Bug Description

When serving a gpt-oss model (e.g. ggml-org/gpt-oss-120b-GGUF) with --reasoning-format auto, the first request succeeds but subsequent multi-turn requests with tool calls fail with HTTP 500:

Jinja Exception: Cannot pass both content and thinking in an assistant message with tool calls!

Regression

Introduced in PR #16937 (commit 87c9efc3b) — "common : move gpt-oss reasoning processing to init params".

That PR moved the thinking field assignment from output serialization (common_chat_msgs_to_json_oaicompat) to input processing (common_chat_params_init_gpt_oss). The new code conditionally adds thinking from reasoning_content when tool calls are present, but does not erase content from the adjusted message.

Root Cause

In common/chat.cpp — common_chat_params_init_gpt_oss():

if (has_reasoning_content && has_tool_calls) {
    auto adjusted_message = msg;
    adjusted_message["thinking"] = msg.at("reasoning_content");
    // BUG: "content" is not erased — template forbids having both
    adjusted_messages.push_back(adjusted_message);
}

The gpt-oss Jinja template (models/templates/openai-gpt-oss-120b.jinja) explicitly checks that tool-call messages have either content or thinking, not both — they render to the same <|channel|>analysis slot. When the client sends back conversation history containing assistant messages with content, reasoning_content, and tool_calls, the adjusted message ends up with {content, thinking, tool_calls} and the template raises an error.

Steps to Reproduce

1. Start llama-server with a gpt-oss model (e.g. llama-server -hf ggml-org/gpt-oss-120b-GGUF --reasoning-format auto --jinja -fa on)
2. Send a chat completion request that triggers tool calls
3. Send a follow-up request including the full conversation history (with the assistant's content, reasoning_content, and tool_calls)
4. Server returns HTTP 500 with the Jinja exception

Fix

Add adjusted_message.erase("content") after setting thinking:

if (has_reasoning_content && has_tool_calls) {
    auto adjusted_message = msg;
    adjusted_message["thinking"] = msg.at("reasoning_content");
    adjusted_message.erase("content");  // template forbids both content and thinking with tool_calls
    adjusted_messages.push_back(adjusted_message);
}