@letta-ai/letta-code 0.29.3 → 0.29.5

package/letta.js

@@ -5462,7 +5462,7 @@ var package_default;
 var init_package = __esm(() => {
   package_default = {
     name: "@letta-ai/letta-code",
-    version: "0.29.3",
+    version: "0.29.5",
     description: "Letta Code is a CLI tool for interacting with stateful Letta agents from the terminal.",
     type: "module",
     packageManager: "bun@1.3.0",
@@ -127383,6 +127383,9 @@ function validatePiProviderRegistration(providerName, config3) {
 }
 
 // src/backend/dev/pi-provider-mod-registry.ts
+function getPiProviderRegistryRevision() {
+  return revisionCounter;
+}
 function bumpProviderRevision(providerName) {
   revisionCounter += 1;
   providerRevisions.set(providerName, revisionCounter);
@@ -345981,6 +345984,7 @@ async function loadLocalMods(options3) {
   const builtinCommandIds = new Set([...options3.builtinCommandIds ?? []]);
   const reservedToolNames = new Set([...options3.reservedToolNames ?? []]);
   const registry2 = createEmptyModRegistry(sources, generation2, capabilities, options3.registerCapabilitiesGlobally !== false);
+  options3.onRegistryCreated?.(registry2);
   for (const source2 of sources) {
     for (const diagnostic of source2.diagnostics ?? []) {
       const owner = createModOwner(diagnostic.path, source2, generation2);
@@ -346191,6 +346195,13 @@ function createModEngine(options3) {
     const nextRegistry = await loadLocalMods({
       ...modOptions,
       generation: loadGeneration,
+      onRegistryCreated: (registry2) => {
+        loadingRegistry = registry2;
+        if (!disposed && loadGeneration === generation2) {
+          activeRegistry = registry2;
+          publish();
+        }
+      },
       onChange: () => {
         if (!disposed && loadingRegistry && loadGeneration === generation2) {
           activeRegistry = loadingRegistry;
@@ -375875,7 +375886,7 @@ class LocalStore {
     const agentsDir = join42(this.storageDir, "agents");
     if (existsSync36(agentsDir)) {
       for (const file3 of readdirSync13(agentsDir)) {
-        if (!file3.endsWith(".json"))
+        if (!file3.endsWith(".json") || file3.startsWith("._"))
           continue;
         const raw = readJsonFile2(join42(agentsDir, file3));
         const agent2 = normalizeAgentRecord(raw, this.defaultAgentModel);
@@ -378334,6 +378345,8 @@ function createLocalEndpointPiProvider(options3) {
     }
   }
   function buildModel(metadata) {
+    const contextWindow = Math.min(metadata.contextLength ?? LOCAL_ENDPOINT_DEFAULT_CONTEXT_WINDOW, LOCAL_ENDPOINT_DEFAULT_CONTEXT_WINDOW);
+    const maxTokens = Math.min(metadata.maxTokens ?? LOCAL_ENDPOINT_DEFAULT_MAX_TOKENS, contextWindow);
     return {
       id: metadata.id,
       name: metadata.id,
@@ -378343,8 +378356,8 @@ function createLocalEndpointPiProvider(options3) {
       reasoning: metadata.thinking === true,
       input: metadata.vision === true ? ["text", "image"] : ["text"],
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-      contextWindow: metadata.contextLength ?? LOCAL_ENDPOINT_DEFAULT_CONTEXT_WINDOW,
-      maxTokens: metadata.maxTokens ?? LOCAL_ENDPOINT_DEFAULT_MAX_TOKENS,
+      contextWindow,
+      maxTokens,
       compat: {
         supportsDeveloperRole: false,
         supportsReasoningEffort: false,
@@ -378516,7 +378529,9 @@ function parseLmStudioModels(data) {
     if (record5.type === "embeddings")
       continue;
     const capabilities = stringArray(record5.capabilities);
-    const contextLength = record5.max_context_length;
+    const loadedContextLength = record5.state === "loaded" && typeof record5.loaded_context_length === "number" && record5.loaded_context_length > 0 ? record5.loaded_context_length : undefined;
+    const maxContextLength = typeof record5.max_context_length === "number" && record5.max_context_length > 0 ? record5.max_context_length : undefined;
+    const contextLength = loadedContextLength ?? maxContextLength;
     models3.push({
       id: record5.id,
       vision: record5.type === "vlm" || capabilities.includes("vision"),
@@ -379024,6 +379039,296 @@ var init_pi_models_runtime = __esm(() => {
   ]);
 });
 
+// src/backend/dev/pi-model-factory.ts
+function isUnselectedLocalModelHandle(model) {
+  return typeof model !== "string" || model.length === 0 || model === "auto" || model === UNSELECTED_LOCAL_MODEL_HANDLE || model.startsWith("letta/");
+}
+function normalizeOpenAICompatibleLocalModelHandle(model) {
+  if (!model?.startsWith("openai/"))
+    return model;
+  const nestedHandle = model.slice("openai/".length);
+  const nestedProvider = resolveProviderFromModelHandle(nestedHandle);
+  if (!nestedProvider)
+    return model;
+  return getPiProviderSpec(nestedProvider).localModelDiscovery === "openai-compatible" ? nestedHandle : model;
+}
+function settingString(value) {
+  return typeof value === "string" && value.length > 0 ? value : undefined;
+}
+function thinkingLevelSetting(value, preserveMax) {
+  const effort = settingString(value);
+  if (effort === "max")
+    return preserveMax ? "max" : "xhigh";
+  return effort === "minimal" || effort === "low" || effort === "medium" || effort === "high" || effort === "xhigh" ? effort : undefined;
+}
+function reasoningForSettings(modelSettings, modelHandle) {
+  const thinking = isRecord(modelSettings.thinking) ? modelSettings.thinking : undefined;
+  if (thinking?.type === "disabled")
+    return;
+  const nestedReasoning = isRecord(modelSettings.reasoning) ? modelSettings.reasoning : undefined;
+  const modelId = modelHandle?.slice(modelHandle.indexOf("/") + 1);
+  const preserveMax = modelId?.startsWith("gpt-5.6") === true;
+  return thinkingLevelSetting(nestedReasoning?.reasoning_effort, preserveMax) ?? thinkingLevelSetting(modelSettings.effort, preserveMax) ?? thinkingLevelSetting(modelSettings.reasoning_effort, preserveMax);
+}
+function applyPiEnvOverrides(overrides) {
+  if (!overrides)
+    return () => {};
+  const previous = new Map;
+  for (const [key, value] of Object.entries(overrides)) {
+    previous.set(key, process.env[key]);
+    if (value === undefined) {
+      delete process.env[key];
+    } else {
+      process.env[key] = value;
+    }
+  }
+  return () => {
+    for (const [key, value] of previous) {
+      if (value === undefined) {
+        delete process.env[key];
+      } else {
+        process.env[key] = value;
+      }
+    }
+  };
+}
+function hasEnvValue2(value) {
+  return typeof value === "string" && value.length > 0;
+}
+function inferDefaultProviderFromStandardKeys() {
+  const hasOpenAIKey = hasEnvValue2(process.env.OPENAI_API_KEY);
+  const hasAnthropicKey = hasEnvValue2(process.env.ANTHROPIC_API_KEY);
+  if (!hasOpenAIKey && hasAnthropicKey)
+    return "anthropic";
+  return DEFAULT_PI_PROVIDER;
+}
+function resolvePiProvider(provider = process.env.LETTA_CODE_DEV_PI_PROVIDER ?? inferDefaultProviderFromStandardKeys()) {
+  if (isPiProvider(provider))
+    return provider;
+  if (getRegisteredPiProvider(provider))
+    return provider;
+  throw new Error(`Unknown pi provider "${provider}". Expected ${expectedPiProviderList()}.`);
+}
+function resolvePiProviderFromAgent(model, modelSettings = {}) {
+  const registeredProvider = resolveRegisteredPiProviderFromModelHandle(model);
+  if (registeredProvider)
+    return registeredProvider;
+  const handleProvider = resolveProviderFromModelHandle(model);
+  if (handleProvider)
+    return handleProvider;
+  const settingsProvider = resolveProviderFromProviderType(modelSettings.provider_type);
+  if (settingsProvider)
+    return settingsProvider;
+  if (model && !isUnselectedLocalModelHandle(model)) {
+    const slashIndex = model.indexOf("/");
+    if (slashIndex > 0) {
+      throw new Error(`Model provider "${model.slice(0, slashIndex)}" is not registered. Load or repair the provider mod, or choose another model with /model.`);
+    }
+  }
+  return resolvePiProvider();
+}
+function resolvePiModelFromAgent(model, provider) {
+  return stripProviderHandlePrefix(model, provider);
+}
+function localProviderRecord(providerNames, storageDir) {
+  for (const providerName of providerNames) {
+    const record5 = getLocalProviderRecordByName(providerName, storageDir);
+    if (record5)
+      return record5;
+  }
+  return null;
+}
+function localProviderConnection(providerNames, storageDir) {
+  const record5 = localProviderRecord(providerNames, storageDir);
+  return {
+    baseURL: record5?.base_url,
+    timeout: resolveLocalProviderTimeout({
+      configuredTimeout: record5?.timeout,
+      providerIds: providerNames
+    }),
+    ...record5 ? { record: record5 } : {}
+  };
+}
+function resolveZaiConnection(options3) {
+  const regularRecord = localProviderRecord(["zai", LOCAL_ZAI_PROVIDER_NAME], options3.storageDir);
+  const codingRecord = localProviderRecord(["zai_coding", LOCAL_ZAI_CODING_PROVIDER_NAME], options3.storageDir);
+  const regularKey = localProviderApiKeyFromRecord(regularRecord) ?? process.env.ZAI_API_KEY ?? process.env.ZHIPU_API_KEY;
+  const codingKey = localProviderApiKeyFromRecord(codingRecord) ?? process.env.ZAI_CODING_API_KEY;
+  const regularConnection = {
+    providerName: "zai",
+    baseURL: regularRecord?.base_url ?? process.env.ZAI_BASE_URL ?? "https://api.z.ai/api/paas/v4",
+    apiKey: regularKey,
+    timeout: resolveLocalProviderTimeout({
+      configuredTimeout: regularRecord?.timeout,
+      providerIds: [LOCAL_ZAI_PROVIDER_NAME, "zai"]
+    })
+  };
+  const codingConnection = {
+    providerName: "zai-coding",
+    baseURL: codingRecord?.base_url ?? process.env.ZAI_CODING_BASE_URL ?? "https://api.z.ai/api/coding/paas/v4",
+    apiKey: codingKey,
+    timeout: resolveLocalProviderTimeout({
+      configuredTimeout: codingRecord?.timeout,
+      providerIds: [LOCAL_ZAI_CODING_PROVIDER_NAME, "zai-coding"]
+    })
+  };
+  if (options3.preferredProviderType === "zai_coding" && codingKey) {
+    return codingConnection;
+  }
+  if (options3.preferredProviderType === "zai" && regularKey) {
+    return regularConnection;
+  }
+  if (codingKey)
+    return codingConnection;
+  if (regularKey)
+    return regularConnection;
+  return codingConnection;
+}
+function fallbackCatalogModelId(provider, modelId) {
+  if (provider !== "openai")
+    return;
+  const withoutReleaseDate = modelId.replace(/-\d{4}-\d{2}-\d{2}$/, "");
+  return withoutReleaseDate === modelId ? undefined : withoutReleaseDate;
+}
+function withOverrides(model, overrides) {
+  return {
+    ...model,
+    ...overrides.baseURL ? { baseUrl: overrides.baseURL } : {},
+    ...overrides.headers ? { headers: { ...model.headers, ...overrides.headers } } : {},
+    ...overrides.contextWindow ? { contextWindow: overrides.contextWindow } : {},
+    ...overrides.maxTokens ? { maxTokens: overrides.maxTokens } : {}
+  };
+}
+function nonNullHeaders(headers) {
+  return Object.fromEntries(Object.entries(headers).filter((entry) => entry[1] !== null));
+}
+function mergeHeaders3(...headers) {
+  const merged = {};
+  for (const header of headers) {
+    if (!header)
+      continue;
+    Object.assign(merged, header);
+  }
+  return Object.keys(merged).length > 0 ? merged : undefined;
+}
+function numericSetting(value) {
+  return typeof value === "number" && Number.isFinite(value) ? value : undefined;
+}
+function bedrockLocalProviderOptions(record5) {
+  if (!record5)
+    return {};
+  const providerOptions = {};
+  const envOverrides = {};
+  if (record5.region) {
+    providerOptions.region = record5.region;
+    envOverrides.AWS_REGION = record5.region;
+    envOverrides.AWS_DEFAULT_REGION = record5.region;
+  }
+  if (record5.profile) {
+    providerOptions.profile = record5.profile;
+    envOverrides.AWS_PROFILE = record5.profile;
+  }
+  if (record5.auth.type === "api" && record5.auth.key) {
+    if (record5.access_key) {
+      envOverrides.AWS_ACCESS_KEY_ID = record5.access_key;
+      envOverrides.AWS_SECRET_ACCESS_KEY = record5.auth.key;
+    } else {
+      providerOptions.bearerToken = record5.auth.key;
+      envOverrides.AWS_BEARER_TOKEN_BEDROCK = record5.auth.key;
+    }
+  }
+  return {
+    ...Object.keys(providerOptions).length > 0 ? { providerOptions } : {},
+    ...Object.keys(envOverrides).length > 0 ? { envOverrides } : {}
+  };
+}
+async function resolvePiModelForAgent(modelHandle, modelSettings = {}, options3 = {}) {
+  const concreteModelHandle = normalizeOpenAICompatibleLocalModelHandle(isUnselectedLocalModelHandle(modelHandle) ? undefined : modelHandle);
+  const provider = options3.provider ? resolvePiProvider(options3.provider) : resolvePiProviderFromAgent(concreteModelHandle, modelSettings);
+  const registeredProvider = getRegisteredPiProvider(provider);
+  const spec = isPiProvider(provider) ? getPiProviderSpec(provider) : undefined;
+  const modelId = options3.model ?? (registeredProvider ? stripRegisteredProviderHandlePrefix(concreteModelHandle, provider) : undefined) ?? (spec ? resolvePiModelFromAgent(concreteModelHandle, spec.id) : undefined) ?? registeredProvider?.config.models?.[0]?.id ?? (spec?.defaultModel ? resolvePiModelFromAgent(spec.defaultModel, spec.id) : undefined) ?? process.env.LETTA_CODE_DEV_PI_MODEL ?? "";
+  const storageDir = options3.localProviderAuthStorageDir;
+  const modelsRuntime = options3.modelsRuntime ?? new LocalPiModelsRuntime({
+    ...storageDir ? { storageDir } : {}
+  });
+  const preferredProviderType = typeof modelSettings.provider_type === "string" ? modelSettings.provider_type : options3.preferredProviderType;
+  const localNames = registeredProvider ? localNamesForProviderId(provider) : spec?.localProviderNames ?? [provider];
+  let connection = localProviderConnection(localNames, storageDir);
+  let baseURL = connection.baseURL ?? spec?.baseUrlEnv?.() ?? spec?.defaultBaseURL ?? registeredProvider?.config.baseUrl;
+  let headers = mergeHeaders3(spec?.headers?.());
+  let providerOptions;
+  let envOverrides;
+  let oauthCredentials;
+  if (!modelId) {
+    throw new Error(`No model selected for provider "${provider}". Choose an available model with /model.`);
+  }
+  const runtimeProviderId = registeredProvider ? provider : spec && modelsRuntime.isRuntimeManagedProvider(spec.id) ? spec.id : spec?.piProvider;
+  if (!runtimeProviderId) {
+    throw new Error(`Unknown model "${modelId}" for provider "${provider}". ` + "Register the provider with models before using it.");
+  }
+  const fallbackModelId = !registeredProvider && spec?.piProvider ? fallbackCatalogModelId(spec.piProvider, modelId) : undefined;
+  const { model: publishedModel, auth: authResult } = await modelsRuntime.resolveTurn(runtimeProviderId, modelId, fallbackModelId);
+  connection = { ...connection, apiKey: authResult?.auth.apiKey };
+  if (authResult?.auth.baseUrl)
+    baseURL = authResult.auth.baseUrl;
+  if (authResult?.auth.headers) {
+    headers = mergeHeaders3(headers, nonNullHeaders(authResult.auth.headers));
+  }
+  if (connection.record?.auth.type === "oauth") {
+    const stored = await modelsRuntime.getStoredCredential(runtimeProviderId);
+    oauthCredentials = stored?.type === "oauth" ? stored : undefined;
+  }
+  if (provider === "zai") {
+    const zai = resolveZaiConnection({
+      storageDir,
+      preferredProviderType: preferredProviderType === "zai" || preferredProviderType === "zai_coding" ? preferredProviderType : undefined
+    });
+    connection = {
+      apiKey: zai.apiKey,
+      baseURL: zai.baseURL,
+      timeout: zai.timeout
+    };
+    baseURL = zai.baseURL;
+  }
+  if (provider === "amazon-bedrock") {
+    const bedrock = bedrockLocalProviderOptions(connection.record);
+    providerOptions = bedrock.providerOptions;
+    envOverrides = bedrock.envOverrides;
+  }
+  if (!publishedModel) {
+    throw new Error(`Unknown model "${modelId}" for provider "${provider}". ` + "Choose an available model with /model.");
+  }
+  const hookedModel = oauthCredentials && registeredProvider?.config.oauth?.modifyModels ? registeredProvider.config.oauth.modifyModels([structuredClone(publishedModel)], oauthCredentials)[0] ?? publishedModel : publishedModel;
+  const contextWindow = numericSetting(modelSettings.context_window_limit);
+  const maxTokens = numericSetting(modelSettings.max_tokens);
+  const allowBaseUrlOverride = !registeredProvider && spec !== undefined && !modelsRuntime.isRuntimeManagedProvider(spec.id);
+  const overrides = {
+    ...allowBaseUrlOverride && baseURL && baseURL !== hookedModel.baseUrl ? { baseURL } : {},
+    ...contextWindow && contextWindow !== hookedModel.contextWindow ? { contextWindow } : {},
+    ...maxTokens && maxTokens !== hookedModel.maxTokens ? { maxTokens } : {}
+  };
+  const model = Object.keys(overrides).length > 0 ? withOverrides(hookedModel, overrides) : hookedModel;
+  return {
+    provider,
+    model,
+    apiKey: connection.apiKey,
+    timeout: connection.timeout,
+    headers,
+    providerOptions,
+    envOverrides
+  };
+}
+var DEFAULT_PI_PROVIDER = "openai", UNSELECTED_LOCAL_MODEL_HANDLE = "local/default";
+var init_pi_model_factory = __esm(() => {
+  init_local_pi_credential_store();
+  init_local_provider_auth_store();
+  init_local_provider_timeout();
+  init_pi_models_runtime();
+  init_pi_provider_mod_registry();
+  init_pi_provider_registry();
+});
+
 // src/backend/local/local-context-estimate.ts
 function positiveUsageNumber(value) {
   return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : undefined;
@@ -379147,561 +379452,6 @@ function estimateLocalContextTokens(messages) {
 }
 var IMAGE_TOKEN_ESTIMATE = 1200;
 
-// src/backend/dev/provider-turn-executor.ts
-var exports_provider_turn_executor = {};
-__export(exports_provider_turn_executor, {
-  providerStreamPart: () => providerStreamPart,
-  providerLocalMessage: () => providerLocalMessage,
-  providerLettaChunk: () => providerLettaChunk,
-  estimateProviderRequestBytes: () => estimateProviderRequestBytes,
-  estimateProviderContextTokens: () => estimateProviderContextTokens,
-  contextTokensFromUsage: () => contextTokensFromUsage,
-  buildProviderTurnInput: () => buildProviderTurnInput,
-  ProviderTurnExecutor: () => ProviderTurnExecutor
-});
-import { randomUUID as randomUUID21 } from "node:crypto";
-function providerStreamPart(part) {
-  return { type: "provider-part", part };
-}
-function providerLocalMessage(message) {
-  return { type: "local-message", message };
-}
-function providerLettaChunk(chunk) {
-  return { type: "letta-chunk", chunk };
-}
-
-class MissingProviderStreamAdapter {
-  async* stream() {
-    yield {
-      type: "error",
-      error: new Error("Provider turn adapter is not configured for this dev backend")
-    };
-  }
-}
-function bodyListField(body, key) {
-  const value = body[key];
-  return Array.isArray(value) ? value : [];
-}
-function buildProviderTurnInput(input) {
-  return {
-    conversationId: input.conversationId,
-    agentId: input.agentId,
-    agent: input.agent,
-    systemPrompt: input.systemPrompt,
-    midConversationSystemPrompt: input.midConversationSystemPrompt,
-    body: input.body,
-    history: input.history,
-    uiMessages: input.uiMessages,
-    clientTools: bodyListField(input.body, "client_tools"),
-    clientSkills: bodyListField(input.body, "client_skills")
-  };
-}
-function stringifyToolInput(input) {
-  if (typeof input === "string")
-    return input;
-  return JSON.stringify(input ?? {});
-}
-function createLocalMessageChunk(message) {
-  return markLocalStateChunkOnly(attachLocalMessage({ message_type: "local_message" }, message));
-}
-function createProviderErrorChunks(error54) {
-  const info = normalizeLocalProviderError(error54);
-  return [
-    {
-      message_type: "error_message",
-      message: info.message,
-      detail: info.detail,
-      error_type: info.error_type,
-      retryable: info.retryable
-    },
-    {
-      message_type: "stop_reason",
-      stop_reason: info.stop_reason
-    }
-  ];
-}
-function contextTokensFromUsage(usage) {
-  return contextTokensFromLocalUsage(usage);
-}
-function estimateSerializedTokens(value) {
-  if (value === undefined || value === null)
-    return 0;
-  try {
-    const serialized = typeof value === "string" ? value : JSON.stringify(value) ?? "";
-    return Math.ceil(serialized.length / 4);
-  } catch {
-    return 0;
-  }
-}
-function estimateProviderContextTokens(input) {
-  const contextEstimate = estimateLocalContextTokens(input.uiMessages);
-  if (contextEstimate.lastUsageIndex !== null) {
-    return contextEstimate.tokens > 0 ? contextEstimate.tokens : undefined;
-  }
-  const systemPromptTokens = estimateSerializedTokens(input.systemPrompt ?? input.agent.system);
-  const messageTokens = contextEstimate.tokens;
-  const toolTokens = estimateSerializedTokens(input.clientTools);
-  const total = systemPromptTokens + messageTokens + toolTokens;
-  return total > 0 ? total : undefined;
-}
-function serializedLength(value) {
-  if (value === undefined || value === null)
-    return 0;
-  try {
-    const serialized = typeof value === "string" ? value : JSON.stringify(value) ?? "";
-    return serialized.length;
-  } catch {
-    return 0;
-  }
-}
-function estimateProviderRequestBytes(input) {
-  const systemPromptBytes = serializedLength(input.systemPrompt ?? input.agent.system);
-  const messageBytes = serializedLength(input.uiMessages);
-  const toolBytes = serializedLength(input.clientTools);
-  const total = systemPromptBytes + messageBytes + toolBytes;
-  return total > 0 ? total : undefined;
-}
-function createUsageStatisticsChunk(usage, contextTokensEstimate) {
-  const promptTokens = usage?.input;
-  const completionTokens = usage?.output;
-  const totalTokens = usage?.totalTokens;
-  const usageContextTokens = usage ? contextTokensFromUsage(usage) : undefined;
-  const contextTokens = usageContextTokens ?? contextTokensEstimate;
-  const cachedInputTokens = usage?.cacheRead;
-  const cacheWriteTokens = usage?.cacheWrite;
-  if (promptTokens === undefined && completionTokens === undefined && totalTokens === undefined && cachedInputTokens === undefined && cacheWriteTokens === undefined && contextTokens === undefined) {
-    return;
-  }
-  return {
-    message_type: "usage_statistics",
-    ...promptTokens !== undefined ? { prompt_tokens: promptTokens } : {},
-    ...completionTokens !== undefined ? { completion_tokens: completionTokens } : {},
-    ...totalTokens !== undefined ? { total_tokens: totalTokens } : {},
-    ...cachedInputTokens !== undefined ? { cached_input_tokens: cachedInputTokens } : {},
-    ...cacheWriteTokens !== undefined ? { cache_write_tokens: cacheWriteTokens } : {},
-    ...contextTokens !== undefined ? { context_tokens: contextTokens } : {}
-  };
-}
-function errorFromAssistantEvent(part) {
-  if (part.type !== "error")
-    return new Error("Unknown provider stream error");
-  return new Error(part.error.errorMessage ?? "Unknown local provider error");
-}
-function contentMatchesMessageType(content, messageType) {
-  if (!content || typeof content !== "object" || !("type" in content)) {
-    return false;
-  }
-  return messageType === "assistant_message" ? content.type === "text" : content.type === "thinking";
-}
-function contiguousContentStartIndex(partial4, contentIndex, messageType) {
-  let startIndex = contentIndex;
-  while (startIndex > 0 && contentMatchesMessageType(partial4.content[startIndex - 1], messageType)) {
-    startIndex -= 1;
-  }
-  return startIndex;
-}
-function otidForContentSegment(otids, prefix, contentIndex, partial4, messageType) {
-  const segmentStartIndex = contiguousContentStartIndex(partial4, contentIndex, messageType);
-  const existing = otids.get(segmentStartIndex);
-  if (existing)
-    return existing;
-  const otid = `${prefix}-${segmentStartIndex}-${randomUUID21()}`;
-  otids.set(segmentStartIndex, otid);
-  return otid;
-}
-function createProviderLettaStream(events, contextTokensEstimate) {
-  const controller = new AbortController;
-  return {
-    controller,
-    async* [Symbol.asyncIterator]() {
-      let sawToolCall = false;
-      let pendingStopReason;
-      let sawUsageStatistics = false;
-      const assistantOtids = new Map;
-      const reasoningOtids = new Map;
-      try {
-        for await (const event2 of events) {
-          if (event2.type === "error") {
-            yield* createProviderErrorChunks(event2.error);
-            return;
-          }
-          if (event2.type === "local-message") {
-            yield createLocalMessageChunk(event2.message);
-            continue;
-          }
-          if (event2.type === "letta-chunk") {
-            yield event2.chunk;
-            continue;
-          }
-          const { part } = event2;
-          if (part.type === "text_delta") {
-            yield {
-              message_type: "assistant_message",
-              otid: otidForContentSegment(assistantOtids, "provider-assistant", part.contentIndex, part.partial, "assistant_message"),
-              content: [{ type: "text", text: part.delta }]
-            };
-            continue;
-          }
-          if (part.type === "thinking_delta") {
-            yield {
-              message_type: "reasoning_message",
-              otid: otidForContentSegment(reasoningOtids, "provider-reasoning", part.contentIndex, part.partial, "reasoning_message"),
-              reasoning: part.delta
-            };
-            continue;
-          }
-          if (part.type === "toolcall_end") {
-            sawToolCall = true;
-            yield {
-              message_type: "approval_request_message",
-              tool_call: {
-                tool_call_id: part.toolCall.id,
-                name: part.toolCall.name,
-                arguments: stringifyToolInput(part.toolCall.arguments)
-              }
-            };
-            continue;
-          }
-          if (part.type === "done") {
-            if (!sawUsageStatistics) {
-              const usageChunk = createUsageStatisticsChunk(part.message.usage, contextTokensEstimate);
-              if (usageChunk) {
-                sawUsageStatistics = true;
-                yield usageChunk;
-              }
-            }
-            pendingStopReason = {
-              message_type: "stop_reason",
-              stop_reason: sawToolCall || part.reason === "toolUse" ? "requires_approval" : part.reason === "length" ? "max_tokens_exceeded" : "end_turn"
-            };
-            continue;
-          }
-          if (part.type === "error") {
-            yield* createProviderErrorChunks(errorFromAssistantEvent(part));
-            return;
-          }
-        }
-        if (pendingStopReason) {
-          yield pendingStopReason;
-        } else if (sawToolCall) {
-          yield {
-            message_type: "stop_reason",
-            stop_reason: "requires_approval"
-          };
-        }
-      } catch (error54) {
-        yield* createProviderErrorChunks(error54);
-      }
-    }
-  };
-}
-
-class ProviderTurnExecutor {
-  adapter;
-  constructor(adapter = new MissingProviderStreamAdapter) {
-    this.adapter = adapter;
-  }
-  async execute(input) {
-    const providerInput = buildProviderTurnInput(input);
-    const events = await this.adapter.stream(providerInput);
-    return createProviderLettaStream(events, estimateProviderContextTokens(providerInput));
-  }
-}
-var init_provider_turn_executor = __esm(() => {
-  init_local_stream_chunks();
-  init_local_provider_errors();
-});
-
-// src/backend/dev/pi-model-factory.ts
-function isUnselectedLocalModelHandle(model) {
-  return typeof model !== "string" || model.length === 0 || model === "auto" || model === UNSELECTED_LOCAL_MODEL_HANDLE || model.startsWith("letta/");
-}
-function normalizeOpenAICompatibleLocalModelHandle(model) {
-  if (!model?.startsWith("openai/"))
-    return model;
-  const nestedHandle = model.slice("openai/".length);
-  const nestedProvider = resolveProviderFromModelHandle(nestedHandle);
-  if (!nestedProvider)
-    return model;
-  return getPiProviderSpec(nestedProvider).localModelDiscovery === "openai-compatible" ? nestedHandle : model;
-}
-function settingString(value) {
-  return typeof value === "string" && value.length > 0 ? value : undefined;
-}
-function thinkingLevelSetting(value, preserveMax) {
-  const effort = settingString(value);
-  if (effort === "max")
-    return preserveMax ? "max" : "xhigh";
-  return effort === "minimal" || effort === "low" || effort === "medium" || effort === "high" || effort === "xhigh" ? effort : undefined;
-}
-function reasoningForSettings(modelSettings, modelHandle) {
-  const thinking = isRecord(modelSettings.thinking) ? modelSettings.thinking : undefined;
-  if (thinking?.type === "disabled")
-    return;
-  const nestedReasoning = isRecord(modelSettings.reasoning) ? modelSettings.reasoning : undefined;
-  const modelId = modelHandle?.slice(modelHandle.indexOf("/") + 1);
-  const preserveMax = modelId?.startsWith("gpt-5.6") === true;
-  return thinkingLevelSetting(nestedReasoning?.reasoning_effort, preserveMax) ?? thinkingLevelSetting(modelSettings.effort, preserveMax) ?? thinkingLevelSetting(modelSettings.reasoning_effort, preserveMax);
-}
-function applyPiEnvOverrides(overrides) {
-  if (!overrides)
-    return () => {};
-  const previous = new Map;
-  for (const [key, value] of Object.entries(overrides)) {
-    previous.set(key, process.env[key]);
-    if (value === undefined) {
-      delete process.env[key];
-    } else {
-      process.env[key] = value;
-    }
-  }
-  return () => {
-    for (const [key, value] of previous) {
-      if (value === undefined) {
-        delete process.env[key];
-      } else {
-        process.env[key] = value;
-      }
-    }
-  };
-}
-function hasEnvValue2(value) {
-  return typeof value === "string" && value.length > 0;
-}
-function inferDefaultProviderFromStandardKeys() {
-  const hasOpenAIKey = hasEnvValue2(process.env.OPENAI_API_KEY);
-  const hasAnthropicKey = hasEnvValue2(process.env.ANTHROPIC_API_KEY);
-  if (!hasOpenAIKey && hasAnthropicKey)
-    return "anthropic";
-  return DEFAULT_PI_PROVIDER;
-}
-function resolvePiProvider(provider = process.env.LETTA_CODE_DEV_PI_PROVIDER ?? inferDefaultProviderFromStandardKeys()) {
-  if (isPiProvider(provider))
-    return provider;
-  if (getRegisteredPiProvider(provider))
-    return provider;
-  throw new Error(`Unknown pi provider "${provider}". Expected ${expectedPiProviderList()}.`);
-}
-function resolvePiProviderFromAgent(model, modelSettings = {}) {
-  const registeredProvider = resolveRegisteredPiProviderFromModelHandle(model);
-  if (registeredProvider)
-    return registeredProvider;
-  const handleProvider = resolveProviderFromModelHandle(model);
-  if (handleProvider)
-    return handleProvider;
-  const settingsProvider = resolveProviderFromProviderType(modelSettings.provider_type);
-  if (settingsProvider)
-    return settingsProvider;
-  if (model && !isUnselectedLocalModelHandle(model)) {
-    const slashIndex = model.indexOf("/");
-    if (slashIndex > 0) {
-      throw new Error(`Model provider "${model.slice(0, slashIndex)}" is not registered. Load or repair the provider mod, or choose another model with /model.`);
-    }
-  }
-  return resolvePiProvider();
-}
-function resolvePiModelFromAgent(model, provider) {
-  return stripProviderHandlePrefix(model, provider);
-}
-function localProviderRecord(providerNames, storageDir) {
-  for (const providerName of providerNames) {
-    const record5 = getLocalProviderRecordByName(providerName, storageDir);
-    if (record5)
-      return record5;
-  }
-  return null;
-}
-function localProviderConnection(providerNames, storageDir) {
-  const record5 = localProviderRecord(providerNames, storageDir);
-  return {
-    baseURL: record5?.base_url,
-    timeout: resolveLocalProviderTimeout({
-      configuredTimeout: record5?.timeout,
-      providerIds: providerNames
-    }),
-    ...record5 ? { record: record5 } : {}
-  };
-}
-function resolveZaiConnection(options3) {
-  const regularRecord = localProviderRecord(["zai", LOCAL_ZAI_PROVIDER_NAME], options3.storageDir);
-  const codingRecord = localProviderRecord(["zai_coding", LOCAL_ZAI_CODING_PROVIDER_NAME], options3.storageDir);
-  const regularKey = localProviderApiKeyFromRecord(regularRecord) ?? process.env.ZAI_API_KEY ?? process.env.ZHIPU_API_KEY;
-  const codingKey = localProviderApiKeyFromRecord(codingRecord) ?? process.env.ZAI_CODING_API_KEY;
-  const regularConnection = {
-    providerName: "zai",
-    baseURL: regularRecord?.base_url ?? process.env.ZAI_BASE_URL ?? "https://api.z.ai/api/paas/v4",
-    apiKey: regularKey,
-    timeout: resolveLocalProviderTimeout({
-      configuredTimeout: regularRecord?.timeout,
-      providerIds: [LOCAL_ZAI_PROVIDER_NAME, "zai"]
-    })
-  };
-  const codingConnection = {
-    providerName: "zai-coding",
-    baseURL: codingRecord?.base_url ?? process.env.ZAI_CODING_BASE_URL ?? "https://api.z.ai/api/coding/paas/v4",
-    apiKey: codingKey,
-    timeout: resolveLocalProviderTimeout({
-      configuredTimeout: codingRecord?.timeout,
-      providerIds: [LOCAL_ZAI_CODING_PROVIDER_NAME, "zai-coding"]
-    })
-  };
-  if (options3.preferredProviderType === "zai_coding" && codingKey) {
-    return codingConnection;
-  }
-  if (options3.preferredProviderType === "zai" && regularKey) {
-    return regularConnection;
-  }
-  if (codingKey)
-    return codingConnection;
-  if (regularKey)
-    return regularConnection;
-  return codingConnection;
-}
-function fallbackCatalogModelId(provider, modelId) {
-  if (provider !== "openai")
-    return;
-  const withoutReleaseDate = modelId.replace(/-\d{4}-\d{2}-\d{2}$/, "");
-  return withoutReleaseDate === modelId ? undefined : withoutReleaseDate;
-}
-function withOverrides(model, overrides) {
-  return {
-    ...model,
-    ...overrides.baseURL ? { baseUrl: overrides.baseURL } : {},
-    ...overrides.headers ? { headers: { ...model.headers, ...overrides.headers } } : {},
-    ...overrides.contextWindow ? { contextWindow: overrides.contextWindow } : {},
-    ...overrides.maxTokens ? { maxTokens: overrides.maxTokens } : {}
-  };
-}
-function nonNullHeaders(headers) {
-  return Object.fromEntries(Object.entries(headers).filter((entry) => entry[1] !== null));
-}
-function mergeHeaders3(...headers) {
-  const merged = {};
-  for (const header of headers) {
-    if (!header)
-      continue;
-    Object.assign(merged, header);
-  }
-  return Object.keys(merged).length > 0 ? merged : undefined;
-}
-function numericSetting(value) {
-  return typeof value === "number" && Number.isFinite(value) ? value : undefined;
-}
-function bedrockLocalProviderOptions(record5) {
-  if (!record5)
-    return {};
-  const providerOptions = {};
-  const envOverrides = {};
-  if (record5.region) {
-    providerOptions.region = record5.region;
-    envOverrides.AWS_REGION = record5.region;
-    envOverrides.AWS_DEFAULT_REGION = record5.region;
-  }
-  if (record5.profile) {
-    providerOptions.profile = record5.profile;
-    envOverrides.AWS_PROFILE = record5.profile;
-  }
-  if (record5.auth.type === "api" && record5.auth.key) {
-    if (record5.access_key) {
-      envOverrides.AWS_ACCESS_KEY_ID = record5.access_key;
-      envOverrides.AWS_SECRET_ACCESS_KEY = record5.auth.key;
-    } else {
-      providerOptions.bearerToken = record5.auth.key;
-      envOverrides.AWS_BEARER_TOKEN_BEDROCK = record5.auth.key;
-    }
-  }
-  return {
-    ...Object.keys(providerOptions).length > 0 ? { providerOptions } : {},
-    ...Object.keys(envOverrides).length > 0 ? { envOverrides } : {}
-  };
-}
-async function resolvePiModelForAgent(modelHandle, modelSettings = {}, options3 = {}) {
-  const concreteModelHandle = normalizeOpenAICompatibleLocalModelHandle(isUnselectedLocalModelHandle(modelHandle) ? undefined : modelHandle);
-  const provider = options3.provider ? resolvePiProvider(options3.provider) : resolvePiProviderFromAgent(concreteModelHandle, modelSettings);
-  const registeredProvider = getRegisteredPiProvider(provider);
-  const spec = isPiProvider(provider) ? getPiProviderSpec(provider) : undefined;
-  const modelId = options3.model ?? (registeredProvider ? stripRegisteredProviderHandlePrefix(concreteModelHandle, provider) : undefined) ?? (spec ? resolvePiModelFromAgent(concreteModelHandle, spec.id) : undefined) ?? registeredProvider?.config.models?.[0]?.id ?? (spec?.defaultModel ? resolvePiModelFromAgent(spec.defaultModel, spec.id) : undefined) ?? process.env.LETTA_CODE_DEV_PI_MODEL ?? "";
-  const storageDir = options3.localProviderAuthStorageDir;
-  const modelsRuntime = options3.modelsRuntime ?? new LocalPiModelsRuntime({
-    ...storageDir ? { storageDir } : {}
-  });
-  const preferredProviderType = typeof modelSettings.provider_type === "string" ? modelSettings.provider_type : options3.preferredProviderType;
-  const localNames = registeredProvider ? localNamesForProviderId(provider) : spec?.localProviderNames ?? [provider];
-  let connection = localProviderConnection(localNames, storageDir);
-  let baseURL = connection.baseURL ?? spec?.baseUrlEnv?.() ?? spec?.defaultBaseURL ?? registeredProvider?.config.baseUrl;
-  let headers = mergeHeaders3(spec?.headers?.());
-  let providerOptions;
-  let envOverrides;
-  let oauthCredentials;
-  if (!modelId) {
-    throw new Error(`No model selected for provider "${provider}". Choose an available model with /model.`);
-  }
-  const runtimeProviderId = registeredProvider ? provider : spec && modelsRuntime.isRuntimeManagedProvider(spec.id) ? spec.id : spec?.piProvider;
-  if (!runtimeProviderId) {
-    throw new Error(`Unknown model "${modelId}" for provider "${provider}". ` + "Register the provider with models before using it.");
-  }
-  const fallbackModelId = !registeredProvider && spec?.piProvider ? fallbackCatalogModelId(spec.piProvider, modelId) : undefined;
-  const { model: publishedModel, auth: authResult } = await modelsRuntime.resolveTurn(runtimeProviderId, modelId, fallbackModelId);
-  connection = { ...connection, apiKey: authResult?.auth.apiKey };
-  if (authResult?.auth.baseUrl)
-    baseURL = authResult.auth.baseUrl;
-  if (authResult?.auth.headers) {
-    headers = mergeHeaders3(headers, nonNullHeaders(authResult.auth.headers));
-  }
-  if (connection.record?.auth.type === "oauth") {
-    const stored = await modelsRuntime.getStoredCredential(runtimeProviderId);
-    oauthCredentials = stored?.type === "oauth" ? stored : undefined;
-  }
-  if (provider === "zai") {
-    const zai = resolveZaiConnection({
-      storageDir,
-      preferredProviderType: preferredProviderType === "zai" || preferredProviderType === "zai_coding" ? preferredProviderType : undefined
-    });
-    connection = {
-      apiKey: zai.apiKey,
-      baseURL: zai.baseURL,
-      timeout: zai.timeout
-    };
-    baseURL = zai.baseURL;
-  }
-  if (provider === "amazon-bedrock") {
-    const bedrock = bedrockLocalProviderOptions(connection.record);
-    providerOptions = bedrock.providerOptions;
-    envOverrides = bedrock.envOverrides;
-  }
-  if (!publishedModel) {
-    throw new Error(`Unknown model "${modelId}" for provider "${provider}". ` + "Choose an available model with /model.");
-  }
-  const hookedModel = oauthCredentials && registeredProvider?.config.oauth?.modifyModels ? registeredProvider.config.oauth.modifyModels([structuredClone(publishedModel)], oauthCredentials)[0] ?? publishedModel : publishedModel;
-  const contextWindow = numericSetting(modelSettings.context_window_limit);
-  const maxTokens = numericSetting(modelSettings.max_tokens);
-  const allowBaseUrlOverride = !registeredProvider && spec !== undefined && !modelsRuntime.isRuntimeManagedProvider(spec.id);
-  const overrides = {
-    ...allowBaseUrlOverride && baseURL && baseURL !== hookedModel.baseUrl ? { baseURL } : {},
-    ...contextWindow && contextWindow !== hookedModel.contextWindow ? { contextWindow } : {},
-    ...maxTokens && maxTokens !== hookedModel.maxTokens ? { maxTokens } : {}
-  };
-  const model = Object.keys(overrides).length > 0 ? withOverrides(hookedModel, overrides) : hookedModel;
-  return {
-    provider,
-    model,
-    apiKey: connection.apiKey,
-    timeout: connection.timeout,
-    headers,
-    providerOptions,
-    envOverrides
-  };
-}
-var DEFAULT_PI_PROVIDER = "openai", UNSELECTED_LOCAL_MODEL_HANDLE = "local/default";
-var init_pi_model_factory = __esm(() => {
-  init_local_pi_credential_store();
-  init_local_provider_auth_store();
-  init_local_provider_timeout();
-  init_pi_models_runtime();
-  init_pi_provider_mod_registry();
-  init_pi_provider_registry();
-});
-
 // src/backend/local/local-model-config.ts
 function localProviderNamesFromRecords(records) {
   return new Set(records.map((record5) => record5.name));
@@ -380358,6 +380108,285 @@ Write in first person as a factual record of what occurred. Be thorough and deta
 Keep your summary under ${SLIDING_WORD_LIMIT} words. Only output the summary.`;
 });
 
+// src/backend/dev/provider-turn-executor.ts
+var exports_provider_turn_executor = {};
+__export(exports_provider_turn_executor, {
+  shouldCompactForContextPressure: () => shouldCompactForContextPressure,
+  providerStreamPart: () => providerStreamPart,
+  providerLocalMessage: () => providerLocalMessage,
+  providerLettaChunk: () => providerLettaChunk,
+  estimateProviderRequestBytes: () => estimateProviderRequestBytes,
+  estimateProviderContextTokens: () => estimateProviderContextTokens,
+  contextTokensFromUsage: () => contextTokensFromUsage,
+  contextCompactionThreshold: () => contextCompactionThreshold,
+  buildProviderTurnInput: () => buildProviderTurnInput,
+  ProviderTurnExecutor: () => ProviderTurnExecutor
+});
+import { randomUUID as randomUUID21 } from "node:crypto";
+function providerStreamPart(part) {
+  return { type: "provider-part", part };
+}
+function providerLocalMessage(message) {
+  return { type: "local-message", message };
+}
+function providerLettaChunk(chunk) {
+  return { type: "letta-chunk", chunk };
+}
+
+class MissingProviderStreamAdapter {
+  async* stream() {
+    yield {
+      type: "error",
+      error: new Error("Provider turn adapter is not configured for this dev backend")
+    };
+  }
+}
+function bodyListField(body, key) {
+  const value = body[key];
+  return Array.isArray(value) ? value : [];
+}
+function buildProviderTurnInput(input) {
+  return {
+    conversationId: input.conversationId,
+    agentId: input.agentId,
+    agent: input.agent,
+    systemPrompt: input.systemPrompt,
+    midConversationSystemPrompt: input.midConversationSystemPrompt,
+    body: input.body,
+    history: input.history,
+    uiMessages: input.uiMessages,
+    clientTools: bodyListField(input.body, "client_tools"),
+    clientSkills: bodyListField(input.body, "client_skills")
+  };
+}
+function stringifyToolInput(input) {
+  if (typeof input === "string")
+    return input;
+  return JSON.stringify(input ?? {});
+}
+function createLocalMessageChunk(message) {
+  return markLocalStateChunkOnly(attachLocalMessage({ message_type: "local_message" }, message));
+}
+function createProviderErrorChunks(error54) {
+  const info = normalizeLocalProviderError(error54);
+  return [
+    {
+      message_type: "error_message",
+      message: info.message,
+      detail: info.detail,
+      error_type: info.error_type,
+      retryable: info.retryable
+    },
+    {
+      message_type: "stop_reason",
+      stop_reason: info.stop_reason
+    }
+  ];
+}
+function contextTokensFromUsage(usage) {
+  return contextTokensFromLocalUsage(usage);
+}
+function estimateSerializedTokens(value) {
+  if (value === undefined || value === null)
+    return 0;
+  try {
+    const serialized = typeof value === "string" ? value : JSON.stringify(value) ?? "";
+    return Math.ceil(serialized.length / 4);
+  } catch {
+    return 0;
+  }
+}
+function estimateProviderContextTokens(input) {
+  const contextEstimate = estimateLocalContextTokens(input.uiMessages);
+  if (contextEstimate.lastUsageIndex !== null) {
+    return contextEstimate.tokens > 0 ? contextEstimate.tokens : undefined;
+  }
+  const systemPromptTokens = estimateSerializedTokens(input.systemPrompt ?? input.agent.system);
+  const messageTokens = contextEstimate.tokens;
+  const toolTokens = estimateSerializedTokens(input.clientTools);
+  const total = systemPromptTokens + messageTokens + toolTokens;
+  return total > 0 ? total : undefined;
+}
+function contextCompactionThreshold(contextWindow) {
+  if (typeof contextWindow !== "number" || !Number.isFinite(contextWindow) || contextWindow <= 0) {
+    return;
+  }
+  const reserveTokens = Math.min(LOCAL_CONTEXT_COMPACTION_RESERVE_TOKENS, Math.max(1, Math.floor(contextWindow * LOCAL_SMALL_CONTEXT_COMPACTION_RESERVE_RATIO)));
+  return Math.max(0, contextWindow - reserveTokens);
+}
+function shouldCompactForContextPressure(input) {
+  const threshold = contextCompactionThreshold(input.contextWindow);
+  return input.contextTokens !== undefined && threshold !== undefined && input.contextTokens > threshold;
+}
+function serializedLength(value) {
+  if (value === undefined || value === null)
+    return 0;
+  try {
+    const serialized = typeof value === "string" ? value : JSON.stringify(value) ?? "";
+    return serialized.length;
+  } catch {
+    return 0;
+  }
+}
+function estimateProviderRequestBytes(input) {
+  const systemPromptBytes = serializedLength(input.systemPrompt ?? input.agent.system);
+  const messageBytes = serializedLength(input.uiMessages);
+  const toolBytes = serializedLength(input.clientTools);
+  const total = systemPromptBytes + messageBytes + toolBytes;
+  return total > 0 ? total : undefined;
+}
+function createUsageStatisticsChunk(usage, contextTokensEstimate) {
+  const promptTokens = usage?.input;
+  const completionTokens = usage?.output;
+  const totalTokens = usage?.totalTokens;
+  const usageContextTokens = usage ? contextTokensFromUsage(usage) : undefined;
+  const contextTokens = usageContextTokens ?? contextTokensEstimate;
+  const cachedInputTokens = usage?.cacheRead;
+  const cacheWriteTokens = usage?.cacheWrite;
+  if (promptTokens === undefined && completionTokens === undefined && totalTokens === undefined && cachedInputTokens === undefined && cacheWriteTokens === undefined && contextTokens === undefined) {
+    return;
+  }
+  return {
+    message_type: "usage_statistics",
+    ...promptTokens !== undefined ? { prompt_tokens: promptTokens } : {},
+    ...completionTokens !== undefined ? { completion_tokens: completionTokens } : {},
+    ...totalTokens !== undefined ? { total_tokens: totalTokens } : {},
+    ...cachedInputTokens !== undefined ? { cached_input_tokens: cachedInputTokens } : {},
+    ...cacheWriteTokens !== undefined ? { cache_write_tokens: cacheWriteTokens } : {},
+    ...contextTokens !== undefined ? { context_tokens: contextTokens } : {}
+  };
+}
+function errorFromAssistantEvent(part) {
+  if (part.type !== "error")
+    return new Error("Unknown provider stream error");
+  return new Error(part.error.errorMessage ?? "Unknown local provider error");
+}
+function contentMatchesMessageType(content, messageType) {
+  if (!content || typeof content !== "object" || !("type" in content)) {
+    return false;
+  }
+  return messageType === "assistant_message" ? content.type === "text" : content.type === "thinking";
+}
+function contiguousContentStartIndex(partial4, contentIndex, messageType) {
+  let startIndex = contentIndex;
+  while (startIndex > 0 && contentMatchesMessageType(partial4.content[startIndex - 1], messageType)) {
+    startIndex -= 1;
+  }
+  return startIndex;
+}
+function otidForContentSegment(otids, prefix, contentIndex, partial4, messageType) {
+  const segmentStartIndex = contiguousContentStartIndex(partial4, contentIndex, messageType);
+  const existing = otids.get(segmentStartIndex);
+  if (existing)
+    return existing;
+  const otid = `${prefix}-${segmentStartIndex}-${randomUUID21()}`;
+  otids.set(segmentStartIndex, otid);
+  return otid;
+}
+function createProviderLettaStream(events, contextTokensEstimate) {
+  const controller = new AbortController;
+  return {
+    controller,
+    async* [Symbol.asyncIterator]() {
+      let sawToolCall = false;
+      let pendingStopReason;
+      let sawUsageStatistics = false;
+      const assistantOtids = new Map;
+      const reasoningOtids = new Map;
+      try {
+        for await (const event2 of events) {
+          if (event2.type === "error") {
+            yield* createProviderErrorChunks(event2.error);
+            return;
+          }
+          if (event2.type === "local-message") {
+            yield createLocalMessageChunk(event2.message);
+            continue;
+          }
+          if (event2.type === "letta-chunk") {
+            yield event2.chunk;
+            continue;
+          }
+          const { part } = event2;
+          if (part.type === "text_delta") {
+            yield {
+              message_type: "assistant_message",
+              otid: otidForContentSegment(assistantOtids, "provider-assistant", part.contentIndex, part.partial, "assistant_message"),
+              content: [{ type: "text", text: part.delta }]
+            };
+            continue;
+          }
+          if (part.type === "thinking_delta") {
+            yield {
+              message_type: "reasoning_message",
+              otid: otidForContentSegment(reasoningOtids, "provider-reasoning", part.contentIndex, part.partial, "reasoning_message"),
+              reasoning: part.delta
+            };
+            continue;
+          }
+          if (part.type === "toolcall_end") {
+            sawToolCall = true;
+            yield {
+              message_type: "approval_request_message",
+              tool_call: {
+                tool_call_id: part.toolCall.id,
+                name: part.toolCall.name,
+                arguments: stringifyToolInput(part.toolCall.arguments)
+              }
+            };
+            continue;
+          }
+          if (part.type === "done") {
+            if (!sawUsageStatistics) {
+              const usageChunk = createUsageStatisticsChunk(part.message.usage, contextTokensEstimate);
+              if (usageChunk) {
+                sawUsageStatistics = true;
+                yield usageChunk;
+              }
+            }
+            pendingStopReason = {
+              message_type: "stop_reason",
+              stop_reason: sawToolCall || part.reason === "toolUse" ? "requires_approval" : part.reason === "length" ? "max_tokens_exceeded" : "end_turn"
+            };
+            continue;
+          }
+          if (part.type === "error") {
+            yield* createProviderErrorChunks(errorFromAssistantEvent(part));
+            return;
+          }
+        }
+        if (pendingStopReason) {
+          yield pendingStopReason;
+        } else if (sawToolCall) {
+          yield {
+            message_type: "stop_reason",
+            stop_reason: "requires_approval"
+          };
+        }
+      } catch (error54) {
+        yield* createProviderErrorChunks(error54);
+      }
+    }
+  };
+}
+
+class ProviderTurnExecutor {
+  adapter;
+  constructor(adapter = new MissingProviderStreamAdapter) {
+    this.adapter = adapter;
+  }
+  async execute(input) {
+    const providerInput = buildProviderTurnInput(input);
+    const events = await this.adapter.stream(providerInput);
+    return createProviderLettaStream(events, estimateProviderContextTokens(providerInput));
+  }
+}
+var LOCAL_CONTEXT_COMPACTION_RESERVE_TOKENS = 16384, LOCAL_SMALL_CONTEXT_COMPACTION_RESERVE_RATIO = 0.2;
+var init_provider_turn_executor = __esm(() => {
+  init_local_stream_chunks();
+  init_local_provider_errors();
+});
+
 // src/backend/dev/pi-image-elision.ts
 function localProviderRequestByteLimit() {
   const raw = process.env[LOCAL_PROVIDER_REQUEST_BYTE_LIMIT_ENV];
@@ -380743,7 +380772,7 @@ class PiStreamAdapter {
   localProviderAuthStorageDir;
   modelsRuntime;
   onContextWindowOverflow;
-  onContextUsage;
+  onContextPressure;
   onLlmStart;
   onLlmEnd;
   constructor(options3 = {}) {
@@ -380754,7 +380783,7 @@ class PiStreamAdapter {
     this.abortSignal = options3.abortSignal;
     this.localProviderAuthStorageDir = options3.localProviderAuthStorageDir;
     this.onContextWindowOverflow = options3.onContextWindowOverflow;
-    this.onContextUsage = options3.onContextUsage;
+    this.onContextPressure = options3.onContextPressure;
     this.onLlmStart = options3.onLlmStart;
     this.onLlmEnd = options3.onLlmEnd;
   }
@@ -380771,6 +380800,33 @@ class PiStreamAdapter {
       ...compaction.stats ? { compaction_stats: compaction.stats } : {}
     });
   }
+  async compactBeforeProviderCall(input) {
+    if (!this.onContextPressure)
+      return null;
+    const contextTokens = estimateProviderContextTokens(input);
+    if (contextTokens === undefined)
+      return null;
+    const localModel = await resolveAvailableLocalModelForTurn({
+      model: input.agent.model,
+      modelSettings: input.agent.model_settings,
+      storageDir: this.localProviderAuthStorageDir,
+      modelsRuntime: this.modelsRuntime
+    });
+    const resolved = await resolvePiModelForAgent(localModel.model, localModel.modelSettings, {
+      localProviderAuthStorageDir: this.localProviderAuthStorageDir,
+      modelsRuntime: this.modelsRuntime
+    });
+    const contextWindow = resolved.model.contextWindow;
+    if (!shouldCompactForContextPressure({ contextTokens, contextWindow })) {
+      return null;
+    }
+    return this.onContextPressure(input, {
+      contextTokens,
+      contextWindow,
+      phase: "preflight",
+      source: "estimate"
+    });
+  }
   async* streamOnce(input) {
     const tools = toPiTools(input.clientTools);
     const localModel = await resolveAvailableLocalModelForTurn({
@@ -380839,6 +380895,7 @@ class PiStreamAdapter {
       const result = this.runStream(resolved.model, context3, options3);
       let streamError;
       let finalMessage;
+      let finalLocalMessage;
       for await (const part of result) {
         if (part.type === "error") {
           const error54 = new PiProviderError(part.error);
@@ -380849,7 +380906,8 @@ class PiStreamAdapter {
         }
         if (part.type === "done") {
           finalMessage = part.message;
-          yield providerLocalMessage(toLocalAssistantMessage(part.message, input));
+          finalLocalMessage = toLocalAssistantMessage(part.message, input);
+          yield providerLocalMessage(finalLocalMessage);
         }
         yield providerStreamPart(part);
       }
@@ -380869,8 +380927,22 @@ class PiStreamAdapter {
       if (finalMessage.stopReason === "error" || finalMessage.stopReason === "aborted") {
         throw new PiProviderError(finalMessage);
       }
-      if (this.onContextUsage) {
-        const compaction = await this.onContextUsage(input, finalMessage.usage);
+      if (this.onContextPressure) {
+        const usageContextTokens = contextTokensFromUsage(finalMessage.usage);
+        const contextTokens = usageContextTokens ?? estimateProviderContextTokens({
+          ...input,
+          uiMessages: [
+            ...input.uiMessages,
+            finalLocalMessage ?? toLocalAssistantMessage(finalMessage, input)
+          ]
+        });
+        const contextWindow = resolved.model.contextWindow;
+        const compaction = shouldCompactForContextPressure({ contextTokens, contextWindow }) && contextTokens !== undefined ? await this.onContextPressure(input, {
+          contextTokens,
+          contextWindow,
+          phase: "post_turn",
+          source: usageContextTokens === undefined ? "estimate" : "usage"
+        }) : null;
         if (compaction) {
           yield* this.emitCompactionChunks(compaction, "context_window_limit");
         }
@@ -380891,9 +380963,19 @@ class PiStreamAdapter {
   }
   async* stream(input) {
     let activeInput = input;
+    let preflightCompactionChecked = false;
     let contextOverflowCompactions = 0;
     let transientRetries = 0;
     while (true) {
+      if (!preflightCompactionChecked) {
+        preflightCompactionChecked = true;
+        const compaction = await this.compactBeforeProviderCall(activeInput);
+        if (compaction) {
+          activeInput = { ...activeInput, uiMessages: compaction.uiMessages };
+          yield* this.emitCompactionChunks(compaction, "context_window_limit");
+          continue;
+        }
+      }
       let emittedModelOutput = false;
       try {
         for await (const event2 of this.streamOnce(activeInput)) {
@@ -380994,7 +381076,7 @@ var init_pi_stream_adapter = __esm(() => {
 });
 
 // src/backend/local/local-executor-factory.ts
-function createLocalExecutor(options3, modelsRuntime, onContextWindowOverflow, onContextUsage, onLlmStart, onLlmEnd) {
+function createLocalExecutor(options3, modelsRuntime, onContextWindowOverflow, onContextPressure, onLlmStart, onLlmEnd) {
   if (options3.executor)
     return options3.executor;
   if (options3.executionMode === "deterministic") {
@@ -381005,7 +381087,7 @@ function createLocalExecutor(options3, modelsRuntime, onContextWindowOverflow, o
     localProviderAuthStorageDir: options3.storageDir,
     modelsRuntime,
     onContextWindowOverflow,
-    onContextUsage,
+    onContextPressure,
     onLlmStart,
     onLlmEnd
   }));
@@ -381455,7 +381537,6 @@ var init_local_backend = __esm(() => {
   init_memory_git();
   init_headless_backend();
   init_pi_models_runtime();
-  init_provider_turn_executor();
   init_compaction();
   init_local_executor_factory();
   init_local_model_config();
@@ -381495,7 +381576,7 @@ var init_local_backend = __esm(() => {
         storedMessageIdPrefix: "letta-msg-",
         localMessageIdPrefix: "ui-msg-"
       };
-      super(options3.defaultAgentId ?? "agent-local-default", createLocalExecutor(options3, runtime, (input, error54) => localBackendRef.current?.compactAfterContextOverflow(input, error54) ?? Promise.resolve(null), (input, usage) => localBackendRef.current?.compactAfterContextUsage(input, usage) ?? Promise.resolve(null), (info) => localBackendRef.current?.emitLlmStart(info) ?? Promise.resolve(), (info) => localBackendRef.current?.emitLlmEnd(info) ?? Promise.resolve()), storeOptions, {
+      super(options3.defaultAgentId ?? "agent-local-default", createLocalExecutor(options3, runtime, (input, error54) => localBackendRef.current?.compactAfterContextOverflow(input, error54) ?? Promise.resolve(null), (input, pressure) => localBackendRef.current?.compactForContextPressure(input, pressure) ?? Promise.resolve(null), (info) => localBackendRef.current?.emitLlmStart(info) ?? Promise.resolve(), (info) => localBackendRef.current?.emitLlmEnd(info) ?? Promise.resolve()), storeOptions, {
         modelHandle: modelConfig.handle,
         runIdPrefix: "local-run-",
         runMetadataBackend: "local"
@@ -381654,12 +381735,7 @@ var init_local_backend = __esm(() => {
         stats: result.stats
       };
     }
-    async compactAfterContextUsage(input, usage) {
-      const contextTokens = contextTokensFromUsage(usage) ?? estimateProviderContextTokens(input);
-      const contextWindow = this.effectiveContextWindow(input.conversationId, input.agentId);
-      if (contextTokens === undefined || contextWindow === undefined || contextTokens <= contextWindow) {
-        return null;
-      }
+    async compactForContextPressure(input, _pressure) {
       const result = await this.compactLocalConversation(input.conversationId, input.agentId, "context_window_limit");
       return {
         uiMessages: this.store.listLocalMessages(input.conversationId, input.agentId),
@@ -470002,11 +470078,6 @@ async function startAppServer(options3 = {}) {
   });
   server2.on("upgrade", (request, socket, head2) => {
     const requestUrl = getRequestUrl(request, listen.host);
-    if (request.headers.origin) {
-      options3.onLog?.(`Rejecting app-server websocket request with Origin header: ${request.url ?? "/"}`);
-      rejectUpgrade(socket, 403, "Forbidden");
-      return;
-    }
     if (requestUrl.pathname !== listen.path && requestUrl.pathname !== "/") {
       rejectUpgrade(socket, 404, "Not Found");
       return;
@@ -470022,6 +470093,11 @@ async function startAppServer(options3 = {}) {
       rejectUpgrade(socket, authError.statusCode, authError.message);
       return;
     }
+    if (request.headers.origin !== undefined && authPolicy.mode === undefined) {
+      options3.onLog?.(`Rejecting unauthenticated app-server websocket request with Origin header: ${request.url ?? "/"}; native clients such as React Native must configure --ws-auth capability-token or --ws-auth signed-bearer-token`);
+      rejectUpgrade(socket, 403, "Forbidden");
+      return;
+    }
     wss.handleUpgrade(request, socket, head2, (websocket) => {
       handleWebSocketConnection(websocket, channel);
     });
@@ -470130,7 +470206,7 @@ Run the local App Server using native v2 WebSocket frames.
 Options:
   --listen [url]  WebSocket listen URL. Defaults to an available loopback port
   --openai-api  Serve OpenAI-compatible /v1/models and /v1/chat/completions routes (each agent is a model)
-  --ws-auth <mode>  WebSocket auth mode for non-loopback listeners. Supported: capability-token, signed-bearer-token
+  --ws-auth <mode>  WebSocket auth for non-loopback listeners and Origin-bearing native clients. Supported: capability-token, signed-bearer-token
   --ws-token-file <path>  Absolute path to the capability-token file
   --ws-token-sha256 <hex>  Hex-encoded SHA-256 digest of the capability token
   --ws-shared-secret-file <path>  Absolute path to the shared secret file for signed JWT bearer tokens
@@ -470250,7 +470326,7 @@ Remote environment options:
 App Server options:
   --listen [url]  Accept App Server connections. If URL is omitted, binds to an available loopback port
   --openai-api  Serve OpenAI-compatible /v1/models and /v1/chat/completions routes (each agent is a model)
-  --ws-auth <mode>  Authentication for non-loopback listeners: capability-token or signed-bearer-token
+  --ws-auth <mode>  Authentication for non-loopback listeners and Origin-bearing native clients: capability-token or signed-bearer-token
   --ws-token-file <path>  Absolute path to the capability-token file
   --ws-token-sha256 <hex>  Hex-encoded SHA-256 digest of the capability token
   --ws-shared-secret-file <path>  Absolute path to the shared secret file for signed JWT bearer tokens
@@ -499449,6 +499525,8 @@ function ModelSelector({
   const [byokProviderAliases, setByokProviderAliases] = import_react91.useState(() => buildByokProviderAliases([]));
   const [openAICompatibleProxyHandles, setOpenAICompatibleProxyHandles] = import_react91.useState(() => getCachedOpenAICompatibleProxyHandles() ?? new Set);
   const [openAICompatibleProxyProviders, setOpenAICompatibleProxyProviders] = import_react91.useState(new Set);
+  const providerRegistryRevision = import_react91.useSyncExternalStore(subscribePiProviderRegistry, getPiProviderRegistryRevision, getPiProviderRegistryRevision);
+  const previousProviderRegistryRevision = import_react91.useRef(providerRegistryRevision);
   const mountedRef = import_react91.useRef(true);
   import_react91.useEffect(() => {
     mountedRef.current = true;
@@ -499511,8 +499589,12 @@ function ModelSelector({
     }
   });
   import_react91.useEffect(() => {
+    if (previousProviderRegistryRevision.current !== providerRegistryRevision) {
+      previousProviderRegistryRevision.current = providerRegistryRevision;
+      clearAvailableModelsCache();
+    }
     loadModels.current(forceRefreshOnMount ?? false);
-  }, [forceRefreshOnMount]);
+  }, [forceRefreshOnMount, providerRegistryRevision]);
   import_react91.useEffect(() => {
     if (localModelCatalog) {
       setByokProviderAliases(buildByokProviderAliases([]));
@@ -500198,6 +500280,7 @@ var init_ModelSelector = __esm(async () => {
   init_available_models();
   init_model();
   init_remote_model_catalog();
+  init_pi_provider_mod_registry();
   init_byok_providers();
   init_settings_manager();
   init_colors();
@@ -500484,8 +500567,9 @@ function ProviderSelector({
   const [awsProfiles, setAwsProfiles] = import_react94.useState([]);
   const [profileIndex, setProfileIndex] = import_react94.useState(0);
   const [isLoadingProfiles, setIsLoadingProfiles] = import_react94.useState(false);
-  const providers = import_react94.useMemo(() => getProviderConfigs(selectedTarget), [selectedTarget]);
-  const filteredProviders = import_react94.useMemo(() => filterProviderConfigs(providers, searchQuery), [providers, searchQuery]);
+  import_react94.useSyncExternalStore(subscribePiProviderRegistry, getPiProviderRegistryRevision, getPiProviderRegistryRevision);
+  const providers = getProviderConfigs(selectedTarget);
+  const filteredProviders = filterProviderConfigs(providers, searchQuery);
   const showProviderStoreTabs = shouldShowProviderStoreTabs(hasCloudCredentials2);
   const connectedProviders = import_react94.useMemo(() => connectedProvidersByTarget[selectedTarget] ?? new Map, [connectedProvidersByTarget, selectedTarget]);
   const isLoading = isProviderTargetLoading({
@@ -501759,6 +501843,7 @@ function ProviderSelector({
 var import_react94, jsx_dev_runtime70, SOLID_LINE17 = "─", VISIBLE_PROVIDERS = 8;
 var init_ProviderSelector = __esm(async () => {
   init_available_models();
+  init_pi_provider_mod_registry();
   init_use_terminal_width();
   init_byok_providers();
   init_chatgpt_usage_service();
@@ -541232,4 +541317,4 @@ function registerBunOAuthFlows() {
 registerBunOAuthFlows();
 await init_src5().then(() => exports_src2);
 
-//# debugId=CFB812BEB8775E6764756E2164756E21
+//# debugId=E5837E6C26403C1164756E2164756E21