agentnova 0.2.2__tar.gz → 0.2.3__tar.gz

{agentnova-0.2.2 → agentnova-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentnova
-Version: 0.2.2
+Version: 0.2.3
 Summary: A minimal, hackable agentic framework for Ollama and BitNet - local-first AI agent toolkit
 Author-email: VTSTech <veritas@vts-tech.org>
 Maintainer-email: VTSTech <veritas@vts-tech.org>
@@ -34,7 +34,7 @@ Requires-Dist: black>=23.0; extra == "dev"
 Requires-Dist: ruff>=0.1.0; extra == "dev"
 Dynamic: license-file
 
-# ⚛️ AgentNova R02
+# ⚛️ AgentNova R02.3
 
 A minimal, hackable agentic framework engineered to run **entirely locally** with [Ollama](https://ollama.com) or [BitNet](https://github.com/microsoft/BitNet).
 
@@ -42,7 +42,11 @@ Inspired by the architecture of OpenClaw, rebuilt from scratch for local-first o
 
 **Written by [VTSTech](https://www.vts-tech.org)** · [GitHub](https://github.com/VTSTech/AgentNova)
 
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/VTSTech/AgentNova/blob/main/AgentNova.ipynb)
 [![GitHub commits](https://badgen.net/github/commits/VTSTech/AgentNova)](https://GitHub.com/VTSTech/AgentNova/commit/) [![GitHub latest commit](https://badgen.net/github/last-commit/VTSTech/AgentNova)](https://GitHub.com/VTSTech/AgentNova/commit/)
+[![pip - agentnova](https://img.shields.io/badge/pip-agentnova-2ea44f?logo=PyPi)](https://pypi.org/project/agentnova/)
+[![License](https://img.shields.io/badge/License-MIT-blue)](#license) [![Go to Python website](https://img.shields.io/badge/dynamic/toml?url=https%3A%2F%2Fraw.githubusercontent.com%2FVTSTech%2FAgentNova%2Frefs%2Fheads%2Fmain%2Fpyproject.toml&query=project.requires-python&label=python&logo=python&logoColor=white)](https://python.org)
+
 ---
 
 ## 📚 Documentation
@@ -195,20 +199,31 @@ agentnova models --tool_support
 
 ### Performance by Tool Support
 
-Recent test results with native tool synthesis:
+R02.3 benchmark results (15-test suite):
+
+| Model | Params | Tool Support | Score | Time |
+|-------|--------|--------------|-------|------|
+| **`granite3.1-moe:1b`** | 1B MoE | react | **93% (14/15)** | 87.8s |
+| **`llama3.2:1b`** | 1.2B | native | **87% (13/15)** | 150.2s |
+| `dolphin3.0-qwen2.5:0.5b` | 500M | none | 73% (11/15) | 27.2s |
+| `qwen3:0.6b` | 600M | react | 67% (10/15) | 189.3s |
+| `qwen2.5-coder:0.5b` | 494M | react | 60% (9/15) | 133.1s |
+
+**Quick Diagnostic (Test 15 - 5 questions, ~30s/model):**
 
-| Model | Params | Tool Support | Calculator | Shell | Python |
-|-------|--------|--------------|------------|-------|--------|
-| `qwen2.5:0.5b` | 494M | native | **100%** | **100%** | **100%** |
-| `qwen2.5-coder:0.5b` | 494M | ReAct | **100%** | **100%** | **100%** |
-| `granite4:350m` | 350M | native | ~90% | ✅ | ✅ |
-| `gemma3:270m` | 270M | none | **64%** | N/A | N/A |
+| Model | Score | Tool Support |
+|-------|-------|--------------|
+| **`qwen3.5:0.8b`** | **100%** | native |
+| **`qwen2.5:0.5b`** | **100%** | react |
+| `functiongemma:270m` | 80% | native |
+| `granite4:350m` | 80% | native |
+| `qwen3:0.6b` | 60% | react |
 
-**Key improvements in R01**:
-- Native tool synthesis extracts expressions from natural language
-- Two-tier retry: hint → synthesize (bypasses confused models)
-- Bare expression wrapping: `2**20` → `print(2**20)`
-- Hallucinated mention detection for models that talk about tools but don't call them
+**Key improvements in R02.3**:
+- **+13%** for granite3.1-moe:1b (80% → 93%) from few-shot fix
+- **+20%** for llama3.2:1b (67% → 87%) from observation role fix
+- **qwen3:0.6b restored** from 0% (broken) with `think=False` API fix
+- **qwen3.5:0.8b** new sub-1B champion with 100% on quick diagnostic
 
 ---
 
@@ -259,7 +274,7 @@ Output shows:
 - **Tool Support** - `✓ native`, `ReAct`, `○ none`, or `untested`
 
 ```
-⚛️ AgentNova R02 Models
+⚛️ AgentNova R02.3 Models
   Model                                      Family       Context    Tool Support
   ──────────────────────────────────────────────────────────────────────────────
   gemma3:270m                                gemma3       32K        ○ none
@@ -276,8 +291,9 @@ Output shows:
 # List all available tests
 agentnova test --list
 
-# Run a quick test suite
-agentnova test quick
+# Quick diagnostic - 5 questions, ~30s/model (NEW in R02.3)
+agentnova test 15 --model granite3.1-moe:1b
+agentnova test 15 --model all --debug
 
 # Run GSM8K benchmark (50 math questions)
 agentnova test 14 --acp --timeout 6400

{agentnova-0.2.2 → agentnova-0.2.3}/README.md

@@ -1,4 +1,4 @@
-# ⚛️ AgentNova R02
+# ⚛️ AgentNova R02.3
 
 A minimal, hackable agentic framework engineered to run **entirely locally** with [Ollama](https://ollama.com) or [BitNet](https://github.com/microsoft/BitNet).
 
@@ -6,7 +6,11 @@ Inspired by the architecture of OpenClaw, rebuilt from scratch for local-first o
 
 **Written by [VTSTech](https://www.vts-tech.org)** · [GitHub](https://github.com/VTSTech/AgentNova)
 
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/VTSTech/AgentNova/blob/main/AgentNova.ipynb)
 [![GitHub commits](https://badgen.net/github/commits/VTSTech/AgentNova)](https://GitHub.com/VTSTech/AgentNova/commit/) [![GitHub latest commit](https://badgen.net/github/last-commit/VTSTech/AgentNova)](https://GitHub.com/VTSTech/AgentNova/commit/)
+[![pip - agentnova](https://img.shields.io/badge/pip-agentnova-2ea44f?logo=PyPi)](https://pypi.org/project/agentnova/)
+[![License](https://img.shields.io/badge/License-MIT-blue)](#license) [![Go to Python website](https://img.shields.io/badge/dynamic/toml?url=https%3A%2F%2Fraw.githubusercontent.com%2FVTSTech%2FAgentNova%2Frefs%2Fheads%2Fmain%2Fpyproject.toml&query=project.requires-python&label=python&logo=python&logoColor=white)](https://python.org)
+
 ---
 
 ## 📚 Documentation
@@ -159,20 +163,31 @@ agentnova models --tool_support
 
 ### Performance by Tool Support
 
-Recent test results with native tool synthesis:
+R02.3 benchmark results (15-test suite):
+
+| Model | Params | Tool Support | Score | Time |
+|-------|--------|--------------|-------|------|
+| **`granite3.1-moe:1b`** | 1B MoE | react | **93% (14/15)** | 87.8s |
+| **`llama3.2:1b`** | 1.2B | native | **87% (13/15)** | 150.2s |
+| `dolphin3.0-qwen2.5:0.5b` | 500M | none | 73% (11/15) | 27.2s |
+| `qwen3:0.6b` | 600M | react | 67% (10/15) | 189.3s |
+| `qwen2.5-coder:0.5b` | 494M | react | 60% (9/15) | 133.1s |
+
+**Quick Diagnostic (Test 15 - 5 questions, ~30s/model):**
 
-| Model | Params | Tool Support | Calculator | Shell | Python |
-|-------|--------|--------------|------------|-------|--------|
-| `qwen2.5:0.5b` | 494M | native | **100%** | **100%** | **100%** |
-| `qwen2.5-coder:0.5b` | 494M | ReAct | **100%** | **100%** | **100%** |
-| `granite4:350m` | 350M | native | ~90% | ✅ | ✅ |
-| `gemma3:270m` | 270M | none | **64%** | N/A | N/A |
+| Model | Score | Tool Support |
+|-------|-------|--------------|
+| **`qwen3.5:0.8b`** | **100%** | native |
+| **`qwen2.5:0.5b`** | **100%** | react |
+| `functiongemma:270m` | 80% | native |
+| `granite4:350m` | 80% | native |
+| `qwen3:0.6b` | 60% | react |
 
-**Key improvements in R01**:
-- Native tool synthesis extracts expressions from natural language
-- Two-tier retry: hint → synthesize (bypasses confused models)
-- Bare expression wrapping: `2**20` → `print(2**20)`
-- Hallucinated mention detection for models that talk about tools but don't call them
+**Key improvements in R02.3**:
+- **+13%** for granite3.1-moe:1b (80% → 93%) from few-shot fix
+- **+20%** for llama3.2:1b (67% → 87%) from observation role fix
+- **qwen3:0.6b restored** from 0% (broken) with `think=False` API fix
+- **qwen3.5:0.8b** new sub-1B champion with 100% on quick diagnostic
 
 ---
 
@@ -223,7 +238,7 @@ Output shows:
 - **Tool Support** - `✓ native`, `ReAct`, `○ none`, or `untested`
 
 ```
-⚛️ AgentNova R02 Models
+⚛️ AgentNova R02.3 Models
   Model                                      Family       Context    Tool Support
   ──────────────────────────────────────────────────────────────────────────────
   gemma3:270m                                gemma3       32K        ○ none
@@ -240,8 +255,9 @@ Output shows:
 # List all available tests
 agentnova test --list
 
-# Run a quick test suite
-agentnova test quick
+# Quick diagnostic - 5 questions, ~30s/model (NEW in R02.3)
+agentnova test 15 --model granite3.1-moe:1b
+agentnova test 15 --model all --debug
 
 # Run GSM8K benchmark (50 math questions)
 agentnova test 14 --acp --timeout 6400

{agentnova-0.2.2 → agentnova-0.2.3}/agentnova/__init__.py

@@ -1,5 +1,5 @@
 """
-⚛️ AgentNova R02 - A minimal, hackable agentic framework for Ollama and BitNet
+⚛️ AgentNova R02.3 - A minimal, hackable agentic framework for Ollama and BitNet
 
 Written by VTSTech
 https://www.vts-tech.org
@@ -184,7 +184,7 @@ __all__ = [
     "create_file_write_action", "create_file_delete_action",
     "create_mkdir_action", "create_shell_action",
     "format_status", "format_progress",
-    # R02: Model Family Configuration
+    # R02.3: Model Family Configuration
     "ModelFamilyConfig", "FAMILY_CONFIGS",
     "get_family_config", "get_stop_tokens", "supports_tools",
     "get_tool_format", "get_preferred_temperature", "should_use_few_shot",

{agentnova-0.2.2 → agentnova-0.2.3}/agentnova/__main__.py

@@ -1,27 +1,27 @@
-#!/usr/bin/env python3
-"""
-⚛️ AgentNova R00 — Command Line Interface
-
-Entry point for: python -m agentnova [command] [options]
-
-Commands:
-  run     Run the agent on a single prompt and exit
-  chat    Interactive multi-turn conversation with memory
-  models  List models available in Ollama
-  tools   List available built-in tools
-  skills  List available Agent Skills
-
-Examples:
-  python -m agentnova run "What is the capital of France?"
-  python -m agentnova chat --model llama3.1:8b --tools calculator,shell
-  python -m agentnova models
-  python -m agentnova tools
-  python -m agentnova skills
-
-Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova
-"""
-
-from agentnova.cli import main
-
-if __name__ == "__main__":
-    main()
+#!/usr/bin/env python3
+"""
+⚛️ AgentNova R00 — Command Line Interface
+
+Entry point for: python -m agentnova [command] [options]
+
+Commands:
+  run     Run the agent on a single prompt and exit
+  chat    Interactive multi-turn conversation with memory
+  models  List models available in Ollama
+  tools   List available built-in tools
+  skills  List available Agent Skills
+
+Examples:
+  python -m agentnova run "What is the capital of France?"
+  python -m agentnova chat --model llama3.1:8b --tools calculator,shell
+  python -m agentnova models
+  python -m agentnova tools
+  python -m agentnova skills
+
+Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova
+"""
+
+from agentnova.cli import main
+
+if __name__ == "__main__":
+    main()

{agentnova-0.2.2 → agentnova-0.2.3}/agentnova/bitnet_client.py

@@ -1,151 +1,151 @@
-#!/usr/bin/env python3
-"""
-⚛️ AgentNova R00 — BitnetClient
-Drop-in replacement for OllamaClient that uses Microsoft's bitnet.cpp
-(llama-server) as the inference backend instead of Ollama.
-
-Architecture:
-  - Wraps bitnet.cpp's llama-server HTTP process
-  - Exposes the same interface as OllamaClient: chat(), list_models(),
-    model_supports_tools(), is_running()
-  - Normalises OpenAI-format responses → Ollama-format so agent.py is
-    completely unmodified
-  - Manages llama-server lifecycle (start/stop/health-check)
-  - Falls back gracefully to ReAct tool-calling (no native function
-    calling in current bitnet models)
-
-Supported models (as of bitnet.cpp 2025):
-  - microsoft/BitNet-b1.58-2B-4T   (~0.4 GB, recommended)
-  - 1bitLLM/bitnet_b1_58-3B        (~0.7 GB)
-  - HF1BitLLM/Llama3-8B-1.58-100B-tokens
-  - tiiuae/Falcon3-1B-Instruct-1.58bit
-  - tiiuae/Falcon3-3B-Instruct-1.58bit
-  - tiiuae/Falcon3-7B-Instruct-1.58bit
-"""
-
-import json
-import urllib.request
-from .config import BITNET_BASE_URL
-
-# Known BitNet model identifiers
-KNOWN_MODELS = [
-    "bitnet-b1.58-2b-4t",
-    "BitNet-b1.58-2B-4T",
-    "bitnet-b1.58-large",
-]
-
-class BitnetClient:
-    def __init__(self, base_url=None, timeout=120):
-        # Prioritize passed URL, fallback to config
-        self.base_url = base_url or BITNET_BASE_URL
-        self.timeout = timeout
-
-    def is_running(self) -> bool:
-        try:
-            with urllib.request.urlopen(f"{self.base_url}/health", timeout=2) as resp:
-                return resp.getcode() == 200
-        except:
-            return False
-
-    def list_models(self):
-        try:
-            with urllib.request.urlopen(f"{self.base_url}/v1/models", timeout=5) as resp:
-                if resp.getcode() == 200:
-                    data = json.loads(resp.read().decode("utf-8"))
-                    # Return the ID of the model(s), cleaned up to show just model folder/filename
-                    models = []
-                    for m in data.get('data', []):
-                        model_id = m['id']
-                        # Strip common prefixes to show just model_dir/filename.gguf
-                        # e.g., /content/BitNet/models/BitNet-b1.58-2B-4T/bitnet_2b_i2_s.gguf
-                        #   -> BitNet-b1.58-2B-4T/bitnet_2b_i2_s.gguf
-                        if '/models/' in model_id:
-                            # Keep everything after /models/
-                            model_id = model_id.split('/models/')[-1]
-                        elif model_id.startswith('/'):
-                            # Just keep last two path components
-                            parts = model_id.strip('/').split('/')
-                            if len(parts) >= 2:
-                                model_id = '/'.join(parts[-2:])
-                        models.append(model_id)
-                    return models
-        except:
-            return []
-
-    def chat(
-        self,
-        model: str,
-        messages: list[dict],
-        options: dict | None = None,
-        stream: bool = False,
-        tools: list | None = None, # Signature fix for agent.py
-        **kwargs,                  # Catch-all for extra agent args
-    ):
-        """
-        Chat completion for BitNet.
-        
-        Note: BitNet llama-server doesn't support streaming in the same way
-        as Ollama. When stream=True, this yields tokens from a streaming
-        response. Otherwise returns a complete response dict.
-        """
-        url = f"{self.base_url}/v1/chat/completions"
-        
-        data = {
-            "model": model,
-            "messages": messages,
-            "stream": stream,
-            "temperature": (options or {}).get("temperature", 0.7),
-        }
-
-        req = urllib.request.Request(
-            url, 
-            data=json.dumps(data).encode("utf-8"),
-            headers={"Content-Type": "application/json"}
-        )
-
-        if stream:
-            # Return a generator for streaming
-            return self._stream_response(req, model)
-        else:
-            # Non-streaming: return complete response
-            with urllib.request.urlopen(req, timeout=self.timeout) as resp:
-                result = json.loads(resp.read().decode("utf-8"))
-                # Normalize OpenAI completion to Ollama format for the Agent
-                return {
-                    "model": model,
-                    "message": {
-                        "role": "assistant",
-                        "content": result["choices"][0]["message"]["content"]
-                    },
-                    "done": True
-                }
-
-    def _stream_response(self, req, model):
-        """
-        Handle SSE streaming response from BitNet server.
-        Yields tokens one at a time.
-        """
-        with urllib.request.urlopen(req, timeout=self.timeout) as resp:
-            buffer = ""
-            for line in resp:
-                line = line.decode("utf-8")
-                if line.startswith("data: "):
-                    data_str = line[6:].strip()
-                    if data_str == "[DONE]":
-                        break
-                    try:
-                        chunk = json.loads(data_str)
-                        if "choices" in chunk and len(chunk["choices"]) > 0:
-                            delta = chunk["choices"][0].get("delta", {})
-                            content = delta.get("content", "")
-                            if content:
-                                yield content
-                    except json.JSONDecodeError:
-                        continue
-
-    def model_supports_tools(self, model: str) -> bool:
-        return False # BitNet models require ReAct fallback in Agent.py
-    
-    def supports_streaming(self) -> bool:
-        """Return True if this client supports streaming."""
+#!/usr/bin/env python3
+"""
+⚛️ AgentNova R00 — BitnetClient
+Drop-in replacement for OllamaClient that uses Microsoft's bitnet.cpp
+(llama-server) as the inference backend instead of Ollama.
+
+Architecture:
+  - Wraps bitnet.cpp's llama-server HTTP process
+  - Exposes the same interface as OllamaClient: chat(), list_models(),
+    model_supports_tools(), is_running()
+  - Normalises OpenAI-format responses → Ollama-format so agent.py is
+    completely unmodified
+  - Manages llama-server lifecycle (start/stop/health-check)
+  - Falls back gracefully to ReAct tool-calling (no native function
+    calling in current bitnet models)
+
+Supported models (as of bitnet.cpp 2025):
+  - microsoft/BitNet-b1.58-2B-4T   (~0.4 GB, recommended)
+  - 1bitLLM/bitnet_b1_58-3B        (~0.7 GB)
+  - HF1BitLLM/Llama3-8B-1.58-100B-tokens
+  - tiiuae/Falcon3-1B-Instruct-1.58bit
+  - tiiuae/Falcon3-3B-Instruct-1.58bit
+  - tiiuae/Falcon3-7B-Instruct-1.58bit
+"""
+
+import json
+import urllib.request
+from .config import BITNET_BASE_URL
+
+# Known BitNet model identifiers
+KNOWN_MODELS = [
+    "bitnet-b1.58-2b-4t",
+    "BitNet-b1.58-2B-4T",
+    "bitnet-b1.58-large",
+]
+
+class BitnetClient:
+    def __init__(self, base_url=None, timeout=120):
+        # Prioritize passed URL, fallback to config
+        self.base_url = base_url or BITNET_BASE_URL
+        self.timeout = timeout
+
+    def is_running(self) -> bool:
+        try:
+            with urllib.request.urlopen(f"{self.base_url}/health", timeout=2) as resp:
+                return resp.getcode() == 200
+        except:
+            return False
+
+    def list_models(self):
+        try:
+            with urllib.request.urlopen(f"{self.base_url}/v1/models", timeout=5) as resp:
+                if resp.getcode() == 200:
+                    data = json.loads(resp.read().decode("utf-8"))
+                    # Return the ID of the model(s), cleaned up to show just model folder/filename
+                    models = []
+                    for m in data.get('data', []):
+                        model_id = m['id']
+                        # Strip common prefixes to show just model_dir/filename.gguf
+                        # e.g., /content/BitNet/models/BitNet-b1.58-2B-4T/bitnet_2b_i2_s.gguf
+                        #   -> BitNet-b1.58-2B-4T/bitnet_2b_i2_s.gguf
+                        if '/models/' in model_id:
+                            # Keep everything after /models/
+                            model_id = model_id.split('/models/')[-1]
+                        elif model_id.startswith('/'):
+                            # Just keep last two path components
+                            parts = model_id.strip('/').split('/')
+                            if len(parts) >= 2:
+                                model_id = '/'.join(parts[-2:])
+                        models.append(model_id)
+                    return models
+        except:
+            return []
+
+    def chat(
+        self,
+        model: str,
+        messages: list[dict],
+        options: dict | None = None,
+        stream: bool = False,
+        tools: list | None = None, # Signature fix for agent.py
+        **kwargs,                  # Catch-all for extra agent args
+    ):
+        """
+        Chat completion for BitNet.
+        
+        Note: BitNet llama-server doesn't support streaming in the same way
+        as Ollama. When stream=True, this yields tokens from a streaming
+        response. Otherwise returns a complete response dict.
+        """
+        url = f"{self.base_url}/v1/chat/completions"
+        
+        data = {
+            "model": model,
+            "messages": messages,
+            "stream": stream,
+            "temperature": (options or {}).get("temperature", 0.7),
+        }
+
+        req = urllib.request.Request(
+            url, 
+            data=json.dumps(data).encode("utf-8"),
+            headers={"Content-Type": "application/json"}
+        )
+
+        if stream:
+            # Return a generator for streaming
+            return self._stream_response(req, model)
+        else:
+            # Non-streaming: return complete response
+            with urllib.request.urlopen(req, timeout=self.timeout) as resp:
+                result = json.loads(resp.read().decode("utf-8"))
+                # Normalize OpenAI completion to Ollama format for the Agent
+                return {
+                    "model": model,
+                    "message": {
+                        "role": "assistant",
+                        "content": result["choices"][0]["message"]["content"]
+                    },
+                    "done": True
+                }
+
+    def _stream_response(self, req, model):
+        """
+        Handle SSE streaming response from BitNet server.
+        Yields tokens one at a time.
+        """
+        with urllib.request.urlopen(req, timeout=self.timeout) as resp:
+            buffer = ""
+            for line in resp:
+                line = line.decode("utf-8")
+                if line.startswith("data: "):
+                    data_str = line[6:].strip()
+                    if data_str == "[DONE]":
+                        break
+                    try:
+                        chunk = json.loads(data_str)
+                        if "choices" in chunk and len(chunk["choices"]) > 0:
+                            delta = chunk["choices"][0].get("delta", {})
+                            content = delta.get("content", "")
+                            if content:
+                                yield content
+                    except json.JSONDecodeError:
+                        continue
+
+    def model_supports_tools(self, model: str) -> bool:
+        return False # BitNet models require ReAct fallback in Agent.py
+    
+    def supports_streaming(self) -> bool:
+        """Return True if this client supports streaming."""
         return True

{agentnova-0.2.2 → agentnova-0.2.3}/agentnova/cli.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-⚛️ AgentNova R02 — CLI
+⚛️ AgentNova R02.3 — CLI
 Entry point: agentnova <command> [options]
 
 Commands:
@@ -924,7 +924,7 @@ def cmd_models(args):
             print(red("✗  bitnet_client.py not found. Copy it into agentnova/."))
             sys.exit(1)
             
-        print(bold("\n⚛️ AgentNova R02 BitNet Model (Remote)"))
+        print(bold("\n⚛️ AgentNova R02.3 BitNet Model (Remote)"))
         
         # We use the client logic to see what is actually running at the URL
         try:
@@ -952,7 +952,7 @@ def cmd_models(args):
         print(yellow("No models found. Pull one with: ollama pull llama3.2:3b"))
         return
 
-    print(bold("\n⚛️ AgentNova R02 Models") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold("\n⚛️ AgentNova R02.3 Models") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     
     # Load tested models from storage
     tested_models = _load_tested_models()
@@ -1091,7 +1091,7 @@ def cmd_models(args):
 
 def cmd_tools(args):
     tools = BUILTIN_REGISTRY.all()
-    print(bold("\n⚛️ AgentNova R02 Tools") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold("\n⚛️ AgentNova R02.3 Tools") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     print(bold(f"{'Tool':<20} Description"))
     print(dim("─" * 70))
     for t in tools:
@@ -1116,7 +1116,7 @@ def cmd_modelfile(args):
         sys.exit(1)
     
     model = args.model
-    print(bold(f"\n⚛️ AgentNova R02 Modelfile") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold(f"\n⚛️ AgentNova R02.3 Modelfile") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     print()
     
     try:
@@ -1181,7 +1181,7 @@ def cmd_skills(args):
     loader = SkillLoader()
     skills = loader.list_skills()
     
-    print(bold("\n⚛️ AgentNova R02 Skills") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold("\n⚛️ AgentNova R02.3 Skills") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     
     if not skills:
         print(yellow("  No skills found."))
@@ -1266,7 +1266,7 @@ def cmd_run(args):
     if acp_plugin:
         bootstrap_result = acp_plugin.bootstrap(claim_primary=False)  # AgentNova is secondary
 
-    print(bold("⚛️ AgentNova R02") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold("⚛️ AgentNova R02.3") + dim(" · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     print(f"Prompt: {args.prompt}")
 
     # Log user message to ACP
@@ -1359,7 +1359,7 @@ def cmd_chat(args):
     parts.append("]")
     status = " ".join(parts)
     
-    print(bold(f"\n⚛️ AgentNova R02 chat") + dim(f"  {status} · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold(f"\n⚛️ AgentNova R02.3 chat") + dim(f"  {status} · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     print(dim("  Type 'exit', 'quit', or Ctrl+C to quit."))
     print(dim("  Type '/help' to see all available commands."))
     print(dim("  Type '/ollama' to manage Ollama models (works with remote Ollama)."))
@@ -1909,7 +1909,7 @@ def cmd_agent(args):
     parts.append("]")
     status = " ".join(parts)
 
-    print(bold(f"\n⚛️ AgentNova R02 agent mode") + dim(f"  {status} · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
+    print(bold(f"\n⚛️ AgentNova R02.3 agent mode") + dim(f"  {status} · Written by VTSTech · https://www.vts-tech.org · https://github.com/VTSTech/AgentNova"))
     print(dim("  Agent mode: Give tasks and the agent will work autonomously."))
     print(dim("  Type '/help' to see available commands."))
     print(dim("  ─────────────────────────────────────"))
@@ -2290,11 +2290,12 @@ def cmd_test(args):
         "13_acp": "Shutdown demo with ACP tracking",
         "14": "GSM8K benchmark - 50 math questions",
         "14_acp": "GSM8K benchmark with ACP tracking",
+        "15": "Quick diagnostic - 5 questions, ~30s/model",
         "backend": "Backend demo - Ollama/BitNet switching",
     }
     
     # Quick test set (skips long-running benchmarks)
-    QUICK_TESTS = ["01", "02", "03", "04", "05", "10", "11"]
+    QUICK_TESTS = ["01", "02", "03", "04", "05", "10", "11", "15"]
     
     if args.list:
         print(bold("\n⚛️ AgentNova Test Examples"))
@@ -2456,7 +2457,7 @@ def build_parser() -> argparse.ArgumentParser:
 
     parser = argparse.ArgumentParser(
         prog="agentnova",
-        description="⚛️ AgentNova R02 - local agentic AI powered by Ollama",
+        description="⚛️ AgentNova R02.3 - local agentic AI powered by Ollama",
     )
 
     sub = parser.add_subparsers(dest="command", metavar="command")