groove-dev 0.27.137 → 0.27.139

package/CLAUDE.md

@@ -1,7 +1,30 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
 # GROOVE — Agent Orchestration Layer
 
 > Spawn fast. Stay aware. Never lose context.
 
+## Build & Development Commands
+
+```bash
+# Build GUI (required after any GUI changes — daemon serves built assets)
+npm run build                    # or: cd packages/gui && npx vite build
+
+# Run all tests (265 tests, 24 suites)
+npm test                         # or: node --test packages/daemon/test/*.test.js
+
+# Run a single test file
+node --test packages/daemon/test/registry.test.js
+
+# Dev servers (not for agents — agents use npm run build)
+npm run dev:gui                  # Vite dev server (HMR)
+npm run dev:daemon               # Daemon in dev mode
+```
+
+**Important:** GUI changes require `npm run build` to take effect (daemon serves static build from `packages/gui/dist/`). Daemon changes require a manual daemon restart — agents must NEVER restart the daemon themselves.
+
 ## What This Is
 
 GROOVE is a lightweight, open-source agent orchestration layer for AI coding tools. It is a **process manager** — it spawns, coordinates, and monitors AI coding agents. It does NOT wrap, proxy, or impersonate any AI API.
@@ -92,7 +115,7 @@ Bundled starter teams: `fullstack.json`, `api-builder.json`, `monorepo.json`
 - Tailwind CSS v4 + Radix UI (GUI styling + accessible primitives)
 - Zustand 5 (GUI state + WebSocket sync)
 - React Flow / @xyflow/react (agent tree visualization)
-- CodeMirror 6 (code editor, 7 languages)
+- CodeMirror 6 + @uiw/codemirror-themes-all (code editor, 7 languages, 38 selectable themes)
 - xterm.js (terminal emulation)
 - Framer Motion (animations)
 - Lucide React (icons)
@@ -201,7 +224,7 @@ React app served by daemon at `http://localhost:31415`. VS Code-style layout. Ta
 - Log files created with 0o600 permissions
 - Command injection prevention (execFileSync with array args in tmux)
 - Scope patterns validated (no absolute paths, no traversal)
-- 137 automated tests across 14 suites
+- 265 automated tests across 24 suites
 
 ## Conventions
 
@@ -211,6 +234,15 @@ React app served by daemon at `http://localhost:31415`. VS Code-style layout. Ta
 - `.groove/` directory in project root for runtime state (gitignored)
 - Generated markdown files: `AGENTS_REGISTRY.md`, `GROOVE_PROJECT_MAP.md`, `GROOVE_DECISIONS.md`
 
+## GUI Styling Rules
+
+- **Tailwind CSS v4 only** — zero inline styles (no `style={{}}` except dynamic values like `width`, `height`)
+- **No absolute-positioned floating buttons** over content — use flex layout rails instead
+- **File tree icons** must be neutral/muted (`text-text-2`/`text-text-3`) — no rainbow colors
+- **Editor themes** are user-selectable via status bar picker — do not hardcode syntax highlight colors
+- **Font sizes:** file trees use `text-xs` (12px), editor uses 12px via theme
+- **No inline styles for colors** — use design token CSS variables (`--color-accent`, `--color-danger`, etc.)
+
 ## Compliance (CRITICAL)
 
 GROOVE is a process manager, NOT a harness. Hard rules:

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

@@ -1,6 +1,6 @@
 {
   "name": "@groove-dev/cli",
-  "version": "0.27.137",
+  "version": "0.27.139",
   "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": "0.27.137",
+  "version": "0.27.139",
   "description": "GROOVE daemon — agent orchestration engine",
   "license": "FSL-1.1-Apache-2.0",
   "type": "module",

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

@@ -98,15 +98,21 @@ export class AdaptiveThresholds {
     return { score, threshold: newThreshold, converged: profile.converged };
   }
 
-  scoreSession(signals) {
+  scoreSession(signals, options = {}) {
     // Score 0-100 based on quality signals.
     // Low score → rotate sooner (threshold decreases)
     // High score → allow more context (threshold increases)
+    //
+    // Self-managing providers (Claude Code) get reduced penalties because
+    // their normal debugging patterns — polling, retrying commands, hitting
+    // errors during investigation — look identical to degradation signals
+    // but are expected productive behavior.
+    const selfManages = options.selfManagesContext ?? false;
     let score = 70; // Baseline: decent session
 
-    // Error rate: each error costs 5 points
+    // Error rate: -5 per error normally, -2 for self-managing (errors during debugging are expected)
     const errorCount = signals.errorCount || 0;
-    score -= errorCount * 5;
+    score -= errorCount * (selfManages ? 2 : 5);
 
     // Repetitions: 3+ Write/Edit to the same file in a sliding window
     const repetitions = signals.repetitions || 0;
@@ -121,7 +127,7 @@ export class AdaptiveThresholds {
     const toolFailures = signals.toolFailures || 0;
     if (toolCalls > 0) {
       const successRate = (toolCalls - toolFailures) / toolCalls;
-      score += Math.round((successRate - 0.8) * 20);
+      score += Math.round((successRate - 0.8) * (selfManages ? 10 : 20));
     }
 
     // File churn: same file written 5+ times — genuine circular refactoring
@@ -130,26 +136,30 @@ export class AdaptiveThresholds {
 
     // Error trend: increasing errors in second half = degradation
     const errorTrend = signals.errorTrend || 0;
-    if (errorTrend > 0) score -= errorTrend * 6;  // getting worse
-    if (errorTrend < 0) score += 3;                // getting better (small bonus)
+    if (errorTrend > 0) score -= errorTrend * (selfManages ? 3 : 6);
+    if (errorTrend < 0) score += 3;
 
     // Files written: productivity signal
     const filesWritten = signals.filesWritten || 0;
     score += Math.min(filesWritten * 2, 10); // Cap at +10
 
-    // Output length decay: assistant responses shrinking dramatically
-    if (signals.outputLengthDecay) score -= 10;
+    // The following signals are skipped for self-managing providers because
+    // they're false positives during debugging:
+    // - outputLengthDecay: shorter responses = narrowing in on a problem
+    // - toolOutputVolume: reading logs/large files is expected
+    // - turnLatencyTrend: thinking longer about hard problems is fine
+    // - bashRepetition: polling (sleep + check) is a normal pattern
+    if (!selfManages) {
+      if (signals.outputLengthDecay) score -= 10;
 
-    // Tool output volume: bloated context from large tool results
-    const toolVol = signals.toolOutputVolume || 0;
-    if (toolVol === 2) score -= 10;
-    else if (toolVol === 1) score -= 5;
+      const toolVol = signals.toolOutputVolume || 0;
+      if (toolVol === 2) score -= 10;
+      else if (toolVol === 1) score -= 5;
 
-    // Turn latency trend: agent slowing down significantly
-    if (signals.turnLatencyTrend) score -= 5;
+      if (signals.turnLatencyTrend) score -= 5;
 
-    // Bash repetition: agent stuck running identical commands
-    if (signals.bashRepetition) score -= 8;
+      if (signals.bashRepetition) score -= 8;
+    }
 
     // Clamp to 0-100
     return Math.max(0, Math.min(100, score));

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

@@ -1693,18 +1693,29 @@ export function createApi(app, daemon) {
   // Rotation = full handoff brief (only for degradation or no session)
   app.post('/api/agents/:id/instruct', async (req, res) => {
     try {
-      const { message } = req.body;
+      const { message, codeContext } = req.body;
       if (!message || typeof message !== 'string' || !message.trim()) {
         return res.status(400).json({ error: 'message is required' });
       }
       const agent = daemon.registry.get(req.params.id);
       if (!agent) return res.status(404).json({ error: 'Agent not found' });
 
+      // Build the final instruction, optionally enriched with code context
+      let finalMessage = message.trim();
+      if (codeContext && typeof codeContext === 'object') {
+        const { filePath, lineStart, lineEnd, selectedCode } = codeContext;
+        if (filePath && typeof filePath === 'string' && selectedCode && typeof selectedCode === 'string') {
+          const start = Number.isFinite(lineStart) ? lineStart : '?';
+          const end = Number.isFinite(lineEnd) ? lineEnd : '?';
+          finalMessage = `${finalMessage}\n\nCode context from ${filePath} (lines ${start}-${end}):\n\`\`\`\n${selectedCode}\n\`\`\``;
+        }
+      }
+
       // Record user feedback so the journalist can include it in future agent context
-      if (daemon.journalist) daemon.journalist.recordUserFeedback(agent, message.trim());
+      if (daemon.journalist) daemon.journalist.recordUserFeedback(agent, finalMessage);
 
       // Agent loop path — send message directly to the running loop
-      const wrappedMessage = wrapWithRoleReminder(agent.role, message.trim());
+      const wrappedMessage = wrapWithRoleReminder(agent.role, finalMessage);
       if (daemon.processes.hasAgentLoop(req.params.id)) {
         const sent = await daemon.processes.sendMessage(req.params.id, wrappedMessage);
         if (sent) {
@@ -1731,7 +1742,7 @@ export function createApi(app, daemon) {
           scope: oldConfig.scope,
           provider: oldConfig.provider,
           model: oldConfig.model,
-          prompt: message.trim(),
+          prompt: finalMessage,
           permission: oldConfig.permission || 'full',
           workingDir: oldConfig.workingDir,
           name: oldConfig.name,
@@ -1754,7 +1765,7 @@ export function createApi(app, daemon) {
           scope: oldConfig.scope,
           provider: oldConfig.provider,
           model: oldConfig.model,
-          prompt: message.trim(),
+          prompt: finalMessage,
           introContext: oldConfig.introContext,
           permission: oldConfig.permission || 'full',
           workingDir: oldConfig.workingDir,
@@ -3340,11 +3351,118 @@ Keep responses concise. Help them think, don't lecture them about the system the
     });
   });
 
+  // Git line status — per-line modification status for editor gutter decorations
+  app.get('/api/files/git-line-status', (req, res) => {
+    const relPath = req.query.path;
+    if (!relPath || typeof relPath !== 'string') {
+      return res.status(400).json({ error: 'path parameter is required' });
+    }
+    if (relPath.includes('\0') || relPath.startsWith('/')) {
+      return res.status(400).json({ error: 'Invalid path' });
+    }
+    const segments = relPath.split(/[/\\]/);
+    if (segments.some(s => s === '..')) {
+      return res.status(400).json({ error: 'Path traversal not allowed' });
+    }
+
+    const rootDir = getEditorRoot();
+    if (!rootDir) return res.status(400).json({ error: 'Editor root not set' });
+
+    const fullPath = resolve(rootDir, relPath);
+    if (!fullPath.startsWith(rootDir + sep) && fullPath !== rootDir) {
+      return res.status(400).json({ error: 'Path outside project' });
+    }
+
+    const result = { lines: { added: [], modified: [], deleted: [] } };
+
+    // Check if file is tracked by git
+    try {
+      execFileSync('git', ['ls-files', '--error-unmatch', '--', relPath], { cwd: rootDir, timeout: 5000, stdio: 'pipe' });
+    } catch {
+      // File not tracked — check if it exists (untracked = all lines added)
+      if (existsSync(fullPath)) {
+        try {
+          const content = readFileSync(fullPath, 'utf8');
+          const lineCount = content.split('\n').length;
+          for (let i = 1; i <= lineCount; i++) result.lines.added.push(i);
+        } catch { /* binary or unreadable */ }
+      }
+      return res.json(result);
+    }
+
+    try {
+      const diffOut = execFileSync('git', ['diff', '--unified=0', '--', relPath], {
+        cwd: rootDir, timeout: 10000, maxBuffer: 5 * 1024 * 1024,
+      }).toString();
+
+      if (!diffOut.trim()) return res.json(result);
+
+      // Check for binary
+      if (diffOut.includes('Binary files')) return res.json(result);
+
+      // Parse unified diff hunks: @@ -oldStart,oldCount +newStart,newCount @@
+      const hunkRe = /^@@ -(\d+)(?:,(\d+))? \+(\d+)(?:,(\d+))? @@/gm;
+      let match;
+      while ((match = hunkRe.exec(diffOut)) !== null) {
+        const oldCount = parseInt(match[2] ?? '1', 10);
+        const newStart = parseInt(match[3], 10);
+        const newCount = parseInt(match[4] ?? '1', 10);
+
+        if (oldCount === 0 && newCount > 0) {
+          // Pure addition
+          for (let i = newStart; i < newStart + newCount; i++) result.lines.added.push(i);
+        } else if (newCount === 0 && oldCount > 0) {
+          // Pure deletion — mark the line where content was removed
+          result.lines.deleted.push(newStart);
+        } else {
+          // Modification
+          for (let i = newStart; i < newStart + newCount; i++) result.lines.modified.push(i);
+        }
+      }
+
+      res.json(result);
+    } catch (err) {
+      if (err.status !== undefined) return res.json(result);
+      res.status(500).json({ error: 'Failed to compute line status' });
+    }
+  });
+
+  // Git branches — list all local branches with current branch marked
+  app.get('/api/files/git-branches', (req, res) => {
+    const rootDir = getEditorRoot();
+    if (!rootDir) return res.status(400).json({ error: 'Editor root not set' });
+
+    const fallback = { current: null, branches: [] };
+
+    try {
+      let current = null;
+      try {
+        current = execFileSync('git', ['rev-parse', '--abbrev-ref', 'HEAD'], {
+          cwd: rootDir, timeout: 5000, stdio: 'pipe',
+        }).toString().trim();
+      } catch { return res.json(fallback); }
+
+      const branchOut = execFileSync('git', ['branch', '--list', '--format=%(refname:short)'], {
+        cwd: rootDir, timeout: 5000, stdio: 'pipe',
+      }).toString();
+
+      const branches = branchOut.split('\n').map(b => b.trim()).filter(Boolean);
+      res.json({ current, branches });
+    } catch {
+      res.json(fallback);
+    }
+  });
+
   // Files touched by an agent during its session
   app.get('/api/agents/:id/files-touched', (req, res) => {
     const agent = daemon.registry.get(req.params.id);
     if (!agent) return res.status(404).json({ error: 'Agent not found' });
-    const files = daemon.registry.getFilesTouched(req.params.id);
+    const rawFiles = daemon.registry.getFilesTouched(req.params.id);
+    const rootDir = agent.workingDir || daemon.projectDir;
+    const files = rawFiles.map(f => {
+      const fullPath = isAbsolute(f.path) ? f.path : resolve(rootDir, f.path);
+      return { ...f, exists: existsSync(fullPath) };
+    });
     res.json({ files, total: files.length });
   });
 

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

@@ -173,12 +173,17 @@ export class Introducer {
         }
       }
 
-      // Project files section — tell the new agent what exists and what to read
+      // Project files section — tell the new agent what exists
+      // When no task is assigned, list files as reference only (not an action prompt)
       if (allTeamFiles.length > 0) {
         lines.push('');
         lines.push(`## Project Files`);
         lines.push('');
-        lines.push(`Your team has created the following files. **Read relevant ones before starting work** to understand what's been built and planned:`);
+        if (hasTask || isRotation) {
+          lines.push(`Your team has created the following files. **Read relevant ones before starting work** to understand what's been built and planned:`);
+        } else {
+          lines.push(`Your team has created the following files (for reference — do NOT read or act on these until you receive a task):`);
+        }
         lines.push('');
 
         // Group by agent for clarity

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

@@ -443,24 +443,47 @@ export class Journalist {
       `Scope: ${agent.scope?.join(', ') || 'unrestricted'}${dir}`,
       `Rotation reason: ${reason}`,
       '',
-      'Analyze the session log below and produce a structured handoff brief.',
+      'Analyze the FULL session log below and produce a structured handoff brief.',
+      'The next agent must be able to continue this work without repeating anything.',
       '',
       'Output EXACTLY these sections, in this priority order:',
       '',
+      '## Investigation Path',
+      '(Chronological narrative of what was explored and why. For each major step:',
+      ' what was tried, what the result was, and whether it was ruled out or remains viable.',
+      ' The next agent must know the full path taken to avoid retreading ground.)',
+      '',
+      '## Current State',
+      '(Where things stand right now. What is the current hypothesis or approach?',
+      ' What was the agent actively working on when rotated? What is the next step?)',
+      '',
+      '## What Was Ruled Out',
+      '(Approaches, files, or theories that were investigated and eliminated.',
+      ' Include WHY each was ruled out so the next agent does not retry them.)',
+      '',
+      '## Key Findings',
+      '(Important discoveries: root causes identified, relevant code locations,',
+      ' configuration details, error patterns, dependencies found. Name files,',
+      ' functions, and line numbers.)',
+      '',
       '## Unresolved Errors',
-      '(What failed, what was tried, and what to avoid repeating.)',
+      '(Errors still occurring. Include exact error messages and context.)',
+      '',
+      '## User Messages',
+      '(EVERY message the user typed, verbatim or near-verbatim. These are critical —',
+      ' the user must never have to repeat themselves after a rotation. Include all',
+      ' instructions, questions, corrections, and context the user provided.)',
       '',
       '## User Constraints',
       '(Explicit user directives, must/never/avoid/use instructions.)',
       '',
-      '## Last 5 Tool Calls',
+      '## Last 15 Tool Calls',
       '(Compact list of the most recent meaningful tool calls and outcomes.)',
       '',
-      '## Accomplishments',
-      '(What was completed. Name files, functions, and line numbers.)',
-      '',
       'Be specific. Name files, functions, and line numbers. Do not summarize vaguely.',
-      'Keep your response under 1500 characters.',
+      'Preserve the investigation narrative — the next agent needs to understand the',
+      'journey, not just the destination.',
+      'Keep your response under 4000 characters.',
       '',
       '---',
       '',
@@ -469,8 +492,8 @@ export class Journalist {
     ];
 
     let totalChars = 0;
-    const cap = 15_000;
-    for (const entry of entries.slice(-200)) {
+    const cap = 40_000;
+    for (const entry of entries.slice(-500)) {
       const line = this.formatEntry(entry);
       if (totalChars + line.length > cap) break;
       parts.push(line);
@@ -861,7 +884,14 @@ export class Journalist {
       ? `- Quality profile: ${specialization.avgQualityScore}/100 across ${specialization.sessionCount} sessions`
       : '';
 
-    const recentChain = this.daemon.memory?.getRecentHandoffMarkdown(agent.role, 1, 1500, agent.workingDir, agent.teamId) || '';
+    const recentChain = this.daemon.memory?.getRecentHandoffMarkdown(agent.role, 3, 3000, agent.workingDir, agent.teamId) || '';
+
+    // User messages from the session log — these are the actual things the user typed.
+    // They must always be preserved: if a user said it, the next agent needs to see it.
+    const userMessages = entries
+      .filter((e) => e.type === 'user' && e.text)
+      .map((e) => `- "${e.text.slice(0, 500)}"`)
+      .join('\n');
 
     const agentFeedback = this.getUserFeedback(agent.id).slice(-5);
     const conversationSummary = agentFeedback.length > 0
@@ -870,7 +900,7 @@ export class Journalist {
 
     const recentTools = entries
       .filter((e) => e.type === 'tool' || e.type === 'error')
-      .slice(-5)
+      .slice(-15)
       .map((e) => `- ${e.type === 'error' ? 'ERROR ' : ''}${e.tool}: ${(e.input || e.text || '').slice(0, 200)}`)
       .join('\n');
 
@@ -907,24 +937,32 @@ export class Journalist {
 
       const fallbackRecentTools = entries
         .filter((e) => e.type === 'tool' || e.type === 'error')
-        .slice(-5)
+        .slice(-15)
         .map((e) => `- ${e.type === 'error' ? 'ERROR ' : ''}${e.tool}: ${(e.input || '').slice(0, 200)}`)
         .join('\n');
 
+      // Build investigation timeline from thinking entries — these capture reasoning and decisions
+      const thinkingEntries = entries
+        .filter((e) => e.type === 'thinking' && e.text && e.text.length > 80)
+        .slice(-10)
+        .map((e) => `- ${e.text.slice(0, 400)}`)
+        .join('\n');
+
       const fallbackParts = [];
+      if (thinkingEntries) fallbackParts.push(`## Investigation Path\n\n${thinkingEntries}`);
       if (errorSummary) fallbackParts.push(`## Unresolved Errors\n\n${errorSummary}`);
-      if (fallbackRecentTools) fallbackParts.push(`## Last 5 Tool Calls\n\n${fallbackRecentTools}`);
+      if (fallbackRecentTools) fallbackParts.push(`## Last 15 Tool Calls\n\n${fallbackRecentTools}`);
       if (fileChangesSummary) fallbackParts.push(`## Files Modified\n\n${fileChangesSummary}`);
       if (resultSummary) fallbackParts.push(`## Accomplishments\n\n${resultSummary}`);
       sessionSummary = fallbackParts.join('\n\n');
     }
 
-    // For quality_degradation rotations, drop user messages (already in session summary)
-    const includeUserMessages = options.reason !== 'quality_degradation';
-
     // 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 ? '…' : '') : '';
 
+    // Combine user messages from session log and feedback API — never drop these
+    const allUserMessages = [userMessages, conversationSummary].filter(Boolean).join('\n');
+
     let brief = [
       `# Handoff Brief — ${agent.name} (${agent.role})`,
       ``,
@@ -933,14 +971,14 @@ export class Journalist {
       `Rotation: ${options.reason || 'manual'}${options.qualityScore ? ` (quality: ${options.qualityScore}/100)` : ''} | Tokens: ${agent.tokensUsed}`,
       specLine,
       ``,
-      // Priority order: session summary (contains unresolved errors) first,
-      // then constraints, then discoveries, then tools — so the most critical
-      // debugging context survives even if the brief hits the hard cap.
+      // Priority order: session summary first, then user messages (never drop —
+      // if the user typed it, the next agent must see it), then constraints,
+      // then discoveries, then tools. Most critical context survives truncation.
       sessionSummary ? `## Session Summary\n\n${sessionSummary}\n` : '',
+      allUserMessages ? `## User Messages (do not ask user to repeat these)\n\n${allUserMessages}\n` : '',
       constraints ? `## Project Constraints (must follow)\n\n${constraints}\n` : '',
       discoveries ? `## Known Issues & Fixes\n\n${discoveries}\n` : '',
-      recentTools ? `## Last 5 Tool Calls\n\n${recentTools}\n` : '',
-      includeUserMessages && conversationSummary ? `## Recent User Messages\n\n${conversationSummary}\n` : '',
+      recentTools ? `## Last 15 Tool Calls\n\n${recentTools}\n` : '',
       recentChain ? `## Rotation History\n\n${recentChain}\n` : '',
       originalTask ? `## Original Task\n\n${originalTask}\n` : '',
       ``,
@@ -948,9 +986,9 @@ export class Journalist {
       `Continue seamlessly — finish the work and deliver the output.`,
     ].filter(Boolean).join('\n');
 
-    // 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]';
+    // Hard cap: 16000 chars — investigation context needs room to preserve the full narrative
+    if (brief.length > 16000) {
+      brief = brief.slice(0, 15950) + '\n\n[Brief truncated — see session logs for full context]';
     }
 
     return brief;

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

@@ -4,8 +4,6 @@
 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';
 
@@ -249,11 +247,14 @@ export class ModelLab {
       throw new Error(errMsg);
     }
 
-    const nodeStream = Readable.fromWeb(resp.body);
+    const reader = resp.body.getReader();
+    const decoder = new TextDecoder();
     let buffer = '';
 
-    for await (const chunk of nodeStream) {
-      buffer += typeof chunk === 'string' ? chunk : chunk.toString('utf8');
+    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() || '';
 

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

@@ -887,23 +887,26 @@ export class ProcessManager {
       }
     }
 
+    // Compute hasTask from actual prompt content — agents spawned without a
+    // prompt should NOT receive handoff history (prevents cross-team contamination).
+    // Discoveries + constraints are always injected (project knowledge).
+    // Handoffs are injected only when the agent has a real task or is a rotation.
+    const hasTask = !!(config.prompt && config.prompt.trim().length > 0);
+    const isRotation = !!(config.isRotation);
+
     // Pre-spawn task negotiation — if same-role agents are running,
-    // query them about current work so the new agent gets a clear assignment
+    // query them about current work so the new agent gets a clear assignment.
+    // Only negotiate when the agent has a task — otherwise the negotiator
+    // output acts as an implicit task and the agent starts working immediately.
     const sameRole = registry.getAll().filter(
       (a) => a.role === config.role && a.id !== agent.id &&
         (a.status === 'running' || a.status === 'starting')
     );
     let taskNegotiation = '';
-    if (sameRole.length > 0) {
+    if (sameRole.length > 0 && (hasTask || isRotation)) {
       taskNegotiation = await this.negotiateTaskSplit(agent, sameRole);
     }
 
-    // Compute hasTask from actual prompt content — agents spawned without a
-    // prompt should NOT receive handoff history (prevents cross-team contamination).
-    // Discoveries + constraints are always injected (project knowledge).
-    // Handoffs are injected only when the agent has a real task or is a rotation.
-    const hasTask = !!(config.prompt && config.prompt.trim().length > 0);
-    const isRotation = !!(config.isRotation);
     let introContext = introducer.generateContext(agent, { taskNegotiation, hasTask, isRotation });
 
     // Intro context size warning and optional truncation (Change 7)

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

@@ -145,15 +145,12 @@ export class Rotator extends EventEmitter {
     }
 
     const signals = signalsEarly;
-    let score = this.daemon.adaptive.scoreSession(signals);
+    const providerForScore = getProvider(agent.provider);
+    const selfManages = providerForScore?.constructor?.managesOwnContext ?? false;
+    let score = this.daemon.adaptive.scoreSession(signals, { selfManagesContext: selfManages });
 
     // 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 (!selfManages) {
       if (ageSec > 1800) score -= 5;
       if (ageSec > 3600) score -= 10;
       if (ageSec > 7200) score -= 15;
@@ -385,6 +382,7 @@ export class Rotator extends EventEmitter {
         contextUsage: agent.contextUsage,
         reason: options.reason || 'manual',
         qualityScore: options.qualityScore || null,
+        qualitySignals: options.signals || null,
         timestamp: new Date().toISOString(),
       };