@bike4mind/cli 2.0.0 → 3.0.0

package/dist/{ImageStore-kVo-oHoS.mjs → ImageStore-DuI3tgwQ.mjs}

@@ -139,12 +139,14 @@ var ImageStore = class ImageStore {
 		const maxDimension = 2048;
 		let width = metadata.width;
 		let height = metadata.height;
-		if (width > maxDimension || height > maxDimension) if (width > height) {
-			height = Math.round(height * maxDimension / width);
-			width = maxDimension;
-		} else {
-			width = Math.round(width * maxDimension / height);
-			height = maxDimension;
+		if (width > maxDimension || height > maxDimension) {
+			if (width > height) {
+				height = Math.round(height * maxDimension / width);
+				width = maxDimension;
+			} else {
+				width = Math.round(width * maxDimension / height);
+				height = maxDimension;
+			}
 		}
 		image = image.resize(width, height);
 		let quality = 85;

package/dist/SandboxRuntimeAdapter-Bt1wPpld.mjs

@@ -0,0 +1,2 @@
+#!/usr/bin/env node
+export { t as createSandboxRuntime, n as detectPlatform, r as expandPath, i as isBinaryAvailable } from "./SandboxRuntimeAdapter-BgLUVTJL.mjs";

package/dist/{WorkItemsClient-Cow6nXx7.mjs → WorkItemsClient-BMY0Qq1s.mjs}

@@ -1,25 +1,73 @@
 #!/usr/bin/env node
+import { kt as tokenEstimateMultiplier } from "./ConfigStore-51UaP9pJ.mjs";
 import { v4 } from "uuid";
 import { get_encoding } from "tiktoken";
+//#region ../../b4m-core/utils/dist/calibratedTokenizer.mjs
+/**
+* Scale a raw cl100k_base count by `multiplier`, rounded up to the nearest whole token. Paired with
+* tokenEstimateMultiplier (@bike4mind/common) to bring a raw count up to a Claude model's tokenizer.
+* The sync counterpart of withTokenEstimateMultiplier, for callers whose counter is synchronous.
+*/
+function scaleTokenEstimate(count, multiplier) {
+	if (multiplier === 1) return count;
+	const hundredths = Math.round(multiplier * 100);
+	return Math.ceil(count * hundredths / 100);
+}
+//#endregion
 //#region src/utils/tokenCounter.ts
 const DEFAULT_CONTEXT_WINDOW = 2e5;
 const IMAGE_BLOCK_TOKEN_ESTIMATE = 1600;
+const TOKEN_CACHE_MIN_CHARS = 256;
+const DEFAULT_TOKEN_CACHE_MAX_CHARS = 4e6;
 /**
 * Token counting utility for context window management.
-* Uses tiktoken (cl100k_base encoding) which works for Claude and GPT-4.
+* Uses tiktoken (cl100k_base encoding), the encoding GPT-4 reports. Claude's tokenizers spend more
+* tokens on the same text, so call `forModel(modelId)` before comparing a count against a model's
+* context window (see tokenEstimateMultiplier in @bike4mind/common).
 */
 var TokenCounter = class {
-	constructor() {
+	constructor(cacheMaxChars = DEFAULT_TOKEN_CACHE_MAX_CHARS) {
+		this.cacheMaxChars = cacheMaxChars;
 		this.encoder = null;
+		this.countCache = /* @__PURE__ */ new Map();
+		this.cachedChars = 0;
 	}
 	getEncoder() {
 		if (!this.encoder) this.encoder = get_encoding("cl100k_base");
 		return this.encoder;
 	}
 	/**
+	* A view of this counter whose counts are scaled to `modelId`'s own tokenizer, or `this` when the
+	* model needs no scaling (non-Claude, or a Claude version tokenEstimateMultiplier leaves at 1). Use
+	* it wherever a count is compared against a model's context window, so a Claude session compacts
+	* and reports against real token usage rather than a cl100k_base undercount.
+	*/
+	forModel(modelId) {
+		const multiplier = tokenEstimateMultiplier(modelId);
+		return multiplier === 1 ? this : new CalibratedTokenCounter(this, multiplier);
+	}
+	/**
 	* Count tokens in a text string
 	*/
 	countTokens(text) {
+		if (text.length < TOKEN_CACHE_MIN_CHARS || text.length > this.cacheMaxChars) return this.encode(text);
+		const cached = this.countCache.get(text);
+		if (cached !== void 0) {
+			this.countCache.delete(text);
+			this.countCache.set(text, cached);
+			return cached;
+		}
+		const count = this.encode(text);
+		this.countCache.set(text, count);
+		this.cachedChars += text.length;
+		for (const [oldest] of this.countCache) {
+			if (this.cachedChars <= this.cacheMaxChars) break;
+			this.countCache.delete(oldest);
+			this.cachedChars -= oldest.length;
+		}
+		return count;
+	}
+	encode(text) {
 		return this.getEncoder().encode_ordinary(text).length;
 	}
 	/**
@@ -74,12 +122,32 @@ var TokenCounter = class {
 	* Free encoder resources when done
 	*/
 	dispose() {
+		this.countCache.clear();
+		this.cachedChars = 0;
 		if (this.encoder) {
 			this.encoder.free();
 			this.encoder = null;
 		}
 	}
 };
+/**
+* A calibrated view over a raw `TokenCounter`. `countTokens` is the one primitive every other count
+* routes through, so overriding it calibrates message, session and tool-schema counts at once. The
+* flat image estimate stays unscaled because it is not a tiktoken count. The view holds no encoder of
+* its own - it reads through `raw` - so `dispose` is a no-op and disposing the view never frees the
+* shared encoder.
+*/
+var CalibratedTokenCounter = class extends TokenCounter {
+	constructor(raw, multiplier) {
+		super();
+		this.raw = raw;
+		this.multiplier = multiplier;
+	}
+	countTokens(text) {
+		return scaleTokenEstimate(this.raw.countTokens(text), this.multiplier);
+	}
+	dispose() {}
+};
 let tokenCounter = null;
 /**
 * Get the singleton TokenCounter instance
@@ -151,7 +219,8 @@ var ConversationContext = class ConversationContext {
 	* rendering each turn's bounded tool trace - plus the system prompt, and
 	* compares against `thresholdRatio` of the context window. Rich-content aware:
 	* replaces the old string-only `content` sum so a session whose weight lives
-	* in tool traces still triggers compaction.
+	* in tool traces still triggers compaction. Counts are calibrated to
+	* `opts.model`, so `systemPromptTokens` must be too (the caller owns it).
 	*/
 	needsCompaction(systemPromptTokens, opts, thresholdRatio = .8) {
 		return this.estimateSessionTokens(systemPromptTokens, opts) >= opts.contextWindow * thresholdRatio;
@@ -162,7 +231,7 @@ var ConversationContext = class ConversationContext {
 	* system prompt. Used by `needsCompaction`.
 	*/
 	estimateSessionTokens(systemPromptTokens, opts) {
-		return systemPromptTokens + this.messages.filter((m) => m.role === "user" || m.role === "assistant").reduce((sum, m) => sum + this.estimateTokens(this.renderHistoryMessage(m, opts)), 0);
+		return systemPromptTokens + this.messages.filter((m) => m.role === "user" || m.role === "assistant").reduce((sum, m) => sum + this.estimateTokens(this.renderHistoryMessage(m, opts), opts), 0);
 	}
 	/** Serialize back to a Session for on-disk persistence (rich content preserved). */
 	toSession() {
@@ -181,12 +250,12 @@ var ConversationContext = class ConversationContext {
 	*/
 	windowedHistory(newInput, opts) {
 		const reserved = opts.reservedTokens ?? 8e3;
-		const currentTokens = this.estimateTokens(this.inputToIMessage(newInput));
+		const currentTokens = this.estimateTokens(this.inputToIMessage(newInput), opts);
 		const history = this.messages.filter((m) => m.role === "user" || m.role === "assistant").map((m) => this.renderHistoryMessage(m, opts));
 		let remaining = opts.contextWindow - reserved - currentTokens;
 		const kept = [];
 		for (let i = history.length - 1; i >= 0; i--) {
-			const cost = this.estimateTokens(history[i]);
+			const cost = this.estimateTokens(history[i], opts);
 			if (cost > remaining) break;
 			remaining -= cost;
 			kept.unshift(history[i]);
@@ -250,7 +319,7 @@ var ConversationContext = class ConversationContext {
 			if (b.type !== "tool_use") continue;
 			const resultText = b.id && results.get(b.id) || "";
 			const line = `[tool ${b.name}] ${this.brief(JSON.stringify(b.input ?? {}))} -> ${this.brief(resultText)}`;
-			const cost = this.counter.countTokens(line);
+			const cost = this.counter.forModel(opts.model).countTokens(line);
 			if (used + cost > budget) {
 				omitted++;
 				continue;
@@ -272,8 +341,8 @@ var ConversationContext = class ConversationContext {
 			content: input
 		};
 	}
-	estimateTokens(message) {
-		return this.counter.countMessageContent(message.content);
+	estimateTokens(message, opts) {
+		return this.counter.forModel(opts.model).countMessageContent(message.content);
 	}
 };
 /**

package/dist/{buildAgent-DOvEp92a.mjs → buildAgent-CrSRKYU_.mjs}

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
-import { I as generateCliTools, J as setWebSocketToolExecutor, L as wrapTools, N as ReActAgent, X as buildSystemPrompt, Y as getPlanModeFilePath, a as loadProjectContext, c as createSseBackend, i as buildProjectAgentStore, it as Logger, l as ServerLlmBackend, n as BackgroundAgentManager, o as McpManager, r as SubagentOrchestrator, t as AgentHistoryStore } from "./AgentHistoryStore-DIMzvf-y.mjs";
-import { $ as hasDeliverablePinnedArtifact, C as ARTIFACT_REMOVED_PLACEHOLDER, E as CONTEXT_WINDOW_SAFETY_BUFFER_TOKENS, J as filterToolArtifactMarkup, K as createThinkMarkerEscaper, M as MODEL_INFO_FIELD_GROUP_OF, O as ChatModels, P as ModelBackend, S as ARTIFACT_DELIVERED_PLACEHOLDER, Z as getQuestErrorCode, ct as stripDeliveredArtifactBlocks, lt as stripToolArtifactMarkup, n as logger, q as escapeThinkMarkers, rt as isUserInitiatedAbort, z as PermissionDeniedError } from "./ConfigStore-DIs-K9oD.mjs";
+import { J as getPlanModeFilePath, L as generateCliTools, N as ReActAgent, R as wrapTools, Y as buildSystemPrompt, a as loadProjectContext, c as createSseBackend, i as buildProjectAgentStore, l as ServerLlmBackend, lt as Logger, n as BackgroundAgentManager, o as McpManager, q as setWebSocketToolExecutor, r as SubagentOrchestrator, t as AgentHistoryStore } from "./AgentHistoryStore-DVxv1kSO.mjs";
+import { D as ARTIFACT_REMOVED_PLACEHOLDER, Dt as stripToolArtifactMarkup, E as ARTIFACT_DELIVERED_PLACEHOLDER, Et as stripDeliveredArtifactBlocks, F as ChatModels, G as ModelBackend, I as CompletionReasoningEffortSchema, N as CONTEXT_WINDOW_SAFETY_BUFFER_TOKENS, U as MODEL_INFO_FIELD_GROUP_OF, X as PermissionDeniedError, _ as logger, ct as escapeThinkMarkers, ft as getQuestErrorCode, lt as filterToolArtifactMarkup, mt as hasDeliverablePinnedArtifact, st as createThinkMarkerEscaper, vt as isUserInitiatedAbort } from "./ConfigStore-51UaP9pJ.mjs";
 import "crypto";
 import { v4 } from "uuid";
 import { z as z$1 } from "zod";
@@ -20,10 +20,28 @@ import WebSocket from "ws";
 * A tool failure that must end the turn rather than be fed back to the model as a
 * recoverable observation: a declined tool (PermissionDeniedError) or an out-of-credits
 * error (tagged via `getQuestErrorCode`), neither of which any retry can satisfy.
+* Also exempts these from the services tool failure guard (toolFailureGuard.ts).
 */
 function isTerminalToolError(error) {
 	return error instanceof PermissionDeniedError || getQuestErrorCode(error) !== void 0;
 }
+/** Whether the named tool is flagged `endsTurnAfterText` in the offered tool list. */
+function isTurnEndingTool(name, tools) {
+	return tools?.find((tool) => tool.toolSchema.name === name)?.endsTurnAfterText === true;
+}
+/**
+* Whether the tool loop should end the turn after this round's tool batch instead of making
+* the follow-up model call. True only when the round already streamed non-empty answer text
+* AND every tool it called is flagged `endsTurnAfterText`: the follow-up round would only
+* repeat the answer. A tool-only round still recurses so the model gets to answer.
+*
+* The single decision point for every adapter's tool loop - keep them from drifting.
+*/
+function shouldEndTurnAfterTools(calledToolNames, tools, roundText) {
+	const names = calledToolNames.filter((name) => Boolean(name));
+	if (names.length === 0 || !roundText?.trim()) return false;
+	return names.every((name) => isTurnEndingTool(name, tools));
+}
 /**
 * Execute an array of async tasks either in parallel (with concurrency limiting)
 * or sequentially, returning outcomes in the original order.
@@ -243,6 +261,8 @@ const stripToolDependentMessages = (messages) => messages.filter((message) => !m
 * WHOLE request with `ValidationException: A maximum of 4 blocks with cache_control may be
 * provided`, which is non-retryable - so an over-budget request loses the turn outright,
 * after the user has already waited for it.
+* See https://docs.claude.com/en/docs/build-with-claude/prompt-caching (tools, system and messages
+* all count toward the same four).
 */
 const MAX_CACHE_CONTROL_BLOCKS = 4;
 /**
@@ -261,18 +281,44 @@ function hasMarker(block) {
 * shareable-prefix breakpoint) - so this adapter's budget is whatever they left, not the full four.
 */
 function censusMarkers(params) {
-	const tools = Array.isArray(params.tools) ? params.tools.filter(hasMarker).length : 0;
-	const system = Array.isArray(params.system) ? params.system.filter(hasMarker).length : 0;
-	let messages = 0;
-	if (Array.isArray(params.messages)) for (const message of params.messages) {
+	const { tools, system, messages } = markerSites(params);
+	return {
+		tools: tools.length,
+		system: system.length,
+		messages: messages.length,
+		total: tools.length + system.length + messages.length
+	};
+}
+function markerSites(params) {
+	const sitesIn = (region, value) => Array.isArray(value) ? value.flatMap((block, index) => hasMarker(block) ? [{
+		region,
+		index
+	}] : []) : [];
+	const messages = [];
+	if (Array.isArray(params.messages)) params.messages.forEach((message, messageIndex) => {
 		const content = message?.content;
-		if (Array.isArray(content)) messages += content.filter(hasMarker).length;
-	}
+		if (!Array.isArray(content)) return;
+		content.forEach((block, index) => {
+			const inner = block?.content;
+			if (Array.isArray(inner)) inner.forEach((innerBlock, nested) => {
+				if (hasMarker(innerBlock)) messages.push({
+					region: "messages",
+					index,
+					message: messageIndex,
+					nested
+				});
+			});
+			if (hasMarker(block)) messages.push({
+				region: "messages",
+				index,
+				message: messageIndex
+			});
+		});
+	});
 	return {
-		tools,
-		system,
-		messages,
-		total: tools + system + messages
+		tools: sitesIn("tools", params.tools),
+		system: sitesIn("system", params.system),
+		messages
 	};
 }
 /**
@@ -371,14 +417,14 @@ var AnthropicCachingAdapter = class {
 			else console.info(message, JSON.stringify(census));
 		} else if (logger) logger.debug("[PromptCache] cache_control census", census);
 		if (outbound.total > MAX_CACHE_CONTROL_BLOCKS) {
-			const message = `[PromptCache] request exceeds the ${MAX_CACHE_CONTROL_BLOCKS}-block cache_control limit on arrival (${outbound.total}); the provider will reject it`;
+			const message = `[PromptCache] request exceeds the ${MAX_CACHE_CONTROL_BLOCKS}-block cache_control limit on arrival (${outbound.total}); capCacheControlMarkers will strip the excess before send`;
 			const detail = {
 				inbound,
 				outbound,
 				limit: MAX_CACHE_CONTROL_BLOCKS
 			};
-			if (logger) logger.error(message, detail);
-			else console.error(message, JSON.stringify(detail));
+			if (logger) logger.warn(message, detail);
+			else console.warn(message, JSON.stringify(detail));
 		} else if (dropped.length > 0) {
 			const message = `[PromptCache] cache_control budget exhausted (limit ${MAX_CACHE_CONTROL_BLOCKS}); skipped breakpoints: ${dropped.join(", ")}`;
 			const detail = {
@@ -479,6 +525,29 @@ var GeminiCachingAdapter = class {
 	}
 };
 /**
+* Cached prompt tokens from a raw provider usage object, across every spelling in use:
+* OpenAI Chat Completions nests them under `prompt_tokens_details`, the OpenAI
+* Responses API under `input_tokens_details`, Moonshot publishes a flat
+* `cached_tokens` alongside the OpenAI-shaped nesting, and DeepSeek its own flat
+* `prompt_cache_hit_tokens`. Reading only one spelling silently bills every cache
+* hit on the other transports at the full input rate.
+*
+* DeepSeek's own spelling leads, because it is the number its invoice is computed
+* from; the OpenAI-shaped ones it also sends are the fallback for a proxy that
+* forwards only those.
+*/
+function cachedTokensFromUsage(usage) {
+	if (!usage) return 0;
+	const candidates = [
+		usage.prompt_cache_hit_tokens,
+		usage.cached_tokens,
+		usage.prompt_tokens_details?.cached_tokens,
+		usage.input_tokens_details?.cached_tokens
+	];
+	for (const value of candidates) if (typeof value === "number" && Number.isFinite(value) && value > 0) return value;
+	return 0;
+}
+/**
 * xAI Grok-specific caching adapter
 * Caching is automatic with optional conversation ID for cache affinity
 */
@@ -495,7 +564,7 @@ var XAICachingAdapter = class {
 		const usage = response.usage;
 		if (!usage) return void 0;
 		const totalInputTokens = usage.prompt_tokens || 0;
-		const cachedTokens = usage.cached_prompt_tokens || 0;
+		const cachedTokens = cachedTokensFromUsage(usage);
 		const cacheHitRate = totalInputTokens > 0 ? cachedTokens / totalInputTokens * 100 : 0;
 		const costSavingsPercent = cacheHitRate * .65;
 		const estimatedLatencyReduction = cacheHitRate * .7;
@@ -553,29 +622,6 @@ var KimiCachingAdapter = class {
 	}
 };
 /**
-* Cached prompt tokens from a raw provider usage object, across every spelling in use:
-* OpenAI Chat Completions nests them under `prompt_tokens_details`, the OpenAI
-* Responses API under `input_tokens_details`, Moonshot publishes a flat
-* `cached_tokens` alongside the OpenAI-shaped nesting, and DeepSeek its own flat
-* `prompt_cache_hit_tokens`. Reading only one spelling silently bills every cache
-* hit on the other transports at the full input rate.
-*
-* DeepSeek's own spelling leads, because it is the number its invoice is computed
-* from; the OpenAI-shaped ones it also sends are the fallback for a proxy that
-* forwards only those.
-*/
-function cachedTokensFromUsage(usage) {
-	if (!usage) return 0;
-	const candidates = [
-		usage.prompt_cache_hit_tokens,
-		usage.cached_tokens,
-		usage.prompt_tokens_details?.cached_tokens,
-		usage.input_tokens_details?.cached_tokens
-	];
-	for (const value of candidates) if (typeof value === "number" && Number.isFinite(value) && value > 0) return value;
-	return 0;
-}
-/**
 * DeepSeek context caching. Automatic, like Moonshot's and xAI's: no parameter,
 * no header, no explicit cache-creation call. The adapter exists only to read
 * the counters back out.
@@ -617,7 +663,8 @@ 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_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;
 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;
@@ -994,6 +1041,20 @@ var OllamaBackend = class OllamaBackend {
 				if (!inheritedArtifactGuard && artifactGuard) await artifactGuard.flush();
 				return;
 			}
+			if (round.toolCalls.length > 0 && shouldEndTurnAfterTools(toolCalls.map((tc) => tc.name), options.tools, round.content)) {
+				this._logger.info("[Tool Execution] Ending turn: answer already produced, only end-of-turn tools ran", {
+					model,
+					toolsExecuted: toolCalls.map((tc) => tc.name)
+				});
+				await (artifactGuard?.callback ?? callback)([round.content], {
+					inputTokens,
+					outputTokens,
+					toolsUsed: executedToolsUsed,
+					stopReason: "tool_use"
+				});
+				if (!inheritedArtifactGuard && artifactGuard) await artifactGuard.flush();
+				return;
+			}
 			await this.complete(model, messages, {
 				...options,
 				_internal: {
@@ -1285,9 +1346,10 @@ var OllamaBackend = class OllamaBackend {
 				const texts = [];
 				const images = [];
 				for (const block of msg.content) if (block.type === "text" && typeof block.text === "string") texts.push(block.text);
-				else if (block.type === "image") if (block.source.type === "base64") images.push(block.source.data);
-				else this._logger.debug("[OllamaBackend] Dropping non-base64 image block; Ollama requires inline base64.");
-				else if (block.type === "image_url") {
+				else if (block.type === "image") {
+					if (block.source.type === "base64") images.push(block.source.data);
+					else this._logger.debug("[OllamaBackend] Dropping non-base64 image block; Ollama requires inline base64.");
+				} else if (block.type === "image_url") {
 					const dataUrl = block.image_url.url.match(/^data:[^,]*;base64,(.+)$/s);
 					if (dataUrl) images.push(dataUrl[1]);
 					else this._logger.debug("[OllamaBackend] Dropping non-data image_url; Ollama requires inline base64.");
@@ -1324,6 +1386,7 @@ ModelBackend.Bedrock, ModelBackend.AWS, ModelBackend.OpenAI, ModelBackend.Anthro
 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;
 //#endregion
 //#region src/llm/NotifyingLlmBackend.ts

package/dist/commands/acpCommand.mjs

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
-import { $ as CustomCommandStore, A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, F as PermissionManager, J as setWebSocketToolExecutor, L as wrapTools, O as createResumeAgentTool, Q as RemoteSkillSource, S as createFindDefinitionTool, T as createSkillTool, b as createGetFileStructureTool, et as CheckpointStore, j as FallbackLlmBackend, k as createBackgroundAgentTools, s as buildSandbox, tt as SessionStore, w as createWriteTodosTool } from "../AgentHistoryStore-DIMzvf-y.mjs";
-import { _ as requireApiUrl, d as isReadOnlyTool, n as logger, t as ConfigStore } from "../ConfigStore-DIs-K9oD.mjs";
-import { a as deferredToolRegistry, i as createToolSearchTool, n as buildSupportingStores, o as NotifyingLlmBackend, r as buildLlmBackend, t as buildAgent } from "../buildAgent-DOvEp92a.mjs";
-import { t as ApiClient } from "../ApiClient-BTI0-m0g.mjs";
+import { A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, F as PermissionManager, I as buildCliToolList, O as createResumeAgentTool, Q as CustomCommandStore, S as createFindDefinitionTool, T as createSkillTool, Z as RemoteSkillSource, b as createGetFileStructureTool, j as FallbackLlmBackend, k as createBackgroundAgentTools, nt as SessionStore, q as setWebSocketToolExecutor, s as buildSandbox, tt as CheckpointStore, w as createWriteTodosTool } from "../AgentHistoryStore-DVxv1kSO.mjs";
+import { _ as logger, o as isReadOnlyTool, t as ConfigStore, u as requireApiUrl } from "../ConfigStore-51UaP9pJ.mjs";
+import { a as deferredToolRegistry, i as createToolSearchTool, n as buildSupportingStores, o as NotifyingLlmBackend, r as buildLlmBackend, t as buildAgent } from "../buildAgent-CrSRKYU_.mjs";
+import { a as ApiClient } from "../dist-BsG2Hlp-.mjs";
 import { randomUUID } from "crypto";
 import { existsSync, realpathSync, statSync } from "fs";
 import { isAbsolute } from "path";
@@ -604,10 +604,7 @@ var AcpServer = class {
 			checkpointStore
 		});
 		const additionalDirectories = await this.resolveAdditionalDirectories();
-		const agentContext = {
-			currentAgent: null,
-			observationQueue: []
-		};
+		const agentContext = { currentAgent: null };
 		const { agentStore, contextResult, loadedB4mTools, orchestrator, backgroundManager, historyStore, mcpManager } = await buildSupportingStores({
 			config,
 			llm,
@@ -629,6 +626,11 @@ var AcpServer = class {
 		});
 		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(() => {
+			if (!agentToolsRef.current) throw new Error("tool_search invoked before agent context was wired");
+			return agentToolsRef.current;
+		}) : null;
 		const cliTools = this.buildCliTools({
 			config,
 			orchestrator,
@@ -640,18 +642,10 @@ var AcpServer = class {
 			agentContext,
 			apiClient,
 			sandboxOrchestrator: sandboxOrchestrator ?? void 0,
-			allowedDirectories: additionalDirectories
+			allowedDirectories: additionalDirectories,
+			toolSearchTool
 		});
-		const agentToolsRef = { current: null };
-		const toolSearchTool = deferredToolRegistry.size() > 0 ? createToolSearchTool(() => {
-			if (!agentToolsRef.current) throw new Error("tool_search invoked before agent context was wired");
-			return agentToolsRef.current;
-		}) : null;
-		const allTools = [
-			...loadedB4mTools,
-			...toolSearchTool ? [toolSearchTool] : [],
-			...cliTools
-		];
+		const allTools = [...loadedB4mTools, ...cliTools];
 		const { agent, buildPromptForMode } = buildAgent({
 			config,
 			modelId: modelInfo.id,
@@ -684,7 +678,7 @@ var AcpServer = class {
 		};
 	}
 	buildCliTools(input) {
-		const { config, orchestrator, agentStore, backgroundManager, historyStore, sessionId, permissionManager, agentContext, apiClient, sandboxOrchestrator, allowedDirectories } = input;
+		const { config, orchestrator, agentStore, backgroundManager, historyStore, sessionId, permissionManager, agentContext, apiClient, sandboxOrchestrator, allowedDirectories, toolSearchTool } = input;
 		const cliWrapDeps = {
 			permissionManager,
 			showPermissionPrompt: this.promptFn,
@@ -694,10 +688,15 @@ var AcpServer = class {
 			sandboxOrchestrator,
 			allowedDirectories
 		};
-		const wrappedSecurityTools = wrapTools([
-			createFindDefinitionTool(allowedDirectories),
-			createGetFileStructureTool(allowedDirectories),
-			...config.preferences.enableSkillTool !== false ? [createSkillTool({
+		return buildCliToolList({
+			toolSearchTool,
+			agentDelegateTool: createAgentDelegateTool(orchestrator, agentStore, sessionId, backgroundManager),
+			backgroundTools: createBackgroundAgentTools(backgroundManager),
+			resumeAgentTool: createResumeAgentTool(orchestrator, historyStore, backgroundManager),
+			writeTodosTool: createWriteTodosTool(createTodoStore()),
+			findDefinitionTool: createFindDefinitionTool(allowedDirectories),
+			getFileStructureTool: createGetFileStructureTool(allowedDirectories),
+			skillTool: config.preferences.enableSkillTool !== false ? createSkillTool({
 				customCommandStore: this.customCommandStore,
 				subagentOrchestrator: orchestrator,
 				sessionId,
@@ -706,17 +705,9 @@ var AcpServer = class {
 					promptFn: this.promptFn
 				},
 				allowedDirectories
-			})] : []
-		], cliWrapDeps);
-		const tools = [
-			createAgentDelegateTool(orchestrator, agentStore, sessionId, backgroundManager),
-			...createBackgroundAgentTools(backgroundManager),
-			createResumeAgentTool(orchestrator, historyStore, backgroundManager),
-			createWriteTodosTool(createTodoStore()),
-			...wrappedSecurityTools
-		];
-		if (config.preferences.enableCoordinatorMode === true) tools.push(createCoordinateTaskTool(orchestrator, agentStore, sessionId));
-		return tools;
+			}) : null,
+			coordinateTaskTool: config.preferences.enableCoordinatorMode === true ? createCoordinateTaskTool(orchestrator, agentStore, sessionId) : null
+		}, cliWrapDeps);
 	}
 	async loadCustomCommands() {
 		this.customCommandStore.setProjectTrusted(this.configStore.isProjectTrusted());

package/dist/commands/apiCommand.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { g as parseApiUrl, t as ConfigStore, v as resolveApiEndpoint } from "../ConfigStore-DIs-K9oD.mjs";
+import { d as resolveApiEndpoint, g as parseApiUrl, t as ConfigStore } from "../ConfigStore-51UaP9pJ.mjs";
 //#region src/commands/apiCommand.ts
 /**
 * External API config command (--api-url / --reset-api)

package/dist/commands/doctorCommand.mjs

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-import { t as version } from "../package-Dl_F2Fah.mjs";
-import { a as fetchLatestVersion, c as isNpmPrefixWritable, i as compareSemver } from "../updateChecker-CQW8bxo6.mjs";
+import { t as version } from "../package-DEsNSP54.mjs";
+import { c as isNpmPrefixWritable, i as fetchLatestVersion, r as compareSemver } from "../updateChecker-CCLVVdH0.mjs";
 import { t as checkRipgrep } from "../ripgrepCheck-BmkyTK2i.mjs";
-import { execSync } from "child_process";
 import { existsSync } from "fs";
 import { homedir } from "os";
 import path from "path";
+import { execSync } from "child_process";
 //#region src/commands/doctorCommand.ts
 /**
 * External doctor command (b4m doctor)

package/dist/commands/envCommand.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as ConfigStore } from "../ConfigStore-DIs-K9oD.mjs";
+import { t as ConfigStore } from "../ConfigStore-51UaP9pJ.mjs";
 //#region src/commands/envCommand.ts
 /**
 * Environment switching for the `--dev` / `--prod` launch flags.

package/dist/commands/headlessCommand.mjs

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
-import { $ as CustomCommandStore, A as createAgentDelegateTool, C as createTodoStore, D as createCoordinateTaskTool, F as PermissionManager, G as classifyCommandRisk, I as generateCliTools, J as setWebSocketToolExecutor, L as wrapTools, N as ReActAgent, O as createResumeAgentTool, Q as RemoteSkillSource, S as createFindDefinitionTool, T as createSkillTool, X as buildSystemPrompt, a as loadProjectContext, b as createGetFileStructureTool, c as createSseBackend, et as CheckpointStore, i as buildProjectAgentStore, j as FallbackLlmBackend, k as createBackgroundAgentTools, n as BackgroundAgentManager, o as McpManager, r as SubagentOrchestrator, s as buildSandbox, t as AgentHistoryStore, tt as SessionStore, w as createWriteTodosTool, x as createWorkItemTools, z as SHELL_LIKE_TOOL_COMMAND_FIELDS } from "../AgentHistoryStore-DIMzvf-y.mjs";
-import { _ as requireApiUrl, d as isReadOnlyTool, n as logger, t as ConfigStore } from "../ConfigStore-DIs-K9oD.mjs";
-import { t as ApiClient } from "../ApiClient-BTI0-m0g.mjs";
-import { r as reconstructTurnBlocks, t as WorkItemsClient } from "../WorkItemsClient-Cow6nXx7.mjs";
+import { A as createAgentDelegateTool, B as SHELL_LIKE_TOOL_COMMAND_FIELDS, C as createTodoStore, D as createCoordinateTaskTool, F as PermissionManager, I as buildCliToolList, K as classifyCommandRisk, L as generateCliTools, N as ReActAgent, O as createResumeAgentTool, Q as CustomCommandStore, R as wrapTools, S as createFindDefinitionTool, T as createSkillTool, Y as buildSystemPrompt, Z as RemoteSkillSource, 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 setWebSocketToolExecutor, r as SubagentOrchestrator, s as buildSandbox, t as AgentHistoryStore, tt as CheckpointStore, w as createWriteTodosTool, x as createWorkItemTools } from "../AgentHistoryStore-DVxv1kSO.mjs";
+import { _ as logger, o as isReadOnlyTool, t as ConfigStore, u as requireApiUrl } from "../ConfigStore-51UaP9pJ.mjs";
+import { r as reconstructTurnBlocks, t as WorkItemsClient } from "../WorkItemsClient-BMY0Qq1s.mjs";
+import { a as ApiClient } from "../dist-BsG2Hlp-.mjs";
 import { randomBytes } from "crypto";
 import { v4 } from "uuid";
 import { readFile } from "fs/promises";
@@ -360,10 +360,7 @@ async function handleHeadlessCommand(options) {
 			checkpointStore,
 			log: headlessSandboxLog
 		});
-		const agentContext = {
-			currentAgent: null,
-			observationQueue: []
-		};
+		const agentContext = { currentAgent: null };
 		const { tools: b4mTools } = await generateCliTools(config.userId, llm, modelInfo.id, permissionManager, promptFn, agentContext, configStore, apiClient, void 0, userQuestionFn, checkpointStore, sandboxOrchestrator, additionalDirectories);
 		const mcpManager = new McpManager(config);
 		const builtinAgentsDir = new URL("../agents/defaults/", import.meta.url).pathname;
@@ -407,11 +404,9 @@ async function handleHeadlessCommand(options) {
 		const resumeAgentTool = createResumeAgentTool(orchestrator, historyStore, backgroundManager);
 		const todoStore = createTodoStore();
 		const writeTodosTool = createWriteTodosTool(todoStore);
-		const findDefinitionTool = wrapTools([createFindDefinitionTool(additionalDirectories)], cliWrapDeps)[0];
-		const getFileStructureTool = wrapTools([createGetFileStructureTool(additionalDirectories)], cliWrapDeps)[0];
 		const workItemTools = config.preferences.enableWorkItemTools ? createWorkItemTools(new WorkItemsClient(apiClient)) : [];
 		const enableSkillTool = config.preferences.enableSkillTool !== false;
-		const skillTool = enableSkillTool ? wrapTools([createSkillTool({
+		const skillTool = enableSkillTool ? createSkillTool({
 			customCommandStore,
 			subagentOrchestrator: orchestrator,
 			sessionId: session.id,
@@ -420,21 +415,18 @@ async function handleHeadlessCommand(options) {
 				promptFn
 			},
 			allowedDirectories: additionalDirectories
-		})], cliWrapDeps)[0] : null;
-		const cliTools = [
+		}) : null;
+		const cliTools = buildCliToolList({
 			agentDelegateTool,
-			...backgroundTools,
+			backgroundTools,
 			resumeAgentTool,
 			writeTodosTool,
-			findDefinitionTool,
-			getFileStructureTool,
-			...workItemTools
-		];
-		if (skillTool) cliTools.push(skillTool);
-		if (config.preferences.enableCoordinatorMode === true) {
-			const coordinateTaskTool = createCoordinateTaskTool(orchestrator, agentStore, session.id);
-			cliTools.push(coordinateTaskTool);
-		}
+			findDefinitionTool: createFindDefinitionTool(additionalDirectories),
+			getFileStructureTool: createGetFileStructureTool(additionalDirectories),
+			workItemTools,
+			skillTool,
+			coordinateTaskTool: config.preferences.enableCoordinatorMode === true ? createCoordinateTaskTool(orchestrator, agentStore, session.id) : null
+		}, cliWrapDeps);
 		const allTools = [
 			...b4mTools,
 			...mcpTools,
@@ -461,7 +453,6 @@ async function handleHeadlessCommand(options) {
 			systemPrompt
 		});
 		agentContext.currentAgent = agent;
-		agent.observationQueue = agentContext.observationQueue;
 		if (streaming) {
 			agent.on("thought", (step) => {
 				emit({