/v1/chat/completions returns 400 for the OpenAI developer role #646

/v1/chat/completions passes message roles through unchanged, so a developer message (sent by Codex and other current OpenAI clients) reaches apply_chat_template. Qwen3.8-Flash-Next's template raises Unexpected message role., which comes back as 400 could not encode request: Unexpected message role.. /v1/responses already folds developer into the system message (server/responses_api.py:263). The same error was reported in #376 for another model; that issue is closed and main still fails.

from transformers import AutoTokenizer
t = AutoTokenizer.from_pretrained("nvidia/Qwen3.8-Flash-Next-NVFP4")
t.apply_chat_template([{"role": "developer", "content": "Answer in French."},
                       {"role": "user", "content": "hi"}], tokenize=False)
# TemplateError: Unexpected message role.   (role "system" renders)

Expected: developer is treated as system on /v1/chat/completions when the template has no developer role, as on /v1/responses.

Environment: main 07813248, transformers 5.15.1, Python 3.14.

Implementation: #391 maps the developer role. On my fork, two follow-ups keep templates that require a leading system message working:

• 6d1f0ad: merge developer messages into the leading system message after mapping them.
• beade07: keep developer ahead of system, and reject images when merging system messages.