Messages carrying both reasoning and reasoning_content are rejected with 400 #46

What happened

/v1/chat/completions rejects a message that carries both reasoning and reasoning_content, with

{"error": "Use one reasoning field per message"}

Either field on its own is accepted. This makes the server unusable for multi-turn clients that replay assistant reasoning on every carrier they know about — such a client works for the first turn and then fails on every turn afterwards.

Steps to reproduce

Server running (rec status → ready, 0.1.1, flint-int4).

Plain request works:

curl -s http://127.0.0.1:11500/v1/chat/completions \
  -H "Authorization: Bearer $(rec auth token)" -H 'Content-Type: application/json' \
  -d '{"model":"flint","messages":[{"role":"user","content":"hi"}],"max_tokens":8}'
→ 200, message.reasoning = "Hmm, the user just said ..."

reasoning only → 200. reasoning_content only → 200. Both → 400:

curl -s http://127.0.0.1:11500/v1/chat/completions \
  -H "Authorization: Bearer $(rec auth token)" -H 'Content-Type: application/json' \
  -d '{"model":"flint","max_tokens":8,"messages":[
        {"role":"user","content":"hi"},
        {"role":"assistant","content":"ok","reasoning":"t","reasoning_content":"t"}]}'
{"error": "Use one reasoning field per message"}

Why it matters

Local OpenAI-compatible servers disagree about the replay carrier: llama.cpp/SGLang read reasoning_content, vLLM/Ollama read reasoning. Harnesses therefore send both (Hermes Agent does exactly this for any localhost endpoint), and every other server tested simply ignores the key it does not read. A 400 here turns a benign extra key into a hard failure that only appears from turn 2 onwards, so it reads as "the model is broken" rather than "the request shape is rejected".

The message also does not name the offending field, so clients that recover from strict-schema 400s by dropping the named key have nothing to match on.

Expected

Per OpenAI-compat convention, accept the message and pick one field (e.g. prefer reasoning, ignore the alias) — or ignore unknown message keys. If you prefer to keep the check, naming the field (reasoning / reasoning_content) would at least let clients drop it and retry.

Environment

• rec 0.1.1, model flint (flint-int4), context 262144
• macOS 27.0.1, Apple M5 Max, 128 GiB
• Client: Hermes Agent 0.22.5 (chat_completions, replays both carriers to localhost endpoints)