fix(openai_compatible): handle None tool.description in function_call format and token counting

[v4mpir0ck]

Bug Description

Two crash points in OAICompatLargeLanguageModel when tool.description is None:

1. Line 601 — "description": tool.description passes None into the function_call payload
2. Line 1130 — _get_num_tokens_by_gpt2(tool.description) crashes with TypeError

TypeError: can only concatenate str (not "NoneType") to str

Root Cause

PromptMessageTool.description is typed as str but receives None in practice when using custom OpenAPI-based tool providers. The tool provider chain (ApiToolBundle.summary → ToolDescription.llm → PromptMessageTool.description) can produce None at several points.

Fix

Add or "" fallback to both lines — minimal, safe, consistent with the guard already used in the azure_openai plugin (line 472):

# Line 601
"description": tool.description or "",

# Line 1130  
num_tokens += self._get_num_tokens_by_gpt2(tool.description or "")

Impact

This bug blocks all agent invocations that use custom API tools with missing descriptions on any plugin that inherits from OAICompatLargeLanguageModel (OpenAI, Azure OpenAI, Ollama, etc).

How to Reproduce

1. Create an OpenAPI tool provider with endpoints that have summary but no description
2. Create an Agent using those tools
3. Invoke the agent → crash before any tool call

Related

• fix(azure_openai): handle None tool.description in _num_tokens_for_tools dify-official-plugins#3095 — same bug in azure_openai plugin's own _num_tokens_for_tools

[gemini-code-assist]

Code Review

This pull request updates the OpenAI-compatible LLM interface to ensure tool descriptions default to an empty string when null during function generation and token calculation. The review feedback identifies a redundant attribute check in the token calculation logic and suggests simplifying the code by removing the unnecessary conditional statement.

[gemini-code-assist]

Since description is a required field in the PromptMessageTool Pydantic model, the hasattr(tool, "description") check is redundant. An instance of PromptMessageTool will always have this attribute.

You can simplify this by removing the conditional check. This would also make the code more consistent with the _generate method (line 601), which accesses tool.description directly.

            num_tokens += self._get_num_tokens_by_gpt2(tool.description or "")