@copilotkit/runtime/dist/v2/runtime/core/channel-manager.cjs.map

<img src="https://github.com/user-attachments/assets/0a6b64d9-e193-4940-a3f6-60334ac34084" alt="banner" style="border-radius: 12px; border: 2px solid #d6d4fa;" />

{"version":3,"file":"channel-manager.cjs","names":["AbstractAgent","EMPTY","EventType","deriveChannelActivationConfig","ChannelConfigError"],"sources":["../../../../src/v2/runtime/core/channel-manager.ts"],"sourcesContent":["import { randomUUID } from \"node:crypto\";\nimport {\n  ChannelConfigError,\n  deriveChannelActivationConfig,\n} from \"./channel-activation-config\";\nimport type { ChannelActivationConfig } from \"./channel-activation-config\";\nimport type { CopilotKitIntelligence } from \"../intelligence-platform\";\nimport { AbstractAgent, EventType } from \"@ag-ui/client\";\nimport type {\n  AgentSubscriber,\n  BaseEvent,\n  Message,\n  RunAgentParameters,\n  RunAgentResult,\n} from \"@ag-ui/client\";\nimport { EMPTY } from \"rxjs\";\nimport type { AgentRunner } from \"../runner/agent-runner\";\n// Type-only: @copilotkit/channels is pure-ESM, so a value import would break this\n// package's CJS output (see `core/runtime.ts` and `channel-activation-config.ts`\n// for the same constraint).\nimport type { Channel, ReplyContinuationOptions } from \"@copilotkit/channels\";\n\n/**\n * Lifecycle status of a single Channel activation, or of the manager overall.\n *\n * - `connecting`: activation in flight, not yet settled.\n * - `online`: activation resolved AND the managed session can currently send.\n *   A drop moves the Channel to `reconnecting` (not `online`); a successful\n *   rejoin restores `online`.\n * - `setup_required`: the Channel is declared but has no managed provider yet —\n *   a valid degraded state, not a failure.\n * - `reconnecting`: the managed session dropped and Phoenix is retrying — not\n *   currently sendable. The manager does NOT re-activate (reconnection is\n *   delegated to the Phoenix connection layer); it only reflects the health the\n *   session reports via its `onStateChange` observer.\n * - `stopped`: {@link ChannelManager.stop} has torn the Channel down.\n * - `error`: activation rejected with a non-setup error, or a previously-online\n *   control link gave up reconnecting after its bounded reconnect window.\n *\n * Both MANAGED (Intelligence-gateway) and DIRECT (developer-supplied adapter)\n * Channels move through these states. A direct Channel is driven by the manager\n * too — it is started via {@link Channel.ɵruntime}`.start()` and reaches `online`\n * once its own transport is up — but it is NOT wired into the Intelligence\n * gateway/canonical/reliability layer: it simply runs its own adapter transport,\n * started and stopped by the manager. A direct Channel has no managed-session\n * drop signal, so it never reports `reconnecting`.\n *\n * That gap is BY DESIGN, not a deferral. Run-correctness (canonical\n * cross-surface history, one outer run and one terminal outcome, durable\n * HITL-resume-across-restart, selection pinning) and the reliability layer are\n * Intelligence-side only — see OSS-599's boundary discipline. A direct Channel's\n * ceiling is the SDK's in-process run loop. Do not \"finish\" this by pulling the\n * canonical/reliability layer down into the SDK: that is explicitly the thing\n * OSS-599 forbids.\n */\nexport type ChannelStatus =\n  | \"connecting\"\n  | \"online\"\n  | \"setup_required\"\n  | \"reconnecting\"\n  | \"stopped\"\n  | \"error\";\n\n/**\n * The lifecycle control surface a Channel host uses to drive and observe\n * managed Channel activation.\n */\nexport interface ChannelsControl {\n  /**\n   * Resolve once every declared Channel has settled to a terminal, non-connecting\n   * state (`online` or `setup_required`). Rejects if any Channel is in `error`,\n   * or — when `timeoutMs` is given — if the whole set has not settled in time.\n   */\n  ready(opts?: { timeoutMs?: number }): Promise<void>;\n  /** Snapshot the overall status and the per-Channel status map. */\n  status(): { overall: ChannelStatus; channels: Record<string, ChannelStatus> };\n  /** Tear down every activated Channel. Idempotent. */\n  stop(): Promise<void>;\n}\n\n/**\n * Signals that a declared Channel cannot be activated because no managed\n * provider exists for it yet. The engine throws this (or any error whose\n * `code === \"SETUP_REQUIRED\"`) to move a Channel to `setup_required` rather\n * than `error` — a declared-but-unprovisioned Channel is a valid degraded\n * state, not a failure.\n */\nexport class ChannelSetupRequiredError extends Error {\n  constructor(message: string) {\n    super(message);\n    this.name = \"ChannelSetupRequiredError\";\n  }\n}\n\n/**\n * The activation engine: given a resolved {@link ChannelActivationConfig} and\n * the declared {@link Channel}, bring the Channel online and return its handle.\n * Injected in tests (a fake engine); defaults to the Realtime Gateway launcher.\n */\nexport type ActivateChannelEngine = (\n  config: ChannelActivationConfig,\n  channel: Channel,\n) => Promise<ChannelsHandle>;\n\n/**\n * Minimal structural view of the `@copilotkit/channels-intelligence`\n * `ChannelsHandle`. Declared locally (not imported) because the runtime is a\n * CJS package that must not take a static dependency on the pure-ESM\n * channels-intelligence package — the default engine reaches its launcher\n * through a dynamic `import()` instead. The manager only ever needs `stop()`.\n */\nexport interface ChannelsHandle {\n  /** Activation metadata declared to Intelligence. Unused by the manager. */\n  metadata: unknown;\n  /** Stop the underlying Channel(s) and release transports. */\n  stop(): Promise<void>;\n  /**\n   * Optional seam: register a callback the handle fires when its managed\n   * session drops. Retained as a per-episode drop breadcrumb; the manager drives\n   * status from {@link ChannelsHandle.onStateChange} instead. Present on the\n   * Realtime Gateway launcher handle; optional for non-gateway/test handles.\n   */\n  onClose?(cb: () => void): void;\n  /**\n   * Optional seam: register a connection-health observer the handle fires as its\n   * managed session moves between `online` (sendable), `reconnecting` (dropped,\n   * Phoenix retrying), and `gave_up` (dead after the bounded reconnect window).\n   * The manager uses this to keep {@link ChannelManager.status} honest — it does\n   * NOT re-activate on a drop (reconnection is delegated to the Phoenix\n   * connection layer; see {@link ChannelManager}). Optional so non-gateway or\n   * test handles that do not implement it are always invoked as\n   * `handle.onStateChange?.(cb)`.\n   */\n  onStateChange?(\n    cb: (\n      state: \"online\" | \"reconnecting\" | \"gave_up\",\n      detail?: { reason?: string; code?: string },\n    ) => void,\n  ): void;\n}\n\n/** Constructor arguments for {@link ChannelManager}. */\nexport interface ChannelManagerArgs {\n  /** The Intelligence runtime client the activation config is derived from. */\n  intelligence: CopilotKitIntelligence;\n  /** The declared framework Channels to activate. */\n  channels: Channel[];\n  /** Standard runtime AgentRunner used by managed Channel executions. */\n  runner?: AgentRunner;\n  /** Standard thread-lock TTL forwarded to Channel AgentRunner heartbeats. */\n  lockTtlSeconds?: number;\n  /** Standard thread-lock heartbeat cadence used by Channel AgentRunner calls. */\n  lockHeartbeatIntervalSeconds?: number;\n  /** Must match web Intelligence runs so channel + HTTP share the same lock key. */\n  lockKeyPrefix?: string;\n  /**\n   * Activation engine. Defaults to a wrapper over the channels-intelligence\n   * Realtime Gateway launcher (`startChannelsOverRealtimeGateway`), reached via\n   * dynamic import so this CJS package keeps no static ESM dependency.\n   */\n  activateChannel?: ActivateChannelEngine;\n  /** Mint a runtime instance id per Channel. Defaults to `rti_{uuid-no-dashes}`. */\n  mintRuntimeInstanceId?: () => string;\n  /** Diagnostic sink. Forwarded to the launcher/transport when the default\n   * activation engine is used, so transport-level drops surface in the managed\n   * path (not just activation-level events). */\n  log?: (msg: string, meta?: unknown) => void;\n  /**\n   * How often (ms) to repeat a \"still down\" log while a managed session is\n   * disconnected. A dropped session was previously silent for as long as the\n   * outage lasted, which made a dead bot indistinguishable from an idle one\n   * (OSS-670). Injectable so tests need no fake timers. Default 30000.\n   */\n  reconnectLogIntervalMs?: number;\n  /** Per-handle deadline (ms) for `handle.stop()` during {@link ChannelManager.stop}\n   * so a wedged stop can't hang SIGTERM shutdown. Default 5000. */\n  stopHandleTimeoutMs?: number;\n}\n\n/** Per-Channel mutable activation entry tracked by the manager. */\ninterface ChannelEntry {\n  status: ChannelStatus;\n  /** Resolves on `online`/`setup_required`; rejects on `error`. Awaited by `ready`. */\n  readonly settled: Promise<void>;\n  handle?: ChannelsHandle;\n  /**\n   * Whether {@link ChannelManager.stopEntry} has already stopped `handle`. Gates\n   * the single-stop guarantee: the success settle handler and `stop()` can both\n   * reach the same entry in the same tick, but the handle is torn down at most\n   * once.\n   */\n  handleStopped: boolean;\n  /** Epoch ms this outage episode began; unset while the session is healthy. */\n  downSince?: number;\n  /** Repeating \"still down\" logger for this outage; cleared on recovery/teardown. */\n  reconnectLogTimer?: ReturnType<typeof setInterval>;\n}\n\n/**\n * Runtime installs this pure-ESM package as a direct dependency, but the\n * specifier must stay non-literal so it never becomes a static dependency of\n * the runtime's CJS build. The packed-consumer contract is enforced by\n * `scripts/release/verify-runtime-package.ts`.\n */\nconst CHANNELS_INTELLIGENCE_SPECIFIER = \"@copilotkit/channels-intelligence\";\n\n/**\n * Structural view of the `@copilotkit/channels-intelligence` module surface the\n * default engine consumes. Declared locally (not imported) for the same\n * CJS/ESM-boundary reason the {@link ChannelsHandle} view is.\n */\nexport interface ChannelsIntelligenceModule {\n  startChannelsOverRealtimeGateway: (\n    channels: Channel[],\n    opts: {\n      wsUrl: string;\n      apiKey: string;\n      scope: { projectId: number; channelName: string };\n      runtimeInstanceId: string;\n      adapter?: string;\n      /** Optional per-Channel override for managed tool-call visibility. */\n      showToolStatus?: boolean;\n      /** Optional per-Channel tuning for continuation messages on long replies. */\n      replyContinuation?: ReplyContinuationOptions;\n      /** Intelligence app-api HTTP base URL, forwarded to the transport so the\n       * managed realtime path enables file/history parity (HTTP-only) — OSS-476. */\n      appApiBaseUrl?: string;\n      /** Diagnostic sink forwarded to the launcher/transport so transport-level\n       * drop diagnostics (e.g. a version-skew missing-leaseToken outage) are not\n       * silent in the managed path. */\n      log?: (msg: string, meta?: unknown) => void;\n      runCanonical(args: {\n        agent: AbstractAgent;\n        threadId: string;\n        runId: string;\n        userId: string;\n        agentId: string;\n        tools: readonly {\n          name: string;\n          description: string;\n          parameters: Record<string, unknown>;\n        }[];\n        context: readonly { description: string; value: string }[];\n        persistedInputMessages: Message[];\n        execute(\n          subscriber: AgentSubscriber,\n          canonicalRun?: { threadId: string; runId: string },\n        ): Promise<{\n          iterations: number;\n          interrupted: boolean;\n          deliveryError?: unknown;\n        }>;\n      }): Promise<{\n        iterations: number;\n        interrupted: boolean;\n        deliveryError?: unknown;\n      }>;\n      loadHistory(args: {\n        threadId: string;\n        appUserId: string;\n      }): Promise<Message[]>;\n    },\n  ) => Promise<ChannelsHandle>;\n}\n\n/**\n * Default engine: wrap the channels-intelligence Realtime Gateway launcher.\n *\n * The module is reached through an injectable importer that defaults to a\n * dynamic `import()` of a non-literal specifier, so the pure-ESM\n * `@copilotkit/channels-intelligence` never becomes a static dependency of this\n * CJS package (mirrors the runtime's other channels seams). The `import`\n * seam is a parameter purely so this function's config→opts mapping and its\n * module-not-found / generic-error branches are unit-testable WITHOUT the real\n * package installed; production always uses the default importer.\n *\n * Passes NO `org`/`channelId` — the launcher's realtime scope treats them as\n * optional.\n *\n * @param config - Resolved activation config for the Channel.\n * @param channel - The Channel to activate.\n * @param importChannelsIntelligence - Test seam; loads the channels-intelligence\n *   module. Defaults to a dynamic import of the real package.\n * @param log - Optional diagnostic sink forwarded to the launcher/transport so\n *   transport-level drop diagnostics are not silent in the managed path.\n * @returns The launcher's {@link ChannelsHandle}.\n */\nexport async function defaultActivateChannel(\n  config: ChannelActivationConfig,\n  channel: Channel,\n  importChannelsIntelligence: () => Promise<ChannelsIntelligenceModule> = () =>\n    import(\n      CHANNELS_INTELLIGENCE_SPECIFIER\n    ) as Promise<ChannelsIntelligenceModule>,\n  log?: (msg: string, meta?: unknown) => void,\n  services?: {\n    runner: AgentRunner;\n    intelligence: CopilotKitIntelligence;\n    lockTtlSeconds?: number;\n    lockHeartbeatIntervalSeconds?: number;\n    lockKeyPrefix?: string;\n  },\n): Promise<ChannelsHandle> {\n  let mod: ChannelsIntelligenceModule;\n  try {\n    mod = await importChannelsIntelligence();\n  } catch (err) {\n    if (isModuleNotFound(err)) {\n      throw new Error(\n        \"Managed Channels require '@copilotkit/channels-intelligence' to be installed. Add it to your app's dependencies.\",\n        { cause: err },\n      );\n    }\n    throw err;\n  }\n  if (!services) {\n    throw new Error(\n      \"Managed Channels require the runtime AgentRunner and Intelligence client\",\n    );\n  }\n  return mod.startChannelsOverRealtimeGateway([channel], {\n    wsUrl: config.wsUrl,\n    apiKey: config.apiKey,\n    scope: { projectId: config.projectId, channelName: config.channelName },\n    runtimeInstanceId: config.runtimeInstanceId,\n    adapter: config.adapter,\n    ...(config.showToolStatus !== undefined\n      ? { showToolStatus: config.showToolStatus }\n      : {}),\n    ...(config.replyContinuation !== undefined\n      ? { replyContinuation: config.replyContinuation }\n      : {}),\n    // Forward the app-api HTTP base URL so the transport wires file/history\n    // (HTTP-only) on the NORMAL managed path — without this, Channels started by\n    // the CopilotRuntime handler run with no history/file support (OSS-476).\n    appApiBaseUrl: config.apiUrl,\n    // Forward the manager's diagnostic sink down to the launcher/transport so a\n    // transport-level drop (e.g. a version-skew missing-leaseToken outage) is\n    // observable in the managed path, not just activation-level events.\n    ...(log ? { log } : {}),\n    runCanonical: (args) =>\n      runCanonicalChannelAgent(\n        services.runner,\n        services.intelligence,\n        services.lockTtlSeconds ?? 20,\n        services.lockHeartbeatIntervalSeconds ?? 15,\n        args,\n        services.lockKeyPrefix,\n      ),\n    loadHistory: async ({ threadId, appUserId }) => {\n      const history = await services.intelligence.getThreadMessages({\n        threadId,\n        userId: appUserId,\n      });\n      return Promise.all(\n        history.messages.map((message) =>\n          toAgentMessage(message, services.intelligence),\n        ),\n      );\n    },\n  });\n}\n\ninterface CanonicalRunArgs {\n  agent: AbstractAgent;\n  threadId: string;\n  runId: string;\n  userId: string;\n  agentId: string;\n  tools: readonly {\n    name: string;\n    description: string;\n    parameters: Record<string, unknown>;\n  }[];\n  context: readonly { description: string; value: string }[];\n  persistedInputMessages: Message[];\n  execute(\n    subscriber: AgentSubscriber,\n    canonicalRun?: { threadId: string; runId: string },\n  ): Promise<{\n    iterations: number;\n    interrupted: boolean;\n    deliveryError?: unknown;\n  }>;\n}\n\n/** One outer agent that lets the standard runner own the whole local tool loop. */\nclass ChannelOuterAgent extends AbstractAgent {\n  constructor(\n    private readonly inner: AbstractAgent,\n    private readonly canonicalThreadId: string,\n    private readonly executeLoop: CanonicalRunArgs[\"execute\"],\n  ) {\n    super({\n      threadId: inner.threadId,\n      initialMessages: inner.messages,\n      initialState: inner.state,\n      ...(inner.agentId ? { agentId: inner.agentId } : {}),\n    });\n  }\n\n  run(): ReturnType<AbstractAgent[\"run\"]> {\n    return EMPTY;\n  }\n\n  override async runAgent(\n    parameters?: RunAgentParameters,\n    subscriber?: AgentSubscriber,\n  ): Promise<RunAgentResult> {\n    if (!parameters?.runId) {\n      throw new Error(\"Canonical Channel run requires a runId\");\n    }\n    const result = await this.executeLoop(subscriber ?? {}, {\n      threadId: this.canonicalThreadId,\n      runId: parameters.runId,\n    });\n    return { result, newMessages: [] };\n  }\n\n  override abortRun(): void {\n    this.inner.abortRun();\n  }\n}\n\n/** Drive one public Channel run through the runtime's existing AgentRunner. */\nasync function runCanonicalChannelAgent(\n  runner: AgentRunner,\n  intelligence: CopilotKitIntelligence,\n  lockTtlSeconds: number,\n  lockHeartbeatIntervalSeconds: number,\n  args: CanonicalRunArgs,\n  lockKeyPrefix?: string,\n): Promise<{\n  iterations: number;\n  interrupted: boolean;\n  deliveryError?: unknown;\n}> {\n  const lock = await intelligence.ɵacquireThreadLock({\n    threadId: args.threadId,\n    runId: args.runId,\n    userId: args.userId,\n    agentId: args.agentId,\n    ttlSeconds: lockTtlSeconds,\n    ...(lockKeyPrefix !== undefined ? { lockKeyPrefix } : {}),\n  });\n  const canonicalThreadId = lock.threadId;\n  const canonicalRunId = lock.runId;\n  let result = { iterations: 0, interrupted: false };\n  const outer = new ChannelOuterAgent(\n    args.agent,\n    canonicalThreadId,\n    async (subscriber, canonicalRun) => {\n      result = await args.execute(subscriber, canonicalRun);\n      return result;\n    },\n  );\n  let stopPromise: Promise<boolean | undefined> | undefined;\n  let heartbeatError: unknown;\n  let heartbeatTimer: ReturnType<typeof setInterval> | undefined;\n  const stopCanonicalRun = (): void => {\n    stopPromise ??= Promise.resolve()\n      .then(() => runner.stop({ threadId: canonicalThreadId }))\n      .catch(() => false);\n  };\n  heartbeatTimer = setInterval(() => {\n    intelligence\n      .ɵrenewThreadLock({\n        threadId: canonicalThreadId,\n        runId: canonicalRunId,\n        ttlSeconds: lockTtlSeconds,\n        ...(lockKeyPrefix !== undefined ? { lockKeyPrefix } : {}),\n      })\n      .catch((error: unknown) => {\n        if (heartbeatTimer === undefined) return;\n        clearInterval(heartbeatTimer);\n        heartbeatTimer = undefined;\n        heartbeatError = error;\n        try {\n          args.agent.abortRun();\n        } catch {\n          // The runner stop below remains the authoritative cancellation path.\n        }\n        stopCanonicalRun();\n      });\n  }, lockHeartbeatIntervalSeconds * 1_000);\n  heartbeatTimer.unref?.();\n\n  try {\n    await new Promise<void>((resolve, reject) => {\n      let terminalError: (Error & { code?: string }) | undefined;\n      const stream = runner.run({\n        threadId: canonicalThreadId,\n        agent: outer,\n        input: {\n          threadId: canonicalThreadId,\n          runId: canonicalRunId,\n          messages: args.agent.messages,\n          state: args.agent.state,\n          tools: [...args.tools],\n          context: [...args.context],\n          forwardedProps: undefined,\n        },\n        persistedInputMessages: args.persistedInputMessages,\n      });\n      stream.subscribe({\n        next: (event: BaseEvent) => {\n          if (event.type !== EventType.RUN_ERROR || terminalError) return;\n          const message =\n            \"message\" in event && typeof event.message === \"string\"\n              ? event.message\n              : \"Canonical Channel agent run failed\";\n          terminalError = new Error(message);\n          terminalError.name = \"ChannelCanonicalRunError\";\n          if (\n            \"code\" in event &&\n            typeof event.code === \"string\" &&\n            event.code.length > 0\n          ) {\n            terminalError.code = event.code;\n          }\n        },\n        error: reject,\n        complete: () => {\n          if (terminalError) {\n            reject(terminalError);\n          } else {\n            resolve();\n          }\n        },\n      });\n    });\n  } finally {\n    if (heartbeatTimer !== undefined) {\n      clearInterval(heartbeatTimer);\n      heartbeatTimer = undefined;\n    }\n    // Always release the product thread lock from the Runtime side. Gateway\n    // may also release on terminal AG-UI ingestion; cleanup is idempotent and\n    // covers runner paths that never stream terminal events (or lose them).\n    await intelligence\n      .ɵcleanupThreadLock({\n        threadId: canonicalThreadId,\n        runId: canonicalRunId,\n      })\n      .catch(() => undefined);\n  }\n\n  if (heartbeatError !== undefined) {\n    await stopPromise;\n    throw heartbeatError;\n  }\n  return result;\n}\n\n/** Convert canonical Intelligence history into AG-UI messages. */\nasync function toAgentMessage(\n  message: {\n    id: string;\n    role: string;\n    activityType?: string;\n    content?: unknown;\n    toolCalls?: Array<{ id: string; name: string; args: string }>;\n    toolCallId?: string;\n  },\n  intelligence: CopilotKitIntelligence,\n): Promise<Message> {\n  const content = await hydrateManagedContent(message.content, intelligence);\n  return {\n    id: message.id,\n    role: message.role as Message[\"role\"],\n    content: content ?? \"\",\n    ...(message.activityType ? { activityType: message.activityType } : {}),\n    ...(message.toolCalls\n      ? {\n          toolCalls: message.toolCalls.map((call) => ({\n            id: call.id,\n            type: \"function\" as const,\n            function: { name: call.name, arguments: call.args },\n          })),\n        }\n      : {}),\n    ...(message.toolCallId ? { toolCallId: message.toolCallId } : {}),\n  } as Message;\n}\n\n/** Resolves managed asset references only at the authorized Runtime boundary. */\nasync function hydrateManagedContent(\n  content: unknown,\n  intelligence: CopilotKitIntelligence,\n): Promise<unknown> {\n  if (Array.isArray(content)) {\n    return Promise.all(\n      content.map(async (part) => {\n        if (\n          typeof part !== \"object\" ||\n          part === null ||\n          !(\"source\" in part) ||\n          typeof part.source !== \"object\" ||\n          part.source === null ||\n          !(\"value\" in part.source) ||\n          typeof part.source.value !== \"string\" ||\n          !part.source.value.startsWith(\"cpki-asset://\")\n        ) {\n          return part;\n        }\n        const assetId = part.source.value.slice(\"cpki-asset://\".length);\n        const asset = await intelligence.ɵgetManagedChannelAsset(assetId);\n        return {\n          ...part,\n          source: {\n            type: \"data\",\n            value: Buffer.from(asset.bytes).toString(\"base64\"),\n            mimeType:\n              asset.mimeType ??\n              (\"mimeType\" in part.source &&\n              typeof part.source.mimeType === \"string\"\n                ? part.source.mimeType\n                : \"application/octet-stream\"),\n          },\n        };\n      }),\n    );\n  }\n\n  if (\n    typeof content === \"object\" &&\n    content !== null &&\n    \"assetId\" in content &&\n    typeof content.assetId === \"string\"\n  ) {\n    const asset = await intelligence.ɵgetManagedChannelAsset(content.assetId);\n    return {\n      ...content,\n      source: {\n        type: \"data\",\n        value: Buffer.from(asset.bytes).toString(\"base64\"),\n        mimeType:\n          asset.mimeType ??\n          (\"mimeType\" in content && typeof content.mimeType === \"string\"\n            ? content.mimeType\n            : \"application/octet-stream\"),\n      },\n    };\n  }\n\n  return content;\n}\n\n/** Whether `err` signals a missing managed provider rather than a hard failure. */\nfunction isSetupRequired(err: unknown): boolean {\n  return (\n    err instanceof ChannelSetupRequiredError ||\n    (typeof err === \"object\" &&\n      err !== null &&\n      (err as { code?: unknown }).code === \"SETUP_REQUIRED\")\n  );\n}\n\n/**\n * Whether `err` is a Node/runtime module-resolution failure — i.e. the error\n * a dynamic `import()` throws when the target package is not installed.\n * Exported so the friendly-error path in {@link defaultActivateChannel} can be\n * unit-tested without forcing a real failing import.\n */\nexport function isModuleNotFound(err: unknown): boolean {\n  if (typeof err !== \"object\" || err === null) {\n    return false;\n  }\n  const code = (err as { code?: unknown }).code;\n  return code === \"ERR_MODULE_NOT_FOUND\" || code === \"MODULE_NOT_FOUND\";\n}\n\n/** Default deadline (ms) for a single `handle.stop()` during teardown. */\nconst DEFAULT_STOP_HANDLE_TIMEOUT_MS = 5_000;\n\n/** Default cadence (ms) for the \"still down\" log while a session is dropped. */\nconst DEFAULT_RECONNECT_LOG_INTERVAL_MS = 30_000;\n\n/**\n * Reject with `timeoutMessage` after `timeoutMs` if `inner` has not settled,\n * otherwise pass `inner` through. When `timeoutMs` is undefined, `inner` is\n * returned unchanged. The timer is `unref`'d so a pending deadline never keeps\n * the process alive, and `inner` always has a settle handler attached, so a\n * timed-out promise that later settles never surfaces as unhandled.\n */\nfunction withTimeout<T>(\n  inner: Promise<T>,\n  timeoutMs: number | undefined,\n  timeoutMessage: string,\n): Promise<T> {\n  if (timeoutMs === undefined) {\n    return inner;\n  }\n  return new Promise<T>((resolve, reject) => {\n    const timer = setTimeout(\n      () => reject(new Error(timeoutMessage)),\n      timeoutMs,\n    );\n    (timer as unknown as { unref?: () => void }).unref?.();\n    inner.then(\n      (value) => {\n        clearTimeout(timer);\n        resolve(value);\n      },\n      (err) => {\n        clearTimeout(timer);\n        reject(err);\n      },\n    );\n  });\n}\n\n/**\n * Drives Channel activation for an Intelligence runtime: lazily activates each\n * declared Channel, tracks per-Channel lifecycle status, exposes readiness, and\n * tears everything down. A MANAGED Channel (empty `adapters`) is activated\n * through the injected engine over the Intelligence gateway; a DIRECT Channel\n * (developer-supplied adapter) is started through its own transport seam\n * (`channel.ɵruntime.start()`) — driven by the manager, but running only its own\n * adapter transport, not the gateway path (direct Channels stay below the\n * canonical/reliability layer by design — see {@link ChannelStatus} and OSS-599).\n * Its very existence means Intelligence is configured, so there is no\n * standalone/self-started path.\n *\n * Activation is lazy and idempotent — constructing the manager does nothing;\n * {@link activate} starts it and a second call is a no-op. Activation throws\n * SYNCHRONOUSLY (a {@link ChannelConfigError}) only for a misconfiguration it\n * can detect up front — a duplicate or missing Channel name. Every OTHER\n * activation failure is recorded as the Channel's status (`error`, or\n * `setup_required` for a missing provider) and surfaced through {@link status}\n * and {@link ready} rather than thrown.\n *\n * Reconnection is NOT handled here — it is delegated to the Phoenix connection\n * layer that backs the launcher. When a managed control socket drops, Phoenix's\n * `Socket` reconnects and rejoins with the same Runtime declaration. Active\n * deliveries request fresh one-use join tokens through that control link. A\n * re-activation here would be both redundant AND broken: re-invoking the engine\n * on an already-started `Channel` throws in `channel.addAdapter` (started=true).\n * The manager therefore never re-activates on a drop.\n *\n * It DOES, however, reflect real connection health through the session's\n * `onStateChange` observer so {@link ChannelManager.status} stays honest rather\n * than reporting `online` forever after a drop: a drop moves the Channel to\n * `reconnecting`, a successful rejoin restores `online`, and a bounded give-up\n * (Phoenix would otherwise retry forever) moves it to `error`.\n */\nexport class ChannelManager implements ChannelsControl {\n  private readonly intelligence: CopilotKitIntelligence;\n  private readonly runner?: AgentRunner;\n  private readonly lockTtlSeconds: number;\n  private readonly lockHeartbeatIntervalSeconds: number;\n  private readonly lockKeyPrefix?: string;\n  private readonly channels: Channel[];\n  private readonly activateChannel: ActivateChannelEngine;\n  private readonly mintRuntimeInstanceId: () => string;\n  private readonly log?: (msg: string, meta?: unknown) => void;\n  private readonly stopHandleTimeoutMs: number;\n  private readonly reconnectLogIntervalMs: number;\n\n  private readonly entries = new Map<string, ChannelEntry>();\n  private activated = false;\n  private stopped = false;\n\n  /** @param args - See {@link ChannelManagerArgs}. */\n  constructor(args: ChannelManagerArgs) {\n    this.intelligence = args.intelligence;\n    this.runner = args.runner;\n    this.lockTtlSeconds = args.lockTtlSeconds ?? 20;\n    this.lockHeartbeatIntervalSeconds = args.lockHeartbeatIntervalSeconds ?? 15;\n    this.lockKeyPrefix = args.lockKeyPrefix;\n    this.channels = args.channels;\n    this.log = args.log;\n    // When using the default engine, forward the manager's log DOWN to the\n    // launcher/transport (via defaultActivateChannel's log param) so a\n    // transport-level drop is observable in the managed path. `this.log` is read\n    // lazily at activation time, so this closure always sees the assigned sink.\n    this.activateChannel =\n      args.activateChannel ??\n      ((config, channel) =>\n        defaultActivateChannel(\n          config,\n          channel,\n          undefined,\n          this.log,\n          this.runner\n            ? {\n                runner: this.runner,\n                intelligence: this.intelligence,\n                lockTtlSeconds: this.lockTtlSeconds,\n                lockHeartbeatIntervalSeconds: this.lockHeartbeatIntervalSeconds,\n                ...(this.lockKeyPrefix !== undefined\n                  ? { lockKeyPrefix: this.lockKeyPrefix }\n                  : {}),\n              }\n            : undefined,\n        ));\n    this.mintRuntimeInstanceId =\n      args.mintRuntimeInstanceId ??\n      (() => `rti_${randomUUID().replace(/-/g, \"\")}`);\n    this.stopHandleTimeoutMs =\n      args.stopHandleTimeoutMs ?? DEFAULT_STOP_HANDLE_TIMEOUT_MS;\n    this.reconnectLogIntervalMs =\n      args.reconnectLogIntervalMs ?? DEFAULT_RECONNECT_LOG_INTERVAL_MS;\n  }\n\n  /**\n   * Start activation of every declared Channel (lazy + idempotent). Mints a\n   * distinct runtime instance id per Channel, derives its activation config,\n   * and calls the engine. Records each Channel as `connecting`, transitioning\n   * to `online`/`setup_required`/`error` as its activation settles.\n   */\n  activate(): void {\n    // Short-circuit on BOTH latches: `activated` makes activation idempotent,\n    // and `stopped` prevents a post-`stop()` activate() from opening transports\n    // on a dead manager. (A late activation self-heals via the post-settle guard,\n    // but never starting it is cheaper and clearer.)\n    if (this.activated || this.stopped) {\n      return;\n    }\n    // Reject duplicate Channel names BEFORE kicking off any engine call. The\n    // manager keys `entries` by name, so a duplicate would let the second\n    // activation's entry silently overwrite the first — leaking the first\n    // Channel's control link out of status()/ready()/stop(). Fail loud here so\n    // nothing is ever activated in that state.\n    this.assertUniqueChannelNames();\n    this.activated = true;\n\n    // Partition declared Channels by transport. A Channel carrying ANY adapter\n    // that is NOT the Intelligence managed adapter (a developer-supplied\n    // slack/discord/... adapter, which lacks `__intelligenceChannel`) is a\n    // DIRECT channel. The manager still owns its lifecycle — but a direct Channel\n    // runs its OWN adapter transport rather than the Intelligence gateway path:\n    // it is started here via `channel.ɵruntime.start()` (not the managed engine)\n    // and torn down via `channel.ɵruntime.stop()`. The distinction is EXCLUSIVE\n    // PER CHANNEL, not per platform — a Channel served by a direct adapter is not\n    // also managed: ANY direct adapter makes the WHOLE Channel direct, regardless\n    // of platform. Attaching the managed adapter alongside a direct one would\n    // double-deliver every turn (and trip the SDK's `assertExclusive` guard,\n    // moving the Channel to `error`). Per the SoT rule, never infer managed intent\n    // from a local direct adapter — a managed-eligible Channel has an empty\n    // `adapters` at declaration time. Managed+direct coexistence on the same\n    // Channel is NOT supported today; it is deferred (OSS-484). Direct Channels\n    // run only their own transport here and stay below the Intelligence\n    // canonical/reliability layer by design, not pending work (OSS-599).\n    for (const channel of this.channels) {\n      const isDirect = channel.adapters.some((a) => !a.__intelligenceChannel);\n      if (isDirect) {\n        this.startDirectChannel(channel);\n        continue;\n      }\n      const name = channel.name!;\n      const runtimeInstanceId = this.mintRuntimeInstanceId();\n\n      let resolveSettled!: () => void;\n      let rejectSettled!: (err: unknown) => void;\n      const settled = new Promise<void>((resolve, reject) => {\n        resolveSettled = resolve;\n        rejectSettled = reject;\n      });\n      // ready() awaits `settled`; if nothing ever handles a rejection there,\n      // Node reports an unhandled rejection. Attach a no-op catch so the\n      // promise is always considered handled — ready() still sees the reason.\n      settled.catch(() => {});\n\n      // Invoke the engine synchronously so activation is observably started the\n      // moment activate() returns (callers assert the engine was called and see\n      // `connecting` before awaiting ready). A synchronous config/engine throw is\n      // turned into a rejected activation so it becomes this channel's status\n      // rather than throwing out of activate().\n      let activation: Promise<ChannelsHandle>;\n      let config: ChannelActivationConfig | undefined;\n      try {\n        config = deriveChannelActivationConfig({\n          intelligence: this.intelligence,\n          channel,\n          runtimeInstanceId,\n        });\n        activation = this.activateChannel(config, channel);\n      } catch (err) {\n        activation = Promise.reject(err);\n      }\n\n      // The deferred `.then` callbacks capture `entry` and run only after the\n      // literal has fully initialized, so referencing it here is safe.\n      const entry: ChannelEntry = {\n        status: \"connecting\",\n        handle: undefined,\n        handleStopped: false,\n        settled,\n      };\n\n      // Anchor the settle handlers. Both branches route every teardown through\n      // the idempotent `stopEntry`, so a late settle can never resurrect a\n      // `stopped` entry and a handle is torn down at most once. The handlers\n      // only mutate state (never throw), so the trailing no-op catch just keeps\n      // the chain from surfacing as an unhandled rejection.\n      activation\n        .then(\n          async (handle) => {\n            entry.handle = handle;\n            if (this.stopped) {\n              // stop() ran before this activation settled, so it could not tear\n              // down a handle that did not exist yet. Release it now (idempotent)\n              // and keep the Channel `stopped`.\n              await this.stopEntry(entry);\n              resolveSettled();\n              return;\n            }\n            entry.status = \"online\";\n            this.registerConnectionObserver(name, entry);\n            resolveSettled();\n          },\n          async (err: unknown) => {\n            if (this.stopped) {\n              // A rejection that arrives AFTER stop() must NOT resurrect the\n              // entry into `error`/`setup_required`: the Channel is already\n              // being torn down. Keep it `stopped` and resolve `settled` so a\n              // subsequent ready() does not reject on a stopped Channel.\n              await this.stopEntry(entry);\n              resolveSettled();\n              return;\n            }\n            if (isSetupRequired(err)) {\n              entry.status = \"setup_required\";\n              this.log?.(`channel \"${name}\" requires setup`, err);\n              resolveSettled();\n            } else {\n              entry.status = \"error\";\n              this.log?.(`channel \"${name}\" failed to activate`, err);\n              rejectSettled(err);\n            }\n          },\n        )\n        .catch(() => {});\n\n      this.entries.set(name, entry);\n    }\n  }\n\n  /**\n   * Start a DIRECT-adapter Channel through its own transport seam\n   * ({@link Channel.ɵruntime}`.start()`), recording a live entry so\n   * {@link ready}/{@link status}/{@link stop} all cover it. A direct Channel is\n   * driven by the manager — but only because the Intelligence runtime constructed\n   * this manager at all — and runs its OWN adapter transport, NOT the Intelligence\n   * gateway path. It is deliberately NOT wired into the gateway/canonical/\n   * reliability layer — that boundary is permanent, not a deferral (OSS-599).\n   *\n   * Mirrors the managed path's settle machinery so teardown resilience is shared:\n   * the entry's `handle` is a synthetic {@link ChannelsHandle} whose `stop()`\n   * calls `channel.ɵruntime.stop()`, so the SAME idempotent, bounded, resilient\n   * {@link stopEntry} that tears down a managed handle tears down a direct Channel\n   * too. The handle is assigned only AFTER `start()` resolves (exactly as the\n   * managed path assigns its handle only on resolve), so a `stop()` during a\n   * still-starting direct Channel returns promptly with nothing to stop and the\n   * post-settle guard tears down the late transport. A direct Channel exposes no\n   * managed-session drop signal, so no connection observer is wired and it never\n   * reaches `reconnecting`.\n   *\n   * @param channel - The direct-adapter Channel to start.\n   */\n  private startDirectChannel(channel: Channel): void {\n    const name = channel.name!;\n    this.log?.(\n      `channel \"${name}\" carries a direct adapter — starting its own transport via channel.ɵruntime.start() (the Intelligence runtime drives its lifecycle; it runs its own adapter transport, NOT the managed gateway path — direct Channels stay below the canonical/reliability layer by design (OSS-599); managed+direct coexistence deferred (OSS-484))`,\n    );\n\n    let resolveSettled!: () => void;\n    let rejectSettled!: (err: unknown) => void;\n    const settled = new Promise<void>((resolve, reject) => {\n      resolveSettled = resolve;\n      rejectSettled = reject;\n    });\n    // ready() awaits `settled`; attach a no-op catch so a rejection is always\n    // considered handled (ready() still sees the reason).\n    settled.catch(() => {});\n\n    // Synthetic handle wrapping the Channel's own stop seam. Assigned to the\n    // entry only on successful start (below) so the shared teardown machinery\n    // (handleStopped guard, withTimeout bounding, resilient allSettled) stops a\n    // direct Channel exactly as it stops a managed handle.\n    const directHandle: ChannelsHandle = {\n      metadata: {},\n      stop: () => channel.ɵruntime.stop(),\n    };\n\n    // Start the direct transport synchronously so it is observably started the\n    // moment activate() returns (callers see `connecting` before awaiting ready).\n    // A synchronous throw becomes this Channel's status rather than throwing out\n    // of activate().\n    let activation: Promise<void>;\n    try {\n      activation = channel.ɵruntime.start();\n    } catch (err) {\n      activation = Promise.reject(err);\n    }\n\n    const entry: ChannelEntry = {\n      status: \"connecting\",\n      handle: undefined,\n      handleStopped: false,\n      settled,\n    };\n\n    activation\n      .then(\n        async () => {\n          entry.handle = directHandle;\n          if (this.stopped) {\n            // stop() ran before start() settled, so it could not tear down a\n            // transport that was not up yet. Release it now (idempotent) and keep\n            // the Channel `stopped`.\n            await this.stopEntry(entry);\n            resolveSettled();\n            return;\n          }\n          entry.status = \"online\";\n          resolveSettled();\n        },\n        async (err: unknown) => {\n          if (this.stopped) {\n            // A start rejection that arrives AFTER stop() must NOT resurrect the\n            // entry into `error`: the Channel is already being torn down. Keep it\n            // `stopped` and resolve `settled` so a later ready() does not reject.\n            entry.handle = directHandle;\n            await this.stopEntry(entry);\n            resolveSettled();\n            return;\n          }\n          entry.status = \"error\";\n          this.log?.(\n            `channel \"${name}\" failed to start its direct transport`,\n            err,\n          );\n          rejectSettled(err);\n        },\n      )\n      .catch(() => {});\n\n    this.entries.set(name, entry);\n  }\n\n  /**\n   * Throw if two declared Channels share a `name`. `entries` is keyed by name,\n   * so a duplicate would overwrite the first Channel's entry and leak its live\n   * session. Called at the very start of {@link activate}, before any engine\n   * call, so a misconfiguration fails loud instead of silently.\n   *\n   * @throws {ChannelConfigError} If any Channel is missing a name, or if any\n   *   name appears more than once.\n   */\n  private assertUniqueChannelNames(): void {\n    const seen = new Set<string>();\n    for (const channel of this.channels) {\n      const name = channel.name;\n      // Check for a missing/empty name FIRST: `channel.name!` on a nameless\n      // Channel keys as the string \"undefined\", which would otherwise report a\n      // spurious duplicate for two nameless Channels before the accurate\n      // missing-name error. Fail with the precise error instead.\n      if (!name) {\n        throw new ChannelConfigError(\n          \"A managed Channel is missing a `name` — every declared Channel must \" +\n            \"have a unique, non-empty name (pass createChannel({ name })).\",\n        );\n      }\n      if (seen.has(name)) {\n        throw new ChannelConfigError(\n          `Duplicate managed Channel name \"${name}\" — every declared Channel ` +\n            `must have a unique name.`,\n        );\n      }\n      seen.add(name);\n    }\n  }\n\n  /**\n   * Resolve when every declared Channel — managed OR direct — has settled to\n   * `online`/`setup_required`.\n   *\n   * A direct-adapter Channel is awaited too: its `settled` resolves once its own\n   * transport is up (`channel.ɵruntime.start()` settling) and rejects if that\n   * start fails, exactly as a managed Channel's `settled` tracks its activation.\n   *\n   * Activates lazily if not already started — so a first call rejects with the\n   * same {@link ChannelConfigError} as the synchronous throw from\n   * {@link activate} for an up-front misconfiguration (duplicate/missing Channel\n   * names). Once activation has been kicked off, all OTHER failures are surfaced\n   * here instead: this rejects with an `AggregateError` if any Channel settled\n   * to `error` OR — when `timeoutMs` is given — did not settle in time. The\n   * `timeoutMs` deadline is applied PER CHANNEL, so the aggregate carries each\n   * failed Channel's real reason AND a named timeout for each Channel still\n   * hanging: a genuine activation error is never masked by a sibling that hangs\n   * (a pre-fix set-wide timeout discarded the real reason in that case).\n   *\n   * A STOPPED manager short-circuits and resolves: a Channel that settled to\n   * `error` BEFORE {@link stop} already rejected its `settled` promise, so\n   * awaiting it here would throw an `AggregateError` even though\n   * {@link status}.overall is `\"stopped\"` — inconsistent with the case where the\n   * Channel was still online at stop() (which resolves). A stopped manager has\n   * nothing left to be ready for, so resolve uniformly.\n   *\n   * `ready()` is ONE-SHOT: it settles on the INITIAL activation outcome. Later\n   * connection-health transitions (a live Channel dropping to `reconnecting`, or\n   * giving up to `error`) are reported through {@link status} — where `online`\n   * means currently-sendable — but do NOT re-arm or re-reject an already-settled\n   * `ready()`.\n   */\n  async ready(opts?: { timeoutMs?: number }): Promise<void> {\n    if (this.stopped) {\n      return;\n    }\n    this.activate();\n    const entries = [...this.entries.entries()];\n    // Apply `timeoutMs` PER CHANNEL rather than to the whole set. A single\n    // set-wide timeout wrapping `allSettled` would, when one channel settles to\n    // `error` while a sibling hangs, reject with only a generic timeout and\n    // DISCARD the erroring channel's real reason. Timing out each channel's\n    // `settled` independently lets `allSettled` collect BOTH a hung channel's\n    // named timeout AND a failed channel's real error into one AggregateError.\n    const results = await Promise.allSettled(\n      entries.map(([name, e]) =>\n        withTimeout(\n          e.settled,\n          opts?.timeoutMs,\n          `channel \"${name}\" did not settle within ${opts?.timeoutMs}ms`,\n        ),\n      ),\n    );\n    const errors = results\n      .filter((r): r is PromiseRejectedResult => r.status === \"rejected\")\n      .map((r) => r.reason);\n    if (errors.length > 0) {\n      throw new AggregateError(\n        errors,\n        `ChannelManager.ready: ${errors.length} channel(s) failed to activate or settle in time`,\n      );\n    }\n  }\n\n  /**\n   * Snapshot status. Every declared Channel — managed OR direct — appears keyed\n   * by name in `channels`; a direct-adapter Channel reads `online` once its own\n   * transport is up, exactly like a managed one.\n   *\n   * `overall` is folded over ALL declared Channels (see {@link computeOverall}),\n   * by precedence `error` > `reconnecting` > `setup_required` > `connecting` >\n   * `online`. `online` means every Channel can currently send. `reconnecting`\n   * outranks `setup_required` because a dropped-but-retrying Channel is an active\n   * outage, louder than a steadily-degraded unprovisioned one (only managed\n   * Channels ever reach `reconnecting`; a direct Channel has no managed-session\n   * drop signal). With no declared Channels at all, `overall` is `online` (nothing\n   * is degraded); once every Channel has been stopped, `overall` is `stopped`.\n   */\n  status(): {\n    overall: ChannelStatus;\n    channels: Record<string, ChannelStatus>;\n  } {\n    const channels: Record<string, ChannelStatus> = {};\n    for (const [name, entry] of this.entries) {\n      channels[name] = entry.status;\n    }\n    // A stopped manager is `stopped` regardless of whether it was ever activated.\n    // stop() before activate() (e.g. SIGTERM during startup) leaves `entries`\n    // empty, and the empty-set fold below returns `online` — a torn-down manager\n    // must never read healthy. Short-circuit before that fold. (After a normal\n    // activate→stop, every entry is already `stopped` and the fold agrees, so\n    // this is also consistent with the populated case.)\n    if (this.stopped) {\n      return { overall: \"stopped\", channels };\n    }\n    // Before activate() has run, `entries` is empty. Folding an empty set gives\n    // `online` — correct for a manager that declares NO channels (nothing is\n    // degraded), but a LIE for one that declares channels and simply has not\n    // opened its socket yet: activation is lazy (deferred to the first\n    // `ready()`), so a not-yet-activated manager must never read `online`.\n    // Report `connecting` (\"not started\") for that case so `status()` is honest\n    // before any `ready()`.\n    if (!this.activated && this.channels.length > 0) {\n      return { overall: \"connecting\", channels };\n    }\n    return { overall: this.computeOverall(Object.values(channels)), channels };\n  }\n\n  /**\n   * Fold per-Channel statuses into a single overall status (see {@link status}).\n   *\n   * Every declared Channel — managed OR direct — participates: a started direct\n   * Channel reads `online` and counts toward health exactly like a managed one,\n   * and its `error` is a real outage that must dominate. Statuses are ranked\n   * `error` > `reconnecting` > `setup_required` > `connecting` > `online`, so a\n   * genuine failure still dominates a healthy sibli