agentnova 0.0__tar.gz → 0.2__tar.gz

{agentnova-0.0 → agentnova-0.2}/LICENSE

@@ -1,21 +1,21 @@
-MIT License
-
-Copyright (c) 2026 Nigel Todman
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+MIT License
+
+Copyright (c) 2026 Nigel Todman
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.

{agentnova-0.0 → agentnova-0.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentnova
-Version: 0.0
+Version: 0.2
 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 R00
+# ⚛️ AgentNova R02
 
 A minimal, hackable agentic framework engineered to run **entirely locally** with [Ollama](https://ollama.com) or [BitNet](https://github.com/microsoft/BitNet).
 
@@ -42,12 +42,9 @@ 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)
 
-[![PyPI version fury.io](https://badge.fury.io/py/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![PyPI status](https://img.shields.io/pypi/status/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![GitHub commits](https://badgen.net/github/commits/VTSTech/AgentNova)](https://GitHub.com/VTSTech/AgentNova/commit/)
-
-
-
-[![PyPI download month](https://img.shields.io/pypi/dm/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![PyPI download week](https://img.shields.io/pypi/dw/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![PyPI download day](https://img.shields.io/pypi/dd/agentnova.svg)](https://pypi.python.org/pypi/agentnova/)
+[![PyPI version fury.io](https://badge.fury.io/py/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/) [![PyPI status](https://img.shields.io/pypi/status/agentnova.svg?style=plastic)](https://pypi.python.org/pypi/agentnova/) [![GitHub commits](https://badgen.net/github/commits/VTSTech/AgentNova?style=plastic)](https://GitHub.com/VTSTech/AgentNova/commit/)
 
+[![PyPI download month](https://img.shields.io/pypi/dm/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/) [![PyPI download week](https://img.shields.io/pypi/dw/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/) [![PyPI download day](https://img.shields.io/pypi/dd/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/)
 
 ---
 
@@ -204,15 +201,20 @@ agentnova models --tool_support
 
 ### Performance by Tool Support
 
-Recent GSM8K benchmark results (50 math questions):
+Recent test results with native tool synthesis:
 
-| Model | Params | Tool Support | Score |
-|-------|--------|--------------|-------|
-| `gemma3:270m` | 270M | none | **64%** |
-| `functiongemma:270m` | 270M | native | 36% |
-| `granite4:350m` | 350M | native | ~40% |
+| 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 |
 
-**Key insight**: Sub-500M models often perform better with `none` (pure reasoning) than with tools!
+**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
 
 ---
 
@@ -263,7 +265,7 @@ Output shows:
 - **Tool Support** - `✓ native`, `ReAct`, `○ none`, or `untested`
 
 ```
-⚛️ AgentNova R00 Models
+⚛️ AgentNova R02 Models
   Model                                      Family       Context    Tool Support
   ──────────────────────────────────────────────────────────────────────────────
   gemma3:270m                                gemma3       32K        ○ none

{agentnova-0.0 → agentnova-0.2}/README.md

@@ -1,4 +1,4 @@
-# ⚛️ AgentNova R00
+# ⚛️ AgentNova R02
 
 A minimal, hackable agentic framework engineered to run **entirely locally** with [Ollama](https://ollama.com) or [BitNet](https://github.com/microsoft/BitNet).
 
@@ -6,12 +6,9 @@ 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)
 
-[![PyPI version fury.io](https://badge.fury.io/py/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![PyPI status](https://img.shields.io/pypi/status/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![GitHub commits](https://badgen.net/github/commits/VTSTech/AgentNova)](https://GitHub.com/VTSTech/AgentNova/commit/)
-
-
-
-[![PyPI download month](https://img.shields.io/pypi/dm/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![PyPI download week](https://img.shields.io/pypi/dw/agentnova.svg)](https://pypi.python.org/pypi/agentnova/) [![PyPI download day](https://img.shields.io/pypi/dd/agentnova.svg)](https://pypi.python.org/pypi/agentnova/)
+[![PyPI version fury.io](https://badge.fury.io/py/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/) [![PyPI status](https://img.shields.io/pypi/status/agentnova.svg?style=plastic)](https://pypi.python.org/pypi/agentnova/) [![GitHub commits](https://badgen.net/github/commits/VTSTech/AgentNova?style=plastic)](https://GitHub.com/VTSTech/AgentNova/commit/)
 
+[![PyPI download month](https://img.shields.io/pypi/dm/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/) [![PyPI download week](https://img.shields.io/pypi/dw/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/) [![PyPI download day](https://img.shields.io/pypi/dd/agentnova.svg?style=plastic)](https://pypi.org/project/agentnova/)
 
 ---
 
@@ -168,15 +165,20 @@ agentnova models --tool_support
 
 ### Performance by Tool Support
 
-Recent GSM8K benchmark results (50 math questions):
+Recent test results with native tool synthesis:
 
-| Model | Params | Tool Support | Score |
-|-------|--------|--------------|-------|
-| `gemma3:270m` | 270M | none | **64%** |
-| `functiongemma:270m` | 270M | native | 36% |
-| `granite4:350m` | 350M | native | ~40% |
+| 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 |
 
-**Key insight**: Sub-500M models often perform better with `none` (pure reasoning) than with tools!
+**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
 
 ---
 
@@ -227,7 +229,7 @@ Output shows:
 - **Tool Support** - `✓ native`, `ReAct`, `○ none`, or `untested`
 
 ```
-⚛️ AgentNova R00 Models
+⚛️ AgentNova R02 Models
   Model                                      Family       Context    Tool Support
   ──────────────────────────────────────────────────────────────────────────────
   gemma3:270m                                gemma3       32K        ○ none

{agentnova-0.0 → agentnova-0.2}/agentnova/__init__.py

@@ -1,5 +1,5 @@
 """
-⚛️ AgentNova R00 - A minimal, hackable agentic framework for Ollama and BitNet
+⚛️ AgentNova R02 - A minimal, hackable agentic framework for Ollama and BitNet
 
 Written by VTSTech
 https://www.vts-tech.org
@@ -15,6 +15,13 @@ from .core.ollama_client import OllamaClient
 from .core.orchestrator import Orchestrator, AgentCard
 from .skills import SkillLoader, Skill, SkillRegistry
 from .acp_plugin import ACPPlugin, create_acp_agent
+from .core.model_family_config import (
+    ModelFamilyConfig, FAMILY_CONFIGS,
+    get_family_config, get_stop_tokens, supports_tools,
+    get_tool_format, get_preferred_temperature, should_use_few_shot,
+    get_few_shot_style, has_known_issues, get_react_system_suffix,
+    get_native_tool_hints,
+)
 
 # Config exports
 from .config import (
@@ -177,6 +184,12 @@ __all__ = [
     "create_file_write_action", "create_file_delete_action",
     "create_mkdir_action", "create_shell_action",
     "format_status", "format_progress",
+    # R02: Model Family Configuration
+    "ModelFamilyConfig", "FAMILY_CONFIGS",
+    "get_family_config", "get_stop_tokens", "supports_tools",
+    "get_tool_format", "get_preferred_temperature", "should_use_few_shot",
+    "get_few_shot_style", "has_known_issues", "get_react_system_suffix",
+    "get_native_tool_hints",
     # Config exports
     "OLLAMA_BASE_URL",
     "BITNET_BASE_URL", 
@@ -191,7 +204,7 @@ __all__ = [
 if _BITNET_AVAILABLE:
     __all__.extend(["BitnetClient", "KNOWN_MODELS"])
 
-__version__ = "0.0"
+__version__ = "0.2"
 __author__ = "VTSTech"
 __author_email__ = "contact@vts-tech.org"
 __url__ = "https://github.com/VTSTech/AgentNova"

{agentnova-0.0 → agentnova-0.2}/agentnova/__main__.py

@@ -1,27 +1,27 @@
-#!/usr/bin/env python3
-"""
-🦞 AgentNova R04 — 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-0.0 → agentnova-0.2}/agentnova/acp_plugin.py

@@ -1,5 +1,5 @@
 """
-🦞 AgentNova ACP Plugin - Bridge to Agent Control Panel
+⚛️ AgentNova ACP Plugin - Bridge to Agent Control Panel
 
 This plugin connects AgentNova agents to an ACP (Agent Control Panel) server,
 enabling real-time monitoring, token tracking, and STOP/Resume control.

{agentnova-0.0 → agentnova-0.2}/agentnova/agent_mode.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-⚛️ AgentNova R04 — Agent Mode
+⚛️ AgentNova R00 — Agent Mode
 
 A goal-driven execution mode where the agent autonomously works through tasks.
 Unlike chat mode (which is user-driven), agent mode:
@@ -29,6 +29,27 @@ from pathlib import Path
 from typing import Any, Callable, Optional
 
 
+# ANSI color helpers
+def dim(text: str) -> str:
+    """Return dimmed text using ANSI escape codes."""
+    return f"\033[2m{text}\033[0m"
+
+
+def green(text: str) -> str:
+    """Return green text using ANSI escape codes."""
+    return f"\033[32m{text}\033[0m"
+
+
+def yellow(text: str) -> str:
+    """Return yellow text using ANSI escape codes."""
+    return f"\033[33m{text}\033[0m"
+
+
+def cyan(text: str) -> str:
+    """Return cyan text using ANSI escape codes."""
+    return f"\033[36m{text}\033[0m"
+
+
 class AgentState(Enum):
     """Agent execution states."""
     IDLE = "idle"           # Ready for new tasks
@@ -622,6 +643,14 @@ Example: [{{"description": "Step 1"}}, {{"description": "Step 2"}}]"""
             # Track the result
             result_msg = run.final_answer
             
+            # Debug output: show step completion info
+            if self.verbose and run:
+                tool_calls = [s for s in run.steps if s.type == "tool_call"]
+                tool_names = list(set(s.tool_name for s in tool_calls if hasattr(s, 'tool_name')))
+                print(dim(f"    ⏱️ {len(run.steps)} steps, {len(tool_calls)} tool calls, {run.total_ms:.0f}ms"))
+                if tool_names:
+                    print(dim(f"    🔧 Tools used: {', '.join(tool_names)}"))
+            
             # Log step completion
             step.status = "done"
             step.completed_at = datetime.now().isoformat()

{agentnova-0.0 → agentnova-0.2}/agentnova/bitnet_client.py

@@ -1,151 +1,151 @@
-#!/usr/bin/env python3
-"""
-🦞 AgentNova R04 — 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