@bike4mind/cli 0.10.0 → 0.10.2

package/dist/index.mjs

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
-import { $ as CommandHistoryStore, A as formatStep, B as DEFAULT_RETRY_CONFIG, C as WebSocketToolExecutor, D as ServerLlmBackend, E as WebSocketLlmBackend, F as PermissionManager, G as OllamaBackend, H as clearFeatureModuleTools, I as generateCliTools, J as buildSkillsPromptSection, K as getPlanModeFilePath, L as ALWAYS_DENIED_FOR_AGENTS, M as loadContextFiles, N as getApiUrl, O as McpManager, P as getEnvironmentName, Q as CheckpointStore, R as DEFAULT_AGENT_MODEL, S as ApiClient, T as FallbackLlmBackend, U as registerFeatureModuleTools, V as DEFAULT_THOROUGHNESS, W as setWebSocketToolExecutor, X as ReActAgent, Y as isReadOnlyTool, Z as CustomCommandStore, _ as createAgentDelegateTool, a as createBlockerTools, at as mergeCommands, b as createSkillTool, c as createDecisionStore, ct as warmFileCache, d as createFindDefinitionTool, et as SessionStore, f as createTodoStore, g as BackgroundAgentManager, h as createBackgroundAgentTools, i as createBlockerStore, it as searchCommands, j as extractCompactInstructions, k as substituteArguments, l as formatDecisionsOutput, m as createCoordinateTaskTool, n as createReviewGateTool, nt as hasFileReferences, o as formatBlockersOutput, ot as formatFileSize, p as createWriteTodosTool, q as buildSystemPrompt, r as formatReviewGatesOutput, rt as processFileReferences, s as createDecisionLogTool, st as searchFiles, t as createReviewGateStore, tt as OAuthClient, u as createGetFileStructureTool, v as AgentStore, w as WebSocketConnectionManager, x as parseAgentConfig, y as SubagentOrchestrator, z as DEFAULT_MAX_ITERATIONS } from "./tools-_X4rUM4L.mjs";
+import { $ as SessionStore, A as formatStep, B as DEFAULT_RETRY_CONFIG, C as WebSocketToolExecutor, D as ServerLlmBackend, E as WebSocketLlmBackend, F as PermissionManager, G as getPlanModeFilePath, H as clearFeatureModuleTools, I as generateCliTools, J as isReadOnlyTool, K as buildSystemPrompt, L as ALWAYS_DENIED_FOR_AGENTS, M as loadContextFiles, N as getApiUrl, O as McpManager, P as getEnvironmentName, Q as CommandHistoryStore, R as DEFAULT_AGENT_MODEL, S as ApiClient, T as FallbackLlmBackend, U as registerFeatureModuleTools, V as DEFAULT_THOROUGHNESS, W as setWebSocketToolExecutor, X as CustomCommandStore, Y as ReActAgent, Z as CheckpointStore, _ as createAgentDelegateTool, a as createBlockerTools, at as formatFileSize, b as createSkillTool, c as createDecisionStore, d as createFindDefinitionTool, et as OAuthClient, f as createTodoStore, g as BackgroundAgentManager, h as createBackgroundAgentTools, i as createBlockerStore, it as mergeCommands, j as extractCompactInstructions, k as substituteArguments, l as formatDecisionsOutput, m as createCoordinateTaskTool, n as createReviewGateTool, nt as processFileReferences, o as formatBlockersOutput, ot as searchFiles, p as createWriteTodosTool, q as buildSkillsPromptSection, r as formatReviewGatesOutput, rt as searchCommands, s as createDecisionLogTool, st as warmFileCache, t as createReviewGateStore, tt as hasFileReferences, u as createGetFileStructureTool, v as AgentStore, w as WebSocketConnectionManager, x as parseAgentConfig, y as SubagentOrchestrator, z as DEFAULT_MAX_ITERATIONS } from "./tools-t-9dtjjl.mjs";
 import { n as useCliStore, t as selectActiveBackgroundAgents } from "./store-DV5s-qni.mjs";
-import { Xt as validateNotebookPath$1, Yt as validateJupyterKernelName, g as CREDIT_DEDUCT_TRANSACTION_TYPES, n as logger, t as ConfigStore, v as ChatModels } from "./ConfigStore-DPWN3-0c.mjs";
-import { a as version, t as checkForUpdate } from "./updateChecker-C3DYG0Gn.mjs";
+import { Ht as validateNotebookPath$1, Vt as validateJupyterKernelName, g as ChatModels, m as CREDIT_DEDUCT_TRANSACTION_TYPES, n as logger, t as ConfigStore } from "./ConfigStore-BX6XcTa1.mjs";
+import { a as version, t as checkForUpdate } from "./updateChecker-DdgysXM8.mjs";
 import React, { useCallback, useEffect, useMemo, useReducer, useRef, useState } from "react";
 import { Box, Static, Text, render, useApp, useInput, usePaste, useStdout } from "ink";
 import { execSync } from "child_process";
@@ -22,6 +22,7 @@ import SelectInput from "ink-select-input";
 import jwt from "jsonwebtoken";
 import open from "open";
 import axios, { isAxiosError } from "axios";
+import { OllamaBackend } from "@bike4mind/llm-adapters";
 import { get_encoding } from "tiktoken";
 import WsWebSocket from "ws";
 //#region src/components/StatusBar.tsx
@@ -3761,6 +3762,196 @@ Unlike agent_delegate (which uses pre-defined agents), this tool lets you compos
 	};
 }
 //#endregion
+//#region src/tools/deferredToolRegistry.ts
+/**
+* Registry of tool schemas that are NOT loaded into the model's initial tool
+* list. The model sees only the names (via the system prompt directory) and
+* must call the `tool_search` meta-tool to load schemas on demand.
+*
+* This mirrors Claude Code's deferred-tool pattern. The win is large for
+* heavy MCP integrations (e.g. 41 GitHub MCP tools at ~250-350 tokens of
+* JSONSchema each = ~10-15k tokens per turn that's now ~1-1.5k of names).
+*/
+var DeferredToolRegistry = class {
+	constructor() {
+		this.byName = /* @__PURE__ */ new Map();
+	}
+	/** Replace registry contents with the supplied tools. Idempotent. */
+	register(tools) {
+		this.byName.clear();
+		for (const tool of tools) this.byName.set(tool.toolSchema.name, tool);
+		logger.debug(`[DeferredToolRegistry] Registered ${tools.length} deferred tool(s)`);
+	}
+	clear() {
+		this.byName.clear();
+	}
+	size() {
+		return this.byName.size;
+	}
+	has(name) {
+		return this.byName.has(name);
+	}
+	get(name) {
+		return this.byName.get(name);
+	}
+	getAll() {
+		return Array.from(this.byName.values());
+	}
+	/** Return tools whose names appear in the supplied list, in input order. */
+	getByNames(names) {
+		const found = [];
+		for (const name of names) {
+			const tool = this.byName.get(name);
+			if (tool) found.push(tool);
+		}
+		return found;
+	}
+	/**
+	* Rank-search deferred tools by query terms. Name matches outrank
+	* description matches; exact substring on name wins ties.
+	*/
+	searchByKeywords(query, maxResults) {
+		const terms = query.toLowerCase().split(/\s+/).filter((t) => t.length > 0);
+		if (terms.length === 0) return [];
+		const scored = [];
+		for (const tool of this.byName.values()) {
+			const name = tool.toolSchema.name.toLowerCase();
+			const desc = (tool.toolSchema.description || "").toLowerCase();
+			let score = 0;
+			for (const term of terms) {
+				if (name.includes(term)) score += 10;
+				if (desc.includes(term)) score += 1;
+			}
+			if (score > 0) scored.push({
+				tool,
+				score
+			});
+		}
+		scored.sort((a, b) => b.score - a.score);
+		return scored.slice(0, maxResults).map((s) => s.tool);
+	}
+	/** Return the directory entries used to render the system-prompt reminder. */
+	getDirectoryNames() {
+		return Array.from(this.byName.keys()).sort();
+	}
+};
+const deferredToolRegistry = new DeferredToolRegistry();
+//#endregion
+//#region src/tools/toolSearchTool.ts
+/**
+* Default number of tools returned for a keyword search. Matches Claude
+* Code's ToolSearch convention. 5 keeps the response payload small while
+* surfacing enough alternatives for the model to refine its query.
+*/
+const DEFAULT_MAX_RESULTS = 5;
+/**
+* Runtime validation for tool_search params. The LLM produces these
+* values, so we validate at this boundary rather than trusting the
+* shape. Coerces `max_results` from string→number for models that emit
+* numeric-looking strings.
+*/
+const ToolSearchParamsSchema = z.object({
+	query: z.string().min(1, "query must be a non-empty string"),
+	max_results: z.coerce.number().int().min(1).max(20).optional()
+});
+/**
+* Parse the query string. Two forms:
+*   - `select:name1,name2,...` — exact-name selection
+*   - free text — keyword search across name + description
+*/
+function parseQuery(query) {
+	const trimmed = query.trim();
+	const selectMatch = trimmed.match(/^select:(.+)$/i);
+	if (selectMatch) return {
+		mode: "select",
+		names: selectMatch[1].split(",").map((n) => n.trim()).filter((n) => n.length > 0)
+	};
+	return {
+		mode: "search",
+		text: trimmed
+	};
+}
+/**
+* Format the loaded-tools response. Mirrors Claude Code's convention:
+* one <function>{...}</function> line per matched tool. The model has
+* already seen this format in its tool-registration system messages, so
+* it parses without additional explanation.
+*
+* Note: the schemas are *also* injected into context.tools by the caller,
+* so on the next iteration the model gets them as native tool definitions.
+* The text response here is for in-turn awareness and audit trail.
+*/
+function renderToolsBlock(tools) {
+	if (tools.length === 0) return "";
+	return `<functions>\n${tools.map((tool) => {
+		const schema = {
+			description: tool.toolSchema.description,
+			name: tool.toolSchema.name,
+			parameters: tool.toolSchema.parameters
+		};
+		return `<function>${JSON.stringify(schema)}</function>`;
+	}).join("\n")}\n</functions>`;
+}
+/**
+* Build the tool_search meta-tool. The returned tool has a closure over
+* the supplied `toolListAccessor`, which it uses to push newly-resolved
+* tool schemas into the live agent context.
+*
+* Idempotent: re-loading a tool that's already in the context is a no-op.
+*/
+function createToolSearchTool(toolListAccessor) {
+	return {
+		toolSchema: {
+			name: "tool_search",
+			description: "Fetches full schema definitions for deferred tools so they can be called. Deferred tools appear by name only in a system reminder; their parameter schemas are NOT loaded by default. Use this tool to load schemas on demand. Query forms: 'select:name1,name2' for exact selection, or free-text keywords to search by name and description. Once a tool's schema is returned, it becomes callable in subsequent turns.",
+			parameters: {
+				type: "object",
+				properties: {
+					query: {
+						type: "string",
+						description: "Either 'select:<comma-separated names>' to fetch specific tools, or free-text keywords (e.g. 'github pull request') to rank-search deferred tools."
+					},
+					max_results: {
+						type: "number",
+						description: `Maximum number of tools to return for keyword search. Defaults to ${DEFAULT_MAX_RESULTS}. Ignored for 'select:' queries.`
+					}
+				},
+				required: ["query"]
+			}
+		},
+		toolFn: async (params) => {
+			const parsedParams = ToolSearchParamsSchema.safeParse(params ?? {});
+			if (!parsedParams.success) {
+				const issue = parsedParams.error.issues[0];
+				return `tool_search: invalid parameters — ${issue.path.join(".") || "params"}: ${issue.message}`;
+			}
+			const { query, max_results } = parsedParams.data;
+			const parsed = parseQuery(query);
+			let matched;
+			let unmatched = [];
+			if (parsed.mode === "select") {
+				matched = deferredToolRegistry.getByNames(parsed.names);
+				const foundNames = new Set(matched.map((t) => t.toolSchema.name));
+				unmatched = parsed.names.filter((n) => !foundNames.has(n));
+			} else {
+				const max = max_results ?? DEFAULT_MAX_RESULTS;
+				matched = deferredToolRegistry.searchByKeywords(parsed.text, max);
+			}
+			if (matched.length === 0) return parsed.mode === "select" ? `tool_search: no deferred tools matched ${parsed.names.join(", ")}. Use a free-text query to search.` : `tool_search: no deferred tools matched query "${parsed.text}".`;
+			const liveTools = toolListAccessor();
+			const liveNames = new Set(liveTools.map((t) => t.toolSchema.name));
+			let added = 0;
+			for (const tool of matched) if (!liveNames.has(tool.toolSchema.name)) {
+				liveTools.push(tool);
+				added++;
+			}
+			logger.debug(`[tool_search] query="${query}" matched=${matched.length} added=${added} alreadyLoaded=${matched.length - added}`);
+			const block = renderToolsBlock(matched);
+			return `${`Loaded ${added} new tool schema(s)${added < matched.length ? ` (${matched.length - added} already loaded)` : ""}. These are now callable in your next message.${unmatched.length > 0 ? `\n\nNot found: ${unmatched.join(", ")}` : ""}`}\n\n${block}`;
+		}
+	};
+}
+//#endregion
 //#region src/features/FeatureModuleRegistry.ts
 /**
 * Manages the lifecycle of opt-in CLI feature modules.
@@ -5976,10 +6167,23 @@ function CliApp() {
 				loadContextFiles(agentProjectDir)
 			]);
 			const mcpTools = mcpManager.getTools();
+			const deferredB4mToolNames = new Set([
+				"math_evaluate",
+				"dice_roll",
+				"current_datetime",
+				"recent_changes",
+				"prompt_enhancement"
+			]);
+			const deferredB4mTools = b4mTools.filter((t) => deferredB4mToolNames.has(t.toolSchema.name));
+			const loadedB4mTools = b4mTools.filter((t) => !deferredB4mToolNames.has(t.toolSchema.name));
+			deferredToolRegistry.register([...mcpTools, ...deferredB4mTools]);
 			if (mcpTools.length > 0) {
 				const serverSummaries = mcpManager.getToolCount().map((s) => `${s.serverName} (${s.count})`).join(", ");
-				startupLog.push(`🛠️ Loaded ${b4mTools.length} B4M + ${mcpTools.length} MCP tool(s): ${serverSummaries}`);
-			} else startupLog.push(`🛠️ Loaded ${b4mTools.length} B4M tool(s), no MCP tools`);
+				startupLog.push(`🛠️ Loaded ${loadedB4mTools.length} B4M + ${mcpTools.length} MCP tool(s, ${deferredB4mTools.length + mcpTools.length} deferred): ${serverSummaries}`);
+			} else {
+				const suffix = deferredB4mTools.length > 0 ? ` (${deferredB4mTools.length} deferred)` : "";
+				startupLog.push(`🛠️ Loaded ${loadedB4mTools.length} B4M tool(s)${suffix}, no MCP tools`);
+			}
 			const agentSummary = agentStore.getSummary();
 			startupLog.push(`🤖 Loaded ${agentSummary.total} agent(s): ${agentSummary.builtin} built-in, ${agentSummary.global} global, ${agentSummary.project} project`);
 			const orchestrator = new SubagentOrchestrator({
@@ -6053,9 +6257,14 @@ function CliApp() {
 				cliTools.push(coordinateTaskTool);
 			}
 			const featureTools = featureRegistry.getAllTools();
+			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 = [
-				...b4mTools,
-				...mcpTools,
+				...loadedB4mTools,
+				...toolSearchTool ? [toolSearchTool] : [],
 				...cliTools,
 				...featureTools
 			];
@@ -6070,7 +6279,7 @@ function CliApp() {
 				const moduleNames = featureRegistry.getModuleNames().join(", ");
 				startupLog.push(`🏰 Feature modules: ${moduleNames} (${featureTools.length} tools)`);
 			}
-			logger.debug(`Total tools available to agent: ${allTools.length} (${b4mTools.length} B4M + ${mcpTools.length} MCP + ${cliTools.length} CLI + ${featureTools.length} feature)`);
+			logger.debug(`Total tools available to agent: ${allTools.length} (${loadedB4mTools.length} B4M loaded + ${cliTools.length} CLI + ${featureTools.length} feature + ${toolSearchTool ? 1 : 0} tool_search, ${deferredB4mTools.length} B4M + ${mcpTools.length} MCP deferred)`);
 			if (contextResult.globalContext) startupLog.push(`📄 Global context: ${contextResult.globalContext.filename}`);
 			if (contextResult.projectContext) startupLog.push(`📄 Project context: ${contextResult.projectContext.filename}`);
 			for (const error of contextResult.errors) startupLog.push(`⚠️  Context file error: ${error}`);
@@ -6084,7 +6293,8 @@ function CliApp() {
 				enableDynamicAgentCreation: config.preferences.enableDynamicAgentCreation === true,
 				additionalDirectories,
 				featureModulePrompts: featureModulePrompts || void 0,
-				planModeFilePath: mode === "plan" ? getPlanModeFilePath(newSession.id) : void 0
+				planModeFilePath: mode === "plan" ? getPlanModeFilePath(newSession.id) : void 0,
+				deferredToolNames: deferredToolRegistry.getDirectoryNames()
 			});
 			const cliSystemPrompt = buildPromptForMode(useCliStore.getState().interactionMode);
 			const maxIterations = config.preferences.maxIterations === null ? 999999 : config.preferences.maxIterations;
@@ -6097,16 +6307,17 @@ function CliApp() {
 				maxIterations,
 				maxTokens: config.preferences.maxTokens,
 				temperature: config.preferences.temperature,
-				systemPrompt: cliSystemPrompt
+				systemPrompt: cliSystemPrompt,
+				unknownToolResolver: async (toolName) => deferredToolRegistry.get(toolName) ?? null
 			});
+			agentToolsRef.current = agent.getTools();
 			agentContext.currentAgent = agent;
-			const agentInternalContext = agent.context;
 			let lastInteractionMode = useCliStore.getState().interactionMode;
 			useCliStore.subscribe((s) => {
 				if (s.interactionMode === lastInteractionMode) return;
 				lastInteractionMode = s.interactionMode;
 				if (agentContext.currentAgent !== agent) return;
-				agentInternalContext.systemPrompt = buildPromptForMode(s.interactionMode);
+				agent.setSystemPrompt(buildPromptForMode(s.interactionMode));
 			});
 			agent.observationQueue = agentContext.observationQueue;
 			const stepHandler = (step) => {
@@ -6562,7 +6773,8 @@ function CliApp() {
 				customCommands: state.customCommandStore.getAllCommands(),
 				enableSkillTool: config?.preferences.enableSkillTool !== false,
 				additionalDirectories: state.additionalDirectories,
-				featureModulePrompts: state.featureRegistry?.getSystemPromptSections() || void 0
+				featureModulePrompts: state.featureRegistry?.getSystemPromptSections() || void 0,
+				deferredToolNames: deferredToolRegistry.getDirectoryNames()
 			});
 			if (tokenCounter.countSessionTokens(activeSession, systemPrompt).totalTokens >= threshold) {
 				console.log("\n⚠️  Context window 80% full. Auto-compacting...\n");
@@ -7753,7 +7965,7 @@ Multi-line Input:
 				const skillsTokens = skillsSection ? tokenCounter.countTokens(skillsSection) : 0;
 				const agentDirectoryTokens = state.agentStore ? tokenCounter.countTokens(state.agentStore.getDirectoryContext()) : 0;
 				const mcpTools = state.mcpManager?.getTools() || [];
-				const mcpToolsTokens = tokenCounter.countToolSchemaTokens(mcpTools);
+				const deferredNames = deferredToolRegistry.getDirectoryNames();
 				const mcpToolCount = state.mcpManager?.getToolCount() || [];
 				const systemPrompt = buildSystemPrompt(variantForCount, {
 					contextContent: state.contextContent,
@@ -7761,10 +7973,13 @@ Multi-line Input:
 					customCommands: commands,
 					enableSkillTool: state.config?.preferences.enableSkillTool !== false,
 					additionalDirectories: state.additionalDirectories,
-					featureModulePrompts: state.featureRegistry?.getSystemPromptSections() || void 0
+					featureModulePrompts: state.featureRegistry?.getSystemPromptSections() || void 0,
+					deferredToolNames: deferredNames
 				});
 				const usage = tokenCounter.countSessionTokens(state.session, systemPrompt);
-				const totalWithTools = usage.totalTokens + mcpToolsTokens;
+				const agentTools = state.agent?.getTools() ?? [];
+				const agentToolsTokens = tokenCounter.countToolSchemaTokens(agentTools);
+				const totalWithTools = usage.totalTokens + agentToolsTokens;
 				const usagePercent = totalWithTools / contextWindow * 100;
 				const BAR_WIDTH = 40;
 				const filledWidth = Math.min(Math.round(usagePercent / 100 * BAR_WIDTH), BAR_WIDTH);
@@ -7781,10 +7996,14 @@ Multi-line Input:
 					console.log(`  Agent Directory:   ${agentDirectoryTokens.toLocaleString()} tokens (${agentCount} agents)`);
 				}
 				if (mcpTools.length > 0) {
-					console.log("\nMCP Tools:");
-					console.log(`  Total:             ${mcpToolsTokens.toLocaleString()} tokens (${mcpTools.length} tools)`);
+					console.log("\nMCP Tools (deferred):");
+					console.log(`  Schemas:           load on demand via tool_search (${mcpTools.length} tools available)`);
 					for (const { serverName, count } of mcpToolCount) console.log(`    ${serverName}: ${count} tools`);
 				}
+				if (agentTools.length > 0) {
+					console.log("\nLoaded Tool Schemas:");
+					console.log(`  Schemas:           ${agentToolsTokens.toLocaleString()} tokens (${agentTools.length} tools active)`);
+				}
 				console.log("\nConversation:");
 				console.log(`  Messages:          ${usage.messageTokens.toLocaleString()} tokens (${state.session.messages.length} messages)`);
 				if (usagePercent >= 80) {
@@ -8320,14 +8539,13 @@ Multi-line Input:
 			const newToolNames = newFeatureRegistry.getAllToolNames();
 			if (newToolNames.length > 0) registerFeatureModuleTools(newToolNames);
 			if (state.wsManager && newFeatureRegistry.hasModules) newFeatureRegistry.registerAllWsHandlers(state.wsManager);
-			const agentContext = state.agent.context;
 			const oldFeatureToolNames = new Set(state.featureRegistry?.getAllToolNames() ?? []);
-			const baseTools = agentContext.tools.filter((t) => !oldFeatureToolNames.has(t.toolSchema.name));
+			const baseTools = state.agent.getTools().filter((t) => !oldFeatureToolNames.has(t.toolSchema.name));
 			const newFeatureTools = newFeatureRegistry.getAllTools();
-			agentContext.tools = [...baseTools, ...newFeatureTools];
+			state.agent.setTools([...baseTools, ...newFeatureTools]);
 			const newFeaturePrompts = newFeatureRegistry.getSystemPromptSections();
 			const planFilePathForRebuild = useCliStore.getState().interactionMode === "plan" && state.session ? getPlanModeFilePath(state.session.id) : void 0;
-			agentContext.systemPrompt = buildSystemPrompt(updatedConfig.preferences.promptVariant ?? "current", {
+			state.agent.setSystemPrompt(buildSystemPrompt(updatedConfig.preferences.promptVariant ?? "current", {
 				contextContent: state.contextContent,
 				agentStore: state.agentStore || void 0,
 				customCommands: state.customCommandStore.getAllCommands(),
@@ -8335,8 +8553,9 @@ Multi-line Input:
 				enableDynamicAgentCreation: updatedConfig.preferences.enableDynamicAgentCreation === true,
 				additionalDirectories: state.additionalDirectories,
 				featureModulePrompts: newFeaturePrompts || void 0,
-				planModeFilePath: planFilePathForRebuild
-			});
+				planModeFilePath: planFilePathForRebuild,
+				deferredToolNames: deferredToolRegistry.getDirectoryNames()
+			}));
 			const moduleNames = newFeatureRegistry.getModuleNames();
 			if (moduleNames.length > 0) console.error(`\n\x1b[36m🏰 Feature modules hot-reloaded: ${moduleNames.join(", ")}\x1b[0m`);
 			else console.error(`\n\x1b[36m🏰 Feature modules disabled\x1b[0m`);