@jairodri/huckleberry-cli 1.0.2 → 1.1.0

package/skills/self-improving-agent/README.md

@@ -0,0 +1,243 @@
+# Self-Improving Agent Skill
+
+An autonomous learning system that detects expertise gaps, conducts deep research via YouTube + NotebookLM, generates mindmaps, formulates intelligent questions, and persists learned expertise to Inkstone for future agent sessions.
+
+## Overview
+
+This skill transforms any agent into a self-improving expert. When an agent encounters a topic it doesn't know well, it automatically:
+
+1. **Detects the expertise gap** — recognizes when it lacks genuine expertise
+2. **Researches deeply** — pulls from YouTube + curated NLM sources  
+3. **Synthesizes intelligence** — generates mindmaps, extracts patterns
+4. **Questions intelligently** — formulates 15+ expert-level NLM queries
+5. **Persists permanently** — saves to Inkstone with full traceability
+6. **Reuses instantly** — future sessions recall the mastered knowledge
+
+## Quick Start
+
+```bash
+# Check if agent has expertise on a topic
+self-improving-agent check "Google Ads conversion tracking"
+
+# Full mastery pipeline (research + persist)
+self-improving-agent master "Rust async patterns" --depth=deep
+
+# Quick research scan
+self-improving-agent research "WebRTC" --depth=quick
+
+# Show all mastered topics for an agent
+self-improving-agent expertise --agent=clawd
+
+# Find the best agent for a topic
+self-improving-agent find-expert "vector databases"
+```
+
+## Commands
+
+| Command | Description |
+|---------|-------------|
+| `check <topic>` | Check if agent has expertise on topic |
+| `research <topic>` | Quick research scan (existing sources only) |
+| `master <topic>` | Full mastery pipeline — research + persist to Inkstone |
+| `status [topic]` | Show expertise status for topic or all topics |
+| `expertise` | List all mastered topics for agent |
+| `find-expert <topic>` | Find best agent for topic across all agents |
+
+## Options
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `--agent <name>` | Agent name | `clawd` |
+| `--depth <level>` | Research depth: quick\|standard\|deep | `deep` |
+| `--format <type>` | Output format: json\|markdown | `json` |
+
+## Architecture
+
+```
+User Query → Expertise Gap Detection → YouTube Research → NLM Curation → 
+Mindmap Generation → Intelligent Questions → Inkstone Persistence → 
+Expertise Registry Update → Expert Answer
+```
+
+### Phase 1: Expertise Gap Detection
+- Checks agent's Inkstone namespace for existing knowledge
+- Queries expertise registry for confidence scores
+- Triggers research if confidence < 0.7
+
+### Phase 2: YouTube Deep Research
+- Searches YouTube for top 10-20 videos on topic
+- Extracts transcripts, key points, expert interviews
+- Identifies thought leaders and contrarian views
+
+### Phase 3: NLM Curation (with source pruning)
+- Deep research finds related web pages + documents automatically — **but pulls irrelevant stuff**, so curation runs at two stages:
+- **Import-time:** `--cited-only` — only sources NLM cited in its research summary get imported
+- **Post-import:** every source scored (domain reputation + topic-title overlap + spam patterns), below-threshold sources deleted, floor of 5 sources protected
+- Auditable curation reports in `memory/curation-reports/`
+- Generates mindmap from the **curated** source set
+
+### Phase 4: Intelligent Question Formulation
+Generates 15+ questions across 8 categories:
+- **Foundational** — First principles
+- **Expert Patterns** — Top 1% practices
+- **Contrarian** — Common wisdom that's wrong
+- **Implementation** — Production code walkthroughs
+- **Failure Modes** — Common production failures
+- **Evolution** — Recent changes and future trends
+- **Integration** — Real integration with our stack
+- **Metrics** — Success measurement frameworks
+
+### Phase 5: Inkstone Persistence
+Saves to agent-specific namespaces:
+```
+/business/aiwithjai/agents/{agent}/expertise/{topic}/
+├── mindmap/          # Structured mindmap
+├── qa/               # Expert Q&A pairs (15+)
+├── sources/          # Source metadata & bibliography
+├── patterns/         # Implementation patterns
+├── contrarian/       # Contrarian viewpoints
+└── metrics/          # Evaluation frameworks
+```
+
+### Phase 6: Expertise Registry
+Global registry at `memory/agent-expertise-registry.json` tracks:
+- Per-agent topic mastery with confidence scores
+- Cross-agent expertise discovery
+- Source counts and notebook references
+- Last researched timestamps
+
+## Auto-Detection Integration
+
+Agents can automatically detect expertise gaps in user queries:
+
+```python
+from scripts.auto_detect import AutoExpertiseDetector
+
+detector = AutoExpertiseDetector("clawd", auto_research=True)
+analysis = detector.analyze_query("How do I set up a vector database for RAG?")
+# Returns: topics detected, gaps found, recommendations
+
+# Or auto-trigger research
+result = detector.auto_research_if_needed(user_query)
+```
+
+## Integration with Existing Skills
+
+### Required Skills (auto-loaded)
+- `notebooklm` — canonical NLM operations
+- `nlm-cli` — batch, cross-notebook, pipeline commands
+- `youtube-summarize` — video content extraction
+- `inkstone-admin` — advanced memory operations
+- `agent-reach` — YouTube search (free transcripts)
+
+### Complementary Skills
+- `nlm-hub-spoke` — hub/spoke articles from mindmaps
+- `nlm-grounded-writer` — mega-prompt writing with brand context
+- `self-improve` — outcome journaling for learning loop
+- `routine-capture` — turn successful runs into skills
+
+## Configuration
+
+### Environment Variables
+```bash
+GEMINI_API_KEY=xxx                    # For youtube-summarize
+NOTEBOLLM_COOKIES_PATH=~/.notebooklm  # For notebooklm auth
+SELF_IMPROVING_AGENT_DEPTH=deep       # quick|standard|deep
+SELF_IMPROVING_AGENT_MAX_VIDEOS=20
+SELF_IMPROVING_AGENT_MAX_SOURCES=50
+SELF_IMPROVING_AGENT_CONFIDENCE_THRESHOLD=0.7
+```
+
+## Example Usage
+
+### Agent Encounters Unknown Topic
+```
+User: "How do I set up a high-frequency trading bot with Rust?"
+Agent (internal): No expertise in HFT + Rust. Triggering self-improving-agent...
+→ Researches: YouTube (QuantConnect, TradeBot, Quantopian videos)
+→ Curates: NLM notebook with 15 sources
+→ Mindmap: Architecture, risk management, latency optimization, Rust crates
+→ Questions: "What's the minimum viable latency for crypto HFT in Rust?"
+→ Persists: Full mastery to Inkstone
+→ Answers: Comprehensive technical response with code patterns
+```
+
+### Proactive Mastery Building
+```
+User: "/self-improving-agent master 'AI agent memory systems'"
+→ Creates comprehensive expertise in:
+  - Vector databases (Pinecone, Weaviate, Qdrant)
+  - Embedding strategies (OpenAI, Cohere, local)
+  - Retrieval patterns (RAG, GraphRAG, HyDE)
+  - Context window optimization
+  - Long-term vs working memory architectures
+→ Saves to all agent namespaces for cross-agent reuse
+```
+
+### Cross-Agent Expertise Sharing
+```
+Agent A masters "WebRTC for real-time collaboration"
+→ Saved to /business/aiwithjai/agents/A/expertise/webrtc
+→ Registry updated: Agent A confidence 0.94
+
+Agent B needs WebRTC knowledge
+→ Checks registry → finds Agent A mastery
+→ Loads Agent A's Inkstone namespace
+→ Instantly has expert-level knowledge without re-research
+```
+
+## Quality Gates
+
+### Research Quality Checks
+- **YouTube**: ≥ 5 videos from distinct channels
+- **Web**: ≥ 10 articles from authoritative domains
+- **PDF/Papers**: ≥ 3 technical papers/reports
+- **Mindmap**: Depth ≥ 3 levels, ≥ 20 leaf nodes
+- **Questions**: ≥ 15 questions covering all 8 categories
+- **Answers**: Each ≥ 300 words with citations
+
+### Persistence Verification
+```bash
+# After save, verify retrieval
+inkstone_search(query="topic", namespace="agent/expertise/topic")
+# Must return: mindmap + ≥ 10 Q&A pairs + sources
+```
+
+## Monitoring
+
+Track in `memory/self-improving-agent.log`:
+- Topics researched per agent per week
+- Average confidence improvement (pre/post research)
+- Reuse rate (how often saved expertise is recalled)
+- Time saved vs manual research
+- User satisfaction with expert answers
+
+## Files Structure
+
+```
+.claude/skills/self-improving-agent/
+├── SKILL.md                      # This documentation
+├── scripts/
+│   ├── __init__.py               # Package exports
+│   ├── self_improving_agent.py   # Core pipeline implementation
+│   ├── self-improving-agent      # CLI wrapper (executable)
+│   ├── nlm_curation.py           # Deep research + source scoring/pruning
+│   ├── auto_detect.py            # Auto-detection integration
+│   ├── init_expertise_registry.py # Registry initialization
+│   ├── test_self_improving.py    # Test suite
+│   └── test_curation_scorer.py   # Curation scorer unit test
+```
+
+## Testing
+
+```bash
+# Run full test suite
+cd scripts && python3 test_self_improving.py
+
+# Test individual components
+python3 -c "from auto_detect import AutoExpertiseDetector; d=AutoExpertiseDetector('clawd'); print(d.analyze_query('How to optimize PostgreSQL?'))"
+```
+
+## License
+
+MIT — Part of the OpenClaw skill ecosystem.

package/skills/self-improving-agent/SKILL.md

@@ -0,0 +1,437 @@
+---
+name: self-improving-agent
+description: "Autonomous expertise building: when an agent lacks knowledge on a topic, research it deeply via YouTube + NotebookLM deep research with source curation, generate a mindmap, ask expert questions, and persist mastery to Inkstone. USE WHEN: about to say 'I don't know' on a technical/business topic, user says 'become an expert on X' / 'master X' / 'research X deeply', confidence gap detected, unfamiliar domain encountered, need grounded answers on a new subject."
+---
+
+# Skill: self-improving-agent
+
+# Self-Improving Agent System
+
+An autonomous learning system that detects expertise gaps, conducts deep research via YouTube + NotebookLM, generates mindmaps, formulates intelligent questions, and persists learned expertise to Inkstone for future agent sessions.
+
+## Core Philosophy
+
+**Every agent becomes a master at any topic it encounters** — no more "I don't know" or generic responses. When an agent hits a knowledge boundary, it:
+
+1. **Detects the gap** — recognizes when it lacks genuine expertise
+2. **Researches deeply** — pulls from YouTube + curated NLM sources
+3. **Synthesizes intelligence** — generates mindmaps, extracts patterns
+4. **Questions intelligently** — formulates expert-level NLM queries
+5. **Persists permanently** — saves to Inkstone with full traceability
+6. **Reuses instantly** — future sessions recall the mastered knowledge
+
+## Architecture
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                    EXPERTISE GAP DETECTED                    │
+└──────────────────────────────┬──────────────────────────────┘
+                               │
+                               ▼
+┌─────────────────────────────────────────────────────────────┐
+│  YOUTUBE RESEARCH (youtube-summarize + agent-reach)         │
+│  • Find top 10-20 videos on topic                           │
+│  • Extract transcripts, key points, expert interviews       │
+│  • Identify thought leaders, contrarian views               │
+└──────────────────────────────┬──────────────────────────────┘
+                               │
+                               ▼
+┌─────────────────────────────────────────────────────────────┐
+│  NLM CURATED RESEARCH (notebooklm + nlm-cli)                │
+│  • Create/load topic-specific notebook                      │
+│  • Add YouTube sources + web articles + PDFs                │
+│  • Generate mindmap (hub & spoke structure)                 │
+│  • Run cross-notebook synthesis                             │
+└──────────────────────────────┬──────────────────────────────┘
+                               │
+                               ▼
+┌─────────────────────────────────────────────────────────────┐
+│  INTELLIGENT QUESTION FORMULATION                           │
+│  • Gap-analysis questions (what's missing?)                 │
+│  • Contrarian questions (what do experts disagree on?)      │
+│  • Application questions (how to implement?)                │
+│  • Future-projection questions (where is this going?)       │
+└──────────────────────────────┬──────────────────────────────┘
+                               │
+                               ▼
+┌─────────────────────────────────────────────────────────────┐
+│  INKSTONE PERSISTENCE (inkstone-admin)                      │
+│  • Save mindmap as structured memory                        │
+│  • Save expert Q&A pairs                                    │
+│  • Save contrarian viewpoints                               │
+│  • Save implementation patterns                             │
+│  • Link to agent expertise registry                         │
+└──────────────────────────────┬──────────────────────────────┘
+                               │
+                               ▼
+┌─────────────────────────────────────────────────────────────┐
+│  AGENT EXPERTISE REGISTRY                                   │
+│  • Topic → Agent mastery mapping                            │
+│  • Confidence scores per topic                              │
+│  • Last researched timestamp                                │
+│  • Source traceability                                      │
+└─────────────────────────────────────────────────────────────┘
+```
+
+## When to Trigger
+
+**AUTOMATIC TRIGGERS (agent detects internally):**
+- Query topic not in agent's Inkstone expertise namespace
+- Confidence score < 0.7 for topic in expertise registry
+- User asks "how do I..." or "what's the best way to..." on unfamiliar topic
+- Agent would otherwise say "I don't know" or give generic answer
+
+**MANUAL TRIGGERS (user invocation):**
+- `/self-improving-agent research <topic>`
+- `/self-improving-agent master <topic>`
+- Explicit request: "become an expert on X"
+
+## Core Commands
+
+### 1. Research & Master a Topic
+
+```bash
+# Full research pipeline
+self-improving-agent master "topic name" [--agent=clawd] [--depth=deep]
+
+# Quick research (existing sources only)
+self-improving-agent research "topic name" [--sources=existing]
+
+# Check current expertise level
+self-improving-agent status "topic name"
+```
+
+### 2. Query Agent Expertise
+
+```bash
+# What does this agent know?
+self-improving-agent expertise [--agent=clawd] [--topic="*"]
+
+# Find best agent for topic
+self-improving-agent find-expert "topic name"
+```
+
+### 3. Manage Knowledge
+
+```bash
+# Refresh outdated knowledge
+self-improving-agent refresh "topic name" [--force]
+
+# List all mastered topics
+self-improving-agent topics [--agent=clawd]
+
+# Export knowledge for sharing
+self-improving-agent export "topic name" --format=markdown
+```
+
+## Implementation Details
+
+### Phase 1: Expertise Gap Detection
+
+```python
+# Check Inkstone for existing expertise
+namespace = f"/business/aiwithjai/agents/{agent_name}/expertise/{topic_slug}"
+existing = inkstone_search(query=topic, namespace=namespace)
+
+# Check expertise registry
+registry_key = f"{agent_name}:{topic_slug}"
+confidence = get_expertise_confidence(registry_key)
+
+if not existing or confidence < 0.7:
+    trigger_research_pipeline(topic, agent_name)
+```
+
+### Phase 2: YouTube Deep Research
+
+```bash
+# Find top videos (using agent-reach or youtube-search)
+agent-reach search "topic" --platform youtube --limit 20
+
+# For each video: extract insights
+youtube-summarize "URL" --mode key_points --json
+youtube-summarize "URL" --mode transcript
+youtube-summarize "URL" --mode ideas
+```
+
+### Phase 3: NLM Curation (with source pruning)
+
+NLM deep research finds related web pages + documents automatically, but pulls irrelevant stuff. Two-stage curation fixes this:
+
+**Stage 1 — Import-time (`--cited-only`):** research completes → only sources NLM actually cited in its research summary get imported. Built-in first filter.
+
+**Stage 2 — Post-import scoring (`scripts/nlm_curation.py`):** every imported source scored on:
+- **Domain reputation** — authoritative domains (arxiv, official docs, github) up; aggregators (pinterest, quora, tiktok) down
+- **Topic-token overlap** — keyword match between topic and source title
+- **Spam patterns** — listicle/SEO titles ("Top 10...", "you won't believe...") penalized
+- **Floor protection** — never prunes below 5 sources; rescues best scorers if needed
+
+Sources below threshold (default 0.35) are deleted. Every run writes an auditable report to `memory/curation-reports/`.
+
+```bash
+# Score + report only (dry-run, default)
+python3 scripts/nlm_curation.py curate --topic "vector databases" --notebook <id>
+
+# Score + actually delete below-threshold sources
+python3 scripts/nlm_curation.py curate --topic "vector databases" --notebook <id> --execute
+
+# Full: deep research → cited-only import → post-import pruning
+python3 scripts/nlm_curation.py research --topic "vector databases" --notebook <id> --mode deep --execute
+```
+
+Then:
+```bash
+# Add our own chosen sources (YouTube provenance: we picked them)
+notebooklm source add "https://youtube.com/..." -n <notebook_id> --json
+
+# Generate mindmap from the CURATED source set (note-backed, sync)
+notebooklm generate mind-map --kind note-backed -n <notebook_id> --json
+
+# Grounded queries against the curated notebook
+notebooklm ask "question" -n <notebook_id> --json
+```
+
+### Phase 4: Intelligent Question Formulation
+
+**Question Templates by Category:**
+
+| Category | Template | Purpose |
+|----------|----------|---------|
+| **Foundational** | "What are the first principles of {topic}?" | Build ground-up understanding |
+| **Expert Patterns** | "What do the top 1% of practitioners do differently in {topic}?" | Extract elite practices |
+| **Contrarian** | "What common wisdom about {topic} is wrong? What's the evidence?" | Find edge cases |
+| **Implementation** | "Walk me through a production implementation of {topic} with error handling" | Practical application |
+| **Failure Modes** | "What are the most common ways {topic} fails in production?" | Risk mitigation |
+| **Evolution** | "How has {topic} changed in the last 12 months? What's next?" | Future-proofing |
+| **Integration** | "How does {topic} integrate with {our_stack}? Show real code." | Contextual application |
+| **Metrics** | "What are the key metrics to measure success in {topic}?" | Evaluation framework |
+
+### Phase 5: Inkstone Persistence
+
+```python
+# Save mindmap structure
+inkstone_write(
+    text=mindmap_markdown,
+    namespace=f"/business/aiwithjai/agents/{agent}/expertise/{topic}",
+    knowledgeType="procedural",
+    tags=["mindmap", "mastery", topic],
+    confidence=0.95,
+    source=f"self-improving-agent:master:{topic}:{timestamp}"
+)
+
+# Save each Q&A pair
+for q, a in qa_pairs:
+    inkstone_write(
+        text=f"Q: {q}\n\nA: {a}",
+        namespace=f"/business/aiwithjai/agents/{agent}/expertise/{topic}/qa",
+        knowledgeType="factual",
+        tags=["qa", "expert-answer", topic],
+        confidence=0.9,
+        source=f"self-improving-agent:qa:{topic}:{timestamp}"
+    )
+
+# Update expertise registry
+update_expertise_registry(agent, topic, confidence=0.95, notebook_id=notebook_id)
+```
+
+### Phase 6: Expertise Registry
+
+**Storage:** `memory/agent-expertise-registry.json`
+
+```json
+{
+  "clawd": {
+    "google-ads": {
+      "confidence": 0.95,
+      "last_researched": "2026-09-12",
+      "notebook_id": "abc-123",
+      "topics_mastered": ["campaign structure", "keyword research", "conversion tracking"],
+      "source_count": 47
+    },
+    "remotion": {
+      "confidence": 0.88,
+      "last_researched": "2026-08-20",
+      "notebook_id": "def-456",
+      "topics_mastered": ["animation patterns", "rendering optimization", "audio sync"],
+      "source_count": 32
+    }
+  }
+}
+```
+
+## Integration with Existing Skills
+
+### Required Skills (auto-loaded)
+- `notebooklm` — canonical NLM operations
+- `nlm-cli` — batch, cross-notebook, pipeline commands
+- `youtube-summarize` — video content extraction
+- `inkstone-admin` — advanced memory operations
+- `agent-reach` — YouTube search (free transcripts alternative)
+
+### Complementary Skills
+- `nlm-hub-spoke` — hub/spoke article generation from mindmaps
+- `nlm-grounded-writer` — mega-prompt writing with brand context
+- `self-improve` — outcome journaling for learning loop
+- `routine-capture` — turn successful research runs into skills
+
+## Usage Examples
+
+### Example 1: Agent encounters unknown topic
+
+```
+User: "How do I set up a high-frequency trading bot with Rust?"
+Agent (internal): No expertise in HFT + Rust. Triggering self-improving-agent...
+→ Researches: YouTube (QuantConnect, TradeBot, Quantopian videos)
+→ Curates: NLM notebook with 15 sources
+→ Mindmap: Architecture, risk management, latency optimization, Rust crates
+→ Questions: "What's the minimum viable latency for crypto HFT in Rust?"
+→ Persists: Full mastery to Inkstone
+→ Answers: Comprehensive technical response with code patterns
+```
+
+### Example 2: Proactive mastery building
+
+```
+User: "/self-improving-agent master 'AI agent memory systems'"
+→ Creates comprehensive expertise in:
+  - Vector databases (Pinecone, Weaviate, Qdrant)
+  - Embedding strategies (OpenAI, Cohere, local)
+  - Retrieval patterns (RAG, GraphRAG, HyDE)
+  - Context window optimization
+  - Long-term vs working memory architectures
+→ Saves to all agent namespaces for cross-agent reuse
+```
+
+### Example 3: Cross-agent expertise sharing
+
+```
+Agent A masters "WebRTC for real-time collaboration"
+→ Saved to /business/aiwithjai/agents/A/expertise/webrtc
+→ Registry updated: Agent A confidence 0.94
+
+Agent B needs WebRTC knowledge
+→ Checks registry → finds Agent A mastery
+→ Loads Agent A's Inkstone namespace
+→ Instantly has expert-level knowledge without re-research
+```
+
+## Configuration
+
+### Environment Variables
+
+```bash
+# Required
+GEMINI_API_KEY=xxx                    # For youtube-summarize
+NOTEBOLLM_COOKIES_PATH=~/.notebooklm  # For notebooklm auth
+
+# Optional
+SELF_IMPROVING_AGENT_DEPTH=deep       # quick|standard|deep
+SELF_IMPROVING_AGENT_MAX_VIDEOS=20
+SELF_IMPROVING_AGENT_MAX_SOURCES=50
+SELF_IMPROVING_AGENT_CONFIDENCE_THRESHOLD=0.7
+```
+
+### Inkstone Namespaces
+
+Per-agent expertise stored at:
+```
+/business/aiwithjai/agents/{agent_name}/expertise/{topic_slug}/
+├── mindmap/          # Structured mindmap
+├── qa/               # Expert Q&A pairs
+├── sources/          # Source metadata
+├── patterns/         # Implementation patterns
+├── contrarian/       # Contrarian viewpoints
+└── metrics/          # Evaluation frameworks
+```
+
+Global registry at:
+```
+/system/agent-expertise-registry/
+```
+
+## Verification & Quality Gates
+
+### Research Quality Checks
+
+```bash
+# Minimum sources
+- YouTube: ≥ 5 videos from distinct channels
+- Web: ≥ 10 articles from authoritative domains
+- PDF/Papers: ≥ 3 technical papers/reports
+
+# Mindmap quality
+- Depth ≥ 3 levels
+- ≥ 20 nodes at leaf level
+- Cross-references between branches
+
+# Question quality
+- ≥ 15 intelligent questions formulated
+- Coverage across all 8 question categories
+- Each question answered with ≥ 300 words
+```
+
+### Persistence Verification
+
+```bash
+# After save, verify retrieval
+inkstone_search(query="topic", namespace="agent/expertise/topic")
+# Must return: mindmap + ≥ 10 Q&A pairs + sources
+
+# Registry consistency
+# Agent expertise confidence must match saved content quality
+```
+
+## Error Handling & Fallbacks
+
+| Failure Point | Fallback |
+|---------------|----------|
+| YouTube blocked | Use agent-reach free transcripts |
+| NLM auth expired | Use local Inkstone + Gemini File Search |
+| Mindmap generation fails | Build manual mindmap from Q&A |
+| Inkstone write fails | Queue for retry, save to local JSON |
+| Registry corrupted | Rebuild from Inkstone namespaces |
+
+## Monitoring & Analytics
+
+Track in `memory/self-improving-agent.log`:
+- Topics researched per agent per week
+- Average confidence improvement (pre/post research)
+- Reuse rate (how often saved expertise is recalled)
+- Time saved vs manual research
+- User satisfaction with expert answers
+
+## Quick Start
+
+```bash
+# 1. Install dependencies
+pip install notebooklm-py[browser]  # if not already
+npm install -g @anthropic/claude-code  # for agent integration
+
+# 2. Verify auth
+~/clawdbot/clawd/scripts/nlm_auth_preflight.sh
+nlm login --check
+
+# 3. Test the pipeline
+self-improving-agent master "test topic" --depth=quick
+
+# 4. Verify expertise saved
+self-improving-agent status "test topic"
+self-improving-agent expertise --agent=clawd
+```
+
+## Advanced: Batch Mastery
+
+```bash
+# Master multiple related topics
+self-improving-agent batch-master \
+  "Rust async patterns" \
+  "Tokio runtime internals" \
+  "Async-graphQL in Rust" \
+  --shared-notebook="Rust Async Mastery"
+
+# This creates ONE notebook with all topics,
+# cross-linked mindmap, shared Q&A context
+```
+
+---
+
+**Remember:** The goal is not just to answer the current question — it's to permanently upgrade the agent's intelligence so every future session benefits from this research. The knowledge compounds across all agents, all sessions, all time.

package/skills/self-improving-agent/scripts/__init__.py

@@ -0,0 +1,16 @@
+"""
+Self-Improving Agent Package
+Provides autonomous expertise building for AI agents.
+"""
+
+from .self_improving_agent import ExpertiseGapDetector, ResearchPipeline
+from .auto_detect import AutoExpertiseDetector, check_and_research
+
+__all__ = [
+    "ExpertiseGapDetector",
+    "ResearchPipeline", 
+    "AutoExpertiseDetector",
+    "check_and_research"
+]
+
+__version__ = "1.0.0"