@axiom-lattice/core 8.0.2 → 8.1.0

package/dist/index.js

@@ -1616,6 +1616,7 @@ __export(index_exports, {
   BUILTIN_PLUGINS: () => BUILTIN_PLUGINS,
   BUILTIN_SKILLS: () => BUILTIN_SKILLS,
   CAPABILITY_TOOL_NAME_CONFLICT: () => CAPABILITY_TOOL_NAME_CONFLICT,
+  COLLECTION_CHANGED_TOPIC: () => COLLECTION_CHANGED_TOPIC,
   CapabilityBundleMergeError: () => CapabilityBundleMergeError,
   CapabilityBundleMerger: () => CapabilityBundleMerger,
   CapabilityBundleValidationError: () => CapabilityBundleValidationError,
@@ -1635,6 +1636,7 @@ __export(index_exports, {
   DEFAULT_DELAY_MS: () => DEFAULT_DELAY_MS,
   DEFAULT_HANDLER_TIMEOUT_MS: () => DEFAULT_HANDLER_TIMEOUT_MS,
   DEFAULT_MAX_ATTEMPTS: () => DEFAULT_MAX_ATTEMPTS,
+  DEFAULT_MODEL_EXTRA: () => DEFAULT_MODEL_EXTRA,
   DEFAULT_RESIDENT_INVOKE_TIMEOUT_MS: () => DEFAULT_RESIDENT_INVOKE_TIMEOUT_MS,
   DaytonaInstance: () => DaytonaInstance,
   DaytonaProvider: () => DaytonaProvider,
@@ -1790,7 +1792,10 @@ __export(index_exports, {
   buildGrepResultsDict: () => buildGrepResultsDict,
   buildHandlerCommand: () => buildHandlerCommand,
   buildInput: () => buildInput,
+  buildModelConfig: () => buildModelConfig,
+  buildModelEmbeddingsConfig: () => buildModelEmbeddingsConfig,
   buildModelLlmConfig: () => buildModelLlmConfig,
+  buildModelSttConfig: () => buildModelSttConfig,
   buildNamedVolumeName: () => buildNamedVolumeName,
   buildRemotePackDirectoryCommand: () => buildRemotePackDirectoryCommand,
   buildSandboxMetadataEnv: () => buildSandboxMetadataEnv,
@@ -1811,6 +1816,7 @@ __export(index_exports, {
   compileSandboxPlugin: () => compileSandboxPlugin,
   compileWorkflow: () => compileWorkflow,
   computeSandboxName: () => computeSandboxName,
+  configureCollectionSync: () => configureCollectionSync,
   configureCoordinationLeaseStore: () => configureCoordinationLeaseStore,
   configurePluginCatalogSync: () => configurePluginCatalogSync,
   configureStores: () => configureStores,
@@ -1988,6 +1994,7 @@ __export(index_exports, {
   projectRoomPlugin: () => projectRoomPlugin,
   projectTaskAdmissionNotifier: () => projectTaskAdmissionNotifier,
   provisionSandboxOpenClient: () => provisionSandboxOpenClient,
+  publishCollectionChanged: () => publishCollectionChanged,
   publishPluginCatalogChanged: () => publishPluginCatalogChanged,
   queueLatticeManager: () => queueLatticeManager,
   readLoopBreakerConfig: () => readLoopBreakerConfig,
@@ -2281,9 +2288,6 @@ var ModelLattice = class extends import_chat_models.BaseChatModel {
   }
 };
 
-// src/model_lattice/ModelLatticeManager.ts
-var import_protocols10 = require("@axiom-lattice/protocols");
-
 // src/store_lattice/StoreLatticeManager.ts
 init_BaseLatticeManager();
 
@@ -3883,6 +3887,9 @@ var InMemoryModelProviderStore = class {
         // explicit flag marks one.
         stale: model.stale ?? false,
         extra: carried && Object.keys(carried).length > 0 ? carried : void 0,
+        // Same rule as `extra`: an omitted kind keeps the stored one, and a new
+        // row — which discovery creates without a kind — starts as chat.
+        kind: model.kind ?? existing?.kind ?? "chat",
         isDefault: model.isDefault ?? (namesDefault ? false : existing?.isDefault) ?? false,
         discoveredAt: now
       };
@@ -7941,7 +7948,136 @@ var SimpleMemoryVectorStore = class extends import_vectorstores.VectorStore {
 };
 
 // src/embeddings_lattice/EmbeddingsLatticeManager.ts
+var import_openai2 = require("@langchain/openai");
 init_BaseLatticeManager();
+
+// src/model_lattice/buildModelLlmConfig.ts
+var ALLOWED_LLM_PROVIDERS = /* @__PURE__ */ new Set(["openai"]);
+var UnsupportedModelProviderError = class extends Error {
+  constructor(llmProvider) {
+    super(`Unsupported LLM provider: ${llmProvider}`);
+    this.llmProvider = llmProvider;
+    this.code = "UNSUPPORTED_PROTOCOL";
+    this.name = "UnsupportedModelProviderError";
+  }
+};
+var DEFAULT_MODEL_EXTRA = {
+  disableThinking: true
+};
+function buildModelLlmConfig(provider2, model) {
+  if (!ALLOWED_LLM_PROVIDERS.has(provider2.llmProvider)) {
+    throw new UnsupportedModelProviderError(provider2.llmProvider);
+  }
+  const extra = { ...model.extra ?? {} };
+  delete extra.apiKeyEnvName;
+  const disableThinking = extra.disableThinking === true;
+  delete extra.disableThinking;
+  if (disableThinking) {
+    extra.modelKwargs = {
+      enable_thinking: false,
+      thinking: { type: "disabled" },
+      ...extra.modelKwargs ?? {}
+    };
+  }
+  return {
+    ...extra,
+    provider: provider2.llmProvider,
+    model: model.modelId,
+    displayName: model.displayName ?? model.modelId,
+    streaming: true,
+    apiKey: provider2.apiKey,
+    // A provider without a key talks to an upstream that needs no
+    // credentials. Saying so explicitly keeps `initChatModel` from falling
+    // back to `process.env[...]` and handing the gateway's own deployment key
+    // to the tenant's host.
+    noAuth: !provider2.apiKey,
+    baseURL: provider2.baseURL,
+    // Chat Completions and the Responses API are different routes with
+    // different request fields; without this the gateway would post
+    // `messages` to an endpoint expecting `input`.
+    ...provider2.apiStyle === "responses" ? { useResponsesApi: true } : {}
+  };
+}
+
+// src/model_lattice/buildModelConfig.ts
+function numberOrUndefined(value) {
+  return typeof value === "number" && Number.isFinite(value) ? value : void 0;
+}
+function buildModelEmbeddingsConfig(provider2, model) {
+  const extra = { ...model.extra ?? {} };
+  delete extra.apiKeyEnvName;
+  const encodingFormat = extra.encodingFormat === "float" || extra.encodingFormat === "base64" ? extra.encodingFormat : void 0;
+  return {
+    ...extra,
+    model: model.modelId,
+    baseURL: provider2.baseURL,
+    apiKey: provider2.apiKey || "EMPTY",
+    dimensions: numberOrUndefined(extra.dimensions),
+    timeout: numberOrUndefined(extra.timeout),
+    batchSize: numberOrUndefined(extra.batchSize),
+    encodingFormat
+  };
+}
+function buildModelSttConfig(provider2, model) {
+  const extra = { ...model.extra ?? {} };
+  delete extra.apiKeyEnvName;
+  return {
+    ...extra,
+    apiMode: extra.apiMode === "chat" ? "chat" : "whisper",
+    model: model.modelId,
+    // Placeholder for the same reason as embeddings: `STTModelLattice` falls
+    // back to `process.env.OPENAI_API_KEY` when this is empty.
+    apiKey: provider2.apiKey || "EMPTY",
+    baseURL: provider2.baseURL
+  };
+}
+function buildModelConfig(provider2, model) {
+  switch (model.kind) {
+    case "embedding":
+      return {
+        kind: "embedding",
+        config: buildModelEmbeddingsConfig(provider2, model)
+      };
+    case "asr":
+      return { kind: "asr", config: buildModelSttConfig(provider2, model) };
+    case "chat":
+    default:
+      return { kind: "chat", config: buildModelLlmConfig(provider2, model) };
+  }
+}
+
+// src/model_lattice/loadTenantModelRow.ts
+var import_protocols10 = require("@axiom-lattice/protocols");
+async function loadTenantModelRow(tenantId2, key5) {
+  try {
+    const { store } = getStoreLattice("default", "modelProvider");
+    const parsed = (0, import_protocols10.parseModelKey)(key5);
+    if (!parsed) {
+      return void 0;
+    }
+    const provider2 = await store.getProviderById(tenantId2, parsed.providerId);
+    if (!provider2 || !provider2.enabled) {
+      return void 0;
+    }
+    const model = await store.getModel(
+      tenantId2,
+      parsed.providerId,
+      parsed.modelId
+    );
+    if (!model || !model.enabled) {
+      return void 0;
+    }
+    return { provider: provider2, model };
+  } catch (error) {
+    console.error(
+      `[model_lattice] Failed to load model "${key5}" for tenant "${tenantId2}":`,
+      error
+    );
+    return void 0;
+  }
+}
+
+// src/embeddings_lattice/EmbeddingsLatticeManager.ts
 var EmbeddingsLatticeManager = class _EmbeddingsLatticeManager extends BaseLatticeManager {
   /**
    * Get EmbeddingsLatticeManager singleton instance
@@ -7973,6 +8109,37 @@ var EmbeddingsLatticeManager = class _EmbeddingsLatticeManager extends BaseLatti
     };
     this.register(key5, embeddingsLattice);
   }
+  /**
+   * Materialise a tenant embedding model that this process has never seen.
+   *
+   * Called by `BaseLatticeManager` on a registry miss. Only `kind: "embedding"`
+   * rows are accepted: a chat or ASR model key must resolve to nothing here, so
+   * a caller asking for an embedding client can never silently get the wrong
+   * kind of model.
+   *
+   * @param tenantId Tenant scope
+   * @param key `${providerId}/${modelId}`
+   */
+  async loadItemFromStore(tenantId2, key5) {
+    const row = await loadTenantModelRow(tenantId2, key5);
+    if (!row || row.model.kind !== "embedding") {
+      return void 0;
+    }
+    const config = buildModelEmbeddingsConfig(row.provider, row.model);
+    return {
+      key: key5,
+      client: new import_openai2.OpenAIEmbeddings({
+        model: config.model,
+        apiKey: config.apiKey,
+        dimensions: config.dimensions,
+        timeout: config.timeout,
+        batchSize: config.batchSize,
+        encodingFormat: config.encodingFormat,
+        configuration: { baseURL: config.baseURL }
+      }),
+      label: row.model.displayName ?? row.model.modelId
+    };
+  }
   /**
    * Get all registered embeddings as { key, label } info list
    */
@@ -7990,6 +8157,53 @@ var EmbeddingsLatticeManager = class _EmbeddingsLatticeManager extends BaseLatti
     }
     return embeddingsLattice;
   }
+  /**
+   * Drop this replica's cached embedding clients for one provider.
+   *
+   * The provider's config (apiKey, baseURL, model `extra`) is baked into the
+   * client at construction, and a resolved entry never re-reads the store. So a
+   * model-provider change reaches this replica only if the entry is evicted
+   * first — the same rule `ModelLatticeManager.evictProviderCache` applies to
+   * chat models, and the same keys the model-provider outbox event covers.
+   *
+   * Unlike the chat manager this deliberately leaves the `"default"` key alone:
+   * here it is a *deployment registration*, not a tenant resolution, so evicting
+   * it would remove the global fallback every tenant relies on.
+   *
+   * @param tenantId Tenant scope
+   * @param providerId Provider whose `${providerId}/` keys are stale
+   */
+  evictProviderCache(tenantId2, providerId) {
+    if (!providerId) {
+      return;
+    }
+    const prefix = `${providerId}/`;
+    for (const key5 of this.keysByTenant(tenantId2)) {
+      if (key5.startsWith(prefix)) {
+        this.removeWithTenant(tenantId2, key5);
+      }
+    }
+  }
+  /**
+   * Resolve an embeddings client for a tenant, loading a miss from the store.
+   *
+   * Mirrors `resolveModelLatticeForTenant`: the tenant's own model first, then
+   * the deployment namespace, which is where boot-time registrations live.
+   * Returns `undefined` rather than throwing, so a caller can fall back.
+   *
+   * @param tenantId Tenant scope
+   * @param key `${providerId}/${modelId}` or a bare deployment key
+   */
+  async getEmbeddingsClientWithTenant(tenantId2, key5) {
+    const own = await this.getOrLoadWithTenant(tenantId2, key5);
+    if (own) {
+      return own.client;
+    }
+    if (tenantId2 !== "default") {
+      return (await this.getOrLoadWithTenant("default", key5))?.client;
+    }
+    return void 0;
+  }
   /**
    * Get all Lattices
    */
@@ -8045,20 +8259,18 @@ var getEmbeddingsClient = (key5) => embeddingsLatticeManager.getEmbeddingsClient
 // src/vectorstore_lattice/VectorStoreProviderManager.ts
 var MemoryVectorStoreProvider = class {
   async create(params) {
-    let embeddings;
-    try {
-      embeddings = embeddingsLatticeManager.getEmbeddingsClient(params.embeddingKey);
-    } catch {
-      try {
-        embeddings = embeddingsLatticeManager.getEmbeddingsClient("default");
-      } catch {
-        throw new Error(
-          `No embedding model registered. Register one first, e.g.:
-  import { registerEmbeddingsLattice } from "@axiom-lattice/core";
-  import { OpenAIEmbeddings } from "@langchain/openai";
-  registerEmbeddingsLattice("default", new OpenAIEmbeddings({ modelName: "text-embedding-3-small" }));`
-        );
-      }
+    const tenantId2 = params.tenantId ?? "default";
+    const embeddings = await embeddingsLatticeManager.getEmbeddingsClientWithTenant(
+      tenantId2,
+      params.embeddingKey
+    ) ?? await embeddingsLatticeManager.getEmbeddingsClientWithTenant(
+      tenantId2,
+      "default"
+    );
+    if (!embeddings) {
+      throw new Error(
+        `No embedding model "${params.embeddingKey}" registered for tenant "${tenantId2}", and no deployment default. Configure one in the model provider panel.`
+      );
     }
     return new SimpleMemoryVectorStore(embeddings);
   }
@@ -12418,42 +12630,6 @@ var InMemoryMenuStore = class {
   }
 };
 
-// src/model_lattice/buildModelLlmConfig.ts
-var ALLOWED_LLM_PROVIDERS = /* @__PURE__ */ new Set(["openai"]);
-var UnsupportedModelProviderError = class extends Error {
-  constructor(llmProvider) {
-    super(`Unsupported LLM provider: ${llmProvider}`);
-    this.llmProvider = llmProvider;
-    this.code = "UNSUPPORTED_PROTOCOL";
-    this.name = "UnsupportedModelProviderError";
-  }
-};
-function buildModelLlmConfig(provider2, model) {
-  if (!ALLOWED_LLM_PROVIDERS.has(provider2.llmProvider)) {
-    throw new UnsupportedModelProviderError(provider2.llmProvider);
-  }
-  const extra = { ...model.extra ?? {} };
-  delete extra.apiKeyEnvName;
-  return {
-    ...extra,
-    provider: provider2.llmProvider,
-    model: model.modelId,
-    displayName: model.displayName ?? model.modelId,
-    streaming: true,
-    apiKey: provider2.apiKey,
-    // A provider without a key talks to an upstream that needs no
-    // credentials. Saying so explicitly keeps `initChatModel` from falling
-    // back to `process.env[...]` and handing the gateway's own deployment key
-    // to the tenant's host.
-    noAuth: !provider2.apiKey,
-    baseURL: provider2.baseURL,
-    // Chat Completions and the Responses API are different routes with
-    // different request fields; without this the gateway would post
-    // `messages` to an endpoint expecting `input`.
-    ...provider2.apiStyle === "responses" ? { useResponsesApi: true } : {}
-  };
-}
-
 // src/model_lattice/ModelLatticeManager.ts
 var DEFAULT_MODEL_KEY = "default";
 var TENANT_DEFAULT_MISS_TTL_MS = 3e4;
@@ -12626,38 +12802,26 @@ var ModelLatticeManager = class _ModelLatticeManager extends BaseLatticeManager
    * @param key Lattice键名（`${providerId}/${modelId}`）
    */
   async loadItemFromStore(tenantId2, key5) {
-    try {
-      const { store } = getStoreLattice("default", "modelProvider");
-      if (key5 === DEFAULT_MODEL_KEY) {
+    if (key5 === DEFAULT_MODEL_KEY) {
+      try {
+        const { store } = getStoreLattice("default", "modelProvider");
         return await this.loadTenantDefault(store, tenantId2);
-      }
-      const parsed = (0, import_protocols10.parseModelKey)(key5);
-      if (!parsed) {
-        return void 0;
-      }
-      const provider2 = await store.getProviderById(tenantId2, parsed.providerId);
-      if (!provider2 || !provider2.enabled) {
-        return void 0;
-      }
-      const model = await store.getModel(
-        tenantId2,
-        parsed.providerId,
-        parsed.modelId
-      );
-      if (!model || !model.enabled) {
+      } catch (error) {
+        console.error(
+          `[ModelLatticeManager] Failed to load model "${key5}" for tenant "${tenantId2}":`,
+          error
+        );
         return void 0;
       }
-      return {
-        key: key5,
-        client: new ModelLattice(buildModelLlmConfig(provider2, model))
-      };
-    } catch (error) {
-      console.error(
-        `[ModelLatticeManager] Failed to load model "${key5}" for tenant "${tenantId2}":`,
-        error
-      );
+    }
+    const row = await loadTenantModelRow(tenantId2, key5);
+    if (!row || row.model.kind !== "chat") {
       return void 0;
     }
+    return {
+      key: key5,
+      client: new ModelLattice(buildModelLlmConfig(row.provider, row.model))
+    };
   }
   /**
    * Forget that a tenant had no default model.
@@ -12689,7 +12853,9 @@ var ModelLatticeManager = class _ModelLatticeManager extends BaseLatticeManager
       return void 0;
     }
     const models = await store.listModels(tenantId2);
-    const marked = models.find((model) => model.isDefault === true);
+    const marked = models.find(
+      (model) => model.isDefault === true && model.kind === "chat"
+    );
     if (!marked || !marked.enabled) {
       this.tenantDefaultMissAt.set(tenantId2, Date.now());
       return void 0;
@@ -12730,6 +12896,37 @@ var ModelLatticeManager = class _ModelLatticeManager extends BaseLatticeManager
   removeLatticeWithTenant(tenantId2, key5) {
     return this.removeWithTenant(tenantId2, key5);
   }
+  /**
+   * 驱逐一个 provider 贡献的缓存，外加本租户已解析的 `default` 哨兵。
+   *
+   * 改了配置（apiKey / baseURL / extra）却还在走旧客户端，根因就是
+   * `BaseLatticeManager.registry` 的**正向条目没有 TTL**：命中一次之后
+   * `loadItemFromStore` 再也不会被调用。驱逐之后，下一次解析会从**共享的**
+   * store 重新物化 —— 这正是多副本「正确性不靠副本之间同步、靠共享库」那条不变式。
+   *
+   * 只驱逐、不重注册：`loadItemFromStore` 已经知道怎么按 key 物化，删掉就够，
+   * 而且这里不需要读库。
+   *
+   * 为什么连哨兵一起删：`default` 是本租户默认模型的**解析结果**，同样被永久缓存。
+   * 不删它，租户换了默认模型就得重启才生效（单副本也一样）。删掉只让下一次解析
+   * 多读一次库。
+   *
+   * @param tenantId 租户ID
+   * @param providerId 该 provider 的 `${providerId}/` 前缀会被清掉；省略则只清
+   *   哨兵（例如只换了默认模型）
+   */
+  evictProviderCache(tenantId2, providerId) {
+    if (providerId) {
+      const prefix = `${providerId}/`;
+      for (const key5 of this.keysByTenant(tenantId2)) {
+        if (key5.startsWith(prefix)) {
+          this.removeWithTenant(tenantId2, key5);
+        }
+      }
+    }
+    this.removeWithTenant(tenantId2, DEFAULT_MODEL_KEY);
+    this.forgetTenantDefaultMiss(tenantId2);
+  }
 };
 var modelLatticeManager = ModelLatticeManager.getInstance();
 var registerModelLattice = (key5, config) => modelLatticeManager.registerLattice(key5, config);
@@ -17732,7 +17929,7 @@ async function describeImage(options) {
 init_BaseLatticeManager();
 
 // src/stt_model_lattice/STTModelLattice.ts
-var import_openai2 = require("@langchain/openai");
+var import_openai3 = require("@langchain/openai");
 var STTModelLattice = class {
   constructor(key5, config) {
     this.key = key5;
@@ -17742,7 +17939,7 @@ var STTModelLattice = class {
     if (!apiKey) {
       throw new Error("No API key configured for STT. Set apiKey, apiKeyEnvName, or OPENAI_API_KEY.");
     }
-    this.openaiClient = new import_openai2.OpenAIClient({
+    this.openaiClient = new import_openai3.OpenAIClient({
       apiKey,
       baseURL: config.baseURL,
       timeout: config.timeout || 3e4
@@ -17764,7 +17961,7 @@ var STTModelLattice = class {
    */
   async transcribeViaWhisper(audio, format) {
     const mimeType = formatToMimeType(format);
-    const file = typeof File !== "undefined" ? new File([audio], `audio.${format}`, { type: mimeType }) : await (0, import_openai2.toFile)(audio, `audio.${format}`, { type: mimeType });
+    const file = typeof File !== "undefined" ? new File([audio], `audio.${format}`, { type: mimeType }) : await (0, import_openai3.toFile)(audio, `audio.${format}`, { type: mimeType });
     const extra = this.config.extra || {};
     const response = await this.openaiClient.audio.transcriptions.create({
       file,
@@ -17858,6 +18055,26 @@ var STTModelLatticeManager = class _STTModelLatticeManager extends BaseLatticeMa
     };
     this.register(key5, sttLattice);
   }
+  /**
+   * Materialise a tenant ASR model that this process has never seen.
+   *
+   * Only `kind: "asr"` rows are accepted, mirroring the embedding loader: a
+   * chat key must never resolve to a speech client.
+   */
+  async loadItemFromStore(tenantId2, key5) {
+    const row = await loadTenantModelRow(tenantId2, key5);
+    if (!row || row.model.kind !== "asr") {
+      return void 0;
+    }
+    const config = buildModelSttConfig(row.provider, row.model);
+    const client = new STTModelLattice(key5, config);
+    return {
+      key: key5,
+      client,
+      config,
+      label: row.model.displayName ?? row.model.modelId
+    };
+  }
   /**
    * Get an STT model lattice by key.
    */
@@ -17889,6 +18106,49 @@ var STTModelLatticeManager = class _STTModelLatticeManager extends BaseLatticeMa
     }
     return lattice.client;
   }
+  /**
+   * Resolve an STT client for a tenant, loading a miss from the store.
+   *
+   * The synchronous `getSTTClientWithTenant` stays as it is for
+   * deployment-registered models; this one is for the tenant catalog, where the
+   * model may exist only in the shared store. Same order as every other
+   * resolver: tenant first, then the deployment namespace.
+   *
+   * @param tenantId Tenant scope
+   * @param key `${providerId}/${modelId}` or a bare deployment key
+   */
+  async getSTTClientWithTenantAsync(tenantId2, key5) {
+    const own = await this.getOrLoadWithTenant(tenantId2, key5);
+    if (own) {
+      return own.client;
+    }
+    if (tenantId2 !== "default") {
+      return (await this.getOrLoadWithTenant("default", key5))?.client;
+    }
+    return void 0;
+  }
+  /**
+   * Drop this replica's cached STT clients for one provider.
+   *
+   * Same rule as the chat and embeddings managers: the provider's key and
+   * address are baked into the client at construction, so a model-provider
+   * change only reaches this replica if the cached entry is evicted first. The
+   * deployment's `"default"` registration is left alone.
+   *
+   * @param tenantId Tenant scope
+   * @param providerId Provider whose `${providerId}/` keys are stale
+   */
+  evictProviderCache(tenantId2, providerId) {
+    if (!providerId) {
+      return;
+    }
+    const prefix = `${providerId}/`;
+    for (const key5 of this.keysByTenant(tenantId2)) {
+      if (key5.startsWith(prefix)) {
+        this.removeWithTenant(tenantId2, key5);
+      }
+    }
+  }
   /**
    * Get all registered STT models as info list.
    */
@@ -17927,13 +18187,16 @@ var getSTTClientWithTenant = (tenantId2, key5) => sttModelLatticeManager.getSTTC
 
 // src/deep_agent_new/backends/transcribeAudio.ts
 async function transcribeAudio(options) {
-  const { audioBuffer, format } = options;
-  if (!sttModelLatticeManager.hasLattice("default")) {
+  const { audioBuffer, format, tenantId: tenantId2 = "default", modelKey = "default" } = options;
+  const client = await sttModelLatticeManager.getSTTClientWithTenantAsync(
+    tenantId2,
+    modelKey
+  );
+  if (!client) {
     throw new Error(
-      "No default STT model registered. Use registerSTTModelLattice('default', { ... }) first."
+      `No STT model "${modelKey}" registered for tenant "${tenantId2}". Configure one in the model provider panel, or register a deployment default.`
     );
   }
-  const client = sttModelLatticeManager.getSTTClient("default");
   return await client.transcribe(audioBuffer, format);
 }
 
@@ -18108,7 +18371,13 @@ The current backend does not support binary reading, unable to process audio.`;
 ${sizeWarning}`;
           }
           const format = detectAudioFormat(file_path);
-          const result = await transcribeAudio({ audioBuffer: buffer2, format });
+          const result = await transcribeAudio({
+            audioBuffer: buffer2,
+            format,
+            // Tenant first; the deployment's `"default"` registration is the
+            // fallback, so existing deployments keep working unchanged.
+            tenantId: runConfig?.tenantId || "default"
+          });
           let output = `[Audio] ${file_path} (${format}, ${(buffer2.length / 1024).toFixed(1)}KB)`;
           if (result.confidence !== void 0) {
             output += `
@@ -33992,6 +34261,28 @@ var VectorStoreLatticeManager = class _VectorStoreLatticeManager extends BaseLat
     const vectorStoreLattice = this.getVectorStoreLattice(key5);
     return vectorStoreLattice.client;
   }
+  /**
+   * Drop every cached vector store belonging to a tenant.
+   *
+   * Stores are registered through the legacy key-only API with a
+   * `${tenantId}:${collection}` key name, so they all live under the `default`
+   * namespace. Matching has to be on the key name, not on the lattice's tenant
+   * prefix — `clearByTenant` would find nothing here.
+   *
+   * The stores are rebuilt lazily on next use. Evicting is required whenever the
+   * embedding client they were built with may have changed: a collection's
+   * `embeddingKey`, or the model provider behind that key.
+   *
+   * @param tenantId Tenant whose vector stores to drop
+   */
+  clearTenantVectorStores(tenantId2) {
+    const prefix = `${tenantId2}:`;
+    for (const lattice of this.getAllLattices()) {
+      if (lattice.key.startsWith(prefix)) {
+        this.removeLattice(lattice.key);
+      }
+    }
+  }
 };
 var vectorStoreLatticeManager = VectorStoreLatticeManager.getInstance();
 var registerVectorStoreLattice = (key5, vectorStore) => vectorStoreLatticeManager.registerLattice(key5, vectorStore);
@@ -34685,6 +34976,26 @@ var SkillLatticeManager = class _SkillLatticeManager extends BaseLatticeManager
 };
 var skillLatticeManager = SkillLatticeManager.getInstance();
 
+// src/services/collectionSync.ts
+var COLLECTION_CHANGED_TOPIC = "collection-changed";
+var outbox;
+var owner = "unknown";
+function configureCollectionSync(next) {
+  outbox = next?.outbox;
+  owner = next?.owner ?? "unknown";
+}
+async function publishCollectionChanged(tenantId2) {
+  if (!outbox || !tenantId2) return;
+  try {
+    await outbox.append({
+      topic: COLLECTION_CHANGED_TOPIC,
+      payload: { tenantId: tenantId2 },
+      originReplicaId: owner
+    });
+  } catch {
+  }
+}
+
 // src/collection_lattice/CollectionLatticeManager.ts
 var CollectionLatticeManager = class _CollectionLatticeManager {
   static getInstance() {
@@ -34697,8 +35008,25 @@ var CollectionLatticeManager = class _CollectionLatticeManager {
     const lattice = storeLatticeManager.getStoreLattice("default", "collection");
     return lattice.store;
   }
+  /**
+   * Drop this replica's cached vector stores for the tenant, and tell peers.
+   *
+   * A vector store is constructed with one embedding client, so any collection
+   * write can invalidate it — creating, renaming, changing `embeddingKey` or
+   * deleting. Clearing the whole tenant is deliberate: the cache is small, and
+   * a per-store match would still miss two collections sharing one client.
+   *
+   * Locally this must happen here because the outbox poller skips events this
+   * replica published itself.
+   */
+  invalidateVectorStores(tenantId2) {
+    vectorStoreLatticeManager.clearTenantVectorStores(tenantId2);
+    void publishCollectionChanged(tenantId2);
+  }
   async registerCollection(tenantId2, data) {
-    return this.getStore().createCollection(tenantId2, data);
+    const collection = await this.getStore().createCollection(tenantId2, data);
+    this.invalidateVectorStores(tenantId2);
+    return collection;
   }
   async getCollection(tenantId2, name) {
     return this.getStore().getCollectionByName(tenantId2, name);
@@ -34707,10 +35035,18 @@ var CollectionLatticeManager = class _CollectionLatticeManager {
     return this.getStore().getAllCollections(tenantId2);
   }
   async removeCollection(tenantId2, name) {
-    return this.getStore().deleteCollection(tenantId2, name);
+    const removed = await this.getStore().deleteCollection(tenantId2, name);
+    if (removed) {
+      this.invalidateVectorStores(tenantId2);
+    }
+    return removed;
   }
   async updateCollection(tenantId2, name, updates) {
-    return this.getStore().updateCollection(tenantId2, name, updates);
+    const updated = await this.getStore().updateCollection(tenantId2, name, updates);
+    if (updated) {
+      this.invalidateVectorStores(tenantId2);
+    }
+    return updated;
   }
 };
 var collectionLatticeManager = CollectionLatticeManager.getInstance();
@@ -34734,7 +35070,11 @@ async function getOrCreateCollectionVectorStore(name, embeddingKey, tenantId2, p
     const initialization = (async () => {
       const provider2 = vectorStoreProviderManager.getProvider(providerKey);
       const tableName = buildTableName(tenantId2, name);
-      const vectorStore = await provider2.create({ tableName, embeddingKey });
+      const vectorStore = await provider2.create({
+        tableName,
+        embeddingKey,
+        tenantId: tenantId2
+      });
       vectorStoreLatticeManager.registerLattice(storeKey, vectorStore);
       return vectorStore;
     })();
@@ -39136,8 +39476,8 @@ var TaskLifecycleNotifier = class {
     this.tails = /* @__PURE__ */ new Map();
   }
   /** Installs (or clears) the durable outbox used to fan notifications to other replicas. */
-  setOutbox(outbox2) {
-    this.outbox = outbox2;
+  setOutbox(outbox3) {
+    this.outbox = outbox3;
   }
   /**
    * Registers an async lifecycle listener.
@@ -39243,8 +39583,8 @@ var ProjectTaskAdmissionNotifier = class {
     this.listenerTimeoutMs = timeout;
   }
   /** Installs (or clears) the durable outbox used to fan notifications to other replicas. */
-  setOutbox(outbox2) {
-    this.outbox = outbox2;
+  setOutbox(outbox3) {
+    this.outbox = outbox3;
   }
   /**
    * Registers a Project task admission listener.
@@ -40091,9 +40431,9 @@ var BELIEF_OWNER_WRITE_ATTEMPTS = 3;
 function isWellFormedImpact(impact) {
   return /^[a-z0-9]+(?:-[a-z0-9]+)*$/.test(impact.key) && Number.isInteger(impact.after) && impact.after >= 0 && impact.after <= 100 && impact.basis.trim().length > 0 && impact.basis.length <= 1e3 && !/[\r\n]/.test(impact.basis);
 }
-async function applyBeliefImpacts(taskStore, owner2, impacts) {
-  if (impacts.length === 0) return { description: owner2.description ?? "" };
-  let current = owner2;
+async function applyBeliefImpacts(taskStore, owner3, impacts) {
+  if (impacts.length === 0) return { description: owner3.description ?? "" };
+  let current = owner3;
   for (let attempt = 0; attempt < BELIEF_OWNER_WRITE_ATTEMPTS; attempt += 1) {
     const parsed = (0, import_protocols32.parseTaskBeliefState)(current.description ?? "");
     const existingTable = parsed.success;
@@ -40222,14 +40562,14 @@ function sameBeliefOwner(left, right) {
 function sameBeliefOwnerIdentity(left, right) {
   return canonicalJson({ ...left, beliefState: void 0 }) === canonicalJson({ ...right, beliefState: void 0 });
 }
-function canonicalBeliefOwner(owner2) {
+function canonicalBeliefOwner(owner3) {
   const beliefState = {
-    entries: owner2.beliefState.entries.map((entry) => ({
+    entries: owner3.beliefState.entries.map((entry) => ({
       ...entry,
       basis: entry.basis.trim().replace(/\s+/g, " ")
     })).sort((left, right) => left.key < right.key ? -1 : left.key > right.key ? 1 : 0)
   };
-  return { ...owner2, beliefState };
+  return { ...owner3, beliefState };
 }
 function eventKeyFor(taskId, result, impacts, beliefOwner) {
   const base = `belief-impact:${taskId}:${buildBeliefCompletionEventKey(taskId, result, impacts)}`;
@@ -41600,8 +41940,8 @@ var TaskLifecycleService = class {
     return { task: existing, parent: parentResult, eventKey, evidence, result, impacts };
   }
   beliefOwnerSnapshot(task, parent) {
-    const owner2 = parent?.ownerType === "agent" ? parent : task;
-    const parsed = (0, import_protocols32.parseTaskBeliefState)(owner2.description ?? "");
+    const owner3 = parent?.ownerType === "agent" ? parent : task;
+    const parsed = (0, import_protocols32.parseTaskBeliefState)(owner3.description ?? "");
     const beliefState = parsed.success ? parsed.state : { entries: [] };
     if (parent?.ownerType === "agent") {
       return {
@@ -43387,8 +43727,8 @@ function createTaskExecutionService(deps) {
 var import_node_crypto14 = require("crypto");
 async function runAsLeader(store, name, fn, options = {}) {
   const key5 = `leader:${name}`;
-  const owner2 = options.owner ?? getReplicaId();
-  const holder = `${owner2}#${(0, import_node_crypto14.randomUUID)()}`;
+  const owner3 = options.owner ?? getReplicaId();
+  const holder = `${owner3}#${(0, import_node_crypto14.randomUUID)()}`;
   const ttlMs = options.ttlMs ?? 6e4;
   if (!await store.tryAcquire(key5, holder, ttlMs)) return { ran: false };
   const heartbeat = setInterval(() => {
@@ -43406,16 +43746,16 @@ async function runAsLeader(store, name, fn, options = {}) {
 
 // src/services/pluginCatalogSync.ts
 var PLUGIN_CATALOG_CHANGED_TOPIC = "plugin-catalog-changed";
-var outbox;
-var owner = "unknown";
+var outbox2;
+var owner2 = "unknown";
 function configurePluginCatalogSync(next) {
-  outbox = next?.outbox;
-  owner = next?.owner ?? "unknown";
+  outbox2 = next?.outbox;
+  owner2 = next?.owner ?? "unknown";
 }
 async function publishPluginCatalogChanged(tenantId2) {
-  if (!outbox || !tenantId2) return;
+  if (!outbox2 || !tenantId2) return;
   try {
-    await outbox.append({ topic: PLUGIN_CATALOG_CHANGED_TOPIC, payload: { tenantId: tenantId2 }, originReplicaId: owner });
+    await outbox2.append({ topic: PLUGIN_CATALOG_CHANGED_TOPIC, payload: { tenantId: tenantId2 }, originReplicaId: owner2 });
   } catch {
   }
 }
@@ -43601,9 +43941,9 @@ var createLoadSkillsTool = ({ skills, readAll, pluginSkills, pluginSkillOwners }
           if (!skillName) continue;
           if (await isPluginSkillManagedInSandbox(sandbox, skillName)) {
             const enabledDefinition = pluginSkills?.[skillName];
-            const owner2 = pluginSkillOwners?.[skillName];
-            const registered = owner2 && enabledDefinition ? PluginRegistry.getSkillForPlugin(owner2, skillName) : void 0;
-            if (!registered || registered.definition !== enabledDefinition || !owner2 || !await isPluginSkillAuthorizedInSandbox(sandbox, skillName, owner2) || await getPluginSkillManifestVersion(sandbox, skillName) !== enabledDefinition.version) {
+            const owner3 = pluginSkillOwners?.[skillName];
+            const registered = owner3 && enabledDefinition ? PluginRegistry.getSkillForPlugin(owner3, skillName) : void 0;
+            if (!registered || registered.definition !== enabledDefinition || !owner3 || !await isPluginSkillAuthorizedInSandbox(sandbox, skillName, owner3) || await getPluginSkillManifestVersion(sandbox, skillName) !== enabledDefinition.version) {
               continue;
             }
           }
@@ -43629,8 +43969,8 @@ var createLoadSkillsTool = ({ skills, readAll, pluginSkills, pluginSkillOwners }
       const seenNames = new Set(filteredSkills.map((skill) => skill.name));
       try {
         for (const [name, definition] of Object.entries(pluginSkills ?? {})) {
-          const owner2 = pluginSkillOwners?.[name];
-          const registered = owner2 ? PluginRegistry.getSkillForPlugin(owner2, name) : void 0;
+          const owner3 = pluginSkillOwners?.[name];
+          const registered = owner3 ? PluginRegistry.getSkillForPlugin(owner3, name) : void 0;
           if (!registered || registered.definition !== definition) continue;
           const { meta } = parseSkillFrontmatter(definition.content);
           if (!meta.name || !meta.description || seenNames.has(meta.name)) continue;
@@ -43667,24 +44007,24 @@ var import_zod52 = __toESM(require("zod"));
 var import_langchain63 = require("langchain");
 
 // src/sandbox_plugin/pluginSkillResolution.ts
-async function resolveTenantPluginSkill(tenantId2, owner2, skillName) {
+async function resolveTenantPluginSkill(tenantId2, owner3, skillName) {
   try {
     const { plugins } = await getSandboxPluginCatalog().ensureLoaded(tenantId2);
-    const plugin = plugins.find((candidate) => candidate.meta.type === owner2);
+    const plugin = plugins.find((candidate) => candidate.meta.type === owner3);
     return plugin?.skills?.[skillName];
   } catch {
     return void 0;
   }
 }
-async function resolveEnabledPluginSkill(owner2, skillName, enabledDefinition, tenantId2) {
-  if (!owner2 || !enabledDefinition) return void 0;
-  const registered = PluginRegistry.getSkillForPlugin(owner2, skillName);
-  if (registered && registered.pluginType === owner2 && registered.definition === enabledDefinition) {
-    return { pluginType: owner2, definition: enabledDefinition };
+async function resolveEnabledPluginSkill(owner3, skillName, enabledDefinition, tenantId2) {
+  if (!owner3 || !enabledDefinition) return void 0;
+  const registered = PluginRegistry.getSkillForPlugin(owner3, skillName);
+  if (registered && registered.pluginType === owner3 && registered.definition === enabledDefinition) {
+    return { pluginType: owner3, definition: enabledDefinition };
   }
   if (!tenantId2) return void 0;
-  const tenantDefinition = await resolveTenantPluginSkill(tenantId2, owner2, skillName);
-  return tenantDefinition ? { pluginType: owner2, definition: tenantDefinition } : void 0;
+  const tenantDefinition = await resolveTenantPluginSkill(tenantId2, owner3, skillName);
+  return tenantDefinition ? { pluginType: owner3, definition: tenantDefinition } : void 0;
 }
 
 // src/tool_lattice/skill/load_skill_content.ts
@@ -43746,10 +44086,10 @@ var createLoadSkillContentTool = (pluginSkills, pluginSkillOwnersOrStaticSelecti
           return result2;
         }
         const enabledDefinition = pluginSkills?.[input.skill_name];
-        const owner2 = pluginSkillOwners?.[input.skill_name];
+        const owner3 = pluginSkillOwners?.[input.skill_name];
         const runConfig = _exe_config?.configurable?.runConfig || {};
         const pluginSkill = await resolveEnabledPluginSkill(
-          owner2,
+          owner3,
           input.skill_name,
           enabledDefinition,
           runConfig.tenantId
@@ -43856,9 +44196,9 @@ var createLoadSkillResourceTool = (pluginSkills, pluginSkillOwners) => {
         validateRelativePath(input.resource_path, "resource");
         const sandbox = await getSandboxFromExeConfig3(_exe_config);
         const definition = pluginSkills?.[input.skill_name];
-        const owner2 = pluginSkillOwners?.[input.skill_name];
+        const owner3 = pluginSkillOwners?.[input.skill_name];
         const tenantId2 = _exe_config?.configurable?.runConfig?.tenantId;
-        const resolved = await resolveEnabledPluginSkill(owner2, input.skill_name, definition, tenantId2);
+        const resolved = await resolveEnabledPluginSkill(owner3, input.skill_name, definition, tenantId2);
         if (definition) {
           if (!resolved || !resolved.definition.resources?.[input.resource_path]) {
             return `Resource "${input.resource_path}" not found for skill "${input.skill_name}".`;
@@ -44222,11 +44562,15 @@ var createSearchCollectionTool = (scope = { collectionKeys: [], connectAll: true
             tenantId2
           );
         }
-        let embeddingsClient;
-        try {
-          embeddingsClient = embeddingsLatticeManager.getEmbeddingsClient(embeddingKey);
-        } catch {
-          embeddingsClient = embeddingsLatticeManager.getEmbeddingsClient("default");
+        const embeddingsClient = await embeddingsLatticeManager.getEmbeddingsClientWithTenant(
+          tenantId2,
+          embeddingKey
+        ) ?? await embeddingsLatticeManager.getEmbeddingsClientWithTenant(
+          tenantId2,
+          "default"
+        );
+        if (!embeddingsClient) {
+          return `No embedding model "${embeddingKey}" is available for this tenant.`;
         }
         const queryVector = await embeddingsClient.embedQuery(query);
         const results = await vectorStore.similaritySearchVectorWithScore(
@@ -52326,6 +52670,7 @@ registerBuiltinPlugins();
   BUILTIN_PLUGINS,
   BUILTIN_SKILLS,
   CAPABILITY_TOOL_NAME_CONFLICT,
+  COLLECTION_CHANGED_TOPIC,
   CapabilityBundleMergeError,
   CapabilityBundleMerger,
   CapabilityBundleValidationError,
@@ -52345,6 +52690,7 @@ registerBuiltinPlugins();
   DEFAULT_DELAY_MS,
   DEFAULT_HANDLER_TIMEOUT_MS,
   DEFAULT_MAX_ATTEMPTS,
+  DEFAULT_MODEL_EXTRA,
   DEFAULT_RESIDENT_INVOKE_TIMEOUT_MS,
   DaytonaInstance,
   DaytonaProvider,
@@ -52500,7 +52846,10 @@ registerBuiltinPlugins();
   buildGrepResultsDict,
   buildHandlerCommand,
   buildInput,
+  buildModelConfig,
+  buildModelEmbeddingsConfig,
   buildModelLlmConfig,
+  buildModelSttConfig,
   buildNamedVolumeName,
   buildRemotePackDirectoryCommand,
   buildSandboxMetadataEnv,
@@ -52521,6 +52870,7 @@ registerBuiltinPlugins();
   compileSandboxPlugin,
   compileWorkflow,
   computeSandboxName,
+  configureCollectionSync,
   configureCoordinationLeaseStore,
   configurePluginCatalogSync,
   configureStores,
@@ -52698,6 +53048,7 @@ registerBuiltinPlugins();
   projectRoomPlugin,
   projectTaskAdmissionNotifier,
   provisionSandboxOpenClient,
+  publishCollectionChanged,
   publishPluginCatalogChanged,
   queueLatticeManager,
   readLoopBreakerConfig,