@bike4mind/cli 3.3.0 → 4.0.0

package/dist/{buildAgent-DyZWK0LW.mjs → buildToolSearch-D9GEK91u.mjs}

@@ -1,19 +1,31 @@
 #!/usr/bin/env node
-import { F as generateCliTools, G as setWebSocketToolExecutor, J as DeferredToolRegistry, U as wrapTools, V as buildSystemPrompt, Y as ReActAgent, a as loadProjectContext, c as createSseBackend, i as buildProjectAgentStore, l as ServerLlmBackend, n as BackgroundAgentManager, o as McpManager, q as getPlanModeFilePath, r as SubagentOrchestrator, t as AgentHistoryStore, ut as Logger } from "./AgentHistoryStore-CLidbOHG.mjs";
-import { A as ARTIFACT_REMOVED_PLACEHOLDER, Ct as isUserInitiatedAbort, I as CONTEXT_WINDOW_SAFETY_BUFFER_TOKENS, Mt as stripToolArtifactMarkup, R as ChatModels, Y as ModelBackend, _t as getQuestErrorCode, b as logger, et as PermissionDeniedError, ft as createThinkMarkerEscaper, jt as stripDeliveredArtifactBlocks, k as ARTIFACT_DELIVERED_PLACEHOLDER, mt as filterToolArtifactMarkup, pt as escapeThinkMarkers, q as MODEL_INFO_FIELD_GROUP_OF, yt as hasDeliverablePinnedArtifact, z as CompletionReasoningEffortSchema } from "./ConfigStore-DUATzLxB.mjs";
+import { F as generateCliTools, G as createToolSearchTool, U as DeferredToolRegistry, V as buildSystemPrompt, W as ReActAgent, Y as setWebSocketToolExecutor, Z as getPlanModeFilePath, a as loadProjectContext, c as createSseBackend, i as buildProjectAgentStore, l as ServerLlmBackend, n as BackgroundAgentManager, o as McpManager, q as wrapTools, r as SubagentOrchestrator, st as Logger, t as AgentHistoryStore } from "./AgentHistoryStore-DgqmcVLN.mjs";
+import { $ as ModelBackend, A as ARTIFACT_REMOVED_PLACEHOLDER, At as isUserInitiatedAbort, B as ChatModels, Et as hasDeliverablePinnedArtifact, Gt as visibleReplyText, P as AudioMimeType, R as CONTEXT_WINDOW_SAFETY_BUFFER_TOKENS, Rt as stripDeliveredArtifactBlocks, V as CompletionReasoningEffortSchema, Z as MODEL_INFO_FIELD_GROUP_OF, b as logger, bt as escapeThinkMarkers, dt as TTS_MAX_INPUT_CHARS, it as PermissionDeniedError, k as ARTIFACT_DELIVERED_PLACEHOLDER, lt as SupportedFabFileMimeTypes, ut as TTS_DEFAULT_MODEL, vt as completionsSupportsToolSearch, wt as getQuestErrorCode, xt as filterToolArtifactMarkup, yt as createThinkMarkerEscaper, zt as stripToolArtifactMarkup } from "./ConfigStore-BwqsQ0U9.mjs";
 import "crypto";
+import "path";
 import { v4 } from "uuid";
 import "axios";
+import "node:crypto";
 import "@aws-sdk/client-bedrock-runtime";
 import "openai";
+import "file-type";
 import "@anthropic-ai/sdk";
 import "@aws-sdk/client-cloudwatch";
+import "@anthropic-ai/sdk/core/streaming";
 import "@google/genai";
 import "lodash/pick.js";
 import "openai/streaming";
 import { Ollama } from "ollama";
 import { Agent } from "undici";
+import "@aws-sdk/client-sqs";
+import "@aws-sdk/client-apigatewaymanagementapi";
+import "@aws-sdk/client-lambda";
 import WebSocket from "ws";
+ChatModels.KIMI_K2_THINKING_BEDROCK, ChatModels.KIMI_K2_5_BEDROCK, ChatModels.DEEPSEEK_R1_BEDROCK, ChatModels.DEEPSEEK_FLASH, ChatModels.DEEPSEEK_V4_PRO, ChatModels.KIMI_K3;
+ChatModels.CLAUDE_4_7_OPUS, ChatModels.CLAUDE_4_7_OPUS_BEDROCK, ChatModels.CLAUDE_4_8_OPUS, ChatModels.CLAUDE_4_8_OPUS_BEDROCK, ChatModels.CLAUDE_5_SONNET, ChatModels.CLAUDE_5_SONNET_BEDROCK, ChatModels.CLAUDE_5_OPUS, ChatModels.CLAUDE_5_OPUS_BEDROCK, ChatModels.CLAUDE_5_5_HAIKU, ChatModels.CLAUDE_5_5_HAIKU_BEDROCK, ChatModels.CLAUDE_5_5_SONNET_BEDROCK;
+ChatModels.CLAUDE_5_SONNET, ChatModels.CLAUDE_5_SONNET_BEDROCK, ChatModels.CLAUDE_5_5_SONNET_BEDROCK, ChatModels.CLAUDE_FABLE_5, ChatModels.CLAUDE_FABLE_5_BEDROCK, ChatModels.CLAUDE_FABLE_5_1_BEDROCK, ChatModels.CLAUDE_5_OPUS, ChatModels.CLAUDE_5_OPUS_BEDROCK, ChatModels.CLAUDE_5_5_OPUS, ChatModels.CLAUDE_5_5_OPUS_BEDROCK, ChatModels.CLAUDE_5_5_HAIKU, ChatModels.CLAUDE_5_5_HAIKU_BEDROCK;
+ModelBackend.Anthropic, ModelBackend.Gemini, ModelBackend.XAI, ModelBackend.Ollama, ModelBackend.BFL, ModelBackend.LocalImage, ModelBackend.AWS;
+//#endregion
 //#region ../../b4m-core/llm-adapters/dist/index.mjs
 /**
 * A tool failure that must end the turn rather than be fed back to the model as a
@@ -38,7 +50,7 @@ function isTurnEndingTool(name, tools) {
 */
 function shouldEndTurnAfterTools(calledToolNames, tools, roundText) {
 	const names = calledToolNames.filter((name) => Boolean(name));
-	if (names.length === 0 || !roundText?.trim()) return false;
+	if (names.length === 0 || !visibleReplyText(roundText).trim()) return false;
 	return names.every((name) => isTurnEndingTool(name, tools));
 }
 /**
@@ -668,12 +680,36 @@ var NoOpCachingAdapter = class {
 	extractCacheStats() {}
 };
 ModelBackend.Anthropic, new AnthropicCachingAdapter(), ModelBackend.OpenAI, new OpenAICachingAdapter(), ModelBackend.Gemini, new GeminiCachingAdapter(), ModelBackend.Bedrock, new AnthropicCachingAdapter(), ModelBackend.XAI, new XAICachingAdapter(), ModelBackend.Kimi, new KimiCachingAdapter(), ModelBackend.DeepSeek, new DeepSeekCachingAdapter(), ModelBackend.Ollama, new NoOpCachingAdapter(), ModelBackend.BFL, new NoOpCachingAdapter(), ModelBackend.VoyageAI, new NoOpCachingAdapter(), ModelBackend.AWS, new NoOpCachingAdapter(), ModelBackend.LocalImage, new NoOpCachingAdapter();
-ChatModels.KIMI_K2_THINKING_BEDROCK, ChatModels.KIMI_K2_5_BEDROCK, ChatModels.DEEPSEEK_R1_BEDROCK, ChatModels.DEEPSEEK_FLASH, ChatModels.DEEPSEEK_V4_PRO, ChatModels.KIMI_K3;
-ChatModels.CLAUDE_4_7_OPUS, ChatModels.CLAUDE_4_7_OPUS_BEDROCK, ChatModels.CLAUDE_4_8_OPUS, ChatModels.CLAUDE_4_8_OPUS_BEDROCK, ChatModels.CLAUDE_5_SONNET, ChatModels.CLAUDE_5_SONNET_BEDROCK, ChatModels.CLAUDE_5_OPUS, ChatModels.CLAUDE_5_OPUS_BEDROCK, ChatModels.CLAUDE_5_5_HAIKU, ChatModels.CLAUDE_5_5_HAIKU_BEDROCK, ChatModels.CLAUDE_5_5_SONNET_BEDROCK;
-ChatModels.CLAUDE_5_SONNET, ChatModels.CLAUDE_5_SONNET_BEDROCK, ChatModels.CLAUDE_5_5_SONNET_BEDROCK, ChatModels.CLAUDE_FABLE_5, ChatModels.CLAUDE_FABLE_5_BEDROCK, ChatModels.CLAUDE_FABLE_5_1_BEDROCK, ChatModels.CLAUDE_5_OPUS, ChatModels.CLAUDE_5_OPUS_BEDROCK, ChatModels.CLAUDE_5_5_OPUS, ChatModels.CLAUDE_5_5_OPUS_BEDROCK, ChatModels.CLAUDE_5_5_HAIKU, ChatModels.CLAUDE_5_5_HAIKU_BEDROCK;
 new Logger();
 ChatModels.CLAUDE_4_5_SONNET, ChatModels.CLAUDE_4_1_OPUS, ChatModels.CLAUDE_4_5_HAIKU, ChatModels.CLAUDE_4_5_OPUS, ChatModels.CLAUDE_4_6_SONNET, ChatModels.CLAUDE_4_6_OPUS;
 ChatModels.CLAUDE_4_5_SONNET_BEDROCK, ChatModels.CLAUDE_4_5_HAIKU_BEDROCK, ChatModels.CLAUDE_4_5_OPUS_BEDROCK, ChatModels.CLAUDE_4_6_SONNET_BEDROCK, ChatModels.CLAUDE_4_6_OPUS_BEDROCK;
+/**
+* Models served by ConverseBedrockBackend. Adding a Converse-compatible Bedrock model is a
+* `ChatModels` entry plus a record here: routing (getLlmByModel's Bedrock branch), the
+* Bedrock listing (UndifferentiatedBedrockBackend) and request shaping all read this table.
+*/
+const CONVERSE_BEDROCK_MODELS = [{
+	id: ChatModels.NOVA_PRO_BEDROCK,
+	type: "text",
+	name: "Amazon Nova Pro",
+	backend: ModelBackend.Bedrock,
+	contextWindow: 3e5,
+	max_tokens: 1e4,
+	can_stream: true,
+	pricing: { 3e5: {
+		input: 8e-7,
+		output: 32e-7
+	} },
+	inlineReasoningTag: "thinking",
+	supportsVision: true,
+	supportsTools: true,
+	supportsImageVariation: false,
+	releaseDate: "2024-12-03",
+	rank: 5,
+	description: "Amazon's Nova Pro via AWS Bedrock - multimodal (text and images), tool use, and a 300K-token context window."
+}];
+new Set(CONVERSE_BEDROCK_MODELS.map((m) => String(m.id)));
+ChatModels.DEEPSEEK_R1_BEDROCK, ModelBackend.Bedrock, ChatModels.DEEPSEEK_V3_1, ModelBackend.Bedrock;
 /** Ollama `done_reason` (chat/generate responses). */
 function normalizeOllamaDoneReason(reason) {
 	if (!reason) return void 0;
@@ -741,8 +777,8 @@ function convertMessageToOpenAIFormat(msg, options = {}) {
 			tool_call_id: block.tool_use_id
 		}));
 	}
-	if (msg.requiresTool !== void 0) {
-		const { requiresTool: _requiresTool, ...providerSafe } = msg;
+	if (msg.requiresTool !== void 0 || msg.cache !== void 0 || msg.perTurn !== void 0) {
+		const { requiresTool: _requiresTool, cache: _cache, perTurn: _perTurn, ...providerSafe } = msg;
 		return [providerSafe];
 	}
 	return [msg];
@@ -818,41 +854,35 @@ var OllamaBackend = class OllamaBackend {
 			fetch: fetchWithTimeout
 		});
 	}
+	/** Rejects when Ollama is unreachable; getAvailableModels needs a down host to differ from an empty one. */
 	async getModelInfo() {
-		try {
-			const models = await this._api.list();
-			const isSelfHost = process.env.B4M_SELF_HOST === "true";
-			return await Promise.all(models.models.map(async (model) => {
-				const { capabilities, contextWindow: reportedContextWindow } = await this.getModelDetailsCached(model.name);
-				const contextWindow = OllamaBackend.effectiveContextWindow(reportedContextWindow);
-				return {
-					id: model.name,
-					type: "text",
-					name: model.name,
-					backend: ModelBackend.Ollama,
-					contextWindow,
-					max_tokens: OllamaBackend.advertisedOutputCap(contextWindow),
-					supportsImageVariation: false,
-					pricing: { [contextWindow]: {
-						input: 0,
-						output: 0
-					} },
-					freeToRun: true,
-					supportsVision: capabilities.includes("vision"),
-					supportsTools: capabilities.includes("tools"),
-					can_think: capabilities.includes("thinking"),
-					can_stream: true,
-					logoFile: "Ollama_Logo.svg",
-					rank: 1,
-					description: isSelfHost ? "Runs locally on your own hardware via Ollama. No API key required, and nothing leaves your machine. Performance and capabilities vary by model." : `This model is served from ${process.env.APP_NAME ? `${process.env.APP_NAME}'s` : "the platform"} Ollama servers using publicly available open-source models. Performance and capabilities vary by model.`
-				};
-			}));
-		} catch (error) {
-			let errorMessage = error instanceof Error ? error.message : String(error);
-			if (error instanceof Error && error.message.includes("503 Service Temporarily Unavailable")) errorMessage = "Ollama server is temporarily unavailable. Please try again later.";
-			this._logger.warn("[OllamaBackend] Error fetching model info from Ollama:", errorMessage);
-			return [];
-		}
+		const models = await this._api.list();
+		const isSelfHost = process.env.B4M_SELF_HOST === "true";
+		return await Promise.all(models.models.map(async (model) => {
+			const { capabilities, contextWindow: reportedContextWindow } = await this.getModelDetailsCached(model.name);
+			const contextWindow = OllamaBackend.effectiveContextWindow(reportedContextWindow);
+			return {
+				id: model.name,
+				type: "text",
+				name: model.name,
+				backend: ModelBackend.Ollama,
+				contextWindow,
+				max_tokens: OllamaBackend.advertisedOutputCap(contextWindow),
+				supportsImageVariation: false,
+				pricing: { [contextWindow]: {
+					input: 0,
+					output: 0
+				} },
+				freeToRun: true,
+				supportsVision: capabilities.includes("vision"),
+				supportsTools: capabilities.includes("tools"),
+				can_think: capabilities.includes("thinking"),
+				can_stream: true,
+				logoFile: "Ollama_Logo.svg",
+				rank: 1,
+				description: isSelfHost ? "Runs locally on your own hardware via Ollama. No API key required, and nothing leaves your machine. Performance and capabilities vary by model." : `This model is served from ${process.env.APP_NAME ? `${process.env.APP_NAME}'s` : "the platform"} Ollama servers using publicly available open-source models. Performance and capabilities vary by model.`
+			};
+		}));
 	}
 	/** Ollama's default when a model doesn't report a context length. */
 	static DEFAULT_CONTEXT_WINDOW = 8192;
@@ -1404,9 +1434,96 @@ ChatModels.GPT5_6_SOL, ChatModels.GPT5_6_LUNA, ChatModels.GPT5_6_TERRA;
 ModelBackend.Bedrock, ModelBackend.AWS, ModelBackend.OpenAI, ModelBackend.Anthropic, ModelBackend.Gemini, ModelBackend.Ollama, ModelBackend.BFL, ModelBackend.XAI, ModelBackend.Kimi, ModelBackend.DeepSeek, ModelBackend.LocalImage, ModelBackend.VoyageAI;
 Object.values(ModelBackend);
 Object.entries(MODEL_INFO_FIELD_GROUP_OF);
-ModelBackend.Anthropic, ModelBackend.Gemini, ModelBackend.XAI, ModelBackend.Ollama, ModelBackend.BFL, ModelBackend.LocalImage, ModelBackend.AWS;
 CompletionReasoningEffortSchema.options;
 ModelBackend.Ollama, ModelBackend.LocalImage;
+Object.assign(Object.create(null), {
+	txt: SupportedFabFileMimeTypes.TXT_PLAIN,
+	ini: SupportedFabFileMimeTypes.TXT_PLAIN,
+	env: SupportedFabFileMimeTypes.TXT_PLAIN,
+	conf: SupportedFabFileMimeTypes.TXT_PLAIN,
+	log: SupportedFabFileMimeTypes.TXT_PLAIN,
+	sql: SupportedFabFileMimeTypes.TXT_PLAIN,
+	text: SupportedFabFileMimeTypes.TXT_PLAIN,
+	md: SupportedFabFileMimeTypes.TXT_MARKDOWN,
+	mdx: SupportedFabFileMimeTypes.TXT_MARKDOWN,
+	mmd: SupportedFabFileMimeTypes.TXT_PLAIN,
+	html: SupportedFabFileMimeTypes.HTML,
+	htm: SupportedFabFileMimeTypes.HTML,
+	shtml: SupportedFabFileMimeTypes.HTML,
+	csv: SupportedFabFileMimeTypes.CSV,
+	jpg: SupportedFabFileMimeTypes.JPG,
+	jpeg: SupportedFabFileMimeTypes.JPG,
+	jfif: SupportedFabFileMimeTypes.JPG,
+	jpe: SupportedFabFileMimeTypes.JPG,
+	png: SupportedFabFileMimeTypes.PNG,
+	gif: SupportedFabFileMimeTypes.GIF,
+	svg: SupportedFabFileMimeTypes.SVG,
+	webp: SupportedFabFileMimeTypes.WEBP,
+	pdf: SupportedFabFileMimeTypes.PDF,
+	json: SupportedFabFileMimeTypes.JSON,
+	xml: SupportedFabFileMimeTypes.XML,
+	docx: SupportedFabFileMimeTypes.DOCX,
+	pptx: SupportedFabFileMimeTypes.PPTX,
+	xlsx: SupportedFabFileMimeTypes.XLSX,
+	xls: SupportedFabFileMimeTypes.XLS,
+	js: SupportedFabFileMimeTypes.JS,
+	mjs: SupportedFabFileMimeTypes.JS,
+	cjs: SupportedFabFileMimeTypes.JS,
+	jsx: SupportedFabFileMimeTypes.JSX,
+	ts: SupportedFabFileMimeTypes.TS,
+	tsx: SupportedFabFileMimeTypes.TS,
+	py: SupportedFabFileMimeTypes.PY,
+	java: SupportedFabFileMimeTypes.JAVA,
+	cpp: SupportedFabFileMimeTypes.CPP,
+	c: SupportedFabFileMimeTypes.CPP,
+	h: SupportedFabFileMimeTypes.CPP,
+	cs: SupportedFabFileMimeTypes.CS,
+	php: SupportedFabFileMimeTypes.PHP,
+	rb: SupportedFabFileMimeTypes.RUBY,
+	go: SupportedFabFileMimeTypes.GO,
+	swift: SupportedFabFileMimeTypes.SWIFT,
+	kt: SupportedFabFileMimeTypes.KOTLIN,
+	rs: SupportedFabFileMimeTypes.RUST,
+	css: SupportedFabFileMimeTypes.CSS,
+	less: SupportedFabFileMimeTypes.LESS,
+	sass: SupportedFabFileMimeTypes.SASS,
+	scss: SupportedFabFileMimeTypes.SCSS,
+	yaml: SupportedFabFileMimeTypes.YAML,
+	yml: SupportedFabFileMimeTypes.YAML,
+	toml: SupportedFabFileMimeTypes.TOML,
+	sh: SupportedFabFileMimeTypes.SH,
+	bash: SupportedFabFileMimeTypes.BASH,
+	mp3: AudioMimeType.MP3,
+	wav: AudioMimeType.WAV,
+	opus: AudioMimeType.OPUS,
+	aac: AudioMimeType.AAC,
+	flac: AudioMimeType.FLAC,
+	pcm: AudioMimeType.PCM,
+	ogg: AudioMimeType.OGG,
+	webm: AudioMimeType.WEBM
+});
+/**
+* Coerce any thrown value into a human-readable string.
+*
+* Handles the three common shapes:
+*   - Error  -> .message
+*   - string -> as-is
+*   - other  -> JSON.stringify with a safe fallback for circular/unstringifiable values
+*/
+function extractErrorMessage(error) {
+	if (error instanceof Error) return error.message;
+	if (typeof error === "string") return error;
+	try {
+		return JSON.stringify(error);
+	} catch {
+		return "Unknown error";
+	}
+}
+ChatModels.GPT5, ChatModels.GPT5_MINI, ChatModels.GPT5_NANO, ChatModels.GPT5_1, ChatModels.GPT5_2, ChatModels.GPT5_4, ChatModels.GPT5_4_MINI, ChatModels.GPT5_4_NANO, ChatModels.GPT5_5, ChatModels.GPT5_6_SOL, ChatModels.GPT5_6_LUNA, ChatModels.GPT5_6_TERRA;
+TTS_DEFAULT_MODEL.openai;
+TTS_MAX_INPUT_CHARS.openai;
+TTS_MAX_INPUT_CHARS.elevenlabs;
+TTS_DEFAULT_MODEL.elevenlabs;
 //#endregion
 //#region src/llm/NotifyingLlmBackend.ts
 /**
@@ -1827,9 +1944,19 @@ async function buildLlmBackend(input, deps = defaultLlmBackendDeps) {
 	let models;
 	if (ollamaHost) {
 		const ollamaBackend = deps.createOllamaBackend(ollamaHost);
-		const [serverModels, ollamaModels] = await Promise.all([llm.getModelInfo(), ollamaBackend.getModelInfo()]);
-		if (serverModels.length === 0 && ollamaModels.length === 0) throw new Error(`No models available from server or Ollama at ${ollamaHost}.\nPull a model: ollama pull qwen3.5`);
-		if (ollamaModels.length === 0) startupLog.push(`⚠️  No models found in Ollama at ${ollamaHost}. Pull one with: ollama pull qwen3.5`);
+		const ollamaListing = ollamaBackend.getModelInfo().then((listed) => ({ models: listed }), (error) => {
+			const message = extractErrorMessage(error);
+			logger.warn(`Error fetching model info from Ollama: ${message}`);
+			return {
+				models: [],
+				error: message
+			};
+		});
+		const [serverModels, ollama] = await Promise.all([llm.getModelInfo(), ollamaListing]);
+		const ollamaModels = ollama.models;
+		if (serverModels.length === 0 && ollamaModels.length === 0) throw new Error(ollama.error !== void 0 ? `No models available from server, and Ollama at ${ollamaHost} is unreachable: ${ollama.error}` : `No models available from server or Ollama at ${ollamaHost}.\nPull a model: ollama pull qwen3.5`);
+		if (ollama.error !== void 0) startupLog.push(`\u26a0\ufe0f  Ollama at ${ollamaHost} is unreachable: ${ollama.error}`);
+		else if (ollamaModels.length === 0) startupLog.push(`⚠️  No models found in Ollama at ${ollamaHost}. Pull one with: ollama pull qwen3.5`);
 		const serverBackend = llm;
 		llm = deps.createMultiBackend(serverBackend, ollamaBackend, serverModels, ollamaModels, config.defaultModel);
 		models = await llm.getModelInfo();
@@ -1988,5 +2115,237 @@ function buildAgent(input) {
 		buildPromptForMode
 	};
 }
+/**
+* The `tool_reference` blocks a `toolReference` tool_search result serializes to, or null for any
+* other text (a no-match hint, a mutateTools result, or a result with an appended warning).
+* Extra keys are tolerated so a field added to the core serializer does not silently disable the
+* native path, and dropped because the provider rejects a `tool_reference` block with any other field.
+*/
+function parseToolReferences(text) {
+	let parsed;
+	try {
+		parsed = JSON.parse(text);
+	} catch {
+		return null;
+	}
+	if (!Array.isArray(parsed) || parsed.length === 0) return null;
+	const isReference = (block) => typeof block === "object" && block !== null && block.type === "tool_reference" && typeof block.tool_name === "string";
+	return parsed.every(isReference) ? parsed.map(({ type, tool_name }) => ({
+		type,
+		tool_name
+	})) : null;
+}
+/**
+* The CLI's `tool_search`: one schema, with the load strategy picked per call so a `/model` switch or a
+* fallback model takes effect on the next search. Both strategies share the registry's search and
+* ranking. With `toolReference` the matched tools are still added to the live list, because the agent
+* executes only tools it holds; ToolSearchLlmBackend keeps them deferred on the wire.
+*/
+function createCliToolSearchTool(options) {
+	const { registry, logger, toolListAccessor, useToolReference } = options;
+	const mutating = createToolSearchTool({
+		registry,
+		logger,
+		toolListAccessor
+	});
+	const referencing = createToolSearchTool({
+		registry,
+		logger,
+		loadStrategy: "toolReference"
+	});
+	return {
+		toolSchema: mutating.toolSchema,
+		toolFn: async (params) => {
+			if (!useToolReference()) return mutating.toolFn(params);
+			const result = await referencing.toolFn(params);
+			const references = parseToolReferences(result);
+			if (references) {
+				const liveTools = toolListAccessor();
+				const liveNames = new Set(liveTools.map((t) => t.toolSchema.name));
+				for (const tool of registry.getByNames(references.map((r) => r.tool_name))) {
+					if (liveNames.has(tool.toolSchema.name)) continue;
+					liveNames.add(tool.toolSchema.name);
+					liveTools.push(tool);
+				}
+			}
+			return result;
+		}
+	};
+}
+//#endregion
+//#region src/llm/ToolSearchLlmBackend.ts
+/**
+* Same gate the completions endpoint applies (cliCompletionsToolSearch.ts), so the CLI never sends
+* `tool_reference` blocks the server would refuse. A model missing from `models` reads false.
+*/
+function modelServesToolReference(models, modelId) {
+	const model = models.find((m) => m.id === modelId);
+	return model !== void 0 && completionsSupportsToolSearch(model);
+}
+function toolSearchCallIds(messages) {
+	const ids = /* @__PURE__ */ new Set();
+	for (const message of messages) {
+		if (message.role !== "assistant" || !Array.isArray(message.content)) continue;
+		for (const block of message.content) if (block?.type === "tool_use" && block.name === "tool_search" && typeof block.id === "string") ids.add(block.id);
+	}
+	return ids;
+}
+/**
+* Rewrites the content of each tool_search result that holds serialized references. `rewrite`
+* returning undefined leaves that result as is. Returns `messages` itself when nothing changed.
+*/
+function rewriteToolSearchResults(messages, rewrite) {
+	const callIds = toolSearchCallIds(messages);
+	if (callIds.size === 0) return messages;
+	let changed = false;
+	const result = messages.map((message) => {
+		if (message.role !== "user" || !Array.isArray(message.content)) return message;
+		let messageChanged = false;
+		const content = message.content.map((block) => {
+			if (block?.type !== "tool_result" || typeof block.content !== "string") return block;
+			if (typeof block.tool_use_id !== "string" || !callIds.has(block.tool_use_id)) return block;
+			const references = parseToolReferences(block.content);
+			const replacement = references ? rewrite(references) : void 0;
+			if (replacement === void 0) return block;
+			messageChanged = true;
+			return {
+				...block,
+				content: replacement
+			};
+		});
+		if (!messageChanged) return message;
+		changed = true;
+		return {
+			...message,
+			content
+		};
+	});
+	return changed ? result : messages;
+}
+function asDeferred(tool) {
+	const { cache_control: _cacheControl, ...toolSchema } = tool.toolSchema;
+	return {
+		...tool,
+		toolSchema: {
+			...toolSchema,
+			defer_loading: true
+		}
+	};
+}
+/**
+* Shapes one request's tools and tool_search results for the model it is sent to. History keeps
+* tool_search results as the serialized text the tool returned, so it stays valid across a model switch.
+*
+* - Native tool search: the live tools other than registry ones, then every registry tool in registry
+*   order with `defer_loading: true`, so the array stays byte-identical across a tool load and the
+*   prompt cache holds. Serialized references become `tool_reference` blocks. A registry tool the live
+*   list holds with no reference left in history (loaded by mutateTools before a model switch, or by
+*   the unknown-tool resolver) is sent non-deferred so it stays callable.
+* - Otherwise: tools unchanged, and any references from an earlier native turn become text, since
+*   the endpoint refuses `tool_reference` there. Referenced tools are already in the live list.
+*
+* Requests without `tool_search` (sub-agents, summaries) pass through untouched.
+*/
+function prepareToolSearchRequest(request, registry, native) {
+	const { messages, tools } = request;
+	if (registry.size() === 0 || !tools.some((t) => t.toolSchema.name === "tool_search")) return request;
+	if (!native) {
+		const textMessages = rewriteToolSearchResults(messages, (references) => `Loaded ${references.length} tool schema(s): ${references.map((r) => r.tool_name).join(", ")}. These are now callable in your next message.`);
+		return textMessages === messages ? request : {
+			messages: textMessages,
+			tools
+		};
+	}
+	const registryTools = registry.getAll();
+	const wireTools = tools.filter((t) => !registry.has(t.toolSchema.name));
+	const wireNames = new Set([...wireTools, ...registryTools].map((t) => t.toolSchema.name));
+	const referenced = /* @__PURE__ */ new Set();
+	const wireMessages = rewriteToolSearchResults(messages, (references) => {
+		const resolved = references.filter((r) => wireNames.has(r.tool_name));
+		if (resolved.length === 0) return void 0;
+		for (const reference of resolved) referenced.add(reference.tool_name);
+		return resolved.length === references.length ? references : resolved;
+	});
+	const liveNames = new Set(tools.map((t) => t.toolSchema.name));
+	for (const tool of registryTools) {
+		const name = tool.toolSchema.name;
+		wireTools.push(liveNames.has(name) && !referenced.has(name) ? tool : asDeferred(tool));
+	}
+	if (wireTools.every((t) => t.toolSchema.defer_loading === true)) return request;
+	return {
+		messages: wireMessages,
+		tools: wireTools
+	};
+}
+/**
+* Decorator that applies prepareToolSearchRequest per call, for the model each call actually goes to.
+* Sits inside FallbackLlmBackend so a fallback model is shaped for itself.
+*/
+var ToolSearchLlmBackend = class {
+	constructor(inner, registry, models) {
+		this.inner = inner;
+		this.registry = registry;
+		this.models = models;
+		this.lastToolSearchRequestNative = false;
+	}
+	get currentModel() {
+		return this.inner.currentModel;
+	}
+	set currentModel(model) {
+		this.inner.currentModel = model;
+	}
+	/** Whether the last request offering tool_search went to a native tool search model. */
+	servesToolReference() {
+		return this.lastToolSearchRequestNative;
+	}
+	async complete(model, messages, options, callback) {
+		const tools = options.tools ?? [];
+		if (!tools.some((t) => t.toolSchema.name === "tool_search")) return this.inner.complete(model, messages, options, callback);
+		const native = modelServesToolReference(this.models, model);
+		this.lastToolSearchRequestNative = native;
+		const prepared = prepareToolSearchRequest({
+			messages,
+			tools
+		}, this.registry, native);
+		const preparedOptions = prepared.tools === tools ? options : {
+			...options,
+			tools: prepared.tools
+		};
+		return this.inner.complete(model, prepared.messages, preparedOptions, callback);
+	}
+	pushToolMessages(messages, tool, result, thinkingBlocks) {
+		this.inner.pushToolMessages(messages, tool, result, thinkingBlocks);
+	}
+	async getModelInfo() {
+		return this.inner.getModelInfo();
+	}
+};
+//#endregion
+//#region src/bootstrap/buildToolSearch.ts
+/**
+* The main agent's tool_search and the backend that shapes its requests per model: native tool search
+* (`toolReference` + `defer_loading`) where the model serves it, `mutateTools` everywhere else. With no
+* deferred tools there is no tool_search and the backend is returned unwrapped.
+*/
+function buildToolSearch(input) {
+	const { llm, models, registry, logger, agentToolsRef } = input;
+	if (registry.size() === 0) return {
+		llm,
+		toolSearchTool: null
+	};
+	const toolSearchLlm = new ToolSearchLlmBackend(llm, registry, models);
+	return {
+		llm: toolSearchLlm,
+		toolSearchTool: createCliToolSearchTool({
+			registry,
+			logger,
+			toolListAccessor: () => {
+				if (!agentToolsRef.current) throw new Error("tool_search invoked before agent context was wired");
+				return agentToolsRef.current;
+			},
+			useToolReference: () => toolSearchLlm.servesToolReference()
+		})
+	};
+}
 //#endregion
-export { NotifyingLlmBackend as a, deferredToolRegistry as i, buildSupportingStores as n, buildLlmBackend as r, buildAgent as t };
+export { deferredToolRegistry as a, buildLlmBackend as i, buildAgent as n, NotifyingLlmBackend as o, buildSupportingStores as r, buildToolSearch as t };

package/dist/commands/acpCommand.mjs

@@ -1,8 +1,9 @@
 #!/usr/bin/env node
-import { $ as CustomCommandStore, A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, G as setWebSocketToolExecutor, N as PermissionManager, O as createResumeAgentTool, P as buildCliToolList, Q as RemoteSkillSource, S as createFindDefinitionTool, T as createSkillTool, X as createToolSearchTool, b as createGetFileStructureTool, j as FallbackLlmBackend, k as createBackgroundAgentTools, nt as CheckpointStore, rt as SessionStore, s as buildSandbox, w as createWriteTodosTool } from "../AgentHistoryStore-CLidbOHG.mjs";
-import { b as logger, i as ToolCatalog, p as requireApiUrl, t as ConfigStore } from "../ConfigStore-DUATzLxB.mjs";
-import { a as NotifyingLlmBackend, i as deferredToolRegistry, n as buildSupportingStores, r as buildLlmBackend, t as buildAgent } from "../buildAgent-DyZWK0LW.mjs";
-import { a as ApiClient } from "../dist-DPP7UmK9.mjs";
+import { A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, N as PermissionManager, O as createResumeAgentTool, P as buildCliToolList, Q as CustomCommandStore, S as createFindDefinitionTool, T as createSkillTool, Y as setWebSocketToolExecutor, b as createGetFileStructureTool, j as FallbackLlmBackend, k as createBackgroundAgentTools, nt as SessionStore, s as buildSandbox, tt as CheckpointStore, w as createWriteTodosTool } from "../AgentHistoryStore-DgqmcVLN.mjs";
+import { b as logger, i as ToolCatalog, p as requireApiUrl, t as ConfigStore } from "../ConfigStore-BwqsQ0U9.mjs";
+import { t as RemoteSkillSource } from "../RemoteSkillSource-qTvRkaNA.mjs";
+import { a as deferredToolRegistry, i as buildLlmBackend, n as buildAgent, o as NotifyingLlmBackend, r as buildSupportingStores, t as buildToolSearch } from "../buildToolSearch-D9GEK91u.mjs";
+import { t as ApiClient } from "../ApiClient-_JrM_xWw.mjs";
 import { randomUUID } from "crypto";
 import { existsSync, realpathSync, statSync } from "fs";
 import { isAbsolute } from "path";
@@ -592,7 +593,7 @@ var AcpServer = class {
 		const tokenGetter = async () => (await this.configStore.getAuthTokens())?.accessToken ?? null;
 		await this.mergeRemoteSkills(config, apiClient);
 		const startupLog = [];
-		const { llm, wsManager, modelInfo } = await buildLlmBackend({
+		const { llm, wsManager, models, modelInfo } = await buildLlmBackend({
 			config,
 			apiClient,
 			tokenGetter,
@@ -628,17 +629,16 @@ var AcpServer = class {
 			onBackgroundStatusChange: () => {},
 			onGroupCompletion: () => {}
 		});
-		const llmWithFallback = config.fallbackModels && config.fallbackModels.length > 0 ? new FallbackLlmBackend(llm, config.fallbackModels, (from, to, error) => logger.debug(`[acp] Model "${from}" failed (${error.message}); falling back to "${to}"`)) : llm;
-		const notifyingLlm = new NotifyingLlmBackend(llmWithFallback, backgroundManager);
 		const agentToolsRef = { current: null };
-		const toolSearchTool = deferredToolRegistry.size() > 0 ? createToolSearchTool({
+		const { llm: toolSearchLlm, toolSearchTool } = buildToolSearch({
+			llm,
+			models,
 			registry: deferredToolRegistry,
 			logger,
-			toolListAccessor: () => {
-				if (!agentToolsRef.current) throw new Error("tool_search invoked before agent context was wired");
-				return agentToolsRef.current;
-			}
-		}) : null;
+			agentToolsRef
+		});
+		const llmWithFallback = config.fallbackModels && config.fallbackModels.length > 0 ? new FallbackLlmBackend(toolSearchLlm, config.fallbackModels, (from, to, error) => logger.debug(`[acp] Model "${from}" failed (${error.message}); falling back to "${to}"`)) : toolSearchLlm;
+		const notifyingLlm = new NotifyingLlmBackend(llmWithFallback, backgroundManager);
 		const cliTools = this.buildCliTools({
 			config,
 			orchestrator,

package/dist/commands/apiCommand.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { m as resolveApiEndpoint, t as ConfigStore, y as parseApiUrl } from "../ConfigStore-DUATzLxB.mjs";
+import { m as resolveApiEndpoint, t as ConfigStore, y as parseApiUrl } from "../ConfigStore-BwqsQ0U9.mjs";
 //#region src/commands/apiCommand.ts
 /**
 * External API config command (--api-url / --reset-api)

package/dist/commands/doctorCommand.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as version } from "../package-DYPVrada.mjs";
+import { t as version } from "../package-CQaam4Dq.mjs";
 import { c as isNpmPrefixWritable, i as fetchLatestVersion, r as compareSemver } from "../updateChecker-CCLVVdH0.mjs";
 import { t as checkRipgrep } from "../ripgrepCheck-BmkyTK2i.mjs";
 import { existsSync } from "fs";

package/dist/commands/envCommand.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as ConfigStore } from "../ConfigStore-DUATzLxB.mjs";
+import { t as ConfigStore } from "../ConfigStore-BwqsQ0U9.mjs";
 //#region src/commands/envCommand.ts
 /**
 * Environment switching for the `--dev` / `--prod` launch flags.

package/dist/commands/headlessCommand.mjs

@@ -1,8 +1,9 @@
 #!/usr/bin/env node
-import { $ as CustomCommandStore, A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, F as generateCliTools, G as setWebSocketToolExecutor, K as classifyCommandRisk, N as PermissionManager, O as createResumeAgentTool, P as buildCliToolList, Q as RemoteSkillSource, S as createFindDefinitionTool, T as createSkillTool, U as wrapTools, V as buildSystemPrompt, Y as ReActAgent, a as loadProjectContext, b as createGetFileStructureTool, c as createSseBackend, i as buildProjectAgentStore, j as FallbackLlmBackend, k as createBackgroundAgentTools, n as BackgroundAgentManager, nt as CheckpointStore, o as McpManager, r as SubagentOrchestrator, rt as SessionStore, s as buildSandbox, t as AgentHistoryStore, w as createWriteTodosTool, x as createWorkItemTools } from "../AgentHistoryStore-CLidbOHG.mjs";
-import { a as commandTextOf, b as logger, c as toolCategory, i as ToolCatalog, p as requireApiUrl, t as ConfigStore } from "../ConfigStore-DUATzLxB.mjs";
-import { r as reconstructTurnBlocks, t as WorkItemsClient } from "../WorkItemsClient-BSoOblGp.mjs";
-import { a as ApiClient } from "../dist-DPP7UmK9.mjs";
+import { A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, F as generateCliTools, N as PermissionManager, O as createResumeAgentTool, P as buildCliToolList, Q as CustomCommandStore, S as createFindDefinitionTool, T as createSkillTool, V as buildSystemPrompt, W as ReActAgent, X as classifyCommandRisk, Y as setWebSocketToolExecutor, a as loadProjectContext, b as createGetFileStructureTool, c as createSseBackend, i as buildProjectAgentStore, j as FallbackLlmBackend, k as createBackgroundAgentTools, n as BackgroundAgentManager, nt as SessionStore, o as McpManager, q as wrapTools, r as SubagentOrchestrator, s as buildSandbox, t as AgentHistoryStore, tt as CheckpointStore, w as createWriteTodosTool, x as createWorkItemTools } from "../AgentHistoryStore-DgqmcVLN.mjs";
+import { a as commandTextOf, b as logger, c as toolCategory, i as ToolCatalog, p as requireApiUrl, t as ConfigStore } from "../ConfigStore-BwqsQ0U9.mjs";
+import { t as RemoteSkillSource } from "../RemoteSkillSource-qTvRkaNA.mjs";
+import { r as reconstructTurnBlocks, t as WorkItemsClient } from "../WorkItemsClient-CERTDrhT.mjs";
+import { t as ApiClient } from "../ApiClient-_JrM_xWw.mjs";
 import { randomBytes } from "crypto";
 import { v4 } from "uuid";
 import { readFile } from "fs/promises";

package/dist/commands/mcpCommand.mjs

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
-import { t as ConfigStore } from "../ConfigStore-DUATzLxB.mjs";
-import { t as version } from "../package-DYPVrada.mjs";
+import { t as ConfigStore } from "../ConfigStore-BwqsQ0U9.mjs";
+import { t as version } from "../package-CQaam4Dq.mjs";
 //#region src/commands/mcpCommand.ts
 /**
 * External MCP commands (b4m mcp list, b4m mcp add, etc.)
@@ -70,12 +70,13 @@ async function handleMcpCommand(subcommand, argv) {
 			await handleDisable(config, argv.name, configStore);
 			break;
 		case "serve": {
-			const { handleMcpServeCommand } = await import("../serve-w6v3rZiR.mjs");
+			const { handleMcpServeCommand } = await import("../serve-CFxUhjml.mjs");
 			await handleMcpServeCommand({
 				http: Boolean(argv.http),
 				port: typeof argv.port === "number" ? argv.port : void 0,
 				apiKey: typeof argv.apiKey === "string" ? argv.apiKey : void 0,
 				apiUrl: typeof argv.apiUrl === "string" ? argv.apiUrl : void 0,
+				token: typeof argv.token === "string" ? argv.token : void 0,
 				version
 			});
 			break;

package/dist/commands/pluginCommand.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as ConfigStore } from "../ConfigStore-DUATzLxB.mjs";
+import { t as ConfigStore } from "../ConfigStore-BwqsQ0U9.mjs";
 import { n as getDefaultPluginsDir, r as isFeatureEnabled, t as PluginStore } from "../PluginStore-DwvOJ-G3.mjs";
 import { promises } from "fs";
 import path from "path";