ChatCompletionsTransport.parse() drops tool_calls when content is non-empty (affects GLM-4.7, other OpenAI-compatible models)

[wgnrai]

Description

ChatCompletionsTransport.parse() in helpers/litellm_transport.py (line ~520) only extracts tool_calls from the response message when content is empty:

if not response_delta:
    tool_calls = _as_list(_get_value(message, "tool_calls"))
    response_delta = ChatCompletionsTransport.tool_calls_text(tool_calls)

Some OpenAI-compatible models (e.g., Zhipu GLM-4.7 via the Z.AI Coding Plan endpoint) return both content and tool_calls simultaneously. When this happens, the parser uses the text content and silently drops the tool calls, breaking the agentic loop.

The same pattern exists in ResponsesTransport.parse_response() (line ~1049).

Steps to Reproduce

1. Configure a model that returns both content + tool_calls (e.g., openai/glm-4.7 via api.z.ai/api/coding/paas/v4)
2. Send a request with tools and tool_choice: "auto"
3. Observe that the response contains both message.content (non-empty text) and message.tool_calls
4. ChatCompletionsTransport.parse() returns the text content and _output_items is missing

Expected Behavior

Tool calls should be preserved regardless of whether content is also present. The finish_reason: "tool_calls" field indicates the model intended a tool call.

Suggested Fix

Check for tool calls first, then fall back to content:

# ChatCompletionsTransport.parse() — line ~518
parsed = {"reasoning_delta": reasoning_delta, "response_delta": response_delta}
tool_calls = _as_list(_get_value(message, "tool_calls"))
tool_calls_text = ChatCompletionsTransport.tool_calls_text(tool_calls) if tool_calls else ""
if tool_calls_text:
    parsed["response_delta"] = tool_calls_text
    parsed["_output_items"] = ChatCompletionsTransport.output_items(tool_calls)
return parsed

Environment

• Agent Zero v2.5
• LiteLLM via litellm_provider: openai with custom api_base
• Z.AI Coding Plan endpoint (api.z.ai/api/coding/paas/v4)
• GLM-4.7 model
• a0_api_mode: chat

Additional Context

Tested all three models on the Z.AI Coding Plan:

• GLM-5.2: Returns empty content + tool_calls (standard format) ✅
• GLM-5-Turbo: Returns empty content + tool_calls (standard format) ✅
• GLM-4.7: Returns both content + tool_calls (non-standard) ❌ — breaks parser

The fix is backward-compatible — when no tool_calls are present, behavior is unchanged.

[wgnrai]

Superseded by PR #1779 which includes the fix.