refactor(agents): make API registry ownership lifecycle-local (#111137)

* refactor(agents): make API registries lifecycle-owned * refactor(agents): keep registry runtime ownership internal * fix(agents): bind session streams to registry runtime * refactor(agents): keep prepared runtime ownership internal * test(agents): model lifecycle runtime fixtures * fix(amazon-bedrock): adapt lifecycle stream types * fix(agents): complete lifecycle runtime migration * test(agents): satisfy lifecycle registry static gates
2026-07-21 02:06:43 +00:00 · 2026-07-18 20:55:07 -07:00
parent b49c5463fe
commit cc57514e68
51 changed files with 900 additions and 223 deletions
@@ -145,7 +145,7 @@ d3102bd433eea32ca2a761030788fe18294aa34266f2fb2d01d74268cbf8fb54  module/inbound
 ee34dd840075bd687624ead228d3aa906fc52965e1708a5a335bb3ad3b671949  module/json-unsafe-integers
 8f37bca66178f4d77303fdd3ded0d445c9358b4f094c07a2f8e32b1721e953e4  module/keyed-async-queue
 d8ca27a737f235e09f1a9f8cd4b82ac0cde96f10c662347b6ac82e27b70b4a40  module/lazy-runtime
-15af70a301d18c36aef7dc3cd75350308f7a020738cc8fe1e118cfc90172729d  module/llm
+11fc218065ff47d3b2c0a2ca36ee66a1effac4e24e2981c9859f17b7b4cb82ae  module/llm
 6ae9b0f1f55ec50fa3e0ebc02052bfd04eca5f34c4a936692634b36a4cfa2035  module/lmstudio
 dd166527916dc5b9694512e3526e774da0b2427ba7fc8316d2fd799c06238e80  module/lmstudio-runtime
 fd00374a91ab5b393152bbbaf90010f9140cf15827bde360d0c6c8724ed34e2e  module/logging-core
@@ -307,6 +307,17 @@ describe("amazon-bedrock provider plugin", () => {
    vi.resetModules();
  });

+  it("publishes its stream through the provider lifecycle", async () => {
+    const provider = await registerSingleProviderPlugin(amazonBedrockPlugin);
+
+    expect(
+      provider.createStreamFn?.({ model: { api: "bedrock-converse-stream" } } as never),
+    ).toBeTypeOf("function");
+    expect(
+      provider.createStreamFn?.({ model: { api: "anthropic-messages" } } as never),
+    ).toBeUndefined();
+  });
+
  it("marks Claude 4.6 Bedrock models as adaptive by default", async () => {
    const provider = await registerSingleProviderPlugin(amazonBedrockPlugin);

@@ -5,7 +5,6 @@
 import type { BedrockClient } from "@aws-sdk/client-bedrock";
 import type { StreamFn } from "openclaw/plugin-sdk/agent-core";
 import type { OpenClawConfig } from "openclaw/plugin-sdk/config-contracts";
-import { registerApiProvider, streamSimple } from "openclaw/plugin-sdk/llm";
 import { resolvePluginConfigObject } from "openclaw/plugin-sdk/plugin-config-runtime";
 import type {
  OpenClawPluginApi,
@@ -25,7 +24,7 @@ import { supportsBedrockPromptCaching } from "./bedrock-options.js";
 import { loadBedrockControlPlaneSdk, runBedrockControlPlaneRequest } from "./control-plane.js";
 import { mergeImplicitBedrockProvider, resolveBedrockConfigApiKey } from "./discovery-shared.js";
 import { bedrockMemoryEmbeddingProviderAdapter } from "./memory-embedding-adapter.js";
-import { streamBedrock, streamSimpleBedrock } from "./stream.runtime.js";
+import { streamSimpleBedrock } from "./stream.runtime.js";
 import {
  isLatestAdaptiveBedrockModelRef,
  isOpus47OrNewerBedrockModelRef,
@@ -95,8 +94,16 @@ function isAnthropicBedrockModel(modelId: string): boolean {
  return false;
 }

+const bedrockStreamFn: StreamFn = (model, context, options) => {
+  if (model.api !== "bedrock-converse-stream") {
+    throw new Error(`Amazon Bedrock stream received unsupported API: ${model.api}`);
+  }
+  // The API check narrows the generic host model to the transport contract.
+  return streamSimpleBedrock(model as Parameters<typeof streamSimpleBedrock>[0], context, options);
+};
+
 function createBedrockNoCacheWrapper(baseStreamFn: StreamFn | undefined): StreamFn {
-  const underlying = baseStreamFn ?? streamSimple;
+  const underlying = baseStreamFn ?? bedrockStreamFn;
  return (model, context, options) =>
    underlying(model, context, {
      ...options,
@@ -382,15 +389,6 @@ export function registerAmazonBedrockPlugin(api: OpenClawPluginApi): void {
  });
  const startupPluginConfig = (api.pluginConfig ?? {}) as AmazonBedrockPluginConfig;

-  registerApiProvider(
-    {
-      api: "bedrock-converse-stream",
-      stream: streamBedrock,
-      streamSimple: streamSimpleBedrock,
-    },
-    `plugin:${providerId}`,
-  );
-
  function resolveCurrentPluginConfig(
    config: OpenClawConfig | undefined,
  ): AmazonBedrockPluginConfig | undefined {
@@ -543,6 +541,8 @@ export function registerAmazonBedrockPlugin(api: OpenClawPluginApi): void {
    },
    resolveConfigApiKey: ({ env }) => resolveBedrockConfigApiKey(env),
    normalizeResolvedModel: normalizeBedrockResolvedModel,
+    createStreamFn: ({ model }) =>
+      model.api === "bedrock-converse-stream" ? bedrockStreamFn : undefined,
    ...anthropicByModelReplayHooks,
    wrapStreamFn: ({ modelId, config, model, streamFn, thinkingLevel, extraParams }) => {
      const currentPluginConfig = resolveCurrentPluginConfig(config);
@@ -6,7 +6,8 @@ import {
 } from "@aws-sdk/client-bedrock-runtime";
 import { onLlmRequestActivity } from "openclaw/plugin-sdk/provider-stream-shared";
 import { afterEach, describe, expect, it, vi } from "vitest";
-import { streamBedrock, streamSimpleBedrock } from "./stream.runtime.js";
+import type { BedrockOptions } from "./bedrock-options.js";
+import { streamSimpleBedrock } from "./stream.runtime.js";
 import { streamTesting as testing } from "./test-support.js";

 function bedrockModel(overrides: Record<string, unknown>) {
@@ -52,9 +53,17 @@ async function* streamEvents(events: unknown[]) {
  }
 }

+function streamBedrockForTest(
+  model: Parameters<typeof streamSimpleBedrock>[0],
+  context: Parameters<typeof streamSimpleBedrock>[1],
+  options: BedrockOptions = {},
+) {
+  return streamSimpleBedrock(model, context, options as never);
+}
+
 async function captureClientRegion(
-  model: Parameters<typeof streamBedrock>[0],
-  options: Parameters<typeof streamBedrock>[2] = {},
+  model: Parameters<typeof streamSimpleBedrock>[0],
+  options: BedrockOptions = {},
 ): Promise<string> {
  const send = vi.spyOn(BedrockRuntimeClient.prototype, "send").mockResolvedValue({
    $metadata: { httpStatusCode: 200 },
@@ -64,7 +73,7 @@ async function captureClientRegion(
    ]),
  } as never);

-  await streamBedrock(
+  await streamBedrockForTest(
    model,
    { messages: [{ role: "user", content: "Hello", timestamp: 0 }] } as never,
    options,
@@ -275,6 +284,7 @@ describe("Bedrock profile endpoint resolution", () => {
  ])(
    "resolves $name to $expectedRegion",
    async ({ modelId, ambientRegion, fallbackRegion, explicitRegion, expectedRegion }) => {
+      vi.stubEnv("AWS_PROFILE", "");
      vi.stubEnv("AWS_REGION", ambientRegion);
      if (fallbackRegion !== undefined) {
        vi.stubEnv("AWS_DEFAULT_REGION", fallbackRegion);
@@ -305,7 +315,7 @@ describe("Bedrock stop reasons", () => {
      ]),
    } as never);

-    const result = await streamBedrock(bedrockModel({}), {
+    const result = await streamBedrockForTest(bedrockModel({}), {
      messages: [{ role: "user", content: "Hello", timestamp: 0 }],
    } as never).result();

@@ -579,7 +589,7 @@ describe("Bedrock Fable contract", () => {
      ]),
    } as never);

-    const stream = streamBedrock(fableModel(), context(), {
+    const stream = streamBedrockForTest(fableModel(), context(), {
      reasoning: "high",
      temperature: 0.2,
      toolChoice: "any",
@@ -614,7 +624,7 @@ describe("Bedrock Fable contract", () => {
      ]),
    } as never);

-    const stream = streamBedrock(fableModel(), context(), {
+    const stream = streamBedrockForTest(fableModel(), context(), {
      reasoning: "high",
      toolChoice: "none",
    });
@@ -126,7 +126,7 @@ function resolveAdaptiveBedrockMaxTokens(
 }

 /** Stream a Bedrock Converse request using Bedrock-specific options. */
-export const streamBedrock: StreamFunction<"bedrock-converse-stream", BedrockOptions> = (
+const streamBedrock: StreamFunction<"bedrock-converse-stream", BedrockOptions> = (
  model: Model<"bedrock-converse-stream">,
  context: Context,
  options: BedrockOptions = {},
@@ -400,7 +400,11 @@ function resolveSimpleBedrockOptions(
  model: Model<"bedrock-converse-stream">,
  options?: SimpleStreamOptions,
 ): BedrockOptions {
-  const base = buildBaseOptions(model, options, undefined);
+  const bedrockOptions = options as BedrockOptions | undefined;
+  const base = {
+    ...bedrockOptions,
+    ...buildBaseOptions(model, options, undefined),
+  };
  if (requiresMandatoryAdaptiveThinking(model)) {
    return {
      ...base,
@@ -0,0 +1,47 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { createApiRegistry } from "../api-registry.js";
+import { createLlmRuntime } from "../stream.js";
+
+const DEFAULT_RUNTIME_KEY = Symbol.for("openclaw.ai.defaultRuntime");
+const globalStore = globalThis as Record<PropertyKey, unknown>;
+const originalDefaultRuntime = globalStore[DEFAULT_RUNTIME_KEY];
+
+afterEach(() => {
+  if (originalDefaultRuntime === undefined) {
+    delete globalStore[DEFAULT_RUNTIME_KEY];
+  } else {
+    globalStore[DEFAULT_RUNTIME_KEY] = originalDefaultRuntime;
+  }
+  vi.resetModules();
+});
+
+describe("default LLM runtime compatibility state", () => {
+  it("keeps opaque legacy registrations out of lifecycle publications", async () => {
+    const registry = createApiRegistry();
+    const stream = () => ({}) as never;
+    registry.registerApiProvider(
+      { api: "test-legacy-plugin", stream, streamSimple: stream },
+      "plugin:test-legacy",
+    );
+    globalStore[DEFAULT_RUNTIME_KEY] = {
+      registry,
+      runtime: createLlmRuntime(registry),
+    };
+    vi.resetModules();
+
+    const runtime = await import("./default-runtime.js");
+
+    expect(runtime.defaultApiRegistry).toBe(registry);
+    expect(runtime.getPublishedApiProviders()).toEqual([]);
+
+    runtime.registerApiProvider(
+      { api: "test-current-plugin", stream, streamSimple: stream },
+      "plugin:test-current",
+    );
+
+    expect(runtime.getApiProvider("test-legacy-plugin")).toBeDefined();
+    expect(runtime.getPublishedApiProviders().map((provider) => provider.api)).toEqual([
+      "test-current-plugin",
+    ]);
+  });
+});
@@ -1,13 +1,20 @@
+import type { Api, StreamOptions } from "@openclaw/llm-core";
 // Process-default registry/runtime retained for the OpenClaw compatibility
 // facade (src/llm). Deliberately not part of the public package API: external
 // consumers create isolated runtimes via createLlmRuntime(); exporting these
 // from the root barrel would reintroduce the mutable process-global registry.
-import { createApiRegistry, type ApiRegistry } from "../api-registry.js";
+import {
+  createApiRegistry,
+  type ApiProvider,
+  type ApiRegistry,
+  type RegisteredApiProvider,
+} from "../api-registry.js";
 import { createLlmRuntime, type LlmRuntime } from "../stream.js";

 type DefaultRuntimeState = {
  registry: ApiRegistry;
  runtime: LlmRuntime;
+  publishedRegistry: ApiRegistry;
 };

 const DEFAULT_RUNTIME_KEY = Symbol.for("openclaw.ai.defaultRuntime");
@@ -15,11 +22,19 @@ const DEFAULT_RUNTIME_KEY = Symbol.for("openclaw.ai.defaultRuntime");
 function resolveDefaultRuntime(): DefaultRuntimeState {
  const globalStore = globalThis as Record<PropertyKey, unknown>;
  if (Object.hasOwn(globalStore, DEFAULT_RUNTIME_KEY)) {
-    return globalStore[DEFAULT_RUNTIME_KEY] as DefaultRuntimeState;
+    const existing = globalStore[DEFAULT_RUNTIME_KEY] as Omit<
+      DefaultRuntimeState,
+      "publishedRegistry"
+    > &
+      Partial<Pick<DefaultRuntimeState, "publishedRegistry">>;
+    // Keep the legacy facade registry intact. Its entries are opaque, so only
+    // registrations made through this host version enter lifecycle snapshots.
+    existing.publishedRegistry ??= createApiRegistry();
+    return existing as DefaultRuntimeState;
  }
  const registry = createApiRegistry();
  const runtime = createLlmRuntime(registry);
-  const state = { registry, runtime };
+  const state = { registry, runtime, publishedRegistry: createApiRegistry() };
  globalStore[DEFAULT_RUNTIME_KEY] = state;
  return state;
 }
@@ -29,12 +44,29 @@ const defaultRuntime = resolveDefaultRuntime();
 export const defaultApiRegistry = defaultRuntime.registry;
 export const defaultLlmRuntime = defaultRuntime.runtime;

-export const {
-  registerApiProvider,
-  getApiProvider,
-  getApiProviders,
-  unregisterApiProviders,
-  clearApiProviders,
-} = defaultApiRegistry;
+export function registerApiProvider<TApi extends Api, TOptions extends StreamOptions>(
+  provider: ApiProvider<TApi, TOptions>,
+  sourceId?: string,
+): void {
+  defaultApiRegistry.registerApiProvider(provider, sourceId);
+  defaultRuntime.publishedRegistry.registerApiProvider(provider, sourceId);
+}
+
+export const { getApiProvider, getApiProviders } = defaultApiRegistry;
+
+/** Returns only explicit compatibility registrations, excluding request-generated aliases. */
+export function getPublishedApiProviders(): RegisteredApiProvider[] {
+  return defaultRuntime.publishedRegistry.getApiProviders();
+}
+
+export function unregisterApiProviders(sourceId: string): void {
+  defaultApiRegistry.unregisterApiProviders(sourceId);
+  defaultRuntime.publishedRegistry.unregisterApiProviders(sourceId);
+}
+
+export function clearApiProviders(): void {
+  defaultApiRegistry.clearApiProviders();
+  defaultRuntime.publishedRegistry.clearApiProviders();
+}

 export const { stream, complete, streamSimple, completeSimple } = defaultLlmRuntime;
@@ -20,6 +20,7 @@ const buildSessionContextMock = vi.fn();
 const ensureOpenClawModelsJsonMock = vi.fn();
 const discoverAuthStorageMock = vi.fn();
 const discoverModelsMock = vi.fn();
+const getModelRegistryRuntimeMock = vi.fn();
 const resolveModelWithRegistryMock = vi.fn();
 const ensureAuthProfileStoreMock = vi.fn();
 const ensureAuthProfileStoreWithoutExternalProfilesMock = vi.fn();
@@ -82,6 +83,10 @@ vi.mock("./agent-model-discovery.js", () => ({
  discoverModels: (...args: unknown[]) => discoverModelsMock(...args),
 }));

+vi.mock("./sessions/model-registry-runtime.js", () => ({
+  getModelRegistryRuntime: (...args: unknown[]) => getModelRegistryRuntimeMock(...args),
+}));
+
 vi.mock("./embedded-agent-runner/model.js", () => ({
  resolveModelAsync: (...args: unknown[]) => resolveModelAsyncMock(...args),
  resolveModelWithRegistry: (...args: unknown[]) => resolveModelWithRegistryMock(...args),
@@ -532,6 +537,11 @@ describe("runBtwSideQuestion", () => {
    ensureOpenClawModelsJsonMock.mockReset();
    discoverAuthStorageMock.mockReset();
    discoverModelsMock.mockReset();
+    getModelRegistryRuntimeMock.mockReset();
+    getModelRegistryRuntimeMock.mockReturnValue({
+      apiRegistry: {},
+      llmRuntime: { streamSimple: streamSimpleMock },
+    });
    resolveModelAsyncMock.mockReset();
    resolveModelWithRegistryMock.mockReset();
    ensureAuthProfileStoreMock.mockReset();
@@ -11,7 +11,6 @@ import type { ChatType } from "../channels/chat-type.js";
 import type { SessionEntry as StoredSessionEntry } from "../config/sessions.js";
 import type { OpenClawConfig } from "../config/types.openclaw.js";
 import { streamWithPayloadPatch } from "../llm/providers/stream-wrappers/stream-payload-utils.js";
-import { streamSimple } from "../llm/stream.js";
 import type {
  AssistantMessageEvent,
  ImageContent,
@@ -81,6 +80,7 @@ import {
 import type { AgentRuntimeAuthPlan } from "./runtime-plan/types.js";
 import { resolveSessionRuntimeOverrideForProvider } from "./session-runtime-compat.js";
 import { stripToolResultDetails } from "./session-transcript-repair.js";
+import { getModelRegistryRuntime } from "./sessions/model-registry-runtime.js";
 import { resolveAgentTimeoutMs } from "./timeout.js";
 import { sanitizeImageBlocks } from "./tool-images.js";

@@ -1145,6 +1145,7 @@ export async function runBtwSideQuestion(
    }
  }
  runtimeModel = applySecretRefHeaderSentinels(runtimeModel, params.cfg);
+  const modelRegistryRuntime = getModelRegistryRuntime(modelRegistry);

  // Use the provider's own stream fn so providers like Ollama (which build
  // `/api/chat` or `/v1/chat/completions` paths based on api mode) construct
@@ -1156,9 +1157,11 @@ export async function runBtwSideQuestion(
    agentDir: params.agentDir,
    workspaceDir,
    env: process.env,
+    apiRegistry: modelRegistryRuntime.apiRegistry,
  });
  const streamFn = resolveEmbeddedAgentStreamFn({
-    currentStreamFn: streamSimple,
+    llmRuntime: modelRegistryRuntime.llmRuntime,
+    currentStreamFn: modelRegistryRuntime.llmRuntime.streamSimple,
    providerStreamFn,
    sessionId,
    signal: params.opts?.abortSignal,
@@ -1,19 +1,15 @@
-import {
-  clearApiProviders,
-  defaultApiRegistry,
-  getApiProvider,
-  registerApiProvider,
-  unregisterApiProviders,
-} from "@openclaw/ai/internal/runtime";
-import { registerBuiltInApiProviders, resetApiProviders } from "@openclaw/ai/providers";
+import { createApiRegistry, type ApiRegistry } from "@openclaw/ai";
+import { resetApiProviders } from "@openclaw/ai/providers";
 // Covers dynamic registration of custom model API providers.
-import { afterEach, describe, expect, it, vi } from "vitest";
+import { beforeEach, describe, expect, it, vi } from "vitest";
 import { createAssistantMessageEventStream } from "../llm/utils/event-stream.js";
 import { ensureCustomApiRegistered } from "./custom-api-registry.js";
 import { buildAssistantMessage, buildUsageWithNoCost } from "./stream-message-shared.js";

+let registry: ApiRegistry;
+
 function getRegisteredTestProvider() {
-  const provider = getApiProvider("test-custom-api");
+  const provider = registry.getApiProvider("test-custom-api");
  if (!provider) {
    throw new Error("expected test-custom-api provider to be registered");
  }
@@ -21,9 +17,8 @@ function getRegisteredTestProvider() {
 }

 describe("ensureCustomApiRegistered", () => {
-  afterEach(() => {
-    clearApiProviders();
-    registerBuiltInApiProviders(defaultApiRegistry);
+  beforeEach(() => {
+    registry = createApiRegistry();
  });

  it("registers a custom api provider once", () => {
@@ -31,8 +26,8 @@ describe("ensureCustomApiRegistered", () => {
    // replace provider entries or create duplicate sources.
    const streamFn = vi.fn(() => createAssistantMessageEventStream());

-    expect(ensureCustomApiRegistered("test-custom-api", streamFn)).toBe(true);
-    expect(ensureCustomApiRegistered("test-custom-api", streamFn)).toBe(false);
+    expect(ensureCustomApiRegistered(registry, "test-custom-api", streamFn)).toBe(true);
+    expect(ensureCustomApiRegistered(registry, "test-custom-api", streamFn)).toBe(false);

    const provider = getRegisteredTestProvider();
    expect(typeof provider.stream).toBe("function");
@@ -42,7 +37,7 @@ describe("ensureCustomApiRegistered", () => {
  it("delegates both stream entrypoints to the provided stream function", () => {
    const stream = createAssistantMessageEventStream();
    const streamFn = vi.fn(() => stream);
-    ensureCustomApiRegistered("test-custom-api", streamFn);
+    ensureCustomApiRegistered(registry, "test-custom-api", streamFn);

    const provider = getRegisteredTestProvider();

@@ -68,7 +63,7 @@ describe("ensureCustomApiRegistered", () => {
      stream.push({ type: "done", reason: "stop", message });
      return stream;
    });
-    ensureCustomApiRegistered("test-custom-api", streamFn);
+    ensureCustomApiRegistered(registry, "test-custom-api", streamFn);

    const provider = getRegisteredTestProvider();
    const stream = provider.stream(
@@ -85,7 +80,7 @@ describe("ensureCustomApiRegistered", () => {
    const streamFn = vi.fn(async () => {
      throw new Error("factory failed");
    });
-    ensureCustomApiRegistered("test-custom-api", streamFn);
+    ensureCustomApiRegistered(registry, "test-custom-api", streamFn);

    const provider = getRegisteredTestProvider();
    const stream = provider.stream(
@@ -107,7 +102,7 @@ describe("ensureCustomApiRegistered", () => {
    const api = "test-reset-plugin-api";
    const streamFn = vi.fn(() => createAssistantMessageEventStream());
    const streamSimpleFn = vi.fn(() => createAssistantMessageEventStream());
-    registerApiProvider(
+    registry.registerApiProvider(
      {
        api,
        stream: streamFn,
@@ -116,11 +111,11 @@ describe("ensureCustomApiRegistered", () => {
      sourceId,
    );

-    resetApiProviders(defaultApiRegistry);
+    resetApiProviders(registry);

-    expect(getApiProvider(api)).toBeDefined();
-    expect(getApiProvider("openai-responses")).toBeDefined();
+    expect(registry.getApiProvider(api)).toBeDefined();
+    expect(registry.getApiProvider("openai-responses")).toBeDefined();

-    unregisterApiProviders(sourceId);
+    registry.unregisterApiProviders(sourceId);
  });
 });
@@ -1,7 +1,7 @@
 /**
 * Registers caller-supplied custom API stream functions with the LLM registry.
 */
-import { getApiProvider, registerApiProvider } from "@openclaw/ai/internal/runtime";
+import type { ApiRegistry } from "@openclaw/ai";
 import type {
  Api,
  AssistantMessageEventStreamContract,
@@ -47,12 +47,16 @@ function adaptCustomStream(
 }

 /** Registers a custom API stream function when no provider already owns it. */
-export function ensureCustomApiRegistered(api: Api, streamFn: StreamFn): boolean {
-  if (getApiProvider(api)) {
+export function ensureCustomApiRegistered(
+  registry: ApiRegistry,
+  api: Api,
+  streamFn: StreamFn,
+): boolean {
+  if (registry.getApiProvider(api)) {
    return false;
  }

-  registerApiProvider(
+  registry.registerApiProvider(
    {
      api,
      stream: (model, context, options) =>
@@ -208,6 +208,10 @@ export const buildEmbeddedSystemPromptMock = vi.fn(() => "");
 export const resolveEmbeddedAgentStreamFnMock: Mock<
  (params?: unknown) => MockEmbeddedAgentStreamFn
 > = vi.fn((_params?: unknown) => vi.fn());
+const getModelRegistryRuntimeMock = vi.fn(() => ({
+  apiRegistry: {},
+  llmRuntime: { streamSimple: vi.fn() },
+}));
 export const getApiKeyForModelMock: Mock<
  (params?: { profileId?: string; allowAuthProfileFallback?: boolean }) => Promise<{
    apiKey: string;
@@ -457,6 +461,11 @@ export function resetCompactSessionStateMocks(): void {
  }));
  resolveEmbeddedAgentStreamFnMock.mockReset();
  resolveEmbeddedAgentStreamFnMock.mockImplementation((_params?: unknown) => vi.fn());
+  getModelRegistryRuntimeMock.mockReset();
+  getModelRegistryRuntimeMock.mockReturnValue({
+    apiRegistry: {},
+    llmRuntime: { streamSimple: vi.fn() },
+  });
  getApiKeyForModelMock.mockReset();
  getApiKeyForModelMock.mockImplementation(async (params?: { profileId?: string }) => ({
    apiKey: "test",
@@ -677,6 +686,10 @@ export async function loadCompactHooksHarness(): Promise<{
    registerProviderStreamForModel: registerProviderStreamForModelMock,
  }));

+  vi.doMock("../sessions/model-registry-runtime.js", () => ({
+    getModelRegistryRuntime: getModelRegistryRuntimeMock,
+  }));
+
  vi.doMock("../../hooks/internal-hooks.js", async () => {
    const actual = await vi.importActual<typeof import("../../hooks/internal-hooks.js")>(
      "../../hooks/internal-hooks.js",
@@ -856,6 +856,7 @@ describe("compactEmbeddedAgentSessionDirect hooks", () => {

    await compactTesting.prepareCompactionSessionAgent({
      session: session as never,
+      llmRuntime: { streamSimple: vi.fn() } as never,
      providerStreamFn: vi.fn(),
      sessionId: "session-1",
      signal: new AbortController().signal,
@@ -921,6 +922,7 @@ describe("compactEmbeddedAgentSessionDirect hooks", () => {
        agent: { streamFn: vi.fn() },
        messages: [{ role: "user", content: "hello" }],
      } as never,
+      llmRuntime: { streamSimple: vi.fn() } as never,
      providerStreamFn: vi.fn(),
      sessionId: "session-1",
      signal: new AbortController().signal,
@@ -2297,6 +2299,7 @@ describe("compactEmbeddedAgentSessionDirect hooks", () => {
      config: undefined,
      agentDir: "/tmp",
      effectiveWorkspace: "/tmp",
+      apiRegistry: {} as never,
    });

    expect(result).toBe(streamFn);
@@ -3,6 +3,7 @@
 */
 import fs from "node:fs/promises";
 import os from "node:os";
+import type { ApiRegistry } from "@openclaw/ai";
 import { isAcpRuntimeSpawnAvailable } from "../../acp/runtime/availability.js";
 import type { ThinkLevel } from "../../auto-reply/thinking.js";
 import { resolveAgentModelFallbackValues } from "../../config/model-input.js";
@@ -158,6 +159,7 @@ import {
  resolveSessionWriteLockOptions,
 } from "../session-write-lock.js";
 import { createAgentSession, estimateTokens, SessionManager } from "../sessions/index.js";
+import { getModelRegistryRuntime } from "../sessions/model-registry-runtime.js";
 import { detectRuntimeShell } from "../shell-utils.js";
 import {
  filterProviderNormalizableTools,
@@ -257,12 +259,14 @@ function resolveCompactionProviderStream(params: {
  config?: OpenClawConfig;
  agentDir: string;
  effectiveWorkspace: string;
+  apiRegistry: ApiRegistry;
 }) {
  return registerProviderStreamForModel({
    model: params.effectiveModel,
    cfg: params.config,
    agentDir: params.agentDir,
    workspaceDir: params.effectiveWorkspace,
+    apiRegistry: params.apiRegistry,
  });
 }

@@ -1499,6 +1503,7 @@ async function compactEmbeddedAgentSessionDirectOnce(
        config: params.config,
        agentDir,
        effectiveWorkspace,
+        apiRegistry: getModelRegistryRuntime(modelRegistry).apiRegistry,
      });
      while (true) {
        // Rebuild the compaction session on retry so provider wrappers, payload
@@ -1527,6 +1532,7 @@ async function compactEmbeddedAgentSessionDirectOnce(
          // through the same transport/payload shaping stack as normal turns.
          await prepareCompactionSessionAgent({
            session,
+            llmRuntime: getModelRegistryRuntime(modelRegistry).llmRuntime,
            providerStreamFn,
            sessionId: params.sessionId,
            signal: runAbortController.signal,
@@ -1,3 +1,4 @@
+import type { LlmRuntime } from "@openclaw/ai";
 import type { ThinkLevel } from "../../auto-reply/thinking.js";
 import type { OpenClawConfig } from "../../config/types.openclaw.js";
 import type { ProviderRuntimeModel } from "../../plugins/provider-runtime-model.types.js";
@@ -14,6 +15,7 @@ import { mapThinkingLevelForProvider } from "./utils.js";

 export async function prepareCompactionSessionAgent(params: {
  session: { agent: { streamFn?: unknown } };
+  llmRuntime: LlmRuntime;
  providerStreamFn: unknown;
  sessionId: string;
  signal: AbortSignal;
@@ -58,6 +60,7 @@ export async function prepareCompactionSessionAgent(params: {
      })
    : params.resolvedApiKey;
  params.session.agent.streamFn = resolveEmbeddedAgentStreamFn({
+    llmRuntime: params.llmRuntime,
    currentStreamFn: resolveEmbeddedAgentBaseStreamFn({ session: params.session as never }),
    providerStreamFn: params.providerStreamFn as never,
    sessionId: params.sessionId,
@@ -1,7 +1,6 @@
 // Coverage for cache-retention defaults and overrides in extra params.
 import type { StreamFn } from "openclaw/plugin-sdk/agent-core";
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
-import { createLlmStreamSimpleMock } from "../../../test/helpers/agents/llm-stream-simple-mock.js";
 import { applyExtraParamsToAgent } from "./extra-params.js";
 import { testing as extraParamsTesting } from "./extra-params.test-support.js";
 import { log } from "./logger.js";
@@ -16,7 +15,7 @@ function applyAndExpectWrapped(params: {
 }) {
  // Wrapping is the observable signal that cache-retention handling was enabled
  // without requiring a real provider stream call.
-  const agent: { streamFn?: StreamFn } = {};
+  const agent: { streamFn?: StreamFn } = { streamFn: vi.fn() as StreamFn };

  applyExtraParamsToAgent(
    agent,
@@ -43,8 +42,6 @@ vi.mock("./logger.js", () => ({
  },
 }));

-vi.mock("../../llm/stream.js", () => createLlmStreamSimpleMock());
-
 beforeEach(() => {
  vi.mocked(log.warn).mockClear();
  extraParamsTesting.setProviderRuntimeDepsForTest({
@@ -20,7 +20,6 @@ import {
 } from "../../llm/providers/stream-wrappers/openai.js";
 import { createOpenRouterSystemCacheWrapper } from "../../llm/providers/stream-wrappers/proxy.js";
 import { streamWithPayloadPatch } from "../../llm/providers/stream-wrappers/stream-payload-utils.js";
-import { streamSimple } from "../../llm/stream.js";
 import type { SimpleStreamOptions } from "../../llm/types.js";
 import {
  createDeepSeekV4OpenAICompatibleThinkingWrapper,
@@ -43,6 +42,13 @@ import type { StreamFn } from "../runtime/index.js";
 import type { SettingsManager } from "../sessions/index.js";
 import { log } from "./logger.js";
 import { parseCacheRetention, resolveCacheRetention } from "./prompt-cache-retention.js";
+
+function requireBaseStreamFn(streamFn: StreamFn | undefined): StreamFn {
+  if (!streamFn) {
+    throw new Error("Cannot apply stream policy without a lifecycle-owned base stream.");
+  }
+  return streamFn;
+}
 import type { ProviderThinkLevel } from "./utils.js";

 const defaultProviderRuntimeDeps = {
@@ -558,7 +564,7 @@ function createStreamFnWithExtraParams(
    log.debug(`creating streamFn wrapper with params: ${JSON.stringify(debugParams)}`);
  }

-  const underlying = baseStreamFn ?? streamSimple;
+  const underlying = requireBaseStreamFn(baseStreamFn);
  const wrappedStreamFn: StreamFn = (callModel, context, options) => {
    const cacheRetention = resolveCacheRetention(
      extraParams,
@@ -661,7 +667,7 @@ function createParallelToolCallsWrapper(
  baseStreamFn: StreamFn | undefined,
  enabled: boolean,
 ): StreamFn {
-  const underlying = baseStreamFn ?? streamSimple;
+  const underlying = requireBaseStreamFn(baseStreamFn);
  return (model, context, options) => {
    if (!supportsGptParallelToolCallsPayload(model.api)) {
      return underlying(model, context, options);
@@ -695,7 +701,7 @@ function shouldStripOpenAICompletionsStore(model: ProviderRuntimeModel): boolean
 }

 function createOpenAICompletionsStoreCompatWrapper(baseStreamFn: StreamFn | undefined): StreamFn {
-  const underlying = baseStreamFn ?? streamSimple;
+  const underlying = requireBaseStreamFn(baseStreamFn);
  return (model, context, options) => {
    if (!shouldStripOpenAICompletionsStore(model as ProviderRuntimeModel)) {
      return underlying(model, context, options);
@@ -751,7 +757,7 @@ function createOpenAICompletionsChatTemplateKwargsWrapper(params: {
  baseStreamFn: StreamFn | undefined;
  configured: Record<string, unknown>;
 }): StreamFn {
-  const underlying = params.baseStreamFn ?? streamSimple;
+  const underlying = requireBaseStreamFn(params.baseStreamFn);
  return (model, context, options) => {
    if (model.api !== "openai-completions") {
      return underlying(model, context, options);
@@ -774,7 +780,7 @@ function createOpenAICompletionsExtraBodyWrapper(
  baseStreamFn: StreamFn | undefined,
  extraBody: Record<string, unknown>,
 ): StreamFn {
-  const underlying = baseStreamFn ?? streamSimple;
+  const underlying = requireBaseStreamFn(baseStreamFn);
  return (model, context, options) => {
    if (model.api !== "openai-completions") {
      return underlying(model, context, options);
@@ -18,6 +18,7 @@ import type {
  IngestResult,
 } from "../../../context-engine/types.js";
 import { formatErrorMessage } from "../../../infra/errors.js";
+import { bindStreamLlmRuntime } from "../../../llm/model-runtime-binding.js";
 import type { Model } from "../../../llm/types.js";
 import type { PluginMetadataSnapshot } from "../../../plugins/plugin-metadata-snapshot.js";
 import { createLazyPromise } from "../../../shared/lazy-runtime.js";
@@ -27,6 +28,10 @@ import type {
  MessagingToolSourceReplyPayload,
 } from "../../embedded-agent-messaging.types.js";
 import type { AgentMessage } from "../../runtime/index.js";
+import {
+  getModelRegistryRuntime,
+  initializeModelRegistryRuntime,
+} from "../../sessions/model-registry-runtime.js";
 import type { WorkspaceBootstrapFile } from "../../workspace.js";

 type SubscribeEmbeddedAgentSessionFn =
@@ -1342,14 +1347,21 @@ export async function createContextEngineAttemptRunner(params: {
    .mockReset()
    .mockReturnValue({ messages: params.sessionMessagesAfterRepair ?? seedMessages });

-  hoisted.createAgentSessionMock.mockImplementation(async () => ({
-    session:
+  const modelRegistry = {};
+  initializeModelRegistryRuntime(modelRegistry);
+  const modelRuntime = getModelRegistryRuntime(modelRegistry).llmRuntime;
+  hoisted.createAgentSessionMock.mockImplementation(async () => {
+    const session =
      params.createSession?.() ??
      createDefaultEmbeddedSession({
        initialMessages: seedMessages,
        prompt: params.sessionPrompt,
-      }),
-  }));
+      });
+    if (session.agent.streamFn) {
+      bindStreamLlmRuntime(session.agent.streamFn, modelRuntime);
+    }
+    return { session };
+  });

  const previousTrajectoryEnv = process.env.OPENCLAW_TRAJECTORY;
  const previousTrajectoryDirEnv = process.env.OPENCLAW_TRAJECTORY_DIR;
@@ -1378,7 +1390,7 @@ export async function createContextEngineAttemptRunner(params: {
      model: testModel,
      authStorage: testAuthStorage as never,
      authProfileStore: { version: 1, profiles: {} },
-      modelRegistry: {} as never,
+      modelRegistry: modelRegistry as never,
      thinkLevel: "off",
      disableTools: true,
      disableMessageTool: true,
@@ -5,6 +5,8 @@ import { streamSimple } from "../../../llm/stream.js";
 vi.mock("../context-engine-capabilities.js", () => ({
  resolveContextEngineCapabilities: async () => ({ llm: undefined }),
 }));
+import type { LlmRuntime } from "@openclaw/ai";
+import { defaultLlmRuntime } from "@openclaw/ai/internal/runtime";
 import { SYSTEM_PROMPT_CACHE_BOUNDARY } from "@openclaw/ai/internal/shared";
 import type { OpenClawConfig } from "../../../config/config.js";
 import { addSession } from "../../bash-process-registry.js";
@@ -15,7 +17,7 @@ import { buildAgentSystemPrompt } from "../../system-prompt.js";
 import type { NormalizedUsage } from "../../usage.js";
 import {
  resolveEmbeddedAgentBaseStreamFn,
-  resolveEmbeddedAgentStreamFn,
+  resolveEmbeddedAgentStreamFn as resolveEmbeddedAgentStreamFnImpl,
 } from "../stream-resolution.js";
 import { buildContextEnginePromptCacheInfo } from "./attempt.context-engine-helpers.js";
 import {
@@ -36,6 +38,17 @@ import {
 } from "./attempt.tool-call-normalization.js";
 import { buildEmbeddedAttemptToolRunContext } from "./attempt.tool-run-context.js";

+const llmRuntime = {
+  ...defaultLlmRuntime,
+  streamSimple,
+} as LlmRuntime;
+
+function resolveEmbeddedAgentStreamFn(
+  params: Omit<Parameters<typeof resolveEmbeddedAgentStreamFnImpl>[0], "llmRuntime">,
+) {
+  return resolveEmbeddedAgentStreamFnImpl({ ...params, llmRuntime });
+}
+
 type FakeWrappedStream = {
  result: () => Promise<unknown>;
  [Symbol.asyncIterator]: () => AsyncIterator<unknown>;
@@ -1,16 +1,18 @@
-import { getApiProvider } from "@openclaw/ai/internal/runtime";
+import type { LlmRuntime } from "@openclaw/ai";
+import { defaultLlmRuntime, getApiProvider } from "@openclaw/ai/internal/runtime";
 import { SYSTEM_PROMPT_CACHE_BOUNDARY } from "@openclaw/ai/internal/shared";
 // Stream resolution tests cover how embedded runs choose provider, boundary,
 // native Codex, or custom stream functions and pass auth/cache/signal options.
 import type { StreamFn } from "openclaw/plugin-sdk/agent-core";
 import { afterEach, describe, expect, it, vi } from "vitest";
+import { bindStreamLlmRuntime } from "../../llm/model-runtime-binding.js";
 import { streamSimple } from "../../llm/stream.js";
 import { mintSecretSentinel } from "../../secrets/sentinel.js";
 import * as providerTransportStream from "../provider-transport-stream.js";
 import {
-  describeEmbeddedAgentStreamStrategy,
+  describeEmbeddedAgentStreamStrategy as describeEmbeddedAgentStreamStrategyImpl,
  resolveEmbeddedAgentApiKey,
-  resolveEmbeddedAgentStreamFn,
+  resolveEmbeddedAgentStreamFn as resolveEmbeddedAgentStreamFnImpl,
 } from "./stream-resolution.js";

 const streamMocks = vi.hoisted(() => ({
@@ -38,6 +40,23 @@ vi.mock("../provider-transport-stream.js", async (importOriginal) => {
  };
 });

+const llmRuntime = {
+  ...defaultLlmRuntime,
+  streamSimple: streamSimple as StreamFn,
+} as LlmRuntime;
+
+function describeEmbeddedAgentStreamStrategy(
+  params: Omit<Parameters<typeof describeEmbeddedAgentStreamStrategyImpl>[0], "llmRuntime">,
+) {
+  return describeEmbeddedAgentStreamStrategyImpl({ ...params, llmRuntime });
+}
+
+function resolveEmbeddedAgentStreamFn(
+  params: Omit<Parameters<typeof resolveEmbeddedAgentStreamFnImpl>[0], "llmRuntime">,
+) {
+  return resolveEmbeddedAgentStreamFnImpl({ ...params, llmRuntime });
+}
+
 const overrideBoundaryAwareStreamFnOnce = (streamFn: StreamFn): void => {
  // Boundary wrapping remains real by default; individual cases replace only
  // the inner stream when they need to inspect forwarded options.
@@ -75,6 +94,21 @@ afterEach(() => {
 });

 describe("describeEmbeddedAgentStreamStrategy", () => {
+  it("recovers the lifecycle owner from a prepared session stream", () => {
+    bindStreamLlmRuntime(streamSimple, llmRuntime);
+
+    expect(
+      describeEmbeddedAgentStreamStrategyImpl({
+        currentStreamFn: streamSimple,
+        model: {
+          api: "openai-responses",
+          provider: "openai",
+          id: "gpt-5.4",
+        } as never,
+      }),
+    ).toBe("boundary-aware:openai-responses");
+  });
+
  it("describes provider-owned stream paths explicitly", () => {
    expect(
      describeEmbeddedAgentStreamStrategy({
@@ -1,9 +1,9 @@
 /**
 * Resolves provider stream functions and API keys for embedded agents.
 */
-import { getApiProvider } from "@openclaw/ai/internal/runtime";
+import type { LlmRuntime } from "@openclaw/ai";
 import { stripSystemPromptCacheBoundary } from "@openclaw/ai/internal/shared";
-import { streamSimple } from "../../llm/stream.js";
+import { getStreamLlmRuntime } from "../../llm/model-runtime-binding.js";
 import { createAnthropicVertexStreamFnForModel } from "../anthropic-vertex-stream.js";
 import { createBoundaryAwareStreamFnForModel } from "../provider-transport-stream.js";
 import type { StreamFn } from "../runtime/index.js";
@@ -18,28 +18,53 @@ type EmbeddedStreamOptions = Parameters<StreamFn>[2] & {

 export function resolveEmbeddedAgentBaseStreamFn(params: {
  session: { agent: { streamFn?: StreamFn } };
-}): StreamFn | undefined {
+}): StreamFn {
  const cached = embeddedAgentBaseStreamFnCache.get(params.session);
  if (cached !== undefined || embeddedAgentBaseStreamFnCache.has(params.session)) {
+    if (!cached) {
+      throw new Error("Agent session has no lifecycle-owned base stream.");
+    }
    return cached;
  }
  const baseStreamFn = params.session.agent.streamFn;
  embeddedAgentBaseStreamFnCache.set(params.session, baseStreamFn);
+  if (!baseStreamFn) {
+    throw new Error("Agent session has no lifecycle-owned base stream.");
+  }
  return baseStreamFn;
 }

+type EmbeddedStreamRuntimeOwner =
+  | {
+      llmRuntime: LlmRuntime;
+      currentStreamFn: StreamFn | undefined;
+    }
+  | {
+      llmRuntime?: never;
+      currentStreamFn: StreamFn;
+    };
+
+function resolveEmbeddedStreamRuntime(owner: EmbeddedStreamRuntimeOwner): LlmRuntime {
+  const runtime = owner.llmRuntime ?? getStreamLlmRuntime(owner.currentStreamFn);
+  if (!runtime) {
+    throw new Error("Embedded stream has no lifecycle runtime owner.");
+  }
+  return runtime;
+}
+
 function isDefaultOpenClawStreamFnForModel(
  model: EmbeddedRunAttemptParams["model"],
  streamFn: StreamFn | undefined,
+  llmRuntime: LlmRuntime,
 ): boolean {
-  if (!streamFn || streamFn === streamSimple) {
+  if (!streamFn || streamFn === llmRuntime.streamSimple) {
    return true;
  }
  const api = typeof model.api === "string" ? model.api.trim() : "";
  if (!api) {
    return false;
  }
-  const provider = getApiProvider(api as never);
+  const provider = llmRuntime.registry.getApiProvider(api as never);
  return streamFn === provider?.streamSimple || streamFn === provider?.stream;
 }

@@ -54,22 +79,25 @@ function isOpenAICodexResponsesModel(model: EmbeddedRunAttemptParams["model"]):
 function resolveOpenClawNativeCodexResponsesStreamFn(params: {
  model: EmbeddedRunAttemptParams["model"];
  currentStreamFn: StreamFn | undefined;
+  llmRuntime: LlmRuntime;
 }): StreamFn | undefined {
  if (!isOpenAICodexResponsesModel(params.model)) {
    return undefined;
  }
-  if (!isDefaultOpenClawStreamFnForModel(params.model, params.currentStreamFn)) {
+  if (!isDefaultOpenClawStreamFnForModel(params.model, params.currentStreamFn, params.llmRuntime)) {
    return undefined;
  }
-  return params.currentStreamFn ?? streamSimple;
+  return params.currentStreamFn ?? params.llmRuntime.streamSimple;
 }

-export function describeEmbeddedAgentStreamStrategy(params: {
-  currentStreamFn: StreamFn | undefined;
-  providerStreamFn?: StreamFn;
-  model: EmbeddedRunAttemptParams["model"];
-  resolvedApiKey?: string;
-}): string {
+export function describeEmbeddedAgentStreamStrategy(
+  params: EmbeddedStreamRuntimeOwner & {
+    providerStreamFn?: StreamFn;
+    model: EmbeddedRunAttemptParams["model"];
+    resolvedApiKey?: string;
+  },
+): string {
+  const llmRuntime = resolveEmbeddedStreamRuntime(params);
  if (params.providerStreamFn) {
    return "provider";
  }
@@ -80,11 +108,12 @@ export function describeEmbeddedAgentStreamStrategy(params: {
    resolveOpenClawNativeCodexResponsesStreamFn({
      model: params.model,
      currentStreamFn: params.currentStreamFn,
+      llmRuntime,
    })
  ) {
    return "openclaw-native-codex-responses";
  }
-  if (isDefaultOpenClawStreamFnForModel(params.model, params.currentStreamFn)) {
+  if (isDefaultOpenClawStreamFnForModel(params.model, params.currentStreamFn, llmRuntime)) {
    return createBoundaryAwareStreamFnForModel(params.model)
      ? `boundary-aware:${params.model.api}`
      : "stream-simple";
@@ -110,18 +139,20 @@ export async function resolveEmbeddedAgentApiKey(params: {
  return params.authStorage ? await params.authStorage.getApiKey(params.provider) : undefined;
 }

-export function resolveEmbeddedAgentStreamFn(params: {
-  currentStreamFn: StreamFn | undefined;
-  providerStreamFn?: StreamFn;
-  sessionId: string;
-  promptCacheKey?: string;
-  signal?: AbortSignal;
-  model: EmbeddedRunAttemptParams["model"];
-  resolvedApiKey?: string;
-  transportAuthAvailable?: boolean;
-  authProfileId?: string;
-  authStorage?: { getApiKey(provider: string): Promise<string | undefined> };
-}): StreamFn {
+export function resolveEmbeddedAgentStreamFn(
+  params: EmbeddedStreamRuntimeOwner & {
+    providerStreamFn?: StreamFn;
+    sessionId: string;
+    promptCacheKey?: string;
+    signal?: AbortSignal;
+    model: EmbeddedRunAttemptParams["model"];
+    resolvedApiKey?: string;
+    transportAuthAvailable?: boolean;
+    authProfileId?: string;
+    authStorage?: { getApiKey(provider: string): Promise<string | undefined> };
+  },
+): StreamFn {
+  const llmRuntime = resolveEmbeddedStreamRuntime(params);
  if (params.providerStreamFn) {
    return wrapEmbeddedAgentStreamFn(params.providerStreamFn, {
      runSignal: params.signal,
@@ -140,7 +171,7 @@ export function resolveEmbeddedAgentStreamFn(params: {
    });
  }

-  const currentStreamFn = params.currentStreamFn ?? streamSimple;
+  const currentStreamFn = params.currentStreamFn ?? llmRuntime.streamSimple;
  if (params.model.provider === "anthropic-vertex") {
    return createAnthropicVertexStreamFnForModel(params.model);
  }
@@ -148,6 +179,7 @@ export function resolveEmbeddedAgentStreamFn(params: {
  const openClawNativeCodexResponsesStreamFn = resolveOpenClawNativeCodexResponsesStreamFn({
    model: params.model,
    currentStreamFn: params.currentStreamFn,
+    llmRuntime,
  });
  if (openClawNativeCodexResponsesStreamFn) {
    return wrapEmbeddedAgentStreamFn(openClawNativeCodexResponsesStreamFn, {
@@ -169,7 +201,7 @@ export function resolveEmbeddedAgentStreamFn(params: {
  }

  if (
-    isDefaultOpenClawStreamFnForModel(params.model, params.currentStreamFn) ||
+    isDefaultOpenClawStreamFnForModel(params.model, params.currentStreamFn, llmRuntime) ||
    hasResolvedRuntimeApiKey(params.resolvedApiKey) ||
    params.transportAuthAvailable ||
    // Proxied anthropic-messages providers (provider !== "anthropic", e.g. pioneer)
@@ -1,7 +1,6 @@
 /**
 * Decodes HTML-entity escaped tool-call arguments in stream wrappers.
 */
-import { streamSimple } from "../../llm/stream.js";
 import { decodeHtmlEntities } from "../../shared/html-entities.js";
 import { visitObjectContentBlocks } from "../../shared/message-content-blocks.js";
 import type { StreamFn } from "../runtime/index.js";
@@ -91,12 +90,9 @@ function wrapStreamMessageObjects(
 }

 /** Wraps a stream function so tool-call arguments are decoded before consumers inspect them. */
-export function createHtmlEntityToolCallArgumentDecodingWrapper(
-  baseStreamFn: StreamFn | undefined,
-): StreamFn {
-  const underlying = baseStreamFn ?? streamSimple;
+export function createHtmlEntityToolCallArgumentDecodingWrapper(baseStreamFn: StreamFn): StreamFn {
  return (model, context, options) => {
-    const maybeStream = underlying(model, context, options);
+    const maybeStream = baseStreamFn(model, context, options);
    if (maybeStream && typeof maybeStream === "object" && "then" in maybeStream) {
      return Promise.resolve(maybeStream).then((stream) =>
        wrapStreamMessageObjects(stream, decodeToolCallArgumentsHtmlEntitiesInMessage),
@@ -1,3 +1,4 @@
+import type { ApiRegistry } from "@openclaw/ai";
 // Verifies the Google simple-completion wrapper and thinking-payload sanitizer hook.
 import { beforeEach, describe, expect, it, vi } from "vitest";
 import type { Model } from "../llm/types.js";
@@ -5,6 +6,9 @@ import type { Model } from "../llm/types.js";
 const streamSimple = vi.fn();
 const sanitizeGoogleThinkingPayload = vi.fn();
 const ensureCustomApiRegistered = vi.fn();
+const apiRegistry = {
+  getApiProvider: vi.fn(() => ({ streamSimple })),
+} as unknown as ApiRegistry;

 vi.mock("../llm/stream.js", () => ({
  streamSimple,
@@ -72,7 +76,7 @@ describe("prepareGoogleSimpleCompletionModel", () => {
      api: "openai-responses",
    } as unknown as Model<"openai-responses">;

-    const result = prepareGoogleSimpleCompletionModel(model);
+    const result = prepareGoogleSimpleCompletionModel(apiRegistry, model);

    expect(result).toBe(model);
    expect(ensureCustomApiRegistered).not.toHaveBeenCalled();
@@ -81,22 +85,23 @@ describe("prepareGoogleSimpleCompletionModel", () => {
  it("registers an OpenClaw-owned Google simple-completion api alias", () => {
    const model = makeGoogleModel();

-    const result = prepareGoogleSimpleCompletionModel(model);
+    const result = prepareGoogleSimpleCompletionModel(apiRegistry, model);

    expect(result).toEqual({
      ...model,
      api: GOOGLE_SIMPLE_COMPLETION_API,
    });
    expect(ensureCustomApiRegistered).toHaveBeenCalledTimes(1);
-    expect(ensureCustomApiRegistered.mock.calls[0]?.[0]).toBe(GOOGLE_SIMPLE_COMPLETION_API);
+    expect(ensureCustomApiRegistered.mock.calls[0]?.[0]).toBe(apiRegistry);
+    expect(ensureCustomApiRegistered.mock.calls[0]?.[1]).toBe(GOOGLE_SIMPLE_COMPLETION_API);
  });

  it.each(["off", "low", "medium", "high", "adaptive"] as const)(
    "sanitizes outbound thinking payload for gemini-flash-latest with reasoning=%s",
    async (reasoning) => {
      const model = makeGoogleModel();
-      const wrapped = prepareGoogleSimpleCompletionModel(model);
-      const streamFn = ensureCustomApiRegistered.mock.calls[0]?.[1] as (
+      const wrapped = prepareGoogleSimpleCompletionModel(apiRegistry, model);
+      const streamFn = ensureCustomApiRegistered.mock.calls[0]?.[2] as (
        ...args: unknown[]
      ) => unknown;

@@ -130,8 +135,8 @@ describe("prepareGoogleSimpleCompletionModel", () => {
      payload.generationConfig.thinkingConfig.thinkingLevel = "MINIMAL";
    });
    const model = makeGoogleModel();
-    prepareGoogleSimpleCompletionModel(model);
-    const streamFn = ensureCustomApiRegistered.mock.calls[0]?.[1] as (
+    prepareGoogleSimpleCompletionModel(apiRegistry, model);
+    const streamFn = ensureCustomApiRegistered.mock.calls[0]?.[2] as (
      ...args: unknown[]
    ) => unknown;

@@ -181,8 +186,8 @@ describe("prepareGoogleSimpleCompletionModel", () => {
          max: null,
        },
      });
-      const wrapped = prepareGoogleSimpleCompletionModel(model);
-      const streamFn = ensureCustomApiRegistered.mock.calls[0]?.[1] as (
+      const wrapped = prepareGoogleSimpleCompletionModel(apiRegistry, model);
+      const streamFn = ensureCustomApiRegistered.mock.calls[0]?.[2] as (
        ...args: unknown[]
      ) => unknown;

@@ -1,3 +1,4 @@
+import type { ApiRegistry } from "@openclaw/ai";
 /**
 * Google simple-completion stream adapter.
 *
@@ -5,7 +6,6 @@
 * backend but sanitizes unsupported thinking payload options for simple models.
 */
 import { clampThinkingLevel } from "@openclaw/ai/internal/runtime";
-import { streamSimple } from "../llm/stream.js";
 import type { Api, Model, ModelThinkingLevel } from "../llm/types.js";
 import {
  sanitizeGoogleThinkingPayload,
@@ -40,11 +40,15 @@ function resolveGoogleSimpleThinkingLevel(
  }
 }

-function buildGoogleSimpleCompletionStreamFn(): StreamFn {
+function buildGoogleSimpleCompletionStreamFn(registry: ApiRegistry): StreamFn {
  return (model, context, options) => {
    const googleModel: Model = { ...model, api: SOURCE_API };
+    const sourceProvider = registry.getApiProvider(SOURCE_API);
+    if (!sourceProvider) {
+      throw new Error(`No API provider registered for api: ${SOURCE_API}`);
+    }
    return streamWithPayloadPatch(
-      streamSimple as unknown as StreamFn,
+      sourceProvider.streamSimple as StreamFn,
      googleModel,
      context,
      options,
@@ -63,10 +67,17 @@ function buildGoogleSimpleCompletionStreamFn(): StreamFn {
 }

 /** Rewrites Google generative-ai models to the simple-completion adapter when needed. */
-export function prepareGoogleSimpleCompletionModel<TApi extends Api>(model: Model<TApi>): Model {
+export function prepareGoogleSimpleCompletionModel<TApi extends Api>(
+  registry: ApiRegistry,
+  model: Model<TApi>,
+): Model {
  if (model.api !== SOURCE_API) {
    return model;
  }
-  ensureCustomApiRegistered(GOOGLE_SIMPLE_COMPLETION_API, buildGoogleSimpleCompletionStreamFn());
+  ensureCustomApiRegistered(
+    registry,
+    GOOGLE_SIMPLE_COMPLETION_API,
+    buildGoogleSimpleCompletionStreamFn(registry),
+  );
  return { ...model, api: GOOGLE_SIMPLE_COMPLETION_API };
 }
@@ -1,5 +1,7 @@
+import { createLlmRuntime, type LlmRuntime } from "@openclaw/ai";
 import type { OpenAICompletionsOptions } from "@openclaw/ai/internal/openai";
 import { getEnvApiKey } from "@openclaw/ai/internal/runtime";
+import { registerBuiltInApiProviders } from "@openclaw/ai/providers";
 import { normalizeProviderId } from "@openclaw/model-catalog-core/provider-id";
 import {
  asDateTimestampMs,
@@ -20,7 +22,7 @@ import { formatErrorMessage } from "../infra/errors.js";
 * Scans remote provider model catalogs for configured providers.
 */
 import { readResponseWithLimit } from "../infra/http-body.js";
-import { complete } from "../llm/stream.js";
+import "../llm/ai-transport-host.js";
 import type { Context, Model, Tool } from "../llm/types.js";
 import { inferParamBFromIdOrName } from "../shared/model-param-b.js";

@@ -299,6 +301,7 @@ async function probeTool(
  model: OpenAIModel,
  apiKey: string,
  timeoutMs: number,
+  complete: LlmRuntime["complete"],
 ): Promise<ProbeResult> {
  const context: Context = {
    messages: [
@@ -345,6 +348,7 @@ async function probeImage(
  model: OpenAIModel,
  apiKey: string,
  timeoutMs: number,
+  complete: LlmRuntime["complete"],
 ): Promise<ProbeResult> {
  const context: Context = {
    messages: [
@@ -433,6 +437,8 @@ export async function scanOpenRouterModels(
  const providerFilter = normalizeProviderId(options.providerFilter ?? "");

  const catalog = await fetchOpenRouterModels(fetchImpl, timeoutMs);
+  const llmRuntime = createLlmRuntime();
+  registerBuiltInApiProviders(llmRuntime.registry);
  const now = Date.now();

  const filtered = catalog.filter((entry) => {
@@ -504,9 +510,9 @@ export async function scanOpenRouterModels(
          reasoning: baseModel.reasoning,
        };

-        const toolResult = await probeTool(model, apiKey, timeoutMs);
+        const toolResult = await probeTool(model, apiKey, timeoutMs, llmRuntime.complete);
        const imageResult = model.input?.includes("image")
-          ? await probeImage(ensureImageInput(model), apiKey, timeoutMs)
+          ? await probeImage(ensureImageInput(model), apiKey, timeoutMs, llmRuntime.complete)
          : { ok: false, latencyMs: null, skipped: true };

        result = buildOpenRouterScanResult({
@@ -1,5 +1,6 @@
 // Live-sweeps discovered model profiles with optional provider/model filters and probes.
 import { writeSync } from "node:fs";
+import { defaultApiRegistry } from "@openclaw/ai/internal/runtime";
 import { normalizeProviderId } from "@openclaw/model-catalog-core/provider-id";
 import { expectDefined } from "@openclaw/normalization-core";
 import { type Api, completeSimple, type Model } from "openclaw/plugin-sdk/llm";
@@ -347,6 +348,7 @@ async function ensureLiveProviderApisRegistered(params: {
  const providerConfig = params.config.models?.providers?.ollama;
  const providerBaseUrl = readConfiguredOllamaBaseUrl(providerConfig) || OLLAMA_DEFAULT_BASE_URL;
  ensureCustomApiRegistered(
+    defaultApiRegistry,
    "ollama",
    createLiveOllamaRuntimeStreamFn({
      createConfiguredOllamaStreamFn,
@@ -1436,6 +1438,7 @@ async function completeSimpleWithTimeout<TApi extends Api>(
  });
  try {
    const completionModel = prepareModelForSimpleCompletion({
+      apiRegistry: defaultApiRegistry,
      model,
      cfg: activeLiveCompletionConfig,
    });
@@ -0,0 +1,39 @@
+import { createApiRegistry, createLlmRuntime } from "@openclaw/ai";
+import { describe, expect, it, vi } from "vitest";
+import { bindModelLlmRuntime } from "../llm/model-runtime-binding.js";
+import { createAssistantMessageEventStream } from "../llm/utils/event-stream.js";
+import { registerProviderStreamForModel } from "./provider-stream.js";
+
+const { providerStream } = vi.hoisted(() => ({
+  providerStream: vi.fn(),
+}));
+
+vi.mock("../plugins/provider-runtime.js", () => ({
+  resolveProviderStreamFn: () => providerStream,
+}));
+
+describe("provider stream lifecycle registration", () => {
+  it("registers provider streams into the prepared model runtime", () => {
+    providerStream.mockReturnValue(createAssistantMessageEventStream());
+    const apiRegistry = createApiRegistry();
+    const llmRuntime = createLlmRuntime(apiRegistry);
+    const model = bindModelLlmRuntime(
+      {
+        api: "test-lifecycle-provider",
+        provider: "test-provider",
+        id: "test-model",
+        name: "Test Model",
+        baseUrl: "https://example.test",
+        reasoning: false,
+        input: ["text"],
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow: 1024,
+        maxTokens: 512,
+      },
+      llmRuntime,
+    );
+
+    expect(registerProviderStreamForModel({ model })).toBeTypeOf("function");
+    expect(apiRegistry.getApiProvider("test-lifecycle-provider")).toBeDefined();
+  });
+});
@@ -3,7 +3,9 @@
 * Resolves plugin-owned or transport-aware stream functions and registers the
 * model API once a concrete stream implementation exists.
 */
+import type { ApiRegistry } from "@openclaw/ai";
 import type { OpenClawConfig } from "../config/types.openclaw.js";
+import { getModelLlmRuntime } from "../llm/model-runtime-binding.js";
 import type { Api, Model } from "../llm/types.js";
 import { resolveProviderStreamFn } from "../plugins/provider-runtime.js";
 import { ensureCustomApiRegistered } from "./custom-api-registry.js";
@@ -23,7 +25,7 @@ export function registerProviderStreamForModel<TApi extends Api>(params: {
  workspaceDir?: string;
  env?: NodeJS.ProcessEnv;
  allowRuntimePluginLoad?: boolean;
-  registerStream?: boolean;
+  apiRegistry?: ApiRegistry;
 }): StreamFn | undefined {
  // Plugin stream factories may capture model headers, so construction is the
  // last safe boundary for providers that do not expose the host fetch seam.
@@ -67,8 +69,9 @@ export function registerProviderStreamForModel<TApi extends Api>(params: {
  }
  // Register custom APIs only after a concrete stream exists, so later callers
  // can route by model.api without reloading provider runtime hooks.
-  if (params.registerStream !== false) {
-    ensureCustomApiRegistered(params.model.api, streamFn);
+  const apiRegistry = params.apiRegistry ?? getModelLlmRuntime(params.model)?.registry;
+  if (apiRegistry) {
+    ensureCustomApiRegistered(apiRegistry, params.model.api, streamFn);
  }
  return streamFn;
 }
@@ -1,5 +1,4 @@
 import { cleanupSessionResources } from "@openclaw/ai/internal/runtime";
-import { streamSimple } from "../../llm/stream.js";
 import type { AssistantMessage, Model } from "../../llm/types.js";
 import type {
  Agent,
@@ -36,6 +35,7 @@ import {
  type TurnStartEvent,
 } from "./extensions/index.js";
 import type { BashExecutionMessage, CustomMessage } from "./messages.js";
+import { getModelRegistryRuntime } from "./model-registry-runtime.js";
 import type { ModelRegistry } from "./model-registry.js";
 import type { PromptTemplate } from "./prompt-templates.js";
 import type { ResourceLoader } from "./resource-loader.js";
@@ -183,7 +183,10 @@ export abstract class AgentSessionBase {
    apiKey?: string;
    headers?: Record<string, string>;
  }> {
-    if (this.agent.streamFn === streamSimple) {
+    if (
+      this.agent.streamFn ===
+      getModelRegistryRuntime(this.sessionModelRegistry).llmRuntime.streamSimple
+    ) {
      return this.getRequiredRequestAuth(model);
    }

@@ -1,5 +1,4 @@
 import { isContextOverflow } from "@openclaw/ai/internal/runtime";
-import { streamSimple } from "../../llm/stream.js";
 import type { AssistantMessage, Model } from "../../llm/types.js";
 import {
  calculateContextTokens,
@@ -14,6 +13,7 @@ import { AgentSessionInspection } from "./agent-session-inspection.js";
 import { unwrapCoreResult } from "./agent-session-utils.js";
 import { formatNoModelSelectedMessage } from "./auth-guidance.js";
 import { preflightManualSessionCompaction } from "./manual-compaction-preflight.js";
+import { getModelRegistryRuntime } from "./model-registry-runtime.js";
 import { getLatestCompactionEntry, type CompactionEntry } from "./session-manager.js";
 import type { SettingsManager } from "./settings-manager.js";

@@ -109,7 +109,10 @@ export abstract class AgentSessionCompaction extends AgentSessionInspection {
      }
    | undefined
  > {
-    if (this.agent.streamFn !== streamSimple) {
+    if (
+      this.agent.streamFn !==
+      getModelRegistryRuntime(this.sessionModelRegistry).llmRuntime.streamSimple
+    ) {
      return this.getCompactionRequestAuth(model);
    }

@@ -1,6 +1,4 @@
 import { basename, dirname } from "node:path";
-import { defaultApiRegistry } from "@openclaw/ai/internal/runtime";
-import { resetApiProviders } from "@openclaw/ai/providers";
 import { AgentSessionCompaction } from "./agent-session-compaction.js";
 import type { ExtensionBindings } from "./agent-session-types.js";
 import { ExtensionRunner, type ToolDefinition, wrapRegisteredTools } from "./extensions/index.js";
@@ -401,8 +399,8 @@ export abstract class AgentSessionExtensions extends AgentSessionCompaction {
    await this.settingsManager.reload();
    this.agent.steeringMode = this.settingsManager.getSteeringMode();
    this.agent.followUpMode = this.settingsManager.getFollowUpMode();
-    resetApiProviders(defaultApiRegistry);
    await this.sessionResourceLoader.reload();
+    this.sessionModelRegistry.refresh();
    this.buildRuntime({
      activeToolNames: this.getActiveToolNames(),
      flagValues: previousFlagValues,
@@ -12,10 +12,6 @@ const streamMocks = vi.hoisted(() => ({
  streamSimple: vi.fn(),
 }));

-vi.mock("../../llm/stream.js", () => ({
-  streamSimple: streamMocks.streamSimple,
-}));
-
 import type { AgentTool } from "../runtime/index.js";
 import type { AgentSessionEvent } from "./agent-session-types.js";
 import { AgentSession } from "./agent-session.js";
@@ -172,6 +168,11 @@ async function createTestSession(
      retry: { enabled: false },
    });
  const sessionManager = options.sessionManager ?? SessionManager.inMemory();
+  const modelRegistry = ModelRegistry.inMemory(authStorage);
+  modelRegistry.registerProvider(model.provider, {
+    api: model.api,
+    streamSimple: streamMocks.streamSimple,
+  });
  const result = await createAgentSession({
    model,
    noTools: "builtin",
@@ -179,7 +180,7 @@ async function createTestSession(
    resourceLoader: options.resourceLoader ?? createResourceLoader(),
    sessionManager,
    settingsManager,
-    modelRegistry: ModelRegistry.inMemory(authStorage),
+    modelRegistry,
  });
  sessions.push(result.session);
  return { ...result, settingsManager, sessionManager };
@@ -0,0 +1,50 @@
+import {
+  createApiRegistry,
+  createLlmRuntime,
+  type ApiRegistry,
+  type LlmRuntime,
+} from "@openclaw/ai";
+import { getPublishedApiProviders } from "@openclaw/ai/internal/runtime";
+import { registerBuiltInApiProviders } from "@openclaw/ai/providers";
+import "../../llm/ai-transport-host.js";
+import { bindStreamLlmRuntime } from "../../llm/model-runtime-binding.js";
+
+type ModelRegistryRuntime = {
+  apiRegistry: ApiRegistry;
+  llmRuntime: LlmRuntime;
+};
+
+const modelRegistryRuntimes = new WeakMap<object, ModelRegistryRuntime>();
+
+function resetApiRegistry(runtime: ModelRegistryRuntime): void {
+  runtime.apiRegistry.clearApiProviders();
+  registerBuiltInApiProviders(runtime.apiRegistry);
+  // The Plugin SDK registry is a shipped compatibility facade. Snapshot it at
+  // lifecycle publication so request-time routing never depends on mutable global state.
+  for (const provider of getPublishedApiProviders()) {
+    runtime.apiRegistry.registerApiProvider(provider);
+  }
+}
+
+/** Creates the runtime facts owned by one model-registry lifecycle. */
+export function initializeModelRegistryRuntime(owner: object): void {
+  const apiRegistry = createApiRegistry();
+  const llmRuntime = createLlmRuntime(apiRegistry);
+  const runtime = { apiRegistry, llmRuntime };
+  bindStreamLlmRuntime(llmRuntime.streamSimple, llmRuntime);
+  resetApiRegistry(runtime);
+  modelRegistryRuntimes.set(owner, runtime);
+}
+
+/** Returns the prepared runtime facts for one model-registry lifecycle. */
+export function getModelRegistryRuntime(owner: object): ModelRegistryRuntime {
+  const runtime = modelRegistryRuntimes.get(owner);
+  if (!runtime) {
+    throw new Error("Model registry runtime is not initialized");
+  }
+  return runtime;
+}
+
+export function resetModelRegistryRuntime(owner: object): void {
+  resetApiRegistry(getModelRegistryRuntime(owner));
+}
@@ -3,9 +3,16 @@
 import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { dirname, join } from "node:path";
-import { afterEach, describe, expect, it } from "vitest";
+import {
+  defaultApiRegistry,
+  getApiProvider,
+  registerApiProvider,
+  unregisterApiProviders,
+} from "@openclaw/ai/internal/runtime";
+import { afterEach, describe, expect, it, vi } from "vitest";
 import { PLUGIN_MODEL_CATALOG_GENERATED_BY } from "../plugin-model-catalog.js";
 import { AuthStorage } from "./auth-storage.js";
+import { getModelRegistryRuntime } from "./model-registry-runtime.js";
 import { ModelRegistry, type ProviderConfigInput } from "./model-registry.js";

 const PLUGIN_MODEL_CATALOG_FILE = "catalog.json";
@@ -595,3 +602,58 @@ describe("ModelRegistry OAuth provider ownership", () => {
    ).toBe("Anthropic (Claude Pro/Max)");
  });
 });
+
+describe("ModelRegistry API provider ownership", () => {
+  it("keeps stream registrations isolated across registry refreshes", () => {
+    const sessionA = ModelRegistry.inMemory(AuthStorage.inMemory());
+    const sessionB = ModelRegistry.inMemory(AuthStorage.inMemory());
+    const streamA = vi.fn(() => ({}) as never);
+    const streamB = vi.fn(() => ({}) as never);
+
+    sessionA.registerProvider("session-a", {
+      api: "test-session-api",
+      streamSimple: streamA,
+    });
+    sessionB.registerProvider("session-b", {
+      api: "test-session-api",
+      streamSimple: streamB,
+    });
+    const runtimeA = getModelRegistryRuntime(sessionA);
+    const runtimeB = getModelRegistryRuntime(sessionB);
+
+    expect(runtimeA.apiRegistry.getApiProvider("test-session-api")?.streamSimple).not.toBe(
+      runtimeB.apiRegistry.getApiProvider("test-session-api")?.streamSimple,
+    );
+    expect(getApiProvider("test-session-api")).toBeUndefined();
+
+    sessionB.unregisterProvider("session-b");
+
+    expect(runtimeA.apiRegistry.getApiProvider("test-session-api")).toBeDefined();
+    expect(runtimeB.apiRegistry.getApiProvider("test-session-api")).toBeUndefined();
+  });
+
+  it("imports published SDK providers without copying request-generated aliases", () => {
+    const publishedSource = "plugin:test-published-api";
+    const requestSource = "custom-api:test-request-api";
+    const stream = vi.fn(() => ({}) as never);
+    registerApiProvider(
+      { api: "test-published-api", stream, streamSimple: stream },
+      publishedSource,
+    );
+    defaultApiRegistry.registerApiProvider(
+      { api: "test-request-api", stream, streamSimple: stream },
+      requestSource,
+    );
+
+    try {
+      const session = ModelRegistry.inMemory(AuthStorage.inMemory());
+      const runtime = getModelRegistryRuntime(session);
+
+      expect(runtime.apiRegistry.getApiProvider("test-published-api")).toBeDefined();
+      expect(runtime.apiRegistry.getApiProvider("test-request-api")).toBeUndefined();
+    } finally {
+      unregisterApiProviders(publishedSource);
+      defaultApiRegistry.unregisterApiProviders(requestSource);
+    }
+  });
+});
@@ -4,8 +4,6 @@

 import { existsSync, readFileSync } from "node:fs";
 import { dirname, join } from "node:path";
-import { defaultApiRegistry, registerApiProvider } from "@openclaw/ai/internal/runtime";
-import { resetApiProviders } from "@openclaw/ai/providers";
 import { type Static, Type } from "typebox";
 import { Compile } from "typebox/compile";
 import type { TLocalizedValidationError } from "typebox/error";
@@ -31,6 +29,11 @@ import {
 } from "../plugin-model-catalog.js";
 import { getAuthStorageOAuthProviderRegistry } from "./auth-storage-oauth-registry.js";
 import type { AuthStatus, AuthStorage } from "./auth-storage.js";
+import {
+  getModelRegistryRuntime,
+  initializeModelRegistryRuntime,
+  resetModelRegistryRuntime,
+} from "./model-registry-runtime.js";
 import { BUILT_IN_PROVIDER_DISPLAY_NAMES } from "./provider-display-names.js";
 import {
  clearConfigValueCache,
@@ -322,6 +325,7 @@ export class ModelRegistry {
    options: ModelRegistryOptions = {},
  ) {
    this.authStorage = authStorage;
+    initializeModelRegistryRuntime(this);
    this.modelsJsonPath = modelsJsonPath;
    this.pluginMetadataSnapshot = resolveModelPluginMetadataSnapshot({
      ...(options.pluginMetadataSnapshot
@@ -354,8 +358,8 @@ export class ModelRegistry {
    this.modelRequestHeaders.clear();
    this.loadError = undefined;

-    // Ensure dynamic API/OAuth registrations are rebuilt from current provider state.
-    resetApiProviders(defaultApiRegistry);
+    // Rebuild this lifecycle's API/OAuth registrations from current provider state.
+    resetModelRegistryRuntime(this);
    getAuthStorageOAuthProviderRegistry(this.authStorage).reset();

    this.loadModels();
@@ -859,7 +863,7 @@ export class ModelRegistry {

    if (config.streamSimple) {
      const streamSimple = config.streamSimple;
-      registerApiProvider(
+      getModelRegistryRuntime(this).apiRegistry.registerApiProvider(
        {
          api: config.api!,
          stream: (model, context, options) =>
@@ -3,6 +3,7 @@ import { createAssistantMessageEventStream, type AssistantMessage } from "opencl
 // session write-lock behavior.
 import { Type } from "typebox";
 import { beforeEach, describe, expect, it, vi } from "vitest";
+import { getStreamLlmRuntime } from "../../llm/model-runtime-binding.js";
 import type { Model, SimpleStreamOptions } from "../../llm/types.js";
 import { createUserTurnTranscriptRecorder } from "../../sessions/user-turn-transcript.js";
 import { createTestUserTurnTranscriptTarget } from "../../sessions/user-turn-transcript.test-support.js";
@@ -24,6 +25,7 @@ import { takeRuntimeUserTurnTranscriptContext } from "../../sessions/user-turn-t
 import { AuthStorage } from "./auth-storage.js";
 import { createExtensionRuntime } from "./extensions/loader.js";
 import type { LoadExtensionsResult, ToolDefinition } from "./extensions/types.js";
+import { getModelRegistryRuntime } from "./model-registry-runtime.js";
 import { ModelRegistry } from "./model-registry.js";
 import type { ResourceLoader } from "./resource-loader.js";
 import { createAgentSession } from "./sdk.js";
@@ -44,6 +46,23 @@ const testModel: Model = {
  maxTokens: 1000,
 };

+describe("createAgentSession runtime ownership", () => {
+  it("binds the installed stream wrapper to the model-registry lifecycle", async () => {
+    const modelRegistry = createTestModelRegistry();
+    const { session } = await createAgentSession({
+      model: testModel,
+      resourceLoader: createEmptyResourceLoader(),
+      sessionManager: SessionManager.inMemory(),
+      settingsManager: SettingsManager.inMemory(),
+      modelRegistry,
+    });
+
+    expect(getStreamLlmRuntime(session.agent.streamFn)).toBe(
+      getModelRegistryRuntime(modelRegistry).llmRuntime,
+    );
+  });
+});
+
 function createModelWithoutBaseUrl(overrides: Partial<Model>): Model {
  const { baseUrl: _baseUrl, ...model } = { ...testModel, ...overrides };
  return model as unknown as Model;
@@ -87,6 +106,17 @@ function createEmptyResourceLoader(): ResourceLoader {
  return createResourceLoaderWithHandlers(new Map());
 }

+function createTestModelRegistry(authStorage = AuthStorage.inMemory()): ModelRegistry {
+  const modelRegistry = ModelRegistry.inMemory(authStorage);
+  for (const api of ["openai-responses", "bedrock-converse-stream"] as const) {
+    modelRegistry.registerProvider(`test-${api}`, {
+      api,
+      streamSimple: streamMocks.streamSimple,
+    });
+  }
+  return modelRegistry;
+}
+
 function createResourceLoaderWithHandlers(
  handlers: Map<string, Array<(...args: unknown[]) => Promise<unknown>>>,
 ): ResourceLoader {
@@ -130,7 +160,7 @@ async function createSessionAndStreamModel(model: Model): Promise<SimpleStreamOp
    resourceLoader: createEmptyResourceLoader(),
    sessionManager: SessionManager.inMemory(),
    settingsManager: SettingsManager.inMemory(),
-    modelRegistry: ModelRegistry.inMemory(AuthStorage.inMemory()),
+    modelRegistry: createTestModelRegistry(),
  });

  await session.agent.streamFn?.(
@@ -735,7 +765,7 @@ describe("AgentSession retry behavior", () => {
      settingsManager: SettingsManager.inMemory({
        retry: retry ?? { baseDelayMs: 0, maxRetries: 1 },
      }),
-      modelRegistry: ModelRegistry.inMemory(authStorage),
+      modelRegistry: createTestModelRegistry(authStorage),
    });
  }

@@ -9,7 +9,7 @@ import {
  resolveThinkingDefaultForModel,
  type ThinkingCatalogEntry,
 } from "../../auto-reply/thinking.js";
-import { streamSimple } from "../../llm/stream.js";
+import { bindStreamLlmRuntime } from "../../llm/model-runtime-binding.js";
 import type { Message, Model } from "../../llm/types.js";
 import { getAgentDir } from "../config.js";
 import {
@@ -30,6 +30,7 @@ import type {
  ToolDefinition,
 } from "./extensions/index.js";
 import { convertToLlm } from "./messages.js";
+import { getModelRegistryRuntime } from "./model-registry-runtime.js";
 import { ModelRegistry } from "./model-registry.js";
 import { findInitialModel } from "./model-resolver.js";
 import { DefaultResourceLoader, type ResourceLoader } from "./resource-loader.js";
@@ -293,6 +294,7 @@ export async function createAgentSession(
  if (!resourceLoader) {
    resourceLoader = new DefaultResourceLoader({ cwd, agentDir, settingsManager });
    await resourceLoader.reload();
+    modelRegistry.refresh();
  }

  // Check if session has existing data to restore
@@ -438,6 +440,7 @@ export async function createAgentSession(
  const runWithSessionWriteLock = async <T>(run: () => Promise<T> | T): Promise<T> =>
    options.withSessionWriteLock ? await options.withSessionWriteLock(run) : await run();

+  const modelRegistryRuntime = getModelRegistryRuntime(modelRegistry);
  const agent: Agent = new Agent({
    initialState: {
      systemPrompt: "",
@@ -453,7 +456,7 @@ export async function createAgentSession(
      }
      const providerRetrySettings = settingsManager.getProviderRetrySettings();
      const attributionHeaders = getAttributionHeaders(modelResult, settingsManager);
-      return streamSimple(modelResult, context, {
+      return modelRegistryRuntime.llmRuntime.streamSimple(modelResult, context, {
        ...optionsLocal,
        apiKey: auth.apiKey,
        timeoutMs: optionsLocal?.timeoutMs ?? providerRetrySettings.timeoutMs,
@@ -506,6 +509,9 @@ export async function createAgentSession(
    thinkingBudgets: settingsManager.getThinkingBudgets(),
    maxRetryDelayMs: settingsManager.getProviderRetrySettings().maxRetryDelayMs,
  });
+  if (agent.streamFn) {
+    bindStreamLlmRuntime(agent.streamFn, modelRegistryRuntime.llmRuntime);
+  }

  // Restore messages if session has existing data
  if (hasExistingSession) {
@@ -31,6 +31,20 @@ vi.mock("../llm/stream.js", () => ({
  completeSimple: hoisted.completeMock,
 }));

+vi.mock("./sessions/model-registry-runtime.js", () => ({
+  getModelRegistryRuntime: () => {
+    const apiRegistry = {};
+    return {
+      apiRegistry,
+      llmRuntime: {
+        registry: apiRegistry,
+        completeSimple: (...args: unknown[]) => hoisted.completeMock(...args),
+        streamSimple: vi.fn(),
+      },
+    };
+  },
+}));
+
 vi.mock("./embedded-agent-runner/model.js", () => ({
  resolveModel: hoisted.resolveModelMock,
  resolveModelAsync: hoisted.resolveModelAsyncMock,
@@ -880,7 +894,11 @@ describe("completeWithPreparedSimpleCompletionModel", () => {
      },
    });

-    expect(hoisted.prepareModelForSimpleCompletionMock).toHaveBeenCalledWith({ model, cfg });
+    expect(hoisted.prepareModelForSimpleCompletionMock).toHaveBeenCalledWith({
+      apiRegistry: expect.anything(),
+      model,
+      cfg,
+    });
    expect(hoisted.completeMock).toHaveBeenCalledWith(
      preparedModel,
      {
@@ -1,4 +1,5 @@
 import { supportsOpenAIReasoningEffort } from "@openclaw/ai/internal/openai";
+import { defaultApiRegistry } from "@openclaw/ai/internal/runtime";
 import { resolveClaudeSonnet5ModelIdentity } from "@openclaw/llm-core";
 /**
 * Simple completion runtime preparation.
@@ -8,6 +9,7 @@ import { resolveClaudeSonnet5ModelIdentity } from "@openclaw/llm-core";
 import type { ThinkLevel } from "../auto-reply/thinking.js";
 import type { OpenClawConfig } from "../config/types.openclaw.js";
 import { formatErrorMessage } from "../infra/errors.js";
+import { bindModelLlmRuntime, getModelLlmRuntime } from "../llm/model-runtime-binding.js";
 import { completeSimple } from "../llm/stream.js";
 import type {
  AssistantMessage,
@@ -49,6 +51,7 @@ import { applyPreparedRuntimeAuthToModel } from "./provider-request-config.js";
 import { protectPreparedProviderRuntimeAuth } from "./provider-secret-egress.js";
 import { buildAgentRuntimeAuthPlan } from "./runtime-plan/auth.js";
 import { materializePreparedRuntimeModel } from "./runtime-plan/materialize-model.js";
+import { getModelRegistryRuntime } from "./sessions/model-registry-runtime.js";
 import { resolveSimpleCompletionModelResolverWorkspace } from "./simple-completion-scope.js";
 import { prepareModelForSimpleCompletion } from "./simple-completion-transport.js";
 import { resolveUtilityModelRefForAgent } from "./utility-model.js";
@@ -445,11 +448,15 @@ export async function prepareSimpleCompletionModel(params: {
        })
      : fingerprintResolvedProviderAuth(auth)
    : undefined;
+  const modelRuntime = getModelRegistryRuntime(resolved.modelRegistry);

  return {
-    model: applySecretRefHeaderSentinels(
-      applyLocalNoAuthHeaderOverride(resolvedModel, resolvedAuth),
-      params.cfg,
+    model: bindModelLlmRuntime(
+      applySecretRefHeaderSentinels(
+        applyLocalNoAuthHeaderOverride(resolvedModel, resolvedAuth),
+        params.cfg,
+      ),
+      modelRuntime.llmRuntime,
    ),
    auth: resolvedAuth,
    ...(sourceAuthFingerprint ? { sourceAuthFingerprint } : {}),
@@ -521,7 +528,17 @@ export async function completeWithPreparedSimpleCompletionModel(params: {
  cfg?: OpenClawConfig;
  options?: SimpleCompletionModelOptions;
 }): Promise<AssistantMessage> {
-  const completionModel = prepareModelForSimpleCompletion({ model: params.model, cfg: params.cfg });
+  const runtime = getModelLlmRuntime(params.model);
+  let completionModel = prepareModelForSimpleCompletion({
+    // Direct SDK callers that did not use the preparation helper keep the shipped
+    // process-default behavior; all prepared host paths carry their lifecycle owner.
+    apiRegistry: runtime?.registry ?? defaultApiRegistry,
+    model: params.model,
+    cfg: params.cfg,
+  });
+  if (runtime) {
+    completionModel = bindModelLlmRuntime(completionModel, runtime);
+  }
  const { reasoning: rawReasoning, ...options } = params.options ?? {};
  const reasoning = normalizeSimpleCompletionReasoning(rawReasoning, completionModel);
  return await completeSimple(completionModel, params.context, {
@@ -1,8 +1,8 @@
-import { registerApiProvider, unregisterApiProviders } from "@openclaw/ai/internal/runtime";
+import { createApiRegistry, type ApiRegistry } from "@openclaw/ai";
 // Simple completion transport tests cover provider-specific stream alias
 // selection before the generic completion helper invokes the LLM layer.
 import type { Model } from "openclaw/plugin-sdk/llm";
-import { afterEach, beforeAll, beforeEach, describe, expect, it, vi } from "vitest";
+import { beforeAll, beforeEach, describe, expect, it, vi } from "vitest";
 import type { OpenClawConfig } from "../config/config.js";
 import { createMoonshotThinkingWrapper } from "../llm/providers/stream-wrappers/moonshot-thinking.js";
 import { mintSecretSentinel } from "../secrets/sentinel.js";
@@ -17,7 +17,7 @@ const createOpenClawTransportStreamFnForModel = vi.fn();
 const createTransportAwareStreamFnForModel = vi.fn();
 const prepareTransportAwareSimpleModel = vi.fn();
 const resolveTransportAwareSimpleApi = vi.fn();
-const prepareGoogleSimpleCompletionModel = vi.fn((model: unknown) => model);
+const prepareGoogleSimpleCompletionModel = vi.fn((_registry: unknown, model: unknown) => model);
 const pluginStreamFn = vi.fn(() => "plugin-stream-result" as never);

 vi.mock("./anthropic-vertex-stream.js", () => ({
@@ -51,17 +51,31 @@ vi.mock("../plugins/provider-runtime.js", async () => {
  };
 });

-let prepareModelForSimpleCompletion: typeof import("./simple-completion-transport.js").prepareModelForSimpleCompletion;
+let prepareModelForSimpleCompletionImpl: typeof import("./simple-completion-transport.js").prepareModelForSimpleCompletion;
+let apiRegistry: ApiRegistry;
 const SIMPLE_COMPLETION_SOURCE_ID = "test:simple-completion-transport";

+function prepareModelForSimpleCompletion(
+  params: Omit<
+    Parameters<
+      typeof import("./simple-completion-transport.js").prepareModelForSimpleCompletion
+    >[0],
+    "apiRegistry"
+  >,
+) {
+  return prepareModelForSimpleCompletionImpl({ ...params, apiRegistry });
+}
+
 describe("prepareModelForSimpleCompletion", () => {
  beforeAll(async () => {
    // Dynamic import lets the mocked transport/provider modules settle before
    // the unit under test captures custom stream registration helpers.
-    ({ prepareModelForSimpleCompletion } = await import("./simple-completion-transport.js"));
+    ({ prepareModelForSimpleCompletion: prepareModelForSimpleCompletionImpl } =
+      await import("./simple-completion-transport.js"));
  });

  beforeEach(() => {
+    apiRegistry = createApiRegistry();
    createAnthropicVertexStreamFnForModel.mockReset();
    ensureCustomApiRegistered.mockReset();
    resolveProviderStreamFn.mockReset();
@@ -81,18 +95,14 @@ describe("prepareModelForSimpleCompletion", () => {
    createTransportAwareStreamFnForModel.mockReturnValue(undefined);
    prepareTransportAwareSimpleModel.mockImplementation((model) => model);
    resolveTransportAwareSimpleApi.mockReturnValue(undefined);
-    prepareGoogleSimpleCompletionModel.mockImplementation((model) => model);
-  });
-
-  afterEach(() => {
-    unregisterApiProviders(SIMPLE_COMPLETION_SOURCE_ID);
+    prepareGoogleSimpleCompletionModel.mockImplementation((_registry, model) => model);
  });

  it("routes provider-owned simple-completion wrappers through an internal API alias", () => {
    const sourceApi = "moonshot-simple-source";
    const sourceResult = { source: true };
    let capturedApi: string | undefined;
-    registerApiProvider(
+    apiRegistry.registerApiProvider(
      {
        api: sourceApi,
        stream: () => sourceResult as never,
@@ -137,7 +147,7 @@ describe("prepareModelForSimpleCompletion", () => {
        }),
      }),
    );
-    const registeredStream = ensureCustomApiRegistered.mock.calls.at(-1)?.[1];
+    const registeredStream = ensureCustomApiRegistered.mock.calls.at(-1)?.[2];
    expect(registeredStream).toBeTypeOf("function");
    const stream = registeredStream(result, { messages: [] }, {});
    expect(stream).toBe(sourceResult);
@@ -193,8 +203,12 @@ describe("prepareModelForSimpleCompletion", () => {
      ...model,
      headers: { Authorization: `Bearer ${secret}` },
    });
-    expect(ensureCustomApiRegistered).toHaveBeenCalledWith("ollama", expect.any(Function));
-    const registeredStream = ensureCustomApiRegistered.mock.calls[0]?.[1] as StreamFn;
+    expect(ensureCustomApiRegistered).toHaveBeenCalledWith(
+      apiRegistry,
+      "ollama",
+      expect.any(Function),
+    );
+    const registeredStream = ensureCustomApiRegistered.mock.calls[0]?.[2] as StreamFn;
    void registeredStream(
      { ...model, headers: { Authorization: `Bearer ${sentinel}` } } as never,
      {} as never,
@@ -228,6 +242,7 @@ describe("prepareModelForSimpleCompletion", () => {

    expect(createAnthropicVertexStreamFnForModel).toHaveBeenCalledWith(model);
    expect(ensureCustomApiRegistered).toHaveBeenCalledWith(
+      apiRegistry,
      "openclaw-anthropic-vertex-simple:https%3A%2F%2Fus-central1-aiplatform.googleapis.com",
      "vertex-stream",
    );
@@ -263,6 +278,7 @@ describe("prepareModelForSimpleCompletion", () => {
    expect(prepareTransportAwareSimpleModel).toHaveBeenCalledWith(model, { cfg: undefined });
    expect(buildTransportAwareSimpleStreamFn).toHaveBeenCalledWith(model, { cfg: undefined });
    expect(ensureCustomApiRegistered).toHaveBeenCalledWith(
+      apiRegistry,
      "openclaw-openai-responses-transport",
      "transport-stream",
    );
@@ -286,7 +302,7 @@ describe("prepareModelForSimpleCompletion", () => {
      maxTokens: 8192,
      headers: {},
    };
-    prepareGoogleSimpleCompletionModel.mockImplementationOnce((m: unknown) => ({
+    prepareGoogleSimpleCompletionModel.mockImplementationOnce((_registry: unknown, m: unknown) => ({
      ...(m as Model<"google-generative-ai">),
      api: "openclaw-google-generative-ai-simple",
    }));
@@ -295,7 +311,7 @@ describe("prepareModelForSimpleCompletion", () => {
    const result = prepareModelForSimpleCompletion({ model });

    expect(prepareTransportAwareSimpleModel).toHaveBeenCalledWith(model, { cfg: undefined });
-    expect(prepareGoogleSimpleCompletionModel).toHaveBeenCalledWith(model);
+    expect(prepareGoogleSimpleCompletionModel).toHaveBeenCalledWith(apiRegistry, model);
    expect(buildTransportAwareSimpleStreamFn).not.toHaveBeenCalled();
    expect(result).toEqual({
      ...model,
@@ -330,6 +346,7 @@ describe("prepareModelForSimpleCompletion", () => {

    expect(buildTransportAwareSimpleStreamFn).toHaveBeenCalledWith(model, { cfg: undefined });
    expect(ensureCustomApiRegistered).toHaveBeenCalledWith(
+      apiRegistry,
      "openclaw-google-generative-ai-transport",
      "google-transport-stream",
    );
@@ -380,6 +397,7 @@ describe("prepareModelForSimpleCompletion", () => {
        { cfg: undefined },
      );
      expect(ensureCustomApiRegistered).toHaveBeenCalledWith(
+        apiRegistry,
        "openclaw-openai-responses-transport",
        "codex-transport-stream",
      );
@@ -1,4 +1,4 @@
-import { getApiProvider } from "@openclaw/ai/internal/runtime";
+import type { ApiRegistry } from "@openclaw/ai";
 /**
 * Simple completion transport preparation.
 *
@@ -65,11 +65,15 @@ function resolveProviderSimpleCompletionApi(model: Model): Api {
    .join(":")}`;
 }

-function applyProviderSimpleCompletionWrapper(model: Model, cfg?: OpenClawConfig): Model {
+function applyProviderSimpleCompletionWrapper(
+  registry: ApiRegistry,
+  model: Model,
+  cfg?: OpenClawConfig,
+): Model {
  if (model.api.startsWith(PROVIDER_SIMPLE_COMPLETION_API_PREFIX)) {
    return model;
  }
-  const sourceProvider = getApiProvider(model.api);
+  const sourceProvider = registry.getApiProvider(model.api);
  if (!sourceProvider) {
    return model;
  }
@@ -93,11 +97,12 @@ function applyProviderSimpleCompletionWrapper(model: Model, cfg?: OpenClawConfig
  }

  const api = resolveProviderSimpleCompletionApi(model);
-  ensureCustomApiRegistered(api, streamFn);
+  ensureCustomApiRegistered(registry, api, streamFn);
  return { ...model, api };
 }

 function prepareCodexSimpleTransportModel<TApi extends Api>(
+  registry: ApiRegistry,
  model: Model<TApi>,
  cfg?: OpenClawConfig,
 ): Model | undefined {
@@ -117,7 +122,7 @@ function prepareCodexSimpleTransportModel<TApi extends Api>(
    return undefined;
  }

-  ensureCustomApiRegistered(api, streamFn);
+  ensureCustomApiRegistered(registry, api, streamFn);
  return {
    ...transportModel,
    api,
@@ -125,38 +130,46 @@ function prepareCodexSimpleTransportModel<TApi extends Api>(
 }

 export function prepareModelForSimpleCompletion<TApi extends Api>(params: {
+  apiRegistry: ApiRegistry;
  model: Model<TApi>;
  cfg?: OpenClawConfig;
 }): Model {
-  const { model, cfg } = params;
+  const { apiRegistry, model, cfg } = params;
  // Only provider-owned custom APIs need runtime stream registration here.
-  if (!getApiProvider(model.api) && registerProviderStreamForModel({ model, cfg })) {
-    return applyProviderSimpleCompletionWrapper(model, cfg);
+  if (
+    !apiRegistry.getApiProvider(model.api) &&
+    registerProviderStreamForModel({ model, cfg, apiRegistry })
+  ) {
+    return applyProviderSimpleCompletionWrapper(apiRegistry, model, cfg);
  }

-  const codexTransportModel = prepareCodexSimpleTransportModel(model, cfg);
+  const codexTransportModel = prepareCodexSimpleTransportModel(apiRegistry, model, cfg);
  if (codexTransportModel) {
-    return applyProviderSimpleCompletionWrapper(codexTransportModel, cfg);
+    return applyProviderSimpleCompletionWrapper(apiRegistry, codexTransportModel, cfg);
  }

  const transportAwareModel = prepareTransportAwareSimpleModel(model, { cfg });
  if (transportAwareModel !== model) {
    const streamFn = buildTransportAwareSimpleStreamFn(model, { cfg });
    if (streamFn) {
-      ensureCustomApiRegistered(transportAwareModel.api, streamFn);
-      return applyProviderSimpleCompletionWrapper(transportAwareModel, cfg);
+      ensureCustomApiRegistered(apiRegistry, transportAwareModel.api, streamFn);
+      return applyProviderSimpleCompletionWrapper(apiRegistry, transportAwareModel, cfg);
    }
  }

  if (model.api === "google-generative-ai") {
-    return applyProviderSimpleCompletionWrapper(prepareGoogleSimpleCompletionModel(model), cfg);
+    return applyProviderSimpleCompletionWrapper(
+      apiRegistry,
+      prepareGoogleSimpleCompletionModel(apiRegistry, model),
+      cfg,
+    );
  }

  if (model.provider === "anthropic-vertex") {
    const api = resolveAnthropicVertexSimpleApi(model.baseUrl);
-    ensureCustomApiRegistered(api, createAnthropicVertexStreamFnForModel(model));
-    return applyProviderSimpleCompletionWrapper({ ...model, api }, cfg);
+    ensureCustomApiRegistered(apiRegistry, api, createAnthropicVertexStreamFnForModel(model));
+    return applyProviderSimpleCompletionWrapper(apiRegistry, { ...model, api }, cfg);
  }

-  return applyProviderSimpleCompletionWrapper(model, cfg);
+  return applyProviderSimpleCompletionWrapper(apiRegistry, model, cfg);
 }
@@ -12,6 +12,10 @@ import * as modelDiscovery from "../agent-model-discovery.js";
 import type { AuthProfileStore } from "../auth-profiles/types.js";
 import * as modelAuth from "../model-auth.js";
 import * as modelsConfig from "../models-config.js";
+import {
+  getModelRegistryRuntime,
+  initializeModelRegistryRuntime,
+} from "../sessions/model-registry-runtime.js";
 import * as pdfNativeProviders from "./pdf-native-providers.js";
 import * as pdfModelConfigModule from "./pdf-tool.model-config.js";
 import { resetPdfToolAuthEnv, withTempPdfAgentDir } from "./pdf-tool.test-support.js";
@@ -19,14 +23,6 @@ import { resetPdfToolAuthEnv, withTempPdfAgentDir } from "./pdf-tool.test-suppor
 const completeMock = vi.hoisted(() => vi.fn());
 const registerProviderStreamForModelMock = vi.hoisted(() => vi.fn());

-vi.mock("../../llm/stream.js", async () => {
-  const actual = await vi.importActual<typeof import("../../llm/stream.js")>("../../llm/stream.js");
-  return {
-    ...actual,
-    complete: completeMock,
-  };
-});
-
 vi.mock("../provider-stream.js", () => ({
  registerProviderStreamForModel: registerProviderStreamForModelMock,
 }));
@@ -148,7 +144,10 @@ async function stubPdfToolInfra(
            maxTokens: 8192,
            input: params?.input ?? ["text", "document"],
          }) as never;
-  vi.spyOn(modelDiscovery, "discoverModels").mockReturnValue({ find } as never);
+  const modelRegistry = { find };
+  initializeModelRegistryRuntime(modelRegistry);
+  getModelRegistryRuntime(modelRegistry).llmRuntime.complete = completeMock;
+  vi.spyOn(modelDiscovery, "discoverModels").mockReturnValue(modelRegistry as never);

  vi.spyOn(modelsConfig, "ensureOpenClawModelsJson").mockResolvedValue({
    agentDir,
@@ -744,6 +743,7 @@ describe("createPdfTool", () => {
      expect(modelAuth.requireApiKey).not.toHaveBeenCalled();
      expect(setRuntimeApiKey).not.toHaveBeenCalled();
      expect(registerProviderStreamForModelMock).toHaveBeenCalledWith({
+        apiRegistry: expect.anything(),
        model: expect.objectContaining({
          provider: "amazon-bedrock",
          api: "bedrock-converse-stream",
@@ -9,7 +9,6 @@ import {
 } from "@openclaw/normalization-core/string-coerce";
 import { Type } from "typebox";
 import type { OpenClawConfig } from "../../config/types.openclaw.js";
-import { complete } from "../../llm/stream.js";
 import type { Context } from "../../llm/types.js";
 import {
  classifyMediaReferenceSource,
@@ -23,6 +22,7 @@ import { applySecretRefHeaderSentinels } from "../model-auth.js";
 import { getModelProviderRequestTransport } from "../provider-request-config.js";
 import { registerProviderStreamForModel } from "../provider-stream.js";
 import { optionalFiniteNumberSchema } from "../schema/typebox.js";
+import { getModelRegistryRuntime } from "../sessions/model-registry-runtime.js";
 import { readFiniteNumberParam, ToolInputError } from "./common.js";
 import { coerceImageModelConfig, type ImageModelConfig } from "./image-tool.helpers.js";
 import {
@@ -245,10 +245,12 @@ async function runPdfPrompt(params: {

      // PDF-only model selections may not have loaded their provider plugin yet.
      // Register before complete() so plugin-owned APIs resolve on first use.
+      const modelRuntime = getModelRegistryRuntime(modelRegistry);
      registerProviderStreamForModel({
        model,
        cfg: effectiveCfg,
        agentDir: params.agentDir,
+        apiRegistry: modelRuntime.apiRegistry,
        ...(params.workspaceDir ? { workspaceDir: params.workspaceDir } : {}),
      });

@@ -266,7 +268,7 @@ async function runPdfPrompt(params: {
          images: [],
        }));
        const context = buildPdfExtractionContext(params.prompt, textOnlyExtractions, model);
-        const message = await complete(model, context, {
+        const message = await modelRuntime.llmRuntime.complete(model, context, {
          apiKey,
          maxTokens: resolvePdfToolMaxTokens(model.maxTokens),
        });
@@ -275,7 +277,7 @@ async function runPdfPrompt(params: {
      }

      const context = buildPdfExtractionContext(params.prompt, extractions, model);
-      const message = await complete(model, context, {
+      const message = await modelRuntime.llmRuntime.complete(model, context, {
        apiKey,
        maxTokens: resolvePdfToolMaxTokens(model.maxTokens),
      });
@@ -13,6 +13,7 @@ import { resolveSimpleCompletionModelResolverWorkspace } from "../../agents/simp
 import type { SessionEntry } from "../../config/sessions.js";
 import type { OpenClawConfig } from "../../config/types.openclaw.js";
 import { onTrustedInternalDiagnosticEvent } from "../../infra/diagnostic-events.js";
+import { bindModelLlmRuntime } from "../../llm/model-runtime-binding.js";
 import type { AssistantMessage, Model, StreamFn, Usage } from "../../llm/types.js";
 import { createAssistantMessageEventStream } from "../../llm/utils/event-stream.js";
 import type { loadManifestMetadataSnapshot } from "../../plugins/manifest-contract-eligibility.js";
@@ -24,6 +25,10 @@ import {
 } from "./inference-runtime.js";
 import { createWorkerToolCallStream } from "./inference-tool-call-stream.js";

+vi.mock("../../agents/sessions/model-registry-runtime.js", () => ({
+  getModelRegistryRuntime: (owner: unknown) => owner,
+}));
+
 type Deps = {
  applyStreamPolicy: typeof applyExtraParamsToAgent;
  loadCatalog: typeof loadModelCatalog;
@@ -172,7 +177,6 @@ function setup(entry: SessionEntry = sessionEntry) {
    authProfile?: string;
    catalogWorkspace?: string;
    prepareWorkspace?: string;
-    registerStream?: boolean;
  } = {};
  const resolveModel = vi.fn<Deps["resolveModel"]>(
    async (_provider, _model, _dir, _cfg, options) => {
@@ -185,8 +189,12 @@ function setup(entry: SessionEntry = sessionEntry) {
      modelParams.modelResolver,
    );
    await modelParams.modelResolver?.(PROVIDER, MODEL, modelParams.agentDir, modelParams.cfg, {});
+    const apiRegistry = {};
    return {
-      model: logicalModel,
+      model: bindModelLlmRuntime(logicalModel, {
+        registry: apiRegistry,
+        streamSimple: fallbackStream,
+      } as never),
      auth: {
        apiKey: AUTH_MARKER,
        profileId: PROFILE,
@@ -201,8 +209,7 @@ function setup(entry: SessionEntry = sessionEntry) {
  const loadManifestSnapshot = vi.fn(
    () => ({ plugins: [] }) as unknown as ReturnType<Deps["loadManifestSnapshot"]>,
  );
-  const resolveProviderStream = vi.fn<Deps["resolveProviderStream"]>((streamParams) => {
-    scope.registerStream = streamParams.registerStream;
+  const resolveProviderStream = vi.fn<Deps["resolveProviderStream"]>(() => {
    return stream;
  });
  const resolveStream = vi.fn<Deps["resolveStream"]>((streamParams) => {
@@ -238,7 +245,6 @@ function setup(entry: SessionEntry = sessionEntry) {
    resolveProviderStream,
    resolveStream,
    applyStreamPolicy,
-    stream: fallbackStream,
    wrapStream: vi.fn((streamFn: StreamFn) => streamFn),
    createTrace: vi.fn(() => ({ traceId: "1".repeat(32), spanId: "2".repeat(16) })),
  };
@@ -364,7 +370,6 @@ describe("worker inference provider runtime", () => {
      authProfile: PROFILE,
      catalogWorkspace: WORKSPACE,
      prepareWorkspace: WORKSPACE,
-      registerStream: false,
    });
    const [streamModel, streamContext, streamOptions] = runtime.stream.mock.calls[0] ?? [];
    expect(streamModel).toMatchObject({ baseUrl: ENDPOINT });
@@ -54,14 +54,13 @@ import {
  freezeDiagnosticTraceContext,
  type DiagnosticTraceContext,
 } from "../../infra/diagnostic-trace-context.js";
-import { streamSimple } from "../../llm/stream.js";
+import { getModelLlmRuntime } from "../../llm/model-runtime-binding.js";
 import type {
  AssistantMessage,
  AssistantMessageEvent,
  Context,
  Model,
  SimpleStreamOptions,
-  StreamFn,
  Tool,
  Usage,
 } from "../../llm/types.js";
@@ -110,7 +109,6 @@ type WorkerInferenceRuntimeDependencies = {
  resolveProviderStream: typeof registerProviderStreamForModel;
  resolveStream: typeof resolveEmbeddedAgentStreamFn;
  applyStreamPolicy: typeof applyExtraParamsToAgent;
-  stream: StreamFn;
  wrapStream: typeof wrapStreamFnWithDiagnosticModelCallEvents;
  createTrace: typeof createDiagnosticTraceContextFromActiveScope;
  recordUsage: (params: WorkerInferenceUsageParams) => void;
@@ -391,7 +389,6 @@ const DEFAULT_DEPENDENCIES: WorkerInferenceRuntimeDependencies = {
  resolveProviderStream: registerProviderStreamForModel,
  resolveStream: resolveEmbeddedAgentStreamFn,
  applyStreamPolicy: applyExtraParamsToAgent,
-  stream: streamSimple as StreamFn,
  wrapStream: wrapStreamFnWithDiagnosticModelCallEvents,
  createTrace: createDiagnosticTraceContextFromActiveScope,
  recordUsage: emitWorkerInferenceUsage,
@@ -635,6 +632,10 @@ export function createWorkerInferenceExecutor(
      model: approved.model,
    };
    const logicalModel = approved.prepared.model;
+    const llmRuntime = getModelLlmRuntime(logicalModel);
+    if (!llmRuntime) {
+      throw new Error("Prepared worker model has no lifecycle runtime owner");
+    }
    const providerModel =
      logicalModel.provider === "openai" && logicalModel.api === "openai-chatgpt-responses"
        ? {
@@ -647,12 +648,12 @@ export function createWorkerInferenceExecutor(
      cfg: config,
      agentDir: approved.agentDir,
      workspaceDir: approved.workspaceDir,
-      registerStream: false,
    });
    const authValue = approved.prepared.auth.apiKey;
    const streamAgent = {
      streamFn: dependencies.resolveStream({
-        currentStreamFn: dependencies.stream,
+        llmRuntime,
+        currentStreamFn: llmRuntime.streamSimple,
        ...(providerStream ? { providerStreamFn: providerStream } : {}),
        sessionId: request.sessionId,
        signal,
@@ -0,0 +1,32 @@
+import type { LlmRuntime } from "@openclaw/ai";
+import type { Model } from "./types.js";
+
+const MODEL_LLM_RUNTIME = Symbol("openclaw.modelLlmRuntime");
+const streamLlmRuntimes = new WeakMap<object, LlmRuntime>();
+
+type RuntimeBoundModel = Model & {
+  [MODEL_LLM_RUNTIME]?: LlmRuntime;
+};
+
+/** Carries the prepared lifecycle runtime without changing the serialized model shape. */
+export function bindModelLlmRuntime(model: Model, runtime: LlmRuntime): Model {
+  const bound = { ...model } as RuntimeBoundModel;
+  Object.defineProperty(bound, MODEL_LLM_RUNTIME, {
+    value: runtime,
+    enumerable: false,
+  });
+  return bound;
+}
+
+export function getModelLlmRuntime(model: Model): LlmRuntime | undefined {
+  return (model as RuntimeBoundModel)[MODEL_LLM_RUNTIME];
+}
+
+/** Associates a prepared stream entry point with the runtime that owns it. */
+export function bindStreamLlmRuntime(streamFn: object, runtime: LlmRuntime): void {
+  streamLlmRuntimes.set(streamFn, runtime);
+}
+
+export function getStreamLlmRuntime(streamFn: object | undefined): LlmRuntime | undefined {
+  return streamFn ? streamLlmRuntimes.get(streamFn) : undefined;
+}
@@ -0,0 +1,37 @@
+import { createApiRegistry, createLlmRuntime } from "@openclaw/ai";
+import { describe, expect, it, vi } from "vitest";
+import { bindModelLlmRuntime } from "./model-runtime-binding.js";
+import { streamSimple } from "./stream.js";
+import { createAssistantMessageEventStream } from "./utils/event-stream.js";
+
+describe("LLM stream facade", () => {
+  it("routes a prepared model through its lifecycle runtime", () => {
+    const registry = createApiRegistry();
+    const runtime = createLlmRuntime(registry);
+    const expected = createAssistantMessageEventStream();
+    const stream = vi.fn(() => expected);
+    registry.registerApiProvider({
+      api: "test-lifecycle-api",
+      stream,
+      streamSimple: stream,
+    });
+    const model = bindModelLlmRuntime(
+      {
+        api: "test-lifecycle-api",
+        provider: "test-provider",
+        id: "test-model",
+        name: "Test Model",
+        baseUrl: "https://example.test",
+        reasoning: false,
+        input: ["text"],
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow: 1024,
+        maxTokens: 512,
+      },
+      runtime,
+    );
+
+    expect(streamSimple(model, { messages: [] })).toBe(expected);
+    expect(stream).toHaveBeenCalledOnce();
+  });
+});
@@ -2,10 +2,54 @@
 // This facade owns the process-default AI runtime wiring: it installs the
 // OpenClaw host policy ports and registers built-in providers exactly once,
 // before any caller imports the stream API.
-import { defaultApiRegistry } from "@openclaw/ai/internal/runtime";
+import { defaultApiRegistry, defaultLlmRuntime } from "@openclaw/ai/internal/runtime";
 import { registerBuiltInApiProviders } from "@openclaw/ai/providers";
+import { getModelLlmRuntime } from "./model-runtime-binding.js";
 import "./ai-transport-host.js";
+import type {
+  Api,
+  AssistantMessage,
+  AssistantMessageEventStreamContract,
+  Context,
+  Model,
+  ProviderStreamOptions,
+  SimpleStreamOptions,
+} from "./types.js";

 registerBuiltInApiProviders(defaultApiRegistry);

-export { complete, completeSimple, stream, streamSimple } from "@openclaw/ai/internal/runtime";
+function resolveRuntime(model: Model) {
+  return getModelLlmRuntime(model) ?? defaultLlmRuntime;
+}
+
+export function stream<TApi extends Api>(
+  model: Model<TApi>,
+  context: Context,
+  options?: ProviderStreamOptions,
+): AssistantMessageEventStreamContract {
+  return resolveRuntime(model).stream(model, context, options);
+}
+
+export function complete<TApi extends Api>(
+  model: Model<TApi>,
+  context: Context,
+  options?: ProviderStreamOptions,
+): Promise<AssistantMessage> {
+  return resolveRuntime(model).complete(model, context, options);
+}
+
+export function streamSimple<TApi extends Api>(
+  model: Model<TApi>,
+  context: Context,
+  options?: SimpleStreamOptions,
+): AssistantMessageEventStreamContract {
+  return resolveRuntime(model).streamSimple(model, context, options);
+}
+
+export function completeSimple<TApi extends Api>(
+  model: Model<TApi>,
+  context: Context,
+  options?: SimpleStreamOptions,
+): Promise<AssistantMessage> {
+  return resolveRuntime(model).completeSimple(model, context, options);
+}
@@ -11,6 +11,8 @@ import { ensureOpenClawModelsJson } from "../agents/models-config.js";
 import { resolveProviderModelMaterializationAuthMode } from "../agents/provider-model-route-auth.js";
 import { protectPreparedProviderRuntimeAuth } from "../agents/provider-secret-egress.js";
 import { providerUsesCredentialScopedModelMetadata } from "../agents/runtime-plan/credential-scoped-model.js";
+import { getModelRegistryRuntime } from "../agents/sessions/model-registry-runtime.js";
+import { bindModelLlmRuntime } from "../llm/model-runtime-binding.js";
 import type { Model } from "../llm/types.js";
 import { prepareProviderRuntimeAuth } from "../plugins/provider-runtime.runtime.js";
 import type { ImageDescriptionRequest } from "./types.js";
@@ -60,6 +62,7 @@ async function prepareResolvedImageRuntime(
  modelRegistry: Awaited<ReturnType<typeof resolveModelAsync>>["modelRegistry"],
 ): Promise<{ apiKey: string; model: Model }> {
  let model = resolvedModel;
+  const modelRuntime = getModelRegistryRuntime(modelRegistry);
  const apiKeyInfo = await getApiKeyForModel({
    model,
    cfg: params.cfg,
@@ -113,7 +116,13 @@ async function prepareResolvedImageRuntime(
    apiKeyInfo.mode === "aws-sdk" &&
    model.api === "bedrock-converse-stream"
  ) {
-    return { apiKey: "", model: applySecretRefHeaderSentinels(model, params.cfg) };
+    return {
+      apiKey: "",
+      model: bindModelLlmRuntime(
+        applySecretRefHeaderSentinels(model, params.cfg),
+        modelRuntime.llmRuntime,
+      ),
+    };
  }
  let apiKey = requireApiKey(apiKeyInfo, model.provider);
  const preparedAuth = protectPreparedProviderRuntimeAuth({
@@ -142,7 +151,13 @@ async function prepareResolvedImageRuntime(
    model = { ...model, baseUrl: runtimeBaseUrl };
  }
  authStorage.setRuntimeApiKey(model.provider, apiKey);
-  return { apiKey, model: applySecretRefHeaderSentinels(model, params.cfg) };
+  return {
+    apiKey,
+    model: bindModelLlmRuntime(
+      applySecretRefHeaderSentinels(model, params.cfg),
+      modelRuntime.llmRuntime,
+    ),
+  };
 }

 export async function resolveImageRuntime(
@@ -153,7 +153,23 @@ vi.mock("../plugins/provider-runtime.runtime.js", () => ({
 }));

 vi.mock("../agents/embedded-agent-runner/model.js", () => ({
-  resolveModelAsync: resolveModelAsyncMock,
+  resolveModelAsync: async (...args: unknown[]) => {
+    const result = await resolveModelAsyncMock(...args);
+    const modelRegistry = (result?.modelRegistry ?? {}) as Record<string, unknown>;
+    return {
+      ...result,
+      modelRegistry: {
+        ...modelRegistry,
+        llmRuntime: modelRegistry.llmRuntime ?? { complete: completeMock },
+      },
+    };
+  },
+}));
+
+vi.mock("../agents/sessions/model-registry-runtime.js", () => ({
+  getModelRegistryRuntime: (owner: { llmRuntime?: unknown }) => ({
+    llmRuntime: owner.llmRuntime ?? { complete: completeMock },
+  }),
 }));

 vi.mock("../plugin-sdk/provider-auth.js", () => ({
@@ -1562,7 +1578,9 @@ describe("describeImageWithModel", () => {
    expect(resolveModelAsyncMock.mock.calls[2]?.[4]).toEqual(
      expect.objectContaining({
        authStorage,
-        modelRegistry,
+        modelRegistry: expect.objectContaining({
+          llmRuntime: expect.anything(),
+        }),
        authProfileId: "github-copilot:backup",
      }),
    );
@@ -435,6 +435,8 @@ async function describeImagesWithModelInternal(
    });
  }

+  // Resolved models carry their lifecycle runtime, so registration targets that
+  // registry before the built-in fallback reaches complete().
  const providerStreamFn = registerProviderStreamForModel({
    model,
    cfg: params.cfg,