groove-dev 0.27.135 → 0.27.137

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

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/cli",
-  "version": "0.27.135",
+  "version": "0.27.137",
   "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": "0.27.135",
+  "version": "0.27.137",
   "description": "GROOVE daemon — agent orchestration engine",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

package/node_modules/@groove-dev/daemon/src/api.js

@@ -124,6 +124,38 @@ export function createApi(app, daemon) {
     res.json({ status: 'ok', uptime: process.uptime() });
   });
 
+  // Debug: test fetch to llama-server from daemon runtime
+  app.get('/api/lab/debug-fetch', async (req, res) => {
+    const target = req.query.url || 'http://localhost:8081/v1/chat/completions';
+    const log = [];
+    try {
+      log.push(`fetch → ${target}`);
+      log.push(`node ${process.version}, electron ${process.versions.electron || 'N/A'}`);
+      const start = Date.now();
+      const r = await fetch(target, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ model: 'Qwen3-0.6B-Q8_0.gguf', messages: [{ role: 'user', content: 'Say ok' }], stream: true, max_tokens: 10 }),
+        signal: AbortSignal.timeout(10000),
+      });
+      log.push(`status=${r.status} in ${Date.now() - start}ms`);
+      const reader = r.body.getReader();
+      let chunks = 0;
+      while (chunks < 5) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        chunks++;
+        log.push(`chunk ${chunks}: ${new TextDecoder().decode(value).slice(0, 120)}`);
+      }
+      reader.cancel();
+      log.push(`total chunks read: ${chunks}`);
+      res.json({ ok: true, log });
+    } catch (err) {
+      log.push(`ERROR: ${err.message}`);
+      res.json({ ok: false, log, error: err.message });
+    }
+  });
+
   // List all agents
   app.get('/api/agents', (req, res) => {
     res.json(daemon.registry.getAll());

package/node_modules/@groove-dev/daemon/src/journalist.js

@@ -853,9 +853,8 @@ export class Journalist {
     const agentLog = filteredLogs[agent.id];
     const entries = agentLog?.entries || [];
 
-    // Layer 7 memory: discoveries (pointer only), constraints, specializations
-    const hasDiscoveries = this.daemon.memory?.getDiscoveriesMarkdown(agent.role, 1, 100);
-    const discoveryPointer = hasDiscoveries ? 'See .groove/memory/agent-discoveries.jsonl for error→fix pairs.' : '';
+    // Layer 7 memory: discoveries (inline, not pointer — agents lose context with pointers), constraints, specializations
+    const discoveries = this.daemon.memory?.getDiscoveriesMarkdown(agent.role, 10, 1500) || '';
     const constraints = this.daemon.memory?.getConstraintsMarkdown(2000) || '';
     const specialization = this.daemon.memory?.getSpecialization(agent.id);
     const specLine = specialization?.avgQualityScore != null
@@ -872,7 +871,7 @@ export class Journalist {
     const recentTools = entries
       .filter((e) => e.type === 'tool' || e.type === 'error')
       .slice(-5)
-      .map((e) => `- ${e.type === 'error' ? 'ERROR ' : ''}${e.tool}: ${(e.input || e.text || '').slice(0, 80)}`)
+      .map((e) => `- ${e.type === 'error' ? 'ERROR ' : ''}${e.tool}: ${(e.input || e.text || '').slice(0, 200)}`)
       .join('\n');
 
     // Try AI-synthesized session summary
@@ -909,7 +908,7 @@ export class Journalist {
       const fallbackRecentTools = entries
         .filter((e) => e.type === 'tool' || e.type === 'error')
         .slice(-5)
-        .map((e) => `- ${e.type === 'error' ? 'ERROR ' : ''}${e.tool}: ${(e.input || '').slice(0, 80)}`)
+        .map((e) => `- ${e.type === 'error' ? 'ERROR ' : ''}${e.tool}: ${(e.input || '').slice(0, 200)}`)
         .join('\n');
 
       const fallbackParts = [];
@@ -923,8 +922,8 @@ export class Journalist {
     // For quality_degradation rotations, drop user messages (already in session summary)
     const includeUserMessages = options.reason !== 'quality_degradation';
 
-    // Cap Original Task to 500 chars
-    const originalTask = agent.prompt ? agent.prompt.slice(0, 500) + (agent.prompt.length > 500 ? '…' : '') : '';
+    // Cap Original Task to 1000 chars — task descriptions for debugging can be long
+    const originalTask = agent.prompt ? agent.prompt.slice(0, 1000) + (agent.prompt.length > 1000 ? '…' : '') : '';
 
     let brief = [
       `# Handoff Brief — ${agent.name} (${agent.role})`,
@@ -934,10 +933,13 @@ export class Journalist {
       `Rotation: ${options.reason || 'manual'}${options.qualityScore ? ` (quality: ${options.qualityScore}/100)` : ''} | Tokens: ${agent.tokensUsed}`,
       specLine,
       ``,
-      discoveryPointer ? `## Known Issues & Fixes\n\n${discoveryPointer}\n` : '',
+      // Priority order: session summary (contains unresolved errors) first,
+      // then constraints, then discoveries, then tools — so the most critical
+      // debugging context survives even if the brief hits the hard cap.
+      sessionSummary ? `## Session Summary\n\n${sessionSummary}\n` : '',
       constraints ? `## Project Constraints (must follow)\n\n${constraints}\n` : '',
+      discoveries ? `## Known Issues & Fixes\n\n${discoveries}\n` : '',
       recentTools ? `## Last 5 Tool Calls\n\n${recentTools}\n` : '',
-      sessionSummary ? `## Session Summary\n\n${sessionSummary}\n` : '',
       includeUserMessages && conversationSummary ? `## Recent User Messages\n\n${conversationSummary}\n` : '',
       recentChain ? `## Rotation History\n\n${recentChain}\n` : '',
       originalTask ? `## Original Task\n\n${originalTask}\n` : '',
@@ -946,9 +948,9 @@ export class Journalist {
       `Continue seamlessly — finish the work and deliver the output.`,
     ].filter(Boolean).join('\n');
 
-    // Hard cap: total brief must not exceed 5000 chars
-    if (brief.length > 5000) {
-      brief = brief.slice(0, 4950) + '\n\n[Brief truncated — see session logs for full context]';
+    // Hard cap: 8000 chars — enough for debugging context without overwhelming the new agent
+    if (brief.length > 8000) {
+      brief = brief.slice(0, 7950) + '\n\n[Brief truncated — see session logs for full context]';
     }
 
     return brief;

package/node_modules/@groove-dev/daemon/src/model-lab.js

@@ -4,6 +4,7 @@
 import { resolve } from 'path';
 import { existsSync, readFileSync, writeFileSync, mkdirSync, readdirSync, unlinkSync } from 'fs';
 import { randomUUID } from 'crypto';
+import { Readable } from 'stream';
 
 const RUNTIME_TYPES = ['ollama', 'vllm', 'llama-cpp', 'tgi', 'openai-compatible'];
 const DEFAULT_OLLAMA_ENDPOINT = 'http://localhost:11434';
@@ -223,10 +224,9 @@ export class ModelLab {
       ...this._buildParameterBody(parameters || {}),
     };
 
-    const endpoint = `${rt.endpoint}/v1/chat/completions`;
-
-    const headers = { 'Content-Type': 'application/json' };
-    if (rt.apiKey) headers['Authorization'] = `Bearer ${rt.apiKey}`;
+    const endpoint = rt.endpoint.replace('localhost', '127.0.0.1');
+    const reqHeaders = { 'Content-Type': 'application/json' };
+    if (rt.apiKey) reqHeaders['Authorization'] = `Bearer ${rt.apiKey}`;
 
     const requestStart = Date.now();
     let ttft = null;
@@ -236,85 +236,61 @@ export class ModelLab {
     let generationStart = null;
     let fullContent = '';
 
-    const resp = await fetch(endpoint, {
+    const resp = await fetch(`${endpoint}/v1/chat/completions`, {
       method: 'POST',
-      headers,
+      headers: reqHeaders,
       body: JSON.stringify(body),
       signal: AbortSignal.timeout(300000),
     });
 
     if (!resp.ok) {
-      let errorMsg;
-      try { errorMsg = (await resp.json()).error?.message || `HTTP ${resp.status}`; } catch { errorMsg = `HTTP ${resp.status}`; }
-      throw new Error(errorMsg);
+      let errMsg = `HTTP ${resp.status}`;
+      try { const e = await resp.json(); errMsg = e.error?.message || errMsg; } catch { /* ignore */ }
+      throw new Error(errMsg);
     }
 
-    const reader = resp.body.getReader();
-    const decoder = new TextDecoder();
+    const nodeStream = Readable.fromWeb(resp.body);
     let buffer = '';
 
-    try {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-
-        buffer += decoder.decode(value, { stream: true });
-        const lines = buffer.split('\n');
-        buffer = lines.pop() || '';
-
-        for (const line of lines) {
-          const trimmed = line.trim();
-          if (!trimmed || !trimmed.startsWith('data: ')) continue;
-          const payload = trimmed.slice(6);
-          if (payload === '[DONE]') continue;
-
-          try {
-            const chunk = JSON.parse(payload);
-            const delta = chunk.choices?.[0]?.delta;
-            if (delta?.reasoning_content) {
-              if (ttft === null) {
-                ttft = Date.now() - requestStart;
-                generationStart = Date.now();
-              }
-              completionTokens++;
-              onEvent({ type: 'reasoning', content: delta.reasoning_content });
-            }
-            if (delta?.content) {
-              if (ttft === null) {
-                ttft = Date.now() - requestStart;
-                generationStart = Date.now();
-              }
-              fullContent += delta.content;
-              completionTokens++;
-              onEvent({ type: 'token', content: delta.content });
-            }
-            if (chunk.usage) {
-              promptTokens = chunk.usage.prompt_tokens || 0;
-              totalTokens = chunk.usage.total_tokens || 0;
-              if (chunk.usage.completion_tokens) {
-                completionTokens = chunk.usage.completion_tokens;
-              }
-            }
-          } catch { /* skip malformed chunk */ }
-        }
+    for await (const chunk of nodeStream) {
+      buffer += typeof chunk === 'string' ? chunk : chunk.toString('utf8');
+      const lines = buffer.split('\n');
+      buffer = lines.pop() || '';
+
+      for (const line of lines) {
+        const trimmed = line.trim();
+        if (!trimmed || !trimmed.startsWith('data: ')) continue;
+        const data = trimmed.slice(6);
+        if (data === '[DONE]') continue;
+
+        try {
+          const parsed = JSON.parse(data);
+          const delta = parsed.choices?.[0]?.delta;
+          if (delta?.reasoning_content) {
+            if (ttft === null) { ttft = Date.now() - requestStart; generationStart = Date.now(); }
+            completionTokens++;
+            onEvent({ type: 'reasoning', content: delta.reasoning_content });
+          }
+          if (delta?.content) {
+            if (ttft === null) { ttft = Date.now() - requestStart; generationStart = Date.now(); }
+            fullContent += delta.content;
+            completionTokens++;
+            onEvent({ type: 'token', content: delta.content });
+          }
+          if (parsed.usage) {
+            promptTokens = parsed.usage.prompt_tokens || 0;
+            totalTokens = parsed.usage.total_tokens || 0;
+            if (parsed.usage.completion_tokens) completionTokens = parsed.usage.completion_tokens;
+          }
+        } catch { /* skip malformed chunk */ }
       }
-    } finally {
-      reader.releaseLock();
     }
 
     const generationTime = generationStart ? Date.now() - generationStart : Date.now() - requestStart;
     const tokensPerSec = generationTime > 0 ? (completionTokens / (generationTime / 1000)) : 0;
 
-    let memoryUsage = null;
-    if (rt.type === 'ollama') {
-      memoryUsage = await this.getOllamaMemoryUsage(rt.endpoint);
-    }
-
     if (sessionId) {
-      this._appendToSession(sessionId, messages, {
-        role: 'assistant',
-        content: fullContent,
-      });
+      this._appendToSession(sessionId, messages, { role: 'assistant', content: fullContent });
     }
 
     onEvent({
@@ -326,9 +302,16 @@ export class ModelLab {
         promptTokens,
         completionTokens,
         generationTime,
-        memoryUsage,
+        memoryUsage: null,
       },
     });
+
+    if (rt.type === 'ollama') {
+      try {
+        const mem = await this.getOllamaMemoryUsage(rt.endpoint);
+        if (mem) onEvent({ type: 'memory', usage: mem });
+      } catch { /* ignore */ }
+    }
   }
 
   _buildParameterBody(params) {

package/node_modules/@groove-dev/daemon/src/rotator.js

@@ -147,11 +147,19 @@ export class Rotator extends EventEmitter {
     const signals = signalsEarly;
     let score = this.daemon.adaptive.scoreSession(signals);
 
-    if (ageSec > 1800) score -= 5;
-    if (ageSec > 3600) score -= 10;
-    if (ageSec > 7200) score -= 15;
-    if (ageSec > 14400) score -= 20;
-    if (ageSec > 28800) score -= 25;
+    // Age penalties: only for providers that don't manage their own context.
+    // Claude Code handles context internally via compaction — long sessions
+    // are normal and productive. Penalizing age causes premature rotation
+    // that destroys active debugging context and creates restart loops.
+    const providerForAge = getProvider(agent.provider);
+    const selfManagesForAge = providerForAge?.constructor?.managesOwnContext ?? false;
+    if (!selfManagesForAge) {
+      if (ageSec > 1800) score -= 5;
+      if (ageSec > 3600) score -= 10;
+      if (ageSec > 7200) score -= 15;
+      if (ageSec > 14400) score -= 20;
+      if (ageSec > 28800) score -= 25;
+    }
 
     score = Math.max(0, Math.min(100, score));
 
@@ -240,16 +248,23 @@ export class Rotator extends EventEmitter {
         }
       }
 
-      // --- Change 4: Truncation-triggered immediate rotation (all providers) ---
-      if (agent.consecutiveTruncations >= 2) {
+      // --- Change 4: Truncation-triggered rotation ---
+      // Self-managing providers need more consecutive truncations — single
+      // truncations can be transient API issues, not session degradation.
+      const truncationThreshold = selfManagesContext ? 4 : 2;
+      if (agent.consecutiveTruncations >= truncationThreshold) {
         console.log(`  Rotator: ${agent.name} consecutiveTruncations=${agent.consecutiveTruncations} — FORCE rotating (incomplete_response)`);
         await this.rotate(agent.id, { reason: 'incomplete_response', qualityScore: 0 });
         continue;
       }
 
-      // --- Change 3: Compaction-aware rotation for self-managing providers ---
+      // --- Change 3: Compaction-aware rotation ---
+      // Only for non-self-managing providers. Claude Code compacts internally
+      // as part of normal operation — it's healthy, not degradation. Counting
+      // compactions toward a ceiling causes premature rotation that destroys
+      // active debugging sessions and creates restart loops.
       const compactions = this.compactionCounts.get(agent.id) || 0;
-      if (compactions >= 5) {
+      if (!selfManagesContext && compactions >= 5) {
         console.log(`  Rotator: ${agent.name} compactions=${compactions} — FORCE rotating (compaction_ceiling)`);
         await this.rotate(agent.id, { reason: 'compaction_ceiling' });
         continue;
@@ -259,18 +274,30 @@ export class Rotator extends EventEmitter {
       // agents don't persist producing bad output for 8-10 minutes
       if (this._isOnCooldown(agent.id, QUALITY_COOLDOWN_MS)) continue;
 
-      // Effective quality threshold: lower for agents showing degradation signals
-      let effectiveQualityThreshold = QUALITY_THRESHOLD;
-      if (compactions >= 3 || agent.truncationSuspected || agent.cacheResetDetected) {
-        effectiveQualityThreshold = 55;
+      // Effective quality threshold depends on provider type.
+      // Self-managing providers (Claude Code): threshold = 15. Only rotate on
+      // truly catastrophic degradation. Normal debugging naturally produces
+      // errors, retries, and bash repetitions — the scoring model treats these
+      // as degradation but they're expected behavior during investigation.
+      // A threshold of 40 (the default) kills debugging sessions at ~8 minutes.
+      let effectiveQualityThreshold;
+      if (selfManagesContext) {
+        effectiveQualityThreshold = 15;
+      } else {
+        effectiveQualityThreshold = QUALITY_THRESHOLD;
+        if (compactions >= 3) effectiveQualityThreshold = 55;
+        if (agent.truncationSuspected || agent.cacheResetDetected) {
+          effectiveQualityThreshold = Math.max(effectiveQualityThreshold, 55);
+        }
       }
 
-      // All providers: quality-based rotation — detects degradation before tokens are wasted
       const quality = this.scoreLiveSession(agent);
       if (quality.hasEnoughData && quality.score < effectiveQualityThreshold) {
-        // Severe degradation (score < 25): rotate immediately regardless of idle state.
-        // The agent is producing bad output — waiting for idle is counterproductive.
-        if (quality.score < 25) {
+        // For self-managing providers, effectiveQualityThreshold IS the severe
+        // threshold (15) — any score below it is catastrophic, rotate immediately.
+        // For others, severe = < 25, moderate = 25-40/55.
+        const severeThreshold = selfManagesContext ? effectiveQualityThreshold : 25;
+        if (quality.score < severeThreshold) {
           console.log(`  Rotator: ${agent.name} quality=${quality.score} — FORCE rotating (severe degradation)`);
           await this.rotate(agent.id, {
             reason: 'quality_degradation',
@@ -279,8 +306,10 @@ export class Rotator extends EventEmitter {
           });
           continue;
         }
-        // Moderate degradation (25-40): rotate when idle
-        if (this._idleMs(agent) > 10_000) {
+        // Moderate degradation (25-40): only for non-self-managing providers.
+        // Claude Code sessions should never be killed for "moderate" quality —
+        // errors during debugging are expected, not degradation.
+        if (!selfManagesContext && this._idleMs(agent) > 10_000) {
           console.log(`  Rotator: ${agent.name} quality=${quality.score} — rotating (quality)`);
           await this.rotate(agent.id, {
             reason: 'quality_degradation',
@@ -343,7 +372,7 @@ export class Rotator extends EventEmitter {
           reason: options.reason || 'manual',
           oldTokens: agent.tokensUsed,
           contextUsage: agent.contextUsage,
-          brief: brief.slice(0, 4000),
+          brief: brief.slice(0, 6000),
         }, agent.workingDir, agent.teamId);
       }