groove-dev 0.27.150 → 0.27.152

package/node_modules/@groove-dev/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/cli",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "GROOVE CLI — manage AI coding agents from your terminal",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

package/node_modules/@groove-dev/daemon/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/daemon",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "GROOVE daemon — agent orchestration engine",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

package/node_modules/@groove-dev/daemon/src/agent-loop.js

@@ -10,6 +10,11 @@ import { existsSync, readFileSync, writeFileSync, mkdirSync, unlinkSync } from '
 import { resolve, dirname } from 'path';
 import { TOOL_DEFINITIONS, ToolExecutor } from './tool-executor.js';
 
+function stripThinkTags(text) {
+  if (!text) return text;
+  return text.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+}
+
 export class AgentLoop extends EventEmitter {
   constructor({ daemon, agent, loopConfig, logStream }) {
     super();
@@ -24,6 +29,11 @@ export class AgentLoop extends EventEmitter {
     this.idle = true;
     this.abortController = null;
 
+    // Tool calling mode: 'native' uses OpenAI function-calling API fields,
+    // 'prompt' injects tool schemas into the system prompt and parses
+    // <tool_call> blocks from the model's text output.
+    this.toolMode = 'native';
+
     // Metrics
     this.totalTokensIn = 0;
     this.totalTokensOut = 0;
@@ -132,7 +142,7 @@ export class AgentLoop extends EventEmitter {
       const response = await this._callApi();
       if (!response || !this.running) break;
 
-      const { content, toolCalls, usage, finishReason } = response;
+      let { content, toolCalls, usage, finishReason } = response;
       consecutiveErrors = 0; // Reset on successful call
 
       // Update token tracking from API response
@@ -140,10 +150,21 @@ export class AgentLoop extends EventEmitter {
         this._updateTokens(usage);
       }
 
-      // Append assistant message to conversation history
+      // Strip thinking tags from display content (keep raw in conversation history)
+      const displayContent = stripThinkTags(content);
+
+      // In prompt-based mode, parse tool calls from the model's text
+      if (this.toolMode === 'prompt' && content) {
+        const parsed = this._parseToolCallsFromText(content);
+        if (parsed.length > 0) {
+          toolCalls = parsed;
+        }
+      }
+
+      // Append assistant message to conversation history (raw content with thinking preserved)
       const assistantMsg = { role: 'assistant' };
       if (content) assistantMsg.content = content;
-      if (toolCalls?.length > 0) {
+      if (this.toolMode === 'native' && toolCalls?.length > 0) {
         assistantMsg.tool_calls = toolCalls.map((tc) => ({
           id: tc.id,
           type: 'function',
@@ -154,17 +175,21 @@ export class AgentLoop extends EventEmitter {
 
       // No tool calls → turn complete, broadcast final text and go idle
       if (!toolCalls || toolCalls.length === 0) {
-        if (content) {
-          this._writeLog({ type: 'assistant', content: content.slice(0, 2000) });
+        if (displayContent) {
+          this._writeLog({ type: 'assistant', content: displayContent.slice(0, 2000) });
         }
-        this.emit('output', { type: 'result', subtype: 'assistant', data: content || 'Turn complete', turns: this.turns });
+        this.emit('output', { type: 'result', subtype: 'assistant', data: displayContent || 'Turn complete', turns: this.turns });
         break;
       }
 
       // Has tool calls — broadcast text before executing tools (if model sent text + tools)
-      if (content) {
-        this._writeLog({ type: 'assistant', content: content.slice(0, 2000) });
-        this.emit('output', { type: 'activity', subtype: 'assistant', data: content });
+      let preToolText = displayContent;
+      if (this.toolMode === 'prompt') {
+        preToolText = stripThinkTags((content || '').replace(/<tool_call>[\s\S]*?<\/tool_call>/g, ''));
+      }
+      if (preToolText) {
+        this._writeLog({ type: 'assistant', content: preToolText.slice(0, 2000) });
+        this.emit('output', { type: 'activity', subtype: 'assistant', data: preToolText });
       }
 
       // Execute each tool call
@@ -172,13 +197,29 @@ export class AgentLoop extends EventEmitter {
         if (!this.running) break;
 
         let args;
+        let parseError = null;
         try {
           args = JSON.parse(call.function.arguments);
-        } catch {
+        } catch (e) {
+          parseError = e.message;
           args = {};
         }
 
         const toolName = call.function.name;
+
+        // Report malformed JSON back to the model instead of silently failing
+        if (parseError) {
+          const errMsg = `Invalid JSON in tool arguments: ${parseError}. Raw: ${call.function.arguments.slice(0, 200)}`;
+          this._writeLog({ type: 'tool_result', tool: toolName, success: false, output: errMsg });
+          this.emit('output', { type: 'activity', subtype: 'tool_result', data: [{ type: 'tool_result', name: toolName, success: false, output: errMsg }] });
+          if (this.toolMode === 'native') {
+            this.messages.push({ role: 'tool', tool_call_id: call.id, content: `Error: ${errMsg}` });
+          } else {
+            this.messages.push({ role: 'user', content: `<tool_result name="${toolName}">\nError: ${errMsg}\n</tool_result>` });
+          }
+          continue;
+        }
+
         const inputSummary = this._summarizeToolInput(toolName, args);
 
         // Log + broadcast tool invocation
@@ -210,12 +251,24 @@ export class AgentLoop extends EventEmitter {
           this.daemon.classifier.addEvent(this.agent.id, { type: 'error', text: result.error });
         }
 
-        // Append tool result to conversation for the model
-        this.messages.push({
-          role: 'tool',
-          tool_call_id: call.id,
-          content: result.success ? (result.result || 'Done.') : `Error: ${result.error}`,
-        });
+        // Append tool result to conversation — cap size to protect context window
+        const MAX_RESULT_CHARS = 30000;
+        let resultContent = result.success ? (result.result || 'Done.') : `Error: ${result.error}`;
+        if (resultContent.length > MAX_RESULT_CHARS) {
+          resultContent = resultContent.slice(0, MAX_RESULT_CHARS) + '\n... (result truncated — use offset/limit for large files, or pipe commands through head/tail)';
+        }
+        if (this.toolMode === 'native') {
+          this.messages.push({
+            role: 'tool',
+            tool_call_id: call.id,
+            content: resultContent,
+          });
+        } else {
+          this.messages.push({
+            role: 'user',
+            content: `<tool_result name="${toolName}">\n${resultContent}\n</tool_result>`,
+          });
+        }
       }
 
       // Context rotation is handled by the Rotator's 15s polling loop
@@ -236,12 +289,15 @@ export class AgentLoop extends EventEmitter {
     const body = {
       model: this.config.model,
       messages: this.messages,
-      tools: TOOL_DEFINITIONS,
-      tool_choice: 'auto',
       temperature: this.config.temperature ?? 0.1,
       max_tokens: this.config.maxResponseTokens || 4096,
     };
 
+    if (this.toolMode === 'native') {
+      body.tools = TOOL_DEFINITIONS;
+      body.tool_choice = 'auto';
+    }
+
     if (this.config.stream !== false) {
       body.stream = true;
       body.stream_options = { include_usage: true };
@@ -283,6 +339,18 @@ export class AgentLoop extends EventEmitter {
         const text = await response.text().catch(() => '');
         const errMsg = `API error ${response.status}: ${text.slice(0, 500)}`;
 
+        // Detect tool_choice rejection (vLLM, TGI, etc. without tool-calling flags)
+        // Fall back to prompt-based tool calling and retry immediately
+        if (response.status === 400 && this.toolMode === 'native' &&
+            (text.includes('tool_choice') || text.includes('tool-call-parser') || text.includes('enable-auto-tool-choice'))) {
+          this._writeLog({ type: 'system', event: 'tool-fallback', reason: 'Runtime rejected native tool calling — switching to prompt-based tools' });
+          this.toolMode = 'prompt';
+          this._injectToolPrompt();
+          delete body.tools;
+          delete body.tool_choice;
+          continue;
+        }
+
         if (response.status === 401 || response.status === 403) {
           this._writeLog({ type: 'error', text: errMsg });
           this.emit('error', { message: errMsg });
@@ -320,6 +388,10 @@ export class AgentLoop extends EventEmitter {
     let finishReason = null;
     let buffer = '';
 
+    // State machine for suppressing <think> blocks during streaming
+    let insideThink = false;
+    let streamBuf = '';
+
     const reader = response.body.getReader();
     const decoder = new TextDecoder();
 
@@ -349,10 +421,56 @@ export class AgentLoop extends EventEmitter {
           if (choice.finish_reason) finishReason = choice.finish_reason;
           const delta = choice.delta || {};
 
-          // Stream text tokens to GUI in real-time
+          // reasoning_content: separate thinking field (vLLM, some OpenAI-compat servers)
+          // Capture for logging but don't stream to GUI
+          if (delta.reasoning_content || delta.reasoning) {
+            // Accumulate in content so it's in the conversation history
+            // but don't stream it to the GUI
+            content += delta.reasoning_content || delta.reasoning;
+          }
+
+          // Stream text tokens to GUI in real-time, suppressing <think> blocks
           if (delta.content) {
             content += delta.content;
-            this.emit('output', { type: 'activity', subtype: 'stream', data: delta.content });
+            streamBuf += delta.content;
+
+            // Process buffer — emit non-think content, suppress think content
+            let safety = 0;
+            while (streamBuf.length > 0 && safety++ < 100) {
+              if (insideThink) {
+                const closeIdx = streamBuf.indexOf('</think>');
+                if (closeIdx >= 0) {
+                  insideThink = false;
+                  streamBuf = streamBuf.slice(closeIdx + 8);
+                } else {
+                  break; // wait for more data
+                }
+              } else {
+                const openIdx = streamBuf.indexOf('<think>');
+                if (openIdx >= 0) {
+                  const before = streamBuf.slice(0, openIdx);
+                  if (before) {
+                    this.emit('output', { type: 'activity', subtype: 'stream', data: before });
+                  }
+                  insideThink = true;
+                  streamBuf = streamBuf.slice(openIdx + 7);
+                } else {
+                  // Hold back bytes that could be the start of a <think> tag
+                  let safeEnd = streamBuf.length;
+                  for (let i = Math.min(6, streamBuf.length); i >= 1; i--) {
+                    if ('<think>'.startsWith(streamBuf.slice(-i))) {
+                      safeEnd = streamBuf.length - i;
+                      break;
+                    }
+                  }
+                  if (safeEnd > 0) {
+                    this.emit('output', { type: 'activity', subtype: 'stream', data: streamBuf.slice(0, safeEnd) });
+                  }
+                  streamBuf = streamBuf.slice(safeEnd);
+                  break;
+                }
+              }
+            }
           }
 
           // Accumulate tool call deltas
@@ -380,6 +498,11 @@ export class AgentLoop extends EventEmitter {
       return null;
     }
 
+    // Flush remaining stream buffer (e.g. unclosed <think> — treat as display content)
+    if (streamBuf) {
+      this.emit('output', { type: 'activity', subtype: 'stream', data: streamBuf });
+    }
+
     return {
       content: content || null,
       toolCalls: toolCalls.size > 0 ? Array.from(toolCalls.values()) : null,
@@ -405,6 +528,65 @@ export class AgentLoop extends EventEmitter {
     };
   }
 
+  // --- Prompt-Based Tool Calling Fallback ---
+
+  _injectToolPrompt() {
+    const toolPrompt = this._buildToolPrompt();
+    const systemIdx = this.messages.findIndex(m => m.role === 'system');
+    if (systemIdx >= 0) {
+      this.messages[systemIdx].content += '\n\n' + toolPrompt;
+    } else {
+      this.messages.unshift({ role: 'system', content: toolPrompt });
+    }
+  }
+
+  _buildToolPrompt() {
+    const toolDefs = TOOL_DEFINITIONS.map(t => {
+      const f = t.function;
+      const params = Object.entries(f.parameters.properties).map(([name, schema]) => {
+        const req = f.parameters.required?.includes(name) ? ' (required)' : ' (optional)';
+        return `  - ${name}: ${schema.type}${req} — ${schema.description}`;
+      }).join('\n');
+      return `### ${f.name}\n${f.description}\nParameters:\n${params}`;
+    }).join('\n\n');
+
+    return `## Available Tools
+
+To use a tool, include a tool_call block in your response:
+
+<tool_call>
+{"name": "tool_name", "arguments": {"param1": "value1"}}
+</tool_call>
+
+You can make multiple tool calls in one response. After each tool call you will receive a <tool_result> with the output.
+
+${toolDefs}
+
+Always use tools to read, write, or search files and to run commands. Do not guess file contents.`;
+  }
+
+  _parseToolCallsFromText(content) {
+    if (!content) return [];
+    const calls = [];
+    const regex = /<tool_call>\s*([\s\S]*?)\s*<\/tool_call>/g;
+    let match;
+    while ((match = regex.exec(content)) !== null) {
+      try {
+        const parsed = JSON.parse(match[1].trim());
+        if (parsed.name) {
+          calls.push({
+            id: `call_${Date.now()}_${calls.length}`,
+            function: {
+              name: parsed.name,
+              arguments: JSON.stringify(parsed.arguments || {}),
+            },
+          });
+        }
+      } catch { /* skip malformed tool call */ }
+    }
+    return calls;
+  }
+
   // --- Token Tracking ---
 
   _updateTokens(usage) {

package/node_modules/@groove-dev/daemon/src/providers/local.js

@@ -137,6 +137,7 @@ export class LocalProvider extends Provider {
     let model = agent.model || 'qwen2.5-coder:7b';
     let apiBase = 'http://localhost:11434/v1';
     let apiKey = agent.apiKey || null;
+    let runtimeType = 'ollama';
 
     if (agent.apiBase) {
       apiBase = agent.apiBase;
@@ -153,6 +154,7 @@ export class LocalProvider extends Provider {
       if (rt) {
         apiBase = rt.endpoint.includes('/v1') ? rt.endpoint : `${rt.endpoint}/v1`;
         if (rt.apiKey) apiKey = rt.apiKey;
+        if (rt.type) runtimeType = rt.type;
         const rtModel = rt.models?.[0];
         model = rtModel?.id || rtModel?.name || ggufId;
       }
@@ -167,6 +169,7 @@ export class LocalProvider extends Provider {
       if (rt) {
         apiBase = rt.endpoint.includes('/v1') ? rt.endpoint : `${rt.endpoint}/v1`;
         if (rt.apiKey) apiKey = rt.apiKey;
+        if (rt.type) runtimeType = rt.type;
         model = modelId;
       }
     }
@@ -176,6 +179,7 @@ export class LocalProvider extends Provider {
     return {
       apiBase,
       model,
+      runtimeType,
       contextWindow,
       temperature: typeof agent.temperature === 'number' ? agent.temperature : 0.1,
       maxResponseTokens: 4096,

package/node_modules/@groove-dev/daemon/src/tool-executor.js

@@ -1,7 +1,7 @@
 // GROOVE — Tool Executor for Local Agent Loop
 // FSL-1.1-Apache-2.0 — see LICENSE
 
-import { readFileSync, writeFileSync, readdirSync, statSync, mkdirSync, existsSync } from 'fs';
+import { readFileSync, writeFileSync, readdirSync, statSync, mkdirSync, existsSync, openSync, readSync, closeSync } from 'fs';
 import { execSync } from 'child_process';
 import { resolve, relative, dirname, sep } from 'path';
 import { minimatch } from 'minimatch';
@@ -190,11 +190,24 @@ export class ToolExecutor {
     if (stat.isDirectory()) {
       return { success: false, error: `Path is a directory, not a file: ${filePath}` };
     }
-    // Guard against huge files
     if (stat.size > 5 * 1024 * 1024) {
       return { success: false, error: `File too large (${formatBytes(stat.size)}). Use offset/limit to read a section.` };
     }
 
+    // Detect binary files — check first 8KB for null bytes
+    if (stat.size > 0) {
+      const probe = Buffer.alloc(Math.min(8192, stat.size));
+      const fd = openSync(resolved, 'r');
+      try {
+        readSync(fd, probe, 0, probe.length, 0);
+      } finally {
+        closeSync(fd);
+      }
+      if (probe.includes(0)) {
+        return { success: false, error: `Binary file (${formatBytes(stat.size)}). Cannot read non-text files.` };
+      }
+    }
+
     const content = readFileSync(resolved, 'utf8');
     let lines = content.split('\n');
     const totalLines = lines.length;
@@ -207,8 +220,19 @@ export class ToolExecutor {
       lines = lines.slice(0, limit);
     }
 
+    // Auto-limit large files when no explicit limit was provided
+    const MAX_LINES = 2000;
+    let autoTruncated = false;
+    if (!limit && lines.length > MAX_LINES) {
+      lines = lines.slice(0, MAX_LINES);
+      autoTruncated = true;
+    }
+
     const numbered = lines.map((line, i) => `${startLine + i}\t${line}`).join('\n');
-    return { success: true, result: numbered, meta: { totalLines } };
+    const result = autoTruncated
+      ? numbered + `\n\n... (showing ${MAX_LINES} of ${totalLines} lines — use offset/limit to read more)`
+      : numbered;
+    return { success: true, result, meta: { totalLines } };
   }
 
   writeFile({ path: filePath, content }) {

package/node_modules/@groove-dev/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/gui",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "GROOVE GUI — visual agent control plane",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "groove-dev",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "Open-source agent orchestration layer — the AI company OS. Local model agent engine (GGUF/Ollama/llama-server), HuggingFace model browser, MCP integrations (Slack, Gmail, Stripe, 15+), agent scheduling (cron), business roles (CMO, CFO, EA). GUI dashboard, multi-agent coordination, zero cold-start, infinite sessions. Works with Claude Code, Codex, Gemini CLI, Ollama, any local model.",
   "license": "FSL-1.1-Apache-2.0",
   "author": "Groove Dev <hello@groovedev.ai> (https://groovedev.ai)",

package/packages/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/cli",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "GROOVE CLI — manage AI coding agents from your terminal",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

package/packages/daemon/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/daemon",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "GROOVE daemon — agent orchestration engine",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

package/packages/daemon/src/agent-loop.js

@@ -10,6 +10,11 @@ import { existsSync, readFileSync, writeFileSync, mkdirSync, unlinkSync } from '
 import { resolve, dirname } from 'path';
 import { TOOL_DEFINITIONS, ToolExecutor } from './tool-executor.js';
 
+function stripThinkTags(text) {
+  if (!text) return text;
+  return text.replace(/<think>[\s\S]*?<\/think>/g, '').trim();
+}
+
 export class AgentLoop extends EventEmitter {
   constructor({ daemon, agent, loopConfig, logStream }) {
     super();
@@ -24,6 +29,11 @@ export class AgentLoop extends EventEmitter {
     this.idle = true;
     this.abortController = null;
 
+    // Tool calling mode: 'native' uses OpenAI function-calling API fields,
+    // 'prompt' injects tool schemas into the system prompt and parses
+    // <tool_call> blocks from the model's text output.
+    this.toolMode = 'native';
+
     // Metrics
     this.totalTokensIn = 0;
     this.totalTokensOut = 0;
@@ -132,7 +142,7 @@ export class AgentLoop extends EventEmitter {
       const response = await this._callApi();
       if (!response || !this.running) break;
 
-      const { content, toolCalls, usage, finishReason } = response;
+      let { content, toolCalls, usage, finishReason } = response;
       consecutiveErrors = 0; // Reset on successful call
 
       // Update token tracking from API response
@@ -140,10 +150,21 @@ export class AgentLoop extends EventEmitter {
         this._updateTokens(usage);
       }
 
-      // Append assistant message to conversation history
+      // Strip thinking tags from display content (keep raw in conversation history)
+      const displayContent = stripThinkTags(content);
+
+      // In prompt-based mode, parse tool calls from the model's text
+      if (this.toolMode === 'prompt' && content) {
+        const parsed = this._parseToolCallsFromText(content);
+        if (parsed.length > 0) {
+          toolCalls = parsed;
+        }
+      }
+
+      // Append assistant message to conversation history (raw content with thinking preserved)
       const assistantMsg = { role: 'assistant' };
       if (content) assistantMsg.content = content;
-      if (toolCalls?.length > 0) {
+      if (this.toolMode === 'native' && toolCalls?.length > 0) {
         assistantMsg.tool_calls = toolCalls.map((tc) => ({
           id: tc.id,
           type: 'function',
@@ -154,17 +175,21 @@ export class AgentLoop extends EventEmitter {
 
       // No tool calls → turn complete, broadcast final text and go idle
       if (!toolCalls || toolCalls.length === 0) {
-        if (content) {
-          this._writeLog({ type: 'assistant', content: content.slice(0, 2000) });
+        if (displayContent) {
+          this._writeLog({ type: 'assistant', content: displayContent.slice(0, 2000) });
         }
-        this.emit('output', { type: 'result', subtype: 'assistant', data: content || 'Turn complete', turns: this.turns });
+        this.emit('output', { type: 'result', subtype: 'assistant', data: displayContent || 'Turn complete', turns: this.turns });
         break;
       }
 
       // Has tool calls — broadcast text before executing tools (if model sent text + tools)
-      if (content) {
-        this._writeLog({ type: 'assistant', content: content.slice(0, 2000) });
-        this.emit('output', { type: 'activity', subtype: 'assistant', data: content });
+      let preToolText = displayContent;
+      if (this.toolMode === 'prompt') {
+        preToolText = stripThinkTags((content || '').replace(/<tool_call>[\s\S]*?<\/tool_call>/g, ''));
+      }
+      if (preToolText) {
+        this._writeLog({ type: 'assistant', content: preToolText.slice(0, 2000) });
+        this.emit('output', { type: 'activity', subtype: 'assistant', data: preToolText });
       }
 
       // Execute each tool call
@@ -172,13 +197,29 @@ export class AgentLoop extends EventEmitter {
         if (!this.running) break;
 
         let args;
+        let parseError = null;
         try {
           args = JSON.parse(call.function.arguments);
-        } catch {
+        } catch (e) {
+          parseError = e.message;
           args = {};
         }
 
         const toolName = call.function.name;
+
+        // Report malformed JSON back to the model instead of silently failing
+        if (parseError) {
+          const errMsg = `Invalid JSON in tool arguments: ${parseError}. Raw: ${call.function.arguments.slice(0, 200)}`;
+          this._writeLog({ type: 'tool_result', tool: toolName, success: false, output: errMsg });
+          this.emit('output', { type: 'activity', subtype: 'tool_result', data: [{ type: 'tool_result', name: toolName, success: false, output: errMsg }] });
+          if (this.toolMode === 'native') {
+            this.messages.push({ role: 'tool', tool_call_id: call.id, content: `Error: ${errMsg}` });
+          } else {
+            this.messages.push({ role: 'user', content: `<tool_result name="${toolName}">\nError: ${errMsg}\n</tool_result>` });
+          }
+          continue;
+        }
+
         const inputSummary = this._summarizeToolInput(toolName, args);
 
         // Log + broadcast tool invocation
@@ -210,12 +251,24 @@ export class AgentLoop extends EventEmitter {
           this.daemon.classifier.addEvent(this.agent.id, { type: 'error', text: result.error });
         }
 
-        // Append tool result to conversation for the model
-        this.messages.push({
-          role: 'tool',
-          tool_call_id: call.id,
-          content: result.success ? (result.result || 'Done.') : `Error: ${result.error}`,
-        });
+        // Append tool result to conversation — cap size to protect context window
+        const MAX_RESULT_CHARS = 30000;
+        let resultContent = result.success ? (result.result || 'Done.') : `Error: ${result.error}`;
+        if (resultContent.length > MAX_RESULT_CHARS) {
+          resultContent = resultContent.slice(0, MAX_RESULT_CHARS) + '\n... (result truncated — use offset/limit for large files, or pipe commands through head/tail)';
+        }
+        if (this.toolMode === 'native') {
+          this.messages.push({
+            role: 'tool',
+            tool_call_id: call.id,
+            content: resultContent,
+          });
+        } else {
+          this.messages.push({
+            role: 'user',
+            content: `<tool_result name="${toolName}">\n${resultContent}\n</tool_result>`,
+          });
+        }
       }
 
       // Context rotation is handled by the Rotator's 15s polling loop
@@ -236,12 +289,15 @@ export class AgentLoop extends EventEmitter {
     const body = {
       model: this.config.model,
       messages: this.messages,
-      tools: TOOL_DEFINITIONS,
-      tool_choice: 'auto',
       temperature: this.config.temperature ?? 0.1,
       max_tokens: this.config.maxResponseTokens || 4096,
     };
 
+    if (this.toolMode === 'native') {
+      body.tools = TOOL_DEFINITIONS;
+      body.tool_choice = 'auto';
+    }
+
     if (this.config.stream !== false) {
       body.stream = true;
       body.stream_options = { include_usage: true };
@@ -283,6 +339,18 @@ export class AgentLoop extends EventEmitter {
         const text = await response.text().catch(() => '');
         const errMsg = `API error ${response.status}: ${text.slice(0, 500)}`;
 
+        // Detect tool_choice rejection (vLLM, TGI, etc. without tool-calling flags)
+        // Fall back to prompt-based tool calling and retry immediately
+        if (response.status === 400 && this.toolMode === 'native' &&
+            (text.includes('tool_choice') || text.includes('tool-call-parser') || text.includes('enable-auto-tool-choice'))) {
+          this._writeLog({ type: 'system', event: 'tool-fallback', reason: 'Runtime rejected native tool calling — switching to prompt-based tools' });
+          this.toolMode = 'prompt';
+          this._injectToolPrompt();
+          delete body.tools;
+          delete body.tool_choice;
+          continue;
+        }
+
         if (response.status === 401 || response.status === 403) {
           this._writeLog({ type: 'error', text: errMsg });
           this.emit('error', { message: errMsg });
@@ -320,6 +388,10 @@ export class AgentLoop extends EventEmitter {
     let finishReason = null;
     let buffer = '';
 
+    // State machine for suppressing <think> blocks during streaming
+    let insideThink = false;
+    let streamBuf = '';
+
     const reader = response.body.getReader();
     const decoder = new TextDecoder();
 
@@ -349,10 +421,56 @@ export class AgentLoop extends EventEmitter {
           if (choice.finish_reason) finishReason = choice.finish_reason;
           const delta = choice.delta || {};
 
-          // Stream text tokens to GUI in real-time
+          // reasoning_content: separate thinking field (vLLM, some OpenAI-compat servers)
+          // Capture for logging but don't stream to GUI
+          if (delta.reasoning_content || delta.reasoning) {
+            // Accumulate in content so it's in the conversation history
+            // but don't stream it to the GUI
+            content += delta.reasoning_content || delta.reasoning;
+          }
+
+          // Stream text tokens to GUI in real-time, suppressing <think> blocks
           if (delta.content) {
             content += delta.content;
-            this.emit('output', { type: 'activity', subtype: 'stream', data: delta.content });
+            streamBuf += delta.content;
+
+            // Process buffer — emit non-think content, suppress think content
+            let safety = 0;
+            while (streamBuf.length > 0 && safety++ < 100) {
+              if (insideThink) {
+                const closeIdx = streamBuf.indexOf('</think>');
+                if (closeIdx >= 0) {
+                  insideThink = false;
+                  streamBuf = streamBuf.slice(closeIdx + 8);
+                } else {
+                  break; // wait for more data
+                }
+              } else {
+                const openIdx = streamBuf.indexOf('<think>');
+                if (openIdx >= 0) {
+                  const before = streamBuf.slice(0, openIdx);
+                  if (before) {
+                    this.emit('output', { type: 'activity', subtype: 'stream', data: before });
+                  }
+                  insideThink = true;
+                  streamBuf = streamBuf.slice(openIdx + 7);
+                } else {
+                  // Hold back bytes that could be the start of a <think> tag
+                  let safeEnd = streamBuf.length;
+                  for (let i = Math.min(6, streamBuf.length); i >= 1; i--) {
+                    if ('<think>'.startsWith(streamBuf.slice(-i))) {
+                      safeEnd = streamBuf.length - i;
+                      break;
+                    }
+                  }
+                  if (safeEnd > 0) {
+                    this.emit('output', { type: 'activity', subtype: 'stream', data: streamBuf.slice(0, safeEnd) });
+                  }
+                  streamBuf = streamBuf.slice(safeEnd);
+                  break;
+                }
+              }
+            }
           }
 
           // Accumulate tool call deltas
@@ -380,6 +498,11 @@ export class AgentLoop extends EventEmitter {
       return null;
     }
 
+    // Flush remaining stream buffer (e.g. unclosed <think> — treat as display content)
+    if (streamBuf) {
+      this.emit('output', { type: 'activity', subtype: 'stream', data: streamBuf });
+    }
+
     return {
       content: content || null,
       toolCalls: toolCalls.size > 0 ? Array.from(toolCalls.values()) : null,
@@ -405,6 +528,65 @@ export class AgentLoop extends EventEmitter {
     };
   }
 
+  // --- Prompt-Based Tool Calling Fallback ---
+
+  _injectToolPrompt() {
+    const toolPrompt = this._buildToolPrompt();
+    const systemIdx = this.messages.findIndex(m => m.role === 'system');
+    if (systemIdx >= 0) {
+      this.messages[systemIdx].content += '\n\n' + toolPrompt;
+    } else {
+      this.messages.unshift({ role: 'system', content: toolPrompt });
+    }
+  }
+
+  _buildToolPrompt() {
+    const toolDefs = TOOL_DEFINITIONS.map(t => {
+      const f = t.function;
+      const params = Object.entries(f.parameters.properties).map(([name, schema]) => {
+        const req = f.parameters.required?.includes(name) ? ' (required)' : ' (optional)';
+        return `  - ${name}: ${schema.type}${req} — ${schema.description}`;
+      }).join('\n');
+      return `### ${f.name}\n${f.description}\nParameters:\n${params}`;
+    }).join('\n\n');
+
+    return `## Available Tools
+
+To use a tool, include a tool_call block in your response:
+
+<tool_call>
+{"name": "tool_name", "arguments": {"param1": "value1"}}
+</tool_call>
+
+You can make multiple tool calls in one response. After each tool call you will receive a <tool_result> with the output.
+
+${toolDefs}
+
+Always use tools to read, write, or search files and to run commands. Do not guess file contents.`;
+  }
+
+  _parseToolCallsFromText(content) {
+    if (!content) return [];
+    const calls = [];
+    const regex = /<tool_call>\s*([\s\S]*?)\s*<\/tool_call>/g;
+    let match;
+    while ((match = regex.exec(content)) !== null) {
+      try {
+        const parsed = JSON.parse(match[1].trim());
+        if (parsed.name) {
+          calls.push({
+            id: `call_${Date.now()}_${calls.length}`,
+            function: {
+              name: parsed.name,
+              arguments: JSON.stringify(parsed.arguments || {}),
+            },
+          });
+        }
+      } catch { /* skip malformed tool call */ }
+    }
+    return calls;
+  }
+
   // --- Token Tracking ---
 
   _updateTokens(usage) {

package/packages/daemon/src/providers/local.js

@@ -137,6 +137,7 @@ export class LocalProvider extends Provider {
     let model = agent.model || 'qwen2.5-coder:7b';
     let apiBase = 'http://localhost:11434/v1';
     let apiKey = agent.apiKey || null;
+    let runtimeType = 'ollama';
 
     if (agent.apiBase) {
       apiBase = agent.apiBase;
@@ -153,6 +154,7 @@ export class LocalProvider extends Provider {
       if (rt) {
         apiBase = rt.endpoint.includes('/v1') ? rt.endpoint : `${rt.endpoint}/v1`;
         if (rt.apiKey) apiKey = rt.apiKey;
+        if (rt.type) runtimeType = rt.type;
         const rtModel = rt.models?.[0];
         model = rtModel?.id || rtModel?.name || ggufId;
       }
@@ -167,6 +169,7 @@ export class LocalProvider extends Provider {
       if (rt) {
         apiBase = rt.endpoint.includes('/v1') ? rt.endpoint : `${rt.endpoint}/v1`;
         if (rt.apiKey) apiKey = rt.apiKey;
+        if (rt.type) runtimeType = rt.type;
         model = modelId;
       }
     }
@@ -176,6 +179,7 @@ export class LocalProvider extends Provider {
     return {
       apiBase,
       model,
+      runtimeType,
       contextWindow,
       temperature: typeof agent.temperature === 'number' ? agent.temperature : 0.1,
       maxResponseTokens: 4096,

package/packages/daemon/src/tool-executor.js

@@ -1,7 +1,7 @@
 // GROOVE — Tool Executor for Local Agent Loop
 // FSL-1.1-Apache-2.0 — see LICENSE
 
-import { readFileSync, writeFileSync, readdirSync, statSync, mkdirSync, existsSync } from 'fs';
+import { readFileSync, writeFileSync, readdirSync, statSync, mkdirSync, existsSync, openSync, readSync, closeSync } from 'fs';
 import { execSync } from 'child_process';
 import { resolve, relative, dirname, sep } from 'path';
 import { minimatch } from 'minimatch';
@@ -190,11 +190,24 @@ export class ToolExecutor {
     if (stat.isDirectory()) {
       return { success: false, error: `Path is a directory, not a file: ${filePath}` };
     }
-    // Guard against huge files
     if (stat.size > 5 * 1024 * 1024) {
       return { success: false, error: `File too large (${formatBytes(stat.size)}). Use offset/limit to read a section.` };
     }
 
+    // Detect binary files — check first 8KB for null bytes
+    if (stat.size > 0) {
+      const probe = Buffer.alloc(Math.min(8192, stat.size));
+      const fd = openSync(resolved, 'r');
+      try {
+        readSync(fd, probe, 0, probe.length, 0);
+      } finally {
+        closeSync(fd);
+      }
+      if (probe.includes(0)) {
+        return { success: false, error: `Binary file (${formatBytes(stat.size)}). Cannot read non-text files.` };
+      }
+    }
+
     const content = readFileSync(resolved, 'utf8');
     let lines = content.split('\n');
     const totalLines = lines.length;
@@ -207,8 +220,19 @@ export class ToolExecutor {
       lines = lines.slice(0, limit);
     }
 
+    // Auto-limit large files when no explicit limit was provided
+    const MAX_LINES = 2000;
+    let autoTruncated = false;
+    if (!limit && lines.length > MAX_LINES) {
+      lines = lines.slice(0, MAX_LINES);
+      autoTruncated = true;
+    }
+
     const numbered = lines.map((line, i) => `${startLine + i}\t${line}`).join('\n');
-    return { success: true, result: numbered, meta: { totalLines } };
+    const result = autoTruncated
+      ? numbered + `\n\n... (showing ${MAX_LINES} of ${totalLines} lines — use offset/limit to read more)`
+      : numbered;
+    return { success: true, result, meta: { totalLines } };
   }
 
   writeFile({ path: filePath, content }) {

package/packages/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/gui",
-  "version": "0.27.150",
+  "version": "0.27.152",
   "description": "GROOVE GUI — visual agent control plane",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",