rlhf-feedback-loop 0.6.5 → 0.6.6

package/README.md

@@ -6,7 +6,19 @@
 [![MCP Ready](https://img.shields.io/badge/MCP-ready-black)](adapters/mcp/server-stdio.js)
 [![DPO Ready](https://img.shields.io/badge/DPO-ready-blue)](scripts/export-dpo-pairs.js)
 
-**Make your AI agent learn from mistakes.** Capture thumbs up/down feedback, block repeated failures, and export DPO training data — across ChatGPT, Claude, Codex, Gemini, and Amp.
+**The complete RLHF data pipeline for AI coding agents.** Capture human feedback, build memory, generate prevention rules, and export DPO training pairs — the full loop from thumbs up/down to model fine-tuning.
+
+## What This Is (and Isn't)
+
+This tool implements the **data collection and preference pipeline** side of RLHF — the part that turns your daily interactions with AI agents into structured training data. Out of the box, it:
+
+- **Captures** thumbs up/down feedback with context, tags, and rubric scores
+- **Remembers** via JSONL logs + LanceDB vector search across sessions
+- **Prevents** repeated mistakes with auto-generated guardrails
+- **Recalls** relevant past feedback mid-conversation (in-session context injection)
+- **Exports** DPO training pairs (prompt/chosen/rejected) for model fine-tuning
+
+It does **not** update model weights in real-time. That's the fine-tuning step, which you do separately using the DPO pairs this tool exports. The full loop: capture feedback here → export DPO pairs → fine-tune with [TRL](https://github.com/huggingface/trl), [OpenPipe](https://openpipe.ai), or any DPO trainer → deploy improved model.
 
 ## Architecture
 
@@ -22,12 +34,12 @@ One command. Pick your platform:
 |----------|---------|
 | **Claude** | `claude mcp add rlhf -- npx -y rlhf-feedback-loop serve` |
 | **Codex** | `codex mcp add rlhf -- npx -y rlhf-feedback-loop serve` |
-| **Gemini** | `gemini mcp add rlhf -- npx -y rlhf-feedback-loop serve` |
-| **Amp** | `cp node_modules/rlhf-feedback-loop/plugins/amp-skill/SKILL.md .amp/skills/rlhf-feedback/SKILL.md` |
+| **Gemini** | `gemini mcp add rlhf "npx -y rlhf-feedback-loop serve"` |
+| **Amp** | `amp mcp add rlhf -- npx -y rlhf-feedback-loop serve` |
 | **Cursor** | `cursor mcp add rlhf -- npx -y rlhf-feedback-loop serve` |
 | **All at once** | `npx add-mcp rlhf-feedback-loop` |
 
-That's it. Your agent can now capture feedback, recall past learnings mid-conversation, and block repeated mistakes.
+That's it. Your agent can now capture feedback, recall past learnings mid-conversation, and block repeated mistakes. Run once per project — the MCP server starts automatically on each session.
 
 ## How It Works
 
@@ -57,15 +69,21 @@ DPO export → fine-tune your model
 
 All data stored locally as **JSONL** files — fully transparent, fully portable, no vendor lock-in. **LanceDB** indexes memories as vector embeddings for semantic search. **ShieldCortex** assembles context packs so your agent starts each task informed.
 
-## Why This Exists
+## Free vs. Cloud Pro
+
+The open-source package is fully functional and free forever. Cloud Pro is for teams that don't want to self-host.
+
+| | Open Source | Cloud Pro ($10/mo) |
+|---|---|---|
+| Feedback capture | Local MCP server | Hosted HTTPS API |
+| Storage | Your machine | Managed cloud |
+| DPO export | CLI command | API endpoint |
+| Setup | `mcp add` one-liner | Provisioned API key |
+| Team sharing | Manual (share JSONL) | Built-in (shared API) |
+| Support | GitHub Issues | Email |
+| Uptime | You manage | We manage (99.9% SLA) |
 
-| Problem | What this does |
-|---------|---------------|
-| Agent keeps making the same mistake | Prevention rules auto-generated from repeated failures |
-| Agent claims "done" without proof | Rubric engine blocks positive feedback without test evidence |
-| Feedback collected but never used | DPO pairs exported for actual model fine-tuning |
-| Different tools, different formats | One MCP server works across 5 platforms |
-| Agent starts every task blank | In-session recall injects past learnings into current conversation |
+[Get Cloud Pro](https://buy.stripe.com/bJe14neyU4r4f0leOD3sI02) | [Live API](https://rlhf-feedback-loop-710216278770.us-central1.run.app)
 
 ## Deep Dive
 

package/adapters/mcp/server-stdio.js

@@ -267,19 +267,52 @@ function formatStats() {
   const pos = entries.filter(e => e.signal === 'positive').length;
   const neg = entries.filter(e => e.signal === 'negative').length;
   const memCount = fs.existsSync(memPath) ? fs.readFileSync(memPath, 'utf8').trim().split('\n').filter(Boolean).length : 0;
-  return [
-    '## Storage Proof',
-    `  Feedback log : ${logPath} (${entries.length} entries)`,
-    `  Memory log   : ${memPath} (${memCount} entries)`,
+
+  // HBR: "Which cases consume disproportionate time?" — top error domains
+  const negEntries = entries.filter(e => e.signal === 'negative');
+  const domainCounts = {};
+  negEntries.forEach(e => {
+    const domain = (e.richContext && e.richContext.domain) || 'general';
+    domainCounts[domain] = (domainCounts[domain] || 0) + 1;
+  });
+  const topDomains = Object.entries(domainCounts).sort((a, b) => b[1] - a[1]).slice(0, 3);
+
+  // HBR: "Glass box" — audit trail of recent decisions
+  const recent = entries.slice(-5).reverse();
+  const auditTrail = recent.map(e => {
+    const sig = e.signal === 'positive' ? 'UP' : 'DN';
+    const ts = (e.timestamp || '').slice(11, 19);
+    const ctx = (e.context || '').slice(0, 60);
+    return `  [${sig}] ${ts} ${ctx}`;
+  });
+
+  const parts = [
+    '## Storage',
+    `  Feedback log : ${entries.length} entries`,
+    `  Memory log   : ${memCount} memories`,
     `  LanceDB      : ${path.join(SAFE_DATA_DIR, 'lancedb/')}`,
     '',
-    '## Cumulative Stats',
-    `  Total feedback  : ${entries.length}`,
-    `  Positive (up)   : ${pos}`,
-    `  Negative (down) : ${neg}`,
-    `  Promoted to mem : ${memCount}`,
-    `  Ratio           : ${pos > 0 ? (pos / (pos + neg) * 100).toFixed(0) + '% positive' : 'n/a'}`,
-  ].join('\n');
+    '## Stats',
+    `  Total     : ${entries.length}`,
+    `  Positive  : ${pos}`,
+    `  Negative  : ${neg}`,
+    `  Promoted  : ${memCount}`,
+    `  Ratio     : ${pos > 0 ? (pos / (pos + neg) * 100).toFixed(0) + '% positive' : 'n/a'}`,
+  ];
+
+  if (topDomains.length > 0) {
+    parts.push('', '## Top Error Domains (where mistakes cluster)');
+    topDomains.forEach(([domain, count]) => {
+      parts.push(`  ${domain}: ${count} failures`);
+    });
+  }
+
+  if (auditTrail.length > 0) {
+    parts.push('', '## Audit Trail (last 5 decisions)');
+    parts.push(...auditTrail);
+  }
+
+  return parts.join('\n');
 }
 
 async function callTool(name, args = {}) {
@@ -321,15 +354,18 @@ async function callToolInner(name, args = {}) {
     const limit = Number(args.limit || 5);
     const parts = [];
 
-    // 1. Vector search for similar past feedback
+    // 1. Vector search for similar past feedback with confidence scores
     try {
       const similar = await searchSimilar(query, limit);
       if (similar.length > 0) {
         parts.push('## Relevant Past Feedback\n');
-        for (const mem of similar) {
+        for (let i = 0; i < similar.length; i++) {
+          const mem = similar[i];
           const signal = mem.signal === 'positive' ? 'GOOD' : 'BAD';
-          parts.push(`**[${signal}]** ${mem.context}`);
+          const confidence = mem._distance != null ? Math.max(0, (1 - mem._distance) * 100).toFixed(0) : '?';
+          parts.push(`**[${signal}]** (${confidence}% match) ${mem.context}`);
           if (mem.tags) parts.push(`  Tags: ${mem.tags}`);
+          if (mem.timestamp) parts.push(`  When: ${mem.timestamp}`);
           parts.push('');
         }
       }
@@ -359,9 +395,13 @@ async function callToolInner(name, args = {}) {
       }
     } catch (_) {}
 
-    const text = parts.length > 0
+    // 4. Append stats + audit trail (glass box)
+    parts.push('');
+    parts.push(formatStats());
+
+    const text = parts.length > 1
       ? parts.join('\n')
-      : 'No past feedback found. This appears to be a fresh start.';
+      : 'No past feedback found. This appears to be a fresh start.\n\n' + formatStats();
 
     return { content: [{ type: 'text', text }] };
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rlhf-feedback-loop",
-  "version": "0.6.5",
+  "version": "0.6.6",
   "description": "Make your AI agent learn from mistakes. Capture thumbs up/down feedback, block repeated failures, export DPO training data. Works with ChatGPT, Claude, Codex, Gemini, Amp.",
   "homepage": "https://github.com/IgorGanapolsky/rlhf-feedback-loop#readme",
   "repository": {