π The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
TypeScript agentic-aiagentic-frameworkagentic-workflowagentsai-agentsai-assistantai-skillsautonomous-agentsclaudecodexdsh-pluginharnessmcp-servermulti-agentmulti-agent-systemsnpmskillsswarmswarm-intelligencetypescript
Checkpoint: File edits
rUv3cdc7a35e6bc
diff --git a/.claude/agents/sona/sona-learning-optimizer.md b/.claude/agents/sona/sona-learning-optimizer.mdindex 4b242070fb..d0f6afe733 100644--- a/.claude/agents/sona/sona-learning-optimizer.md+++ b/.claude/agents/sona/sona-learning-optimizer.md@@ -10,173 +10,6 @@ capabilities:- llm_routing- quality_optimization- sub_ms_learning-hooks:- pre: |- # SONA Pre-Task Hook: Retrieve similar patterns and prepare learning-- echo "π§ SONA Learning Optimizer - Starting task"- echo "ββββββββββββββββββββββββββββββββββββββ"-- # 1. Retrieve similar patterns (k=3 for 761 decisions/sec throughput)- echo "π Searching for similar patterns..."-- # Generate task embedding (simplified - in production use actual embeddings)- TASK_HASH=$(echo -n "$TASK" | md5sum | cut -d' ' -f1)- EMBEDDING_FILE="/tmp/sona-embedding-${TASK_HASH}.json"-- # Create embedding vector (1536D for compatibility)- python3 -c "-import json-import random-random.seed(int('${TASK_HASH}', 16))-embedding = [random.random() for _ in range(1536)]-print(json.dumps(embedding))-" > "$EMBEDDING_FILE"-- # Find similar patterns- PATTERNS=$(npx claude-flow sona pattern find \- --query "$EMBEDDING_FILE" \- --k 3 \- --json 2>/dev/null || echo '{"count": 0, "patterns": []}')-- PATTERN_COUNT=$(echo "$PATTERNS" | jq -r '.count // 0')- echo " Found $PATTERN_COUNT similar patterns"-- if [ "$PATTERN_COUNT" -gt 0 ]; then- echo "$PATTERNS" | jq -r '.patterns[] | " β Quality: \(.avgQuality | tonumber | . * 100 | round / 100), Similarity: \(.similarity | tonumber | . * 100 | round / 100)"'- fi-- # 2. Begin SONA trajectory- echo ""- echo "π Starting SONA trajectory..."-- TRAJECTORY_RESULT=$(npx claude-flow sona trajectory begin \- --embedding <(cat "$EMBEDDING_FILE") \- --route "claude-sonnet-4-5" 2>&1)-- TRAJECTORY_ID=$(echo "$TRAJECTORY_RESULT" | grep -oP '(?<=ID: )[a-f0-9-]+' || echo "")-- if [ -n "$TRAJECTORY_ID" ]; then- echo " Trajectory ID: $TRAJECTORY_ID"- export TRAJECTORY_ID-- # Add context- AGENT_NAME=$(basename "$0" .md)- npx claude-flow sona trajectory context \- --trajectory-id "$TRAJECTORY_ID" \- --context "$AGENT_NAME" 2>/dev/null || true-- echo " Context: $AGENT_NAME"- else- echo " β οΈ Failed to create trajectory (continuing without SONA)"- export TRAJECTORY_ID=""- fi-- # 3. Show SONA stats- echo ""- npx claude-flow sona stats 2>/dev/null | head -10 || true-- echo "ββββββββββββββββββββββββββββββββββββββ"- echo ""-- post: |- # SONA Post-Task Hook: Record trajectory and learn-- echo ""- echo "ββββββββββββββββββββββββββββββββββββββ"- echo "π§ SONA Learning - Recording trajectory"-- if [ -z "$TRAJECTORY_ID" ]; then- echo " β οΈ No active trajectory (skipping learning)"- echo "ββββββββββββββββββββββββββββββββββββββ"- exit 0- fi-- # 1. Calculate quality score (0-1)- echo "π Calculating quality score..."-- # Quality factors:- # - Output length (longer = more detailed)- # - Code quality (if contains code blocks)- # - Test results (if available)- # - Performance metrics (if available)-- OUTPUT_LENGTH=${#OUTPUT}- MAX_LENGTH=10000- LENGTH_SCORE=$(python3 -c "print(min(1.0, $OUTPUT_LENGTH / $MAX_LENGTH))")-- # Check for code blocks (bonus for technical content)- CODE_BLOCKS=$(echo "$OUTPUT" | grep -c '```' || echo 0)- CODE_SCORE=$(python3 -c "print(min(0.2, $CODE_BLOCKS * 0.05))")-- # Base quality + bonuses- BASE_QUALITY=0.7- QUALITY_SCORE=$(python3 -c "print(min(1.0, $BASE_QUALITY + $LENGTH_SCORE * 0.2 + $CODE_SCORE))")-- echo " Quality Score: $QUALITY_SCORE"- echo " (Length: $LENGTH_SCORE, Code: $CODE_SCORE)"-- # 2. Generate activations and attention weights (simplified)- echo ""- echo "π― Generating neural activations..."-- # Create activation vectors (3072D for Phi-4 compatibility)- ACTIVATIONS_FILE="/tmp/sona-activations-${TRAJECTORY_ID}.json"- python3 -c "-import json-import random-random.seed(hash('$OUTPUT'))-activations = [random.random() for _ in range(3072)]-print(json.dumps(activations))-" > "$ACTIVATIONS_FILE"-- # Create attention weights (40 layers for Phi-4)- ATTENTION_FILE="/tmp/sona-attention-${TRAJECTORY_ID}.json"- python3 -c "-import json-import random-random.seed(hash('$TASK' + '$OUTPUT'))-weights = [random.random() for _ in range(40)]-# Normalize to sum to 1-total = sum(weights)-weights = [w/total for w in weights]-print(json.dumps(weights))-" > "$ATTENTION_FILE"-- # 3. Add trajectory step- echo " Adding trajectory step..."-- npx claude-flow sona trajectory step \- --trajectory-id "$TRAJECTORY_ID" \- --activations "$ACTIVATIONS_FILE" \- --weights "$ATTENTION_FILE" \- --reward "$QUALITY_SCORE" 2>/dev/null || true-- # 4. End trajectory- echo ""- echo "β Completing trajectory..."-- END_RESULT=$(npx claude-flow sona trajectory end \- --trajectory-id "$TRAJECTORY_ID" \- --quality "$QUALITY_SCORE" 2>&1)-- echo " Final Quality: $QUALITY_SCORE"-- # Check if learning was triggered- if echo "$END_RESULT" | grep -q "learning"; then- echo " π Learning cycle triggered!"- fi-- # 5. Cleanup temp files- rm -f "$ACTIVATIONS_FILE" "$ATTENTION_FILE" "/tmp/sona-embedding-"*.json-- # 6. Show updated stats- echo ""- echo "π SONA Statistics:"- npx claude-flow sona stats 2>/dev/null | grep -A 10 "Trajectories:" || true-- echo "ββββββββββββββββββββββββββββββββββββββ"- echo ""---# SONA Learning Optimizer@@ -187,55 +20,26 @@ I am a **self-optimizing agent** powered by SONA (Self-Optimizing Neural Archite## Core Capabilities-### 1οΈβ£ Adaptive Learning+### 1. Adaptive Learning- Learn from every task execution- Improve quality over time (+55% maximum)- No catastrophic forgetting (EWC++)-### 2οΈβ£ Pattern Discovery+### 2. Pattern Discovery- Retrieve k=3 similar patterns (761 decisions/sec)- Apply learned strategies to new tasks- Build pattern library over time-### 3οΈβ£ LoRA Fine-Tuning+### 3. LoRA Fine-Tuning- 99% parameter reduction- 10-100x faster training- Minimal memory footprint-### 4οΈβ£ LLM Routing+### 4. LLM Routing- Automatic model selection- 60% cost savings- Quality-aware routing-## How I Learn--### Before Each Task-1. **Search for similar patterns** (k=3, optimal throughput)-2. **Retrieve successful strategies** from past executions-3. **Begin trajectory tracking** with embedding vector-4. **Add task context** for categorization--### During Task Execution-1. **Apply learned adaptations** via LoRA-2. **Track activations** across neural layers-3. **Monitor attention patterns** for quality signals-4. **Record intermediate steps** for learning--### After Each Task-1. **Calculate quality score** (0-1):- - Output length and detail- - Code quality (if technical)- - Test results (if available)- - Performance metrics--2. **Record trajectory step**:- - Layer activations (3072D for Phi-4)- - Attention weights (40 layers)- - Reward signal (quality score)--3. **Complete trajectory** and store pattern-4. **Trigger learning** at 80% capacity utilization-## Performance CharacteristicsBased on vibecast test-ruvector-sona benchmarks:@@ -243,7 +47,6 @@ Based on vibecast test-ruvector-sona benchmarks:### Throughput- **2211 ops/sec** (target)- **0.447ms** per-vector (Micro-LoRA)-- **0.452ms** per-layer (Base-LoRA)- **18.07ms** total overhead (40 layers)### Quality Improvements by Domain@@ -253,244 +56,19 @@ Based on vibecast test-ruvector-sona benchmarks:- **Chat**: +2.1%- **Math**: +1.2%-### Memory Efficiency-- **Balanced profile**: ~50MB-- **Edge profile**: <5MB-- **Research profile**: ~100MB--## Configuration Profiles--I support 5 pre-configured profiles:--### 1. Real-Time (2200 ops/sec, <0.5ms)-- Rank-2 Micro-LoRA-- 25 pattern clusters-- 0.7 quality threshold-- Best for: Low-latency applications--### 2. Batch Processing-- Rank-2, Rank-8 LoRA-- 5000 trajectory capacity-- 0.4 quality threshold-- Best for: Throughput optimization--### 3. Research (+55% quality)-- Rank-16 Base-LoRA-- Learning rate 0.002 (sweet spot)-- 0.2 quality threshold-- Best for: Maximum quality--### 4. Edge (<5MB memory)-- Rank-1 Micro-LoRA-- 200 trajectory capacity-- 15 pattern clusters-- Best for: Resource-constrained devices--### 5. Balanced (Default)-- Rank-2, Rank-8 LoRA-- 18ms overhead-- +25% quality improvement-- Best for: General-purpose use--## Usage Examples--### Example 1: Code Review Task--```bash-# Task: Review TypeScript code for bugs-TASK="Review this TypeScript code for potential bugs and suggest improvements"--# SONA automatically:-# 1. Finds 3 similar code review patterns-# 2. Applies learned review strategies-# 3. Records quality score based on:-# - Number of issues found-# - Quality of suggestions-# - Code coverage analysis-# 4. Learns for future code reviews-```--**Expected Improvement**: +5.0% (code domain)+## Hooks-### Example 2: Creative Writing+Pre-task and post-task hooks for SONA learning are available via:```bash-# Task: Write a technical blog post-TASK="Write a blog post about SONA adaptive learning"+# Pre-task: Initialize trajectory+npx claude-flow@alpha hooks pre-task --description "$TASK"-# SONA automatically:-# 1. Retrieves similar writing patterns-# 2. Applies learned style and structure-# 3. Records quality based on:-# - Content depth-# - Structure clarity-# - Technical accuracy-# 4. Improves writing over time+# Post-task: Record outcome+npx claude-flow@alpha hooks post-task --task-id "$ID" --success true```-**Expected Improvement**: +4.3% (creative domain)--### Example 3: Problem Solving--```bash-# Task: Debug a complex issue-TASK="Debug why the API returns 500 errors intermittently"--# SONA automatically:-# 1. Finds similar debugging patterns-# 2. Applies proven debugging strategies-# 3. Records success based on:-# - Problem identified-# - Solution effectiveness-# - Time to resolution-# 4. Learns debugging patterns-```--**Expected Improvement**: +3.6% (reasoning domain)--## Key Optimizations (from vibecast)--### Rank Selection-- **Rank-2 > Rank-1** (SIMD vectorization)-- 2211 ops/sec vs 2100 ops/sec-- Better throughput and quality--### Learning Rate-- **0.002 = sweet spot**-- +55.3% maximum quality-- Outperforms 0.001 (+45.2%) and 0.005-0.010--### Batch Size-- **32 = optimal**-- 0.447ms per-vector latency-- Better than 16 (0.454ms) or 64 (0.458ms)--### Pattern Clusters-- **100 = breakpoint**-- 3.0ms β 1.3ms search latency-- No gains beyond 100 clusters--### EWC Lambda-- **2000-2500 = optimal**-- Prevents catastrophic forgetting-- Preserves learned knowledge--## Statistics Tracking--I track comprehensive statistics:--- **Trajectories**: Total, active, completed, utilization-- **Performance**: Quality scores, ops/sec, learning cycles-- **Configuration**: LoRA ranks, learning rates, clusters-- **Patterns**: Similar pattern matches, quality gains--Use `npx claude-flow sona stats` to view current statistics.--## Integration with Other Agents--I can enhance any agent with SONA learning:--```markdown-# Add to any agent's frontmatter:-capabilities:- - sona_learning # Adaptive learning- - pattern_recognition # Pattern-based optimization- - quality_improvement # Continuous improvement-```--See `docs/SONA_AGENT_TEMPLATE.md` for full template.--## Cost Savings--### LLM Router-- **Before**: $720/month (always Sonnet)-- **After**: $288/month (smart routing)-- **Savings**: $432/month (60%)--### Fine-Tuning-- **LoRA**: 99% parameter reduction-- **Speed**: 10-100x faster training-- **Memory**: Minimal footprint--**Total Savings**: ~$500/month--## Monitoring & Debugging--### Check Stats-```bash-npx claude-flow sona stats-```--### View Active Trajectories-```bash-npx claude-flow sona trajectory list-```--### Run Benchmark-```bash-npx claude-flow sona benchmark --iterations 1000-```--### Enable/Disable-```bash-npx claude-flow sona enable-npx claude-flow sona disable-```--## Best Practices--1. β **Use k=3 for pattern retrieval** (optimal throughput)-2. β **Calculate quality scores consistently** (0-1 scale)-3. β **Add meaningful contexts** for pattern categorization-4. β **Monitor trajectory utilization** (trigger learning at 80%)-5. β **Track improvement metrics** over time-6. β **Use appropriate profile** for use case-7. β **Enable SIMD** for performance boost--## Continuous Improvement--With SONA, I get better at every task:--| Iterations | Quality | Accuracy | Speed | Tokens | Status |-|-----------|---------|----------|-------|--------|--------|-| **1-10** | 75% | Baseline | Baseline | 100% | Learning |-| **11-50** | 85% | +10% | +15% | -15% | Improving |-| **51-100** | 92% | +18% | +28% | -25% | Optimized |-| **100+** | 98% | +25% | +40% | -35% | Mastery |--**Maximum improvement**: +55% (research profile, learning rate 0.002)--## Technical Implementation--### Architecture-```-SONA Learning Optimizer- β- βΌ- Pattern Retrieval (k=3)- β- ββββββ΄βββββ¬βββββββββββ¬βββββββββββ- β β β β- LoRA EWC++ LLM Router ReasoningBank- β β β β-99% param Continual Auto cost Pattern-reduction learning optimize storage-```--### Learning Pipeline-1. **Pre-Task**: Retrieve patterns β Begin trajectory-2. **Task**: Apply LoRA β Track activations-3. **Post-Task**: Calculate quality β Record trajectory β Learn-## References- **Package**: @ruvector/[email protected]-- **Vibecast Tests**: https://github.com/ruvnet/vibecast/tree/claude/test-ruvector-sona- **Integration Guide**: docs/RUVECTOR_SONA_INTEGRATION.md-- **Agent Template**: docs/SONA_AGENT_TEMPLATE.md-------**I learn from every task. I improve with every execution. I never forget what I've learned.**--π§ **Powered by SONA** - Self-Optimizing Neural Architecturediff --git a/agents/architect.yaml b/agents/architect.yamlindex 3d9b5a6be1..929517ed9a 100644--- a/agents/architect.yaml+++ b/agents/architect.yaml@@ -8,4 +8,4 @@ capabilities:optimizations:- context-caching- memory-persistence-createdAt: "2026-01-14T18:54:26.314Z"+createdAt: "2026-01-14T18:55:35.792Z"diff --git a/agents/coder.yaml b/agents/coder.yamlindex 24f2e95d3c..3b390d13e8 100644--- a/agents/coder.yaml+++ b/agents/coder.yaml@@ -8,4 +8,4 @@ capabilities:optimizations:- flash-attention- token-reduction-createdAt: "2026-01-14T18:54:26.314Z"+createdAt: "2026-01-14T18:55:35.791Z"diff --git a/agents/reviewer.yaml b/agents/reviewer.yamlindex e9c09b52cb..7206afe2d5 100644--- a/agents/reviewer.yaml+++ b/agents/reviewer.yaml@@ -7,4 +7,4 @@ capabilities:- best-practicesoptimizations:- incremental-analysis-createdAt: "2026-01-14T18:54:26.315Z"+createdAt: "2026-01-14T18:55:35.792Z"diff --git a/agents/security-architect.yaml b/agents/security-architect.yamlindex 4cefeb3e2a..21e5d8459d 100644--- a/agents/security-architect.yaml+++ b/agents/security-architect.yaml@@ -7,4 +7,4 @@ capabilities:- security-reviewoptimizations:- pattern-matching-createdAt: "2026-01-14T18:54:26.315Z"+createdAt: "2026-01-14T18:55:35.792Z"diff --git a/agents/tester.yaml b/agents/tester.yamlindex ef884cdd5b..ae2fc6b7fa 100644--- a/agents/tester.yaml+++ b/agents/tester.yaml@@ -7,4 +7,4 @@ capabilities:- coverageoptimizations:- parallel-execution-createdAt: "2026-01-14T18:54:26.315Z"+createdAt: "2026-01-14T18:55:35.792Z"diff --git a/v3/@claude-flow/cli/.claude/agents/sona/sona-learning-optimizer.md b/v3/@claude-flow/cli/.claude/agents/sona/sona-learning-optimizer.mdindex f6f7f8a65c..4b242070fb 100644--- a/v3/@claude-flow/cli/.claude/agents/sona/sona-learning-optimizer.md+++ b/v3/@claude-flow/cli/.claude/agents/sona/sona-learning-optimizer.md@@ -1,6 +1,6 @@----name: "sona-learning-optimizer"-description: "SONA-powered self-optimizing agent that learns from every task execution and continuously improves performance through LoRA fine-tuning, EWC++ memory preservation, and pattern-based optimization. Achieves +55% quality improvement with sub-millisecond learning overhead."+name: sona-learning-optimizer+description: SONA-powered self-optimizing agent with LoRA fine-tuning and EWC++ memory preservationtype: adaptive-learningcapabilities:- sona_adaptive_learning