@kernel.chat/kbot 1.3.1 → 2.3.0

package/dist/agent.js

@@ -1,23 +1,37 @@
-// K:BOT Agent Loop — ReAct reasoning cycle
-// Message → API → Tool Calls → Execute Locally → Continue or Return
+// K:BOT Agent Loop v2 — Autonomous Reasoning Engine
+// Message → Think → Plan → Execute → Verify → Learn → Return
 //
-// EFFICIENCY-FIRST DESIGN:
+// INTELLIGENCE ARCHITECTURE:
 // 1. Local-first: handle simple tasks (file reads, git, ls) without any API call
-// 2. Context batching: gather all relevant context before the first API call
-// 3. One-shot prompting: send rich context so the agent gets it right in one try
-// 4. Tool execution is always local and free — only reasoning burns tokens
-// 5. Smart model selection: use fast model for simple tasks, default for complex
-import { getApiKey, getApiBase, isByokEnabled, getByokKey, getByokProvider, getProviderModel, getProvider, estimateCost, } from './auth.js';
+// 2. Deep context: project memory + learned patterns + user knowledge + corrections
+// 3. Plan-then-execute: think step by step, then use tools aggressively
+// 4. Self-correction: if a tool fails, analyze the error and try a different approach
+// 5. Auto-research: if you don't know how, search GitHub/web FIRST, then execute
+// 6. Verify: always confirm the output exists and is correct
+// 7. Learn: extract knowledge from every interaction for future use
+import { getByokKey, getByokProvider, getProviderModel, getProvider, estimateCost, isLocalProvider, warmOllamaModelCache, } from './auth.js';
 import { executeTool, getTool, getToolDefinitionsForApi, } from './tools/index.js';
 import { formatContextForPrompt } from './context.js';
+import { getMatrixSystemPrompt } from './matrix.js';
+import { buildFullLearningContext, findPattern, recordPattern, cacheSolution, updateProfile, classifyTask, extractKeywords, learnFromExchange, updateProjectMemory, shouldAutoTrain, selfTrain, } from './learning.js';
 import { getMemoryPrompt, addTurn, getPreviousMessages } from './memory.js';
-import { createSpinner, printToolCall, printToolResult, printResponse, printError, printInfo } from './ui.js';
-const MAX_TOOL_LOOPS = 15;
-const KBOT_VERSION = '1.2.0';
+import { createSpinner, printToolCall, printToolResult, printResponse, printError, printInfo, printWarn } from './ui.js';
+import { parseMultimodalMessage, toAnthropicContent } from './multimodal.js';
+import { streamAnthropicResponse, streamOpenAIResponse } from './streaming.js';
+import { checkPermission } from './permissions.js';
+import { runPreToolHook, runPostToolHook } from './hooks.js';
+const MAX_TOOL_LOOPS = 75;
+/** Maximum cumulative cost (USD) before auto-stopping tool loops */
+const MAX_COST_CEILING = 1.00;
+/** Cached system prompt — rebuilt only when context changes, not every iteration */
+let _cachedSystemPrompt = null;
+let _cachedSystemPromptKey = null;
 // ── Local-first execution ──
 async function tryLocalFirst(message) {
     const lower = message.toLowerCase().trim();
-    const readMatch = lower.match(/^(?:read|show|cat|view|open)\s+(.+)$/i);
+    // Only match file-like paths — avoid intercepting "open chrome" or "show me how to..."
+    const readMatch = lower.match(/^(?:read|cat|view)\s+(.+)$/i)
+        || lower.match(/^(?:show|open)\s+((?:\.{0,2}\/|~\/|\w+\.\w+).+)$/i);
     if (readMatch) {
         const tool = getTool('read_file');
         if (tool)
@@ -44,7 +58,10 @@ async function tryLocalFirst(message) {
         if (tool)
             return tool.execute({});
     }
-    const grepMatch = lower.match(/^(?:grep|search|find)\s+['"""]?(.+?)['"""]?\s+(?:in\s+)?(.+)$/i);
+    // Only match explicit grep-like patterns: "grep PATTERN in PATH" or "search for PATTERN in PATH"
+    // Avoid intercepting natural language like "search the web for..."
+    const grepMatch = lower.match(/^(?:grep)\s+['"""]?(.+?)['"""]?\s+(?:in\s+)?(.+)$/i)
+        || lower.match(/^(?:search|find)\s+(?:for\s+)?['"""]?(.+?)['"""]?\s+in\s+(.+)$/i);
     if (grepMatch) {
         const tool = getTool('grep');
         if (tool)
@@ -56,7 +73,26 @@ async function tryLocalFirst(message) {
     return null;
 }
 /** Anthropic Messages API (Claude) */
-async function callAnthropic(apiKey, apiUrl, model, systemContext, messages) {
+async function callAnthropic(apiKey, apiUrl, model, systemContext, messages, tools, options) {
+    // Build messages — use multimodal content blocks if images are present
+    const apiMessages = messages.map((m, i) => {
+        // Only the first user message might have images
+        if (i === messages.length - 1 && m.role === 'user' && options?.multimodal?.isMultimodal) {
+            return { role: m.role, content: toAnthropicContent(options.multimodal) };
+        }
+        return { role: m.role, content: m.content };
+    });
+    const body = {
+        model,
+        max_tokens: options?.thinking ? 16384 : 8192,
+        system: systemContext || undefined,
+        messages: apiMessages,
+    };
+    if (tools && tools.length > 0)
+        body.tools = tools;
+    if (options?.thinking) {
+        body.thinking = { type: 'enabled', budget_tokens: options.thinkingBudget || 10000 };
+    }
     const res = await fetch(apiUrl, {
         method: 'POST',
         headers: {
@@ -64,46 +100,142 @@ async function callAnthropic(apiKey, apiUrl, model, systemContext, messages) {
             'x-api-key': apiKey,
             'anthropic-version': '2023-06-01',
         },
-        body: JSON.stringify({
-            model,
-            max_tokens: 8192,
-            system: systemContext || undefined,
-            messages: messages.map(m => ({ role: m.role, content: m.content })),
-        }),
+        body: JSON.stringify(body),
     });
     if (!res.ok) {
         const err = await res.json().catch(() => ({ error: { message: `HTTP ${res.status}` } }));
         throw new Error(err.error?.message || `Anthropic error: ${res.status}`);
     }
     const data = await res.json();
-    const text = (data.content || []).filter((b) => b.type === 'text').map((b) => b.text).join('');
+    const contentBlocks = data.content || [];
+    const text = contentBlocks.filter((b) => b.type === 'text').map((b) => b.text).join('');
+    const thinkingText = contentBlocks.filter((b) => b.type === 'thinking').map((b) => b.thinking).join('');
+    const toolUseBlocks = contentBlocks.filter((b) => b.type === 'tool_use');
     const u = data.usage || {};
-    return { content: text, model: data.model, usage: { input_tokens: u.input_tokens || 0, output_tokens: u.output_tokens || 0 } };
+    const result = {
+        content: text,
+        thinking: thinkingText || undefined,
+        model: data.model,
+        usage: { input_tokens: u.input_tokens || 0, output_tokens: u.output_tokens || 0 },
+        stop_reason: data.stop_reason,
+    };
+    if (toolUseBlocks.length > 0) {
+        result.tool_calls = toolUseBlocks.map((b) => ({
+            id: b.id,
+            name: b.name,
+            arguments: b.input || {},
+        }));
+    }
+    return result;
 }
 /** OpenAI-compatible Chat Completions API
- *  Works with: OpenAI, Mistral, xAI, DeepSeek, Groq, Together, Fireworks, Perplexity
+ *  Works with: OpenAI, Mistral, xAI, DeepSeek, Groq, Together, Fireworks, Perplexity, Ollama, OpenClaw
  */
-async function callOpenAICompat(apiKey, apiUrl, model, systemContext, messages) {
+async function callOpenAICompat(apiKey, apiUrl, model, systemContext, messages, tools) {
     const apiMessages = [];
     if (systemContext)
         apiMessages.push({ role: 'system', content: systemContext });
     apiMessages.push(...messages.map(m => ({ role: m.role, content: m.content })));
+    const body = { model, max_tokens: 8192, messages: apiMessages };
+    if (tools && tools.length > 0) {
+        body.tools = tools.map(t => ({
+            type: 'function',
+            function: { name: t.name, description: t.description, parameters: t.input_schema },
+        }));
+    }
+    // Local providers (Ollama, OpenClaw) may not need auth headers
+    const headers = { 'Content-Type': 'application/json' };
+    if (apiKey && apiKey !== 'local') {
+        headers['Authorization'] = `Bearer ${apiKey}`;
+    }
     const res = await fetch(apiUrl, {
         method: 'POST',
-        headers: {
-            'Content-Type': 'application/json',
-            'Authorization': `Bearer ${apiKey}`,
-        },
-        body: JSON.stringify({ model, max_tokens: 8192, messages: apiMessages }),
+        headers,
+        body: JSON.stringify(body),
+        signal: AbortSignal.timeout(300_000), // 5 min timeout for local models
     });
     if (!res.ok) {
         const err = await res.json().catch(() => ({ error: { message: `HTTP ${res.status}` } }));
         throw new Error(err.error?.message || `API error: ${res.status}`);
     }
     const data = await res.json();
-    const content = data.choices?.[0]?.message?.content || '';
+    const choice = data.choices?.[0] || {};
+    let content = choice.message?.content || '';
     const u = data.usage || {};
-    return { content, model: data.model || model, usage: { input_tokens: u.prompt_tokens || 0, output_tokens: u.completion_tokens || 0 } };
+    const result = {
+        content,
+        model: data.model || model,
+        usage: { input_tokens: u.prompt_tokens || 0, output_tokens: u.completion_tokens || 0 },
+        stop_reason: choice.finish_reason,
+    };
+    // Standard tool_calls from the API
+    if (choice.message?.tool_calls && choice.message.tool_calls.length > 0) {
+        result.tool_calls = choice.message.tool_calls.map((tc) => ({
+            id: tc.id,
+            name: tc.function.name,
+            arguments: typeof tc.function.arguments === 'string'
+                ? JSON.parse(tc.function.arguments)
+                : tc.function.arguments || {},
+        }));
+    }
+    // Fallback: Small local models (7B) sometimes emit tool calls as raw JSON in content
+    // instead of structured tool_calls. Parse these so tools still work with Ollama.
+    if (!result.tool_calls && content && tools && tools.length > 0) {
+        const toolNames = tools.map(t => t.function?.name || t.name).filter(Boolean);
+        const parsed = tryParseInlineToolCalls(content, toolNames);
+        if (parsed.length > 0) {
+            result.tool_calls = parsed;
+            // Remove the raw JSON from the displayed content
+            result.content = content.replace(/```(?:json)?\s*\{[\s\S]*?\}\s*```/g, '').replace(/\{[\s\S]*?"name"\s*:\s*"[a-z_]+[\s\S]*?\}/g, '').trim();
+        }
+    }
+    return result;
+}
+/** Try to parse tool calls that local models emit as raw JSON in their text output */
+function tryParseInlineToolCalls(content, knownTools) {
+    const calls = [];
+    // Pattern 1: JSON in code blocks ```json { "name": "tool_name", ... } ```
+    const codeBlockPattern = /```(?:json)?\s*(\{[\s\S]*?\})\s*```/g;
+    let match;
+    while ((match = codeBlockPattern.exec(content)) !== null) {
+        const parsed = tryParseToolJson(match[1], knownTools);
+        if (parsed)
+            calls.push(parsed);
+    }
+    if (calls.length > 0)
+        return calls;
+    // Pattern 2: Raw JSON objects with "name" field matching known tools
+    const jsonPattern = /\{[^{}]*"name"\s*:\s*"([a-z_]+)"[^{}]*\}/g;
+    while ((match = jsonPattern.exec(content)) !== null) {
+        if (knownTools.includes(match[1])) {
+            const parsed = tryParseToolJson(match[0], knownTools);
+            if (parsed)
+                calls.push(parsed);
+        }
+    }
+    return calls;
+}
+function tryParseToolJson(json, knownTools) {
+    try {
+        const obj = JSON.parse(json);
+        const name = obj.name || obj.function?.name;
+        if (!name || !knownTools.includes(name))
+            return null;
+        const args = obj.arguments || obj.parameters || obj.input || obj;
+        // Remove meta fields to get clean arguments
+        const cleanArgs = { ...args };
+        delete cleanArgs.name;
+        delete cleanArgs.function;
+        delete cleanArgs.type;
+        return {
+            id: `inline_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`,
+            name,
+            arguments: cleanArgs,
+        };
+    }
+    catch {
+        return null;
+    }
 }
 /** Google Gemini API */
 async function callGemini(apiKey, apiUrl, model, systemContext, messages) {
@@ -155,123 +287,381 @@ async function callCohere(apiKey, apiUrl, model, systemContext, messages) {
     const u = data.usage?.tokens || {};
     return { content, model, usage: { input_tokens: u.input_tokens || 0, output_tokens: u.output_tokens || 0 } };
 }
+/** Streaming provider call — tokens appear progressively in terminal */
+async function callProviderStreaming(provider, apiKey, model, systemContext, messages, tools, options) {
+    const p = getProvider(provider);
+    let state;
+    if (p.apiStyle === 'anthropic') {
+        state = await streamAnthropicResponse(apiKey, p.apiUrl, model, systemContext, messages.map(m => ({ role: m.role, content: m.content })), tools, { thinking: options?.thinking, thinkingBudget: options?.thinkingBudget });
+    }
+    else {
+        state = await streamOpenAIResponse(apiKey, p.apiUrl, model, systemContext, messages.map(m => ({ role: m.role, content: m.content })), tools);
+    }
+    const result = {
+        content: state.content,
+        thinking: state.thinking || undefined,
+        model: state.model || model,
+        usage: state.usage,
+        stop_reason: state.stopReason,
+    };
+    if (state.toolCalls.length > 0) {
+        result.tool_calls = state.toolCalls
+            .filter(tc => tc.name)
+            .map(tc => {
+            let args = {};
+            if (tc.partialJson) {
+                try {
+                    args = JSON.parse(tc.partialJson);
+                }
+                catch { /* malformed JSON from stream — use empty args */ }
+            }
+            return {
+                id: tc.id || `stream_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`,
+                name: tc.name,
+                arguments: args,
+            };
+        });
+    }
+    // Fallback: if local model streamed tool calls as raw JSON in content,
+    // try to parse them (won't fix the displayed output, but tools will work)
+    if (!result.tool_calls && result.content && tools && tools.length > 0) {
+        const toolNames = tools.map(t => t.name);
+        const parsed = tryParseInlineToolCalls(result.content, toolNames);
+        if (parsed.length > 0) {
+            result.tool_calls = parsed;
+            result.content = result.content
+                .replace(/```(?:json)?\s*\{[\s\S]*?\}\s*```/g, '')
+                .replace(/\{[\s\S]*?"name"\s*:\s*"[a-z_]+[\s\S]*?\}/g, '')
+                .trim();
+        }
+    }
+    return result;
+}
+/** Detect if a message is casual conversation that doesn't need tools */
+function isCasualMessage(message) {
+    const lower = message.toLowerCase().trim();
+    // Very short messages are usually conversational
+    if (lower.length < 20 && !/\b(fix|create|build|run|deploy|install|delete|remove|write|edit|read|find|search|open|show|list|git|npm|pip|cargo)\b/.test(lower)) {
+        return true;
+    }
+    // Greetings and chitchat
+    const casualPatterns = [
+        /^(hey|hi|hello|yo|sup|what's up|whats up|howdy|hola)\b/,
+        /^(how are you|how's it going|what's good|how do you do)\b/,
+        /^(thanks|thank you|thx|ty|cool|nice|great|awesome|perfect|ok|okay|sure|got it|understood)\b/,
+        /^(do you|can you|are you|what are you|who are you|what is|what's your)\b/,
+        /^(tell me about|explain|what do you think|how does|why does|why is|what if)\b/,
+        /^(good morning|good night|good evening|good afternoon|gm|gn)\b/,
+        /^(bye|goodbye|see you|later|peace|quit|exit)\b/,
+        /^(yes|no|maybe|probably|nah|nope|yep|yeah)\b/,
+        /^(lol|lmao|haha|bruh|wow|damn|dang|omg|wtf)\b/,
+        /\?$/, // Questions are usually conversational unless they contain action words
+    ];
+    // If it matches a casual pattern AND doesn't contain action words, it's casual
+    const isCasualPattern = casualPatterns.some(p => p.test(lower));
+    const hasActionWords = /\b(fix|create|build|run|deploy|install|delete|remove|write|edit|make|generate|scaffold|refactor|update|add|implement|set up|configure|debug|test)\b/.test(lower);
+    if (isCasualPattern && !hasActionWords)
+        return true;
+    // Questions that end with ? and don't have action words
+    if (lower.endsWith('?') && !hasActionWords && lower.length < 100)
+        return true;
+    return false;
+}
+/** Core tools that small local models can handle without getting confused */
+const CORE_TOOLS = new Set([
+    'read_file', 'write_file', 'list_directory', 'bash',
+    'git_status', 'git_diff', 'git_commit', 'git_log',
+    'grep', 'web_search',
+]);
+/** Detect if a message describes a complex multi-step task */
+function isComplexTask(message) {
+    const lower = message.toLowerCase();
+    const complexSignals = [
+        /\b(refactor|migrate|convert|rewrite|restructure|reorganize)\b/,
+        /\b(all files|every file|across the|codebase|entire project)\b/,
+        /\b(first|then|after that|finally|step \d|phase \d)\b/,
+        /\b(add (?:a |an )?new (?:feature|system|module|layer))\b/,
+        /\b(build and deploy|ci\s*\/?\s*cd|set up|configure)\b.*\b(pipeline|workflow|infrastructure)\b/,
+    ];
+    const signalCount = complexSignals.filter(r => r.test(lower)).length;
+    const isLong = message.length > 500;
+    return signalCount >= 2 || (signalCount >= 1 && isLong);
+}
 /** Universal provider call — routes to the right API format */
-async function callProvider(provider, apiKey, model, systemContext, messages) {
+async function callProvider(provider, apiKey, model, systemContext, messages, tools, options) {
     const p = getProvider(provider);
-    switch (p.apiStyle) {
-        case 'anthropic': return callAnthropic(apiKey, p.apiUrl, model, systemContext, messages);
-        case 'google': return callGemini(apiKey, p.apiUrl, model, systemContext, messages);
-        case 'cohere': return callCohere(apiKey, p.apiUrl, model, systemContext, messages);
-        case 'openai': return callOpenAICompat(apiKey, p.apiUrl, model, systemContext, messages);
-        default: return callOpenAICompat(apiKey, p.apiUrl, model, systemContext, messages);
+    try {
+        switch (p.apiStyle) {
+            case 'anthropic': return await callAnthropic(apiKey, p.apiUrl, model, systemContext, messages, tools, options);
+            case 'google': return await callGemini(apiKey, p.apiUrl, model, systemContext, messages);
+            case 'cohere': return await callCohere(apiKey, p.apiUrl, model, systemContext, messages);
+            case 'openai': return await callOpenAICompat(apiKey, p.apiUrl, model, systemContext, messages, tools);
+            default: return await callOpenAICompat(apiKey, p.apiUrl, model, systemContext, messages, tools);
+        }
+    }
+    catch (err) {
+        // Auto-retry with fallback model for local providers
+        if (isLocalProvider(provider) && model !== p.fastModel) {
+            const errMsg = err instanceof Error ? err.message : String(err);
+            // Only retry on model-specific errors, not connection errors
+            if (errMsg.includes('not found') || errMsg.includes('does not exist') || errMsg.includes('model')) {
+                printWarn(`Model ${model} unavailable, falling back to ${p.fastModel}...`);
+                switch (p.apiStyle) {
+                    case 'openai': return callOpenAICompat(apiKey, p.apiUrl, p.fastModel, systemContext, messages, tools);
+                    default: return callOpenAICompat(apiKey, p.apiUrl, p.fastModel, systemContext, messages, tools);
+                }
+            }
+        }
+        throw err;
     }
 }
 // ── Main agent loop ──
 export async function runAgent(message, options = {}) {
-    const byokMode = isByokEnabled();
-    const apiKey = byokMode ? getByokKey() : getApiKey();
-    if (!apiKey) {
-        throw new Error(byokMode
-            ? 'No LLM API key configured. Run `kbot byok` to set up.'
-            : 'No API key configured. Run `kbot auth` to set up.');
-    }
-    // Step 1: Local-first
-    const localResult = await tryLocalFirst(message);
-    if (localResult !== null) {
-        addTurn({ role: 'user', content: message });
-        addTurn({ role: 'assistant', content: localResult });
-        printInfo('(handled locally — 0 tokens used)');
-        return { content: localResult, agent: 'local', model: 'none', toolCalls: 0 };
+    const apiKey = getByokKey();
+    const byokProvider = getByokProvider();
+    const isLocal = byokProvider ? isLocalProvider(byokProvider) : false;
+    if (!apiKey && !isLocal) {
+        throw new Error('No LLM API key configured. Run `kbot byok` to set up, or `kbot ollama` for local models.');
+    }
+    // Step 0a: Warm Ollama model cache if using local provider
+    if (isLocal && byokProvider === 'ollama') {
+        warmOllamaModelCache().catch(() => { }); // non-blocking
+    }
+    // Step 0: Parse multimodal content (images in message)
+    const parsed = options.multimodal || parseMultimodalMessage(message);
+    if (parsed.isMultimodal) {
+        printInfo(`(${parsed.imageCount} image${parsed.imageCount > 1 ? 's' : ''} attached)`);
+    }
+    // Step 1: Local-first (skip if multimodal — needs AI to interpret)
+    if (!parsed.isMultimodal) {
+        const localResult = await tryLocalFirst(message);
+        if (localResult !== null) {
+            addTurn({ role: 'user', content: message });
+            addTurn({ role: 'assistant', content: localResult });
+            printInfo('(handled locally — 0 tokens used)');
+            return { content: localResult, agent: 'local', model: 'none', toolCalls: 0 };
+        }
+    }
+    // Step 1.5: Complexity detection — auto-plan complex tasks
+    if (isComplexTask(message) && !message.startsWith('/plan')) {
+        printInfo('Complex task detected. Using autonomous planner...');
+        try {
+            const { autonomousExecute, formatPlanSummary } = await import('./planner.js');
+            const plan = await autonomousExecute(message, {
+                ...options,
+                agent: options.agent || 'coder',
+            }, { autoApprove: false, onApproval: async () => true });
+            const summary = formatPlanSummary(plan);
+            addTurn({ role: 'user', content: message });
+            addTurn({ role: 'assistant', content: summary });
+            return {
+                content: summary,
+                agent: options.agent || 'coder',
+                model: 'planner',
+                toolCalls: plan.steps.filter(s => s.status === 'done').length,
+            };
+        }
+        catch {
+            // Planner failed — fall through to regular agent loop
+            printWarn('Planner failed, falling back to direct execution...');
+        }
     }
     const tier = options.tier || 'free';
-    const tools = getToolDefinitionsForApi(tier);
-    // Step 2: Build context
+    const allTools = getToolDefinitionsForApi(tier);
+    const casual = isCasualMessage(message);
+    // Smart tool filtering:
+    // 1. Casual messages → no tools (just chat)
+    // 2. Local small models → core tools only (10 instead of 60+, prevents confusion)
+    // 3. Everything else → full tool set
+    let tools;
+    if (casual) {
+        tools = []; // No tools for casual conversation
+    }
+    else if (isLocal) {
+        tools = allTools.filter(t => CORE_TOOLS.has(t.name));
+    }
+    else {
+        tools = allTools;
+    }
+    // Step 2: Build context (cached — only rebuilt when inputs change)
+    const matrixPrompt = options.agent ? getMatrixSystemPrompt(options.agent) : null;
     const contextSnippet = options.context ? formatContextForPrompt(options.context) : '';
     const memorySnippet = getMemoryPrompt();
-    const efficiencyNote = `[K:BOT Efficiency Note]
-You have local tools (file read/write/edit, grep, git, bash) that execute for FREE on the user's machine.
-Prefer local tools over asking the user to do things manually.
-Be thorough in your FIRST response — aim to solve the task in one shot.
-Only use web_search when the user explicitly asks for current information.`;
-    const systemContext = [contextSnippet, memorySnippet, efficiencyNote].filter(Boolean).join('\n');
+    const learningContext = buildFullLearningContext(message, process.cwd());
+    // Cache key to avoid rebuilding identical system prompts
+    const cacheKey = `${options.agent || ''}:${contextSnippet.length}:${memorySnippet.length}`;
+    let systemContext;
+    if (_cachedSystemPromptKey === cacheKey && _cachedSystemPrompt) {
+        // Reuse cached base prompt, only update learning context (changes per message)
+        systemContext = learningContext ? `${learningContext}\n\n${_cachedSystemPrompt}` : _cachedSystemPrompt;
+    }
+    else {
+        const preContext = `You are K:BOT, an AI that lives in the user's terminal. Talk naturally — be direct, concise, and conversational. You're like a skilled colleague, not a corporate chatbot.
+
+Conversation style:
+- Be casual and natural. Use short sentences. Don't over-explain.
+- When the user asks a question, answer it directly. Lead with the answer, not the reasoning.
+- When chatting casually, just chat. Not everything needs tools or code.
+- If the user says "hey" or "what's up", respond naturally — don't launch into capabilities.
+- Match the user's energy. Short question → short answer. Detailed question → detailed response.
+- Never say "I'd be happy to help with that" or "Certainly!" or "Great question!" — just do it.
+- Don't repeat back what the user said. Don't restate the problem before solving it.
+
+How you work with tools:
+- When asked to create, fix, or build something — do it directly with your tools. Don't describe what you would do, just do it.
+- If something fails, read the error and try a different approach. Don't give up.
+- Make reasonable decisions autonomously — pick good defaults.
+- After completing a task, briefly say what you did and where the output is.
+- You run commands yourself. You never tell the user to run something.
+
+Always quote file paths that contain spaces. Never reference internal system names.`;
+        const matrixPrefix = matrixPrompt ? `[Agent Persona]\n${matrixPrompt}\n\nIMPORTANT: Stay in character as defined above. Your responses should reflect this agent's expertise and perspective.\n` : '';
+        _cachedSystemPrompt = [matrixPrefix, contextSnippet, memorySnippet, preContext].filter(Boolean).join('\n');
+        _cachedSystemPromptKey = cacheKey;
+        systemContext = learningContext ? `${learningContext}\n\n${_cachedSystemPrompt}` : _cachedSystemPrompt;
+    }
     let toolCallCount = 0;
     let lastResponse = null;
-    let pendingToolResults = [];
+    const toolSequenceLog = [];
+    const originalMessage = message;
+    let cumulativeCostUsd = 0;
+    // Loop messages track the full conversation within a multi-tool execution.
+    // This includes assistant responses (with tool-use reasoning) and tool results,
+    // so the AI maintains context across tool iterations.
+    const loopMessages = [];
     for (let i = 0; i < MAX_TOOL_LOOPS; i++) {
-        const spinner = createSpinner(i === 0 ? 'Thinking...' : `Running tools (${toolCallCount})...`);
-        spinner.start();
+        // Cost ceiling — stop burning money on runaway loops
+        if (cumulativeCostUsd > MAX_COST_CEILING) {
+            printWarn(`Cost ceiling reached ($${cumulativeCostUsd.toFixed(2)} > $${MAX_COST_CEILING}). Stopping tool loop.`);
+            break;
+        }
+        // Don't use spinner when streaming (conflicts with stdout)
+        const useSpinner = !options.stream;
+        const spinner = useSpinner ? createSpinner(i === 0 ? 'Thinking...' : `Running tools (${toolCallCount})...`) : null;
+        spinner?.start();
         try {
-            if (byokMode) {
-                // ── BYOK: Call provider directly ──
-                const provider = getByokProvider();
-                const speed = options.model === 'haiku' || options.model === 'fast' ? 'fast' : 'default';
-                const model = getProviderModel(provider, speed);
-                const messages = [
-                    ...getPreviousMessages(),
-                    { role: 'user', content: i === 0 ? (systemContext ? `${systemContext}\n\n${message}` : message) : message },
-                ];
-                if (pendingToolResults.length > 0) {
-                    for (const tr of pendingToolResults) {
-                        messages.push({ role: 'user', content: `[Tool Result for ${tr.tool_call_id}]: ${tr.result}` });
-                    }
-                    pendingToolResults = [];
-                }
-                spinner.stop();
-                const result = await callProvider(provider, apiKey, model, systemContext, messages);
+            // ── BYOK: Call provider directly with tool-use support ──
+            const provider = byokProvider || 'anthropic';
+            const speed = options.model === 'haiku' || options.model === 'fast' ? 'fast' : 'default';
+            const model = getProviderModel(provider, speed, originalMessage);
+            const byokTools = tools.map(t => ({
+                name: t.name,
+                description: t.description,
+                input_schema: t.input_schema,
+            }));
+            // Build messages: session history + original user message + loop context
+            const messages = [
+                ...getPreviousMessages(),
+                { role: 'user', content: message },
+                ...loopMessages,
+            ];
+            spinner?.stop();
+            // Use streaming if requested and provider supports it
+            // Disable streaming for local models when tools are active — local models
+            // often emit tool calls as raw JSON text, which can't be cleaned up after streaming
+            const p = getProvider(provider);
+            const canStream = options.stream
+                && p.apiStyle !== 'google'
+                && p.apiStyle !== 'cohere'
+                && !(isLocal && byokTools.length > 0); // Don't stream local + tools (inline tool parsing needs full response)
+            const result = canStream
+                ? await callProviderStreaming(provider, apiKey || 'local', model, systemContext, messages, byokTools, {
+                    thinking: options.thinking,
+                    thinkingBudget: options.thinkingBudget,
+                })
+                : await callProvider(provider, apiKey || 'local', model, systemContext, messages, byokTools, {
+                    multimodal: i === 0 ? parsed : undefined,
+                    thinking: options.thinking,
+                    thinkingBudget: options.thinkingBudget,
+                });
+            const iterationCost = estimateCost(provider, result.usage.input_tokens, result.usage.output_tokens);
+            cumulativeCostUsd += iterationCost;
+            if (result.tool_calls && result.tool_calls.length > 0) {
                 lastResponse = {
-                    type: 'text',
+                    type: 'tool_calls',
                     agent: options.agent || 'kernel',
                     model: result.model,
                     content: result.content,
+                    tool_calls: result.tool_calls.map(tc => ({
+                        id: tc.id,
+                        name: tc.name,
+                        arguments: tc.arguments,
+                    })),
                     usage: {
                         input_tokens: result.usage.input_tokens,
                         output_tokens: result.usage.output_tokens,
-                        cost_usd: estimateCost(provider, result.usage.input_tokens, result.usage.output_tokens),
+                        cost_usd: iterationCost,
                     },
                 };
             }
             else {
-                // ── Kernel API: Route through Matrix ──
-                const apiBase = getApiBase();
-                const body = {
-                    message: i === 0 ? (systemContext ? `${systemContext}\n\n${message}` : message) : message,
-                    mode: 'json',
-                    max_tokens: 8192,
-                    previous_messages: getPreviousMessages(),
-                    tools: tools.length > 0 ? tools : undefined,
-                };
-                if (options.agent && options.agent !== 'auto')
-                    body.agent = options.agent;
-                if (pendingToolResults.length > 0) {
-                    body.tool_results = pendingToolResults;
-                    pendingToolResults = [];
-                }
-                const res = await fetch(`${apiBase}/chat`, {
-                    method: 'POST',
-                    headers: {
-                        'Content-Type': 'application/json',
-                        Authorization: `Bearer ${apiKey}`,
-                        'X-Kbot-Version': KBOT_VERSION,
+                lastResponse = {
+                    type: 'text',
+                    agent: options.agent || 'kernel',
+                    model: result.model,
+                    content: result.content,
+                    streamed: canStream, // flag to skip re-printing
+                    usage: {
+                        input_tokens: result.usage.input_tokens,
+                        output_tokens: result.usage.output_tokens,
+                        cost_usd: iterationCost,
                     },
-                    body: JSON.stringify(body),
-                });
-                spinner.stop();
-                if (!res.ok) {
-                    const err = await res.json().catch(() => ({ error: `HTTP ${res.status}` }));
-                    throw new Error(err.error || `API error: ${res.status}`);
-                }
-                lastResponse = await res.json();
+                };
             }
             // Text response → done
             if (lastResponse.type === 'text' || !lastResponse.tool_calls || lastResponse.tool_calls.length === 0) {
                 const content = lastResponse.content || '';
-                addTurn({ role: 'user', content: message });
+                addTurn({ role: 'user', content: originalMessage });
                 addTurn({ role: 'assistant', content });
+                // ── Recursive Learning: record what worked (async — non-blocking) ──
+                const totalTokens = lastResponse.usage
+                    ? (lastResponse.usage.input_tokens || 0) + (lastResponse.usage.output_tokens || 0)
+                    : 0;
+                // Run all learning asynchronously to avoid blocking the response
+                setImmediate(() => {
+                    try {
+                        // Record pattern if tools were used successfully
+                        if (toolCallCount > 0 && toolSequenceLog.length > 0) {
+                            const patternHint = findPattern(originalMessage);
+                            const savedTokens = patternHint ? patternHint.avgTokensSaved : 0;
+                            recordPattern(originalMessage, toolSequenceLog, savedTokens);
+                        }
+                        // Cache solution for reuse (only meaningful responses)
+                        if (content.length > 50 && toolCallCount <= 5) {
+                            cacheSolution(originalMessage, content.slice(0, 2000));
+                        }
+                        // Update user profile
+                        updateProfile({
+                            tokens: totalTokens,
+                            tokensSaved: findPattern(originalMessage)?.avgTokensSaved || 0,
+                            agent: lastResponse.agent || 'kernel',
+                            taskType: classifyTask(originalMessage),
+                            techTerms: extractKeywords(originalMessage),
+                        });
+                        // Deep learning — extract knowledge, detect corrections, update project memory
+                        learnFromExchange(originalMessage, content, toolSequenceLog, process.cwd());
+                        // Track project context
+                        if (toolSequenceLog.length > 0) {
+                            updateProjectMemory(process.cwd(), {
+                                stack: extractKeywords(originalMessage),
+                            });
+                        }
+                        // Auto self-training trigger
+                        if (shouldAutoTrain()) {
+                            try {
+                                selfTrain();
+                            }
+                            catch { /* silent */ }
+                        }
+                    }
+                    catch { /* learning failures are non-critical */ }
+                });
                 return {
                     content,
                     agent: lastResponse.agent || 'kernel',
                     model: lastResponse.model || 'unknown',
                     toolCalls: toolCallCount,
+                    streamed: lastResponse.streamed || false,
                     usage: lastResponse.usage,
                 };
             }
@@ -280,16 +670,43 @@ Only use web_search when the user explicitly asks for current information.`;
             const results = [];
             for (const call of toolCalls) {
                 toolCallCount++;
+                toolSequenceLog.push(call.name);
                 printToolCall(call.name, call.arguments || {});
+                // Permission check — confirm destructive operations
+                const permitted = await checkPermission(call.name, call.arguments || {});
+                if (!permitted) {
+                    results.push({ tool_call_id: call.id, result: 'Denied by user — operation skipped.', error: true });
+                    printToolResult('Denied by user', true);
+                    continue;
+                }
+                // Pre-tool hook
+                const preHook = runPreToolHook(call.name, call.arguments || {}, options.agent || 'kernel');
+                if (preHook.blocked) {
+                    results.push({ tool_call_id: call.id, result: `Blocked by hook: ${preHook.blockReason}`, error: true });
+                    printToolResult(`Blocked by hook: ${preHook.blockReason}`, true);
+                    continue;
+                }
                 const result = await executeTool(call);
                 results.push(result);
                 printToolResult(result.result, result.error);
+                // Post-tool hook
+                runPostToolHook(call.name, call.arguments || {}, result.result, options.agent || 'kernel');
             }
-            pendingToolResults = results;
-            message = `Here are the tool results:\n${results.map(r => `[${r.tool_call_id}]: ${r.result}`).join('\n\n')}\n\nContinue based on these results.`;
+            // ── Maintain conversation context across tool iterations ──
+            // 1. Include the assistant's response (its reasoning + which tools it chose)
+            const assistantSummary = lastResponse.content
+                ? lastResponse.content
+                : `Using tools: ${toolCalls.map(tc => tc.name).join(', ')}`;
+            loopMessages.push({ role: 'assistant', content: assistantSummary });
+            // 2. Include tool results so the AI sees what happened
+            const toolResultSummary = results.map(r => {
+                const status = r.error ? '[ERROR] ' : '';
+                return `${r.tool_call_id} (${toolCalls.find(tc => tc.id === r.tool_call_id)?.name || 'unknown'}): ${status}${r.result}`;
+            }).join('\n\n');
+            loopMessages.push({ role: 'user', content: `Tool results:\n${toolResultSummary}` });
         }
         catch (err) {
-            spinner.stop();
+            spinner?.stop();
             throw err;
         }
     }
@@ -306,14 +723,54 @@ Only use web_search when the user explicitly asks for current information.`;
 export async function runAndPrint(message, options = {}) {
     try {
         const response = await runAgent(message, options);
-        printResponse(response.agent, response.content);
+        // Skip re-printing if content was already streamed to stdout
+        if (!response.streamed) {
+            printResponse(response.agent, response.content);
+        }
+        // Usage footer — subtle, one line (stderr — status, not content)
         if (response.usage) {
-            const { input_tokens, output_tokens, cost_usd } = response.usage;
-            process.stdout.write(`  \x1b[2m${response.agent} · ${response.model} · ${input_tokens + output_tokens} tokens · $${cost_usd.toFixed(4)}\x1b[0m\n`);
+            const tokens = response.usage.input_tokens + response.usage.output_tokens;
+            const cost = response.usage.cost_usd === 0 ? 'free' : `$${response.usage.cost_usd.toFixed(4)}`;
+            process.stderr.write(`\n  \x1b[2m${tokens} tokens · ${cost}\x1b[0m\n`);
         }
     }
     catch (err) {
-        printError(err instanceof Error ? err.message : String(err));
+        const errMsg = err instanceof Error ? err.message : String(err);
+        // Ollama connection errors — friendly, actionable
+        if (errMsg.includes('fetch failed') || errMsg.includes('ECONNREFUSED')) {
+            const config = await import('./auth.js').then(m => m.loadConfig());
+            if (config?.byok_provider === 'ollama') {
+                printError('Ollama isn\'t running.');
+                printInfo('Open the Ollama app or run: ollama serve');
+                return;
+            }
+            if (config?.byok_provider === 'openclaw') {
+                printError('OpenClaw gateway isn\'t running.');
+                printInfo('Start it: openclaw-cmd start');
+                return;
+            }
+            printError('Can\'t reach the AI provider.');
+            printInfo('Check your internet connection and try again.');
+            return;
+        }
+        // Model not found
+        if (errMsg.includes('model') && (errMsg.includes('not found') || errMsg.includes('does not exist'))) {
+            printError('That model isn\'t installed yet.');
+            printInfo('Download it: ollama pull <model-name>');
+            return;
+        }
+        // Rate limiting
+        if (errMsg.includes('rate') || errMsg.includes('429') || errMsg.includes('too many')) {
+            printError('Too many requests. Wait a moment and try again.');
+            return;
+        }
+        // Auth errors
+        if (errMsg.includes('401') || errMsg.includes('403') || errMsg.includes('invalid') && errMsg.includes('key')) {
+            printError('API key issue. Your key may be expired or invalid.');
+            printInfo('Update it: kbot auth');
+            return;
+        }
+        printError(errMsg);
         process.exit(1);
     }
 }