[Frontend] Merge developer instructions into tool block for Harmony prompt rendering

[beom115]

Purpose

Current Harmony chat implementation renders the system/developer instructions and tool definitions in two separate <|start|>developer blocks. However, models that follow the OpenAI gpt-oss-20b/120b chat template (based on the chat_template.jinja) expect a single developer block containing both # Instructions and # Tools sections.

This PR modifies _make_request_with_harmony to:

Extract the initial system or developer message.

Merge it into the same developer block as the tool definitions under the # Instructions header.

This ensures compatibility with models that require a unified developer message block.

Test Plan

import json
from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")

tools = [{
    "type": "function",
    "function": {
        "name": "get_current_weather",
        "description": "Get weather information for a given location.",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {"type": "string", "description": "Name city and state"},
            },
            "required": ["location"],
        },
    }
}]

messages = [
    {"role": "developer", "content": "Please Call Tool."},
    {"role": "user", "content": "What's the weather like in New York?"}
]

response = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=messages,
    tools=tools,
)

Test Result

Before (Redundant blocks):

<|start|>system<|message|>...<|end|>
<|start|>developer<|message|># Tools\n\n...<|end|>
<|start|>developer<|message|>Please Call Tool.<|end|>
<|start|>user<|message|>...<|end|>

After (Merged block - Correct for gpt-oss):

<|start|>system<|message|>...<|end|>
<|start|>developer<|message|># Instructions\n\nPlease Call Tool.\n\n# Tools\n\n...<|end|>
<|start|>user<|message|>...<|end|>

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

The pull request correctly identifies the need to merge system/developer instructions into the tool block for Harmony-based models to ensure compatibility with specific chat templates. However, the current implementation has a critical logic flaw that leads to the loss of user-provided instructions when the VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS environment variable is enabled. Additionally, the code uses unsafe dictionary access on message objects which could lead to runtime errors if non-dictionary types are encountered.

[gemini-code-assist]

Accessing ["role"] directly is unsafe here. ChatCompletionMessageParam is a union that includes OpenAIHarmonyMessage (a class), and even for dict-based messages, it is safer to use .get("role") to avoid potential KeyError or TypeError. Given that harmony_utils.py explicitly handles non-dict inputs, this part should be equally robust to prevent crashes if objects are passed programmatically.

[gemini-code-assist]

This implementation causes data loss when VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS is enabled. The get_developer_message function in harmony_utils.py ignores the instructions argument if that environment variable is set. Since get_system_message was already called (at line 1946) without these instructions, the user-provided system/developer message content will be completely discarded from the final prompt. To fix this, the extraction logic should be moved before the get_system_message call so instructions can be passed to the appropriate block based on the environment configuration.