Native tool calling not working (ctx.tools is always None)

[cizekmilan]

Description

When using mcs-driver-rest with models that support native tool/function calling, ctx.tools is always None. Instead, the driver falls back to embedding tools into system_message as text.

This happens even for models that clearly support tool calling.

---

Tested models

• gpt-4
• gpt-4.5
• gpt-5
• gpt-5.2
• local model (gpt-oss-120b via vLLM with --tool-call-parser openai)

---

Tested OpenAPI sources

• Custom FastAPI app (/openapi.json)
• Petstore example (https://petstore3.swagger.io/api/v3/openapi.json)
• Context7 (https://mcsd.io/context7.json)

---

Expected behavior

When passing a model that supports tool calling:

ctx = driver.get_driver_context(model_name=LLM_MODEL)

ctx.tools should contain tool definitions in OpenAI format so they can be passed to:

client.chat.completions.create(..., tools=ctx.tools)

---

Actual behavior

ctx.tools is always:

None

Tools are instead embedded into ctx.system_message as text instructions.

As a result:

• Native tool calling cannot be used
• The system falls back to JSON-in-text parsing

---

Minimal reproducible example

from openai import OpenAI
from mcs.driver.rest import RestDriver

LLM_API_KEY = "..."
LLM_MODEL = "gpt-5.2"

client = OpenAI(api_key=LLM_API_KEY)

driver = RestDriver(
    url="http://127.0.0.1:8000/openapi.json"
)

ctx = driver.get_driver_context(model_name=LLM_MODEL)

print("TOOLS:", ctx.tools)  # always None

messages = [
    {"role": "system", "content": "You are a helpful assistant. Use tools when appropriate."},
    #{"role": "system", "content": ctx.system_message},
    {"role": "user", "content": "Find customer Kralova and show tickets and payments."},
]

completion = client.chat.completions.create(
    model=LLM_MODEL,
    messages=messages,
    tools=ctx.tools,
)

print(completion.choices[0].message)

---

Notes

• The OpenAPI schema is valid and tools are correctly present in system_message
• The issue appears to be in model capability detection inside the driver
• It seems to rely on a hardcoded allowlist or model name matching
• Newer models (e.g. gpt-5.x) are not recognized as supporting tools

---

Suggested fix

• Do not rely on static model name detection

• Either:

  • always return tools when OpenAPI is valid
  • or allow forcing native tool mode (e.g. force_tools=True)
  • or detect capability dynamically

---

Impact

This prevents using native tool calling with modern models and forces fallback behavior, even when unnecessary.