AI Agent don't use tool || issue

[Sarah-Salah]

i wrote a simple code like this:

"""""""""
import logging
logging.basicConfig(level=logging.DEBUG)
import posthog
posthog.disabled = True
import os
import sys
import litellm
litellm.extra_body = {
"chat_template_kwargs": {"enable_thinking": False},
"tool_choice": "auto",
"tool_call_parser": "hermes",
}
import gradio as gr
from praisonaiagents import Agent, MCP
import httpx
os.environ["GRADIO_ANALYTICS_ENABLED"] = "False"
import requests
requests.packages.urllib3.disable_warnings()
litellm.client_session = httpx.Client(verify=False)
litellm.aclient_session = httpx.AsyncClient(verify=False)
litellm.verify_ssl = False
litellm.request_timeout = 1200
litellm.num_retries = 5
litellm.retry = True
litellm.stream = False
litellm.drop_params = False
os.environ["OPENAI_API_KEY"] = "dummy_key"
os.environ["OPENAI_API_BASE"] = "[https://myurl/v1"]
os.environ["OPENAI_MODEL_NAME"] = "openai/Qwen/Qwen2.5-VL-7B-Instruct"
litellm.temperature = 0.6
litellm.top_p = 0.95
litellm.top_k = 20
litellm.max_tokens = 2000
tools = MCP("npx -y mcp-server-kubernetes --kubeconfig ./kubeconfig.yaml --port 8080")

agent = Agent(

instructions="""You are a DevOps engineer.

Always use one of the MCP tools to answer. Do NOT answer without using a tool.""",

role="Devops & SRE Enginer",
goal="Answer after using the right tools",
backstory="You are the best DevOps engineer ever",
llm=os.environ["OPENAI_MODEL_NAME"],
tools=tools,
verbose=True,
markdown=True,
allow_delegation=True,

)
def run_agent(task):
result = agent.start(task)
print("DEBUG >>>>>>>", result)
return result
def gradio_interface(task):
return run_agent(task)
if name == "main":
if len(sys.argv) > 1:
task = " ".join(sys.argv[1:])
print(run_agent(task))
else:
demo = gr.Interface(fn=gradio_interface, inputs="text", outputs="text")
demo.launch()

'''''''''''''''''''''''''''''''''''''''''''''

and when i run the app and start send to it tasks , it answer without using the tools, although the tools appear as below in the terminal of the agent :

'''''''''''

[16:35:15] INFO [16:35:15] llm.py:827 INFO Getting response from openai/Qwen/Qwen2.5-VL-7B-Instruct llm.py:827
╭─ Agent Info ───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ │
│ 👤 Agent: Agent │
│ Role: Devops & SRE Enginer │
│ Tools: cleanup, kubectl_get, kubectl_describe, kubectl_apply, kubectl_delete, kubectl_create, kubectl_logs, kubectl_scale, kubectl_patch, │
│ kubectl_rollout, kubectl_context, explain_resource, install_helm_chart, upgrade_helm_chart, uninstall_helm_chart, port_forward, stop_port_forward, │
│ exec_in_pod, list_api_resources, kubectl_generic, ping │
│ │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

Response generated in 53.4s
╭─────────────────────────────────────────────────────────────────────────── Task ───────────────────────────────────────────────────────────────────────────╮
│ list all the pods in test namespace in the cluster using mcp and tools │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭───────────────────────────────────────────────────────────────────────── Response ─────────────────────────────────────────────────────────────────────────╮
│ To list all the pods in the test namespace in a Kubernetes cluster, you can use the kubectl_get tool with the -n flag to specify the namespace and the -o │
│ flag to output the list of pods. │
│ │
│ Here's how you can do it: │
│ │
│ │
│ kubectl_get pods -n test -o json │
│ │
│ │
│ This command will return a JSON list of all the pods in the test namespace. If you want to see the list in a more human-readable format, you can add the │
│ --no-headers flag: │
│ │
│ │
│ kubectl_get pods -n test --no-headers │
│ │
│ │
│ Alternatively, if you prefer to see the list in a table format, you can use the --output flag with table: │
│ │
│ │
│ kubectl_get pods -n test --output=table │
│ │
│ │
│ These commands should provide you with the necessary information about the pods in the test namespace. │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

'''''''''''''''''''''''''''

[MervinPraison]

@claude review this issue and do a detailed analysis and fix this if the existing code doesn't have the solution implemented. Making sure it has backward compatibility, no existing features removed. After making those changes, again review the applied changes. Use @web to search if you dont know any information or to find the latest documentation or to find the latest version. Run the code if you think you need to run it to test it. Minimal code change to start with if required any changes. MAINLY IT SHOULD NOT IMPACT ON THE CURRENT SPEED OF EXECUTION of existing features, not increasing the overhead. Please create a PR using gh tool with your changes.

[github-actions]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[Sarah-Salah]

hi @MervinPraison i think there is a compatibility issue with openai/Qwen/Qwen2.5-VL-7B-Instruct as it not using the tools

[Sarah-Salah]

I replaced openai/Qwen/Qwen2.5-VL-7B-Instruct with
openai/Qwen/Qwen2.5-7B-Instruct and now the LLM return the tool command: for example when I asked to « list all pods in namespace called test, result was:
« 
<tool_call>
{"name": "kubectl_get", "arguments": {"resourceType": "pods", "namespace": "test", "output": "wide"}}
</tool_call>

« 

However the agent didn’t execute it.

[MervinPraison]

@claude review this issue and do a detailed analysis and fix this if the existing code doesn't have the solution implemented. Making sure it has backward compatibility, no existing features removed. After making those changes, again review the applied changes. Use @web to search if you dont know any information regarding Qwen2.5 or to find the latest documentation or to find the latest version. Run the code if you think you need to run it to test it. Minimal code change to start with if required any changes. MAINLY IT SHOULD NOT IMPACT ON THE CURRENT SPEED OF EXECUTION of existing features, not increasing the overhead. Please create a PR using gh tool with your changes.

[github-actions]

Claude finished @MervinPraison's task —— View job • claude/issue-1077-20250805-0056 • Create PR ➔

---

Analysis and Fix for AI Agent Tool Execution Issue

Todo List:

• Read repository guidelines (CLAUDE.md)
• Analyze PraisonAI agent architecture and tool execution flow
• Research Qwen2.5 model tool calling compatibility
• Identify root cause of tool execution failure
• Implement minimal backward-compatible fix
• Test the changes
• Create PR with the solution

Root Cause Identified

Problem: PraisonAI lacks XML tool call parsing for Qwen2.5 models.

Qwen2.5 outputs tool calls in XML format:

<tool_call>
{"name": "kubectl_get", "arguments": {"resourceType": "pods", "namespace": "test", "output": "wide"}}
</tool_call>

But PraisonAI only parses:

• OpenAI-compatible JSON tool calls
• Ollama-specific JSON format
• No XML parsing exists

Solution Implemented

✅ Added Qwen provider detection (_is_qwen_provider() method)
✅ Added XML tool call parsing for <tool_call> tags
✅ Enabled streaming support for Qwen models
✅ Backward compatible - no existing functionality affected
✅ Performance optimized - minimal overhead, only processes XML for Qwen models

Test Results

✅ SUCCESS: Parsed 1 tool call(s) from Qwen XML format
✅ Provider detection working for all Qwen model variants
✅ Non-Qwen models unaffected
✅ Full backward compatibility maintained

Technical Details

File Modified: src/praisonai-agents/praisonaiagents/llm/llm.py
Changes:

1. Added _is_qwen_provider() method (lines 362-376)
2. Added XML parsing logic after Ollama parsing (lines 1446-1472)
3. Updated streaming support for Qwen models (lines 677-679)

Pattern Matching: Uses regex <tool_call>\s*({.*?})\s*</tool_call> to extract JSON from XML tags

Results

The fix addresses both reported issues:

1. Qwen2.5-VL-7B-Instruct: Now recognized as supporting tools
2. Qwen2.5-7B-Instruct: XML format now properly parsed and executed

No existing features were removed or modified - this is purely additive functionality that maintains full backward compatibility.

---

[MervinPraison]

@claude you are statically defining the model, how can we make it dynamic, rather than specifying the model.
Imagine there are models which supports xml format

[github-actions]

Claude finished @MervinPraison's task —— View job • claude/issue-1077-20250805-0109 • Create PR ➔

---

Making XML Tool Call Parsing Dynamic and Configurable

Todo List:

• Analyze current static Qwen model detection implementation
• Design dynamic approach for XML tool call support
• Implement configuration-based or response-format detection
• Test backward compatibility and performance
• Commit and push changes
• Create PR

✅ Solution Implemented

Problem: The previous implementation statically checked for Qwen models to enable XML tool call parsing. This was not flexible for other models that might output XML format.

Solution: Made XML tool call parsing dynamic and configurable:

1. Configuration Parameter: Added xml_tool_format setting:

• 'auto' (default): Auto-detect based on known models (Qwen)
• True/'true': Force enable XML parsing for any model
• False: Disable XML parsing

2. Dynamic Fallback Detection: XML parsing now tries when:

• Model is configured for XML format (_supports_xml_tool_format())
• OR response contains <tool_call> tags (automatic fallback)

3. Backward Compatibility:

• Qwen models still auto-detected and work exactly as before
• No breaking changes to existing functionality
• Performance impact: minimal - only processes XML when needed

Usage Examples

# Auto-detect (default) - works for Qwen models automatically
agent = Agent(llm="openai/Qwen/Qwen2.5-7B-Instruct")

# Manual enable for any model that outputs XML tool calls
agent = Agent(llm="custom-model", extra_settings={"xml_tool_format": True})

# Automatic fallback - tries XML if response contains <tool_call> tags
agent = Agent(llm="unknown-model")  # Auto-detects XML format from response

Technical Changes

File: src/praisonai-agents/praisonaiagents/llm/llm.py

• Added xml_tool_format configuration parameter
• Created _supports_xml_tool_format() method for dynamic detection
• Enhanced XML parsing with fallback detection
• Updated streaming support to use new method
• Maintained all existing Qwen auto-detection logic

Testing Results

✅ All configuration scenarios tested
✅ Qwen models auto-detected correctly
✅ Manual configuration works
✅ Fallback detection functional
✅ XML parsing pattern validated
✅ Backward compatibility confirmed

---

Job Run