatStartSessionEndpointContext = { endpoint: "sessions", chatId: args.chatId };
  const url = `${resolveChatStartBaseURL("sessions", args.chatId, args.baseURLOption)}/api/v1/sessions`;
  const init: RequestInit = {
    method: "POST",
    headers: overrideRequestHeaders(accessToken),
    body: JSON.stringify(args.body),
  };
  const response = args.fetchOverride
    ? await args.fetchOverride(url, init, ctx)
    : await fetch(url, init);
  if (!response.ok) {
    const text = await response.text().catch(() => "");
    throw new Error(`sessions.start failed: ${response.status} ${text}`);
  }
  const json = (await response.json()) as { id: string; runId: string; publicAccessToken: string };
  return json;
}

async function mintPublicTokenWithOverride(args: {
  chatId: string;
  expirationTime: string | number | Date;
  baseURLOption: string | ChatStartSessionBaseURLResolver | undefined;
  fetchOverride: ChatStartSessionFetchOverride | undefined;
}): Promise<string> {
  const accessToken = apiClientManager.accessToken;
  if (!accessToken) {
    throw new Error(
      "chat.createStartSessionAction: no API access token configured for JWT mint."
    );
  }
  const ctx: ChatStartSessionEndpointContext = { endpoint: "auth", chatId: args.chatId };
  const url = `${resolveChatStartBaseURL("auth", args.chatId, args.baseURLOption)}/api/v1/auth/jwt/claims`;
  const init: RequestInit = {
    method: "POST",
    headers: overrideRequestHeaders(accessToken),
  };
  const response = args.fetchOverride
    ? await args.fetchOverride(url, init, ctx)
    : await fetch(url, init);
  if (!response.ok) {
    const text = await response.text().catch(() => "");
    throw new Error(`auth.createPublicToken failed: ${response.status} ${text}`);
  }
  const claims = (await response.json()) as Record<string, unknown>;
  return generateJWT({
    secretKey: accessToken,
    payload: {
      ...claims,
      scopes: [`read:sessions:${args.chatId}`, `write:sessions:${args.chatId}`],
    },
    expirationTime: args.expirationTime,
  });
}

export const chat = {
  /** Create a chat agent. See {@link chatAgent}. */
  agent: chatAgent,
  /** Create a custom agent with manual lifecycle control. See {@link chatCustomAgent}. */
  customAgent: chatCustomAgent,
  /** Create a chat task with a fixed {@link UIMessage} subtype and optional default stream options. See {@link withUIMessage}. */
  withUIMessage,
  /** Create a chat task with a fixed client data schema. See {@link withClientData}. */
  withClientData,
  /** Create a server-side helper for starting (or resuming) a Session for a chatId. See {@link createChatStartSessionAction}. */
  createStartSessionAction: createChatStartSessionAction,
  /** Pipe a stream to the chat transport. See {@link pipeChat}. */
  pipe: pipeChat,
  /** Create a per-run typed local. See {@link chatLocal}. */
  local: chatLocal,
  /** Create a public access token for a chat task. See {@link createChatAccessToken}. */
  createAccessToken: createChatAccessToken,
  /** Override the turn timeout at runtime (duration string). See {@link setTurnTimeout}. */
  setTurnTimeout,
  /** Override the turn timeout at runtime (seconds). See {@link setTurnTimeoutInSeconds}. */
  setTurnTimeoutInSeconds,
  /** Override the idle timeout at runtime. See {@link setIdleTimeoutInSeconds}. */
  setIdleTimeoutInSeconds,
  /** Override toUIMessageStream() options for the current turn. See {@link setUIMessageStreamOptions}. */
  setUIMessageStreamOptions,
  /** Check if the current turn was stopped by the user. See {@link isStopped}. */
  isStopped,
  /** Request that the run exits after the current turn so the next message starts on the latest version. See {@link requestUpgrade}. */
  requestUpgrade,
  /** Exit the run after the current turn completes, without any upgrade signal. See {@link endRun}. */
  endRun,
  /** Clean up aborted parts from a UIMessage. See {@link cleanupAbortedParts}. */
  cleanupAbortedParts,
  /** Register background work that runs in parallel with streaming. See {@link chatDefer}. */
  defer: chatDefer,
  /** Queue model messages for injection at the next `prepareStep` boundary. See {@link injectBackgroundContext}. */
  inject: injectBackgroundContext,
  /** Typed chat output stream for writing custom chunks or piping from subtasks. */
  stream: chatStream,
  /** Write data parts that persist to the response message. See {@link chatResponse}. */
  response: chatResponse,
  /** Pre-built input stream for receiving messages from the transport. */
  messages: messagesInput,
  /** Create a managed stop signal wired to the stop input stream. See {@link createStopSignal}. */
  createStopSignal,
  /** Signal the frontend that the current turn is complete. See {@link chatWriteTurnComplete}. */
  writeTurnComplete: chatWriteTurnComplete,
  /** Pipe a stream and capture the response message. See {@link pipeChatAndCapture}. */
  pipeAndCapture: pipeChatAndCapture,
  /** Message accumulator class for raw task chat. See {@link ChatMessageAccumulator}. */
  MessageAccumulator: ChatMessageAccumulator,
  /** Create a chat session (async iterator). See {@link createChatSession}. */
  createSession: createChatSession,
  /**
   * Store and retrieve a resolved prompt for the current run.
   *
   * - `chat.prompt.set(resolved)` — store a `ResolvedPrompt` or plain string
   * - `chat.prompt()` — read the stored prompt (throws if not set)
   */
  prompt: Object.assign(getChatPrompt, { set: setChatPrompt }),
  /**
   * Store and retrieve resolved agent skills for the current run.
   *
   * - `chat.skills.set([...])` — store an array of `ResolvedSkill`s
   * - `chat.skills()` — read the stored skills (returns undefined if none)
   *
   * Skills set here are automatically injected into `streamText` by
   * `chat.toStreamTextOptions()`: skill descriptions land in the system
   * prompt and `loadSkill` / `readFile` / `bash` tools are added to the
   * tool set.
   */
  skills: Object.assign(getChatSkills, { set: setChatSkills }),
  /**
   * Returns an options object ready to spread into `streamText()`.
   * Reads the stored prompt and returns `{ system, experimental_telemetry, ...config }`.
   * Returns `{}` if no prompt has been set.
   */
  toStreamTextOptions,
  /**
   * Replace the accumulated conversation messages for compaction.
   * Call from `onTurnStart` or `onTurnComplete`. Takes `UIMessage[]` and
   * converts to `ModelMessage[]` internally.
   */
  setMessages: setChatMessages,
  /**
   * Imperative API for modifying the accumulated message history.
   * Supports rollback, remove, replace, slice, and full replacement.
   * Can be called from any hook or `run()`.
   */
  history: chatHistory,
  /** Check if it's safe to compact messages (no in-flight tool calls). */
  isCompactionSafe,
  /** Returns a `prepareStep` function that handles context compaction automatically. */
  compactionStep: chatCompactionStep,
  /** Low-level compaction for use inside a custom `prepareStep`. */
  compact: chatCompact,
  /** Read the current compaction state (summary + base message count). */
  getCompactionState,
  /**
   * The friendlyId (`session_*`) of the backing Session for the current chat.agent run.
   * Useful for persisting alongside `runId` so reloads can resume the same session.
   * Throws if called outside a chat.agent `run()` or hook.
   */
  get sessionId(): string {
    return getChatSession().id;
  },
};

/**
 * Writes a `turn-complete` control record to the chat output stream and,
 * if we have a prior turn-complete's seq_num, appends an S2 `trim` command
 * record back to it — keeping `session.out` bounded to roughly one turn
 * at steady state.
 *
 * The control record's body is empty; `trigger-control: turn-complete`
 * plus an optional `public-access-token` ride on the headers (see
 * `docs/ai-chat/client-protocol.mdx`). SDK transports filter it from the
 * consumer chunk stream and surface it via `onControl` / `onTurnComplete`.
 *
 * Trim is opportunistic and monotonic at S2's layer. A failed trim is
 * logged and swallowed; the next turn will retry against a fresher
 * target seq_num.
 *
 * @internal
 */
async function writeTurnCompleteChunk(
  _chatId?: string,
  publicAccessToken?: string
): Promise<StreamWriteResult> {
  const session = getChatSession();

  // 1. Write the turn-complete control record. The ack's `lastEventId` is
  //    this record's seq_num — that's the trim target for the NEXT turn.
  //
  //    Sibling headers:
  //    - `public-access-token` (optional): refresh token surfaced to
  //      browser-side transports via `onTurnComplete`.
  //    - `session-in-event-id` (optional): the committed-consume cursor
  //      on `.in` as of this turn-complete. On next worker boot, the
  //      boot scan reads this back and seeds the `.in` subscription so
  //      already-processed user messages aren't re-delivered.
  const extraHeaders: Array<[string, string]> = [];
  if (publicAccessToken) {
    extraHeaders.push(["public-access-token", publicAccessToken]);
  }
  const inCursor = session.in.lastDispatchedSeqNum();
  if (inCursor !== undefined) {
    extraHeaders.push([SESSION_IN_EVENT_ID_HEADER, String(inCursor)]);
  }
  const result = await session.out.writeControl(
    TRIGGER_CONTROL_SUBTYPE.TURN_COMPLETE,
    extraHeaders
  );
  const T_N = result.lastEventId ? Number.parseInt(result.lastEventId, 10) : undefined;

  // 2. Trim back to the previous turn-complete, if we have one. Skipping on
  //    first-turn-ever (or first turn post-OOM without a snapshot seed) is
  //    fine — the chain catches up next turn.
  const slot = locals.get(lastTurnCompleteSeqNumKey);
  const prev = slot?.value;
  if (slot && prev !== undefined) {
    try {
      await session.out.trimTo(prev);
    } catch (err) {
      logger.warn("chat.agent: trim failed; will retry next turn", {
        error: err instanceof Error ? err.message : String(err),
        prev,
      });
    }
  }

  // 3. Advance the slot so the next turn-complete trims back to this one.
  if (slot && T_N !== undefined && Number.isFinite(T_N)) {
    slot.value = T_N;
  }

  return result;
}

/**
 * Hand off the session to a fresh run on the latest version and emit a
 * telemetry chunk on `.out` so the transport can hide it from the
 * consumer.
 *
 * Server-side flow (in `POST /sessions/:id/end-and-continue`):
 *   1. Trigger a new run with the session's `triggerConfig`
 *   2. Atomically swap `Session.currentRunId` to the new run's id
 *      (via optimistic claim keyed on the calling run's id)
 *   3. Return the new runId
 *
 * The transport keeps its `.out` SSE open across the swap — v1's last
 * chunks land, v2's new chunks land on the same stream (S2 keys on
 * the session, not the run). The transport filters
 * `trigger:upgrade-required` for cleanliness; consumers see no gap.
 *
 * If the swap fails (no current run, no env auth, etc.) we still emit
 * the chunk and exit. The next `.in/append` will trigger a new run via
 * the probe path; it just won't be quite as seamless.
 *
 * @internal
 */
async function writeUpgradeRequiredChunk(): Promise<StreamWriteResult> {
  const ctx = taskContext.ctx;
  const chatId = ctx?.run.id ? getChatIdFromContext() : undefined;
  const callingRunId = ctx?.run.id;

  if (chatId && callingRunId) {
    const apiClient = apiClientManager.clientOrThrow();
    try {
      await apiClient.endAndContinueSession(chatId, {
        callingRunId,
        reason: "upgrade",
      });
    } catch (error) {
      // Non-fatal: the next `.in/append` re-triggers via the probe.
      // Swallow rather than throw so we still emit the chunk + exit.
      logger.warn("end-and-continue failed; falling back to probe-on-append", {
        chatId,
        callingRunId,
        error,
      });
    }
  }

  const session = getChatSession();
  return session.out.writeControl(TRIGGER_CONTROL_SUBTYPE.UPGRADE_REQUIRED);
}

/**
 * Resolves the current chat's `chatId` (used as session externalId) from
 * the bound session handle. Returns `undefined` if no agent is bound —
 * shouldn't happen at the call sites that invoke
 * `writeUpgradeRequiredChunk`, but defensive against misuse.
 * @internal
 */
function getChatIdFromContext(): string | undefined {
  return locals.get(chatSessionHandleKey)?.id;
}

/**
 * Extracts the text content of the last user message from a UIMessage array.
 * Returns undefined if no user message is found.
 * @internal
 */
function extractLastUserMessageText(messages: UIMessage[]): string | undefined {
  for (let i = messages.length - 1; i >= 0; i--) {
    const msg = messages[i]!;
    if (msg.role !== "user") continue;

    // UIMessage uses parts array
    if (msg.parts) {
      const textParts = msg.parts
        .filter((p: any) => p.type === "text" && p.text)
        .map((p: any) => p.text as string);
      if (textParts.length > 0) {
        return textParts.join("\n");
      }
    }

    break;
  }

  return undefined;
}

/**
 * Strips ephemeral OpenAI Responses API `itemId` from a UIMessage's parts.
 *
 * The OpenAI Responses provider attaches `itemId` to message parts via
 * `providerMetadata.openai.itemId`. These IDs are ephemeral — sending them
 * back in a subsequent `streamText` call causes 404s because the provider
 * can't find the referenced item (especially for stopped/partial responses).
 *
 * @internal
 */
function stripProviderMetadata(message: UIMessage): UIMessage {
  if (!message.parts) return message;
  return {
    ...message,
    parts: message.parts.map((part: any) => {
      const openai = part.providerMetadata?.openai;
      if (!openai?.itemId) return part;

      const { itemId, ...restOpenai } = openai;
      const { openai: _, ...restProviders } = part.providerMetadata;
      return {
        ...part,
        providerMetadata: {
          ...restProviders,
          ...(Object.keys(restOpenai).length > 0 ? { openai: restOpenai } : {}),
        },
      };
    }),
  };
}