* claim: #3798 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * fix(ai-chat): replay once with reasoning_effort none when tools are refused (#3798) Newer OpenAI models answer HTTP 400 to function tools on /v1/chat/completions unless reasoning_effort is "none". On that exact refusal the request is sent again once with it, and later requests of the chat carry it directly. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -75,7 +75,17 @@ impl ApiBackend {
|
||||
let stop = cancel.clone();
|
||||
let wake: Arc<dyn Fn() + Send + Sync> = Arc::from(ctx.wake);
|
||||
std::thread::spawn(move || {
|
||||
let mut w = Worker { cfg, mcp, next: 0, tools: Vec::new(), history: Vec::new(), events: ev_tx, wake, cancel: stop };
|
||||
let mut w = Worker {
|
||||
cfg,
|
||||
mcp,
|
||||
next: 0,
|
||||
tools: Vec::new(),
|
||||
history: Vec::new(),
|
||||
events: ev_tx,
|
||||
wake,
|
||||
cancel: stop,
|
||||
no_reasoning: Default::default(),
|
||||
};
|
||||
w.run(cmd_rx);
|
||||
});
|
||||
Ok(ApiBackend { cmds: Some(cmd_tx), events, cancel, mcp: handle })
|
||||
@@ -137,6 +147,8 @@ struct Worker {
|
||||
events: Sender<Event>,
|
||||
wake: Arc<dyn Fn() + Send + Sync>,
|
||||
cancel: Arc<AtomicBool>,
|
||||
/// Set once the endpoint refused tools together with a reasoning effort: later requests ask for none.
|
||||
no_reasoning: std::cell::Cell<bool>,
|
||||
}
|
||||
|
||||
impl Worker {
|
||||
@@ -322,7 +334,28 @@ impl Worker {
|
||||
let mut body = json!({"model": self.cfg.model, "messages": self.history});
|
||||
if !self.tools.is_empty() {
|
||||
body["tools"] = Value::Array(self.tools.clone());
|
||||
if self.no_reasoning.get() {
|
||||
body["reasoning_effort"] = json!("none");
|
||||
}
|
||||
}
|
||||
let (status, v) = self.post(&body)?;
|
||||
// Some OpenAI models refuse function tools on this endpoint unless the reasoning effort is
|
||||
// "none" (HTTP 400 naming `reasoning_effort`): ask again once with it, and keep it for the
|
||||
// next requests.
|
||||
let refused = status == 400 && !self.tools.is_empty() && !self.no_reasoning.get() && http_error(status, &v).contains("reasoning_effort");
|
||||
if refused {
|
||||
self.no_reasoning.set(true);
|
||||
body["reasoning_effort"] = json!("none");
|
||||
let (status, v) = self.post(&body)?;
|
||||
return if status == 200 { Ok(v) } else { Err(http_error(status, &v)) };
|
||||
}
|
||||
if status != 200 {
|
||||
return Err(http_error(status, &v));
|
||||
}
|
||||
Ok(v)
|
||||
}
|
||||
|
||||
fn post(&self, body: &Value) -> Result<(u16, Value), String> {
|
||||
let mut headers = vec![("Content-Type".to_string(), "application/json".to_string())];
|
||||
if let Some(k) = &self.cfg.key {
|
||||
headers.push(("Authorization".into(), format!("Bearer {k}")));
|
||||
@@ -351,11 +384,7 @@ impl Worker {
|
||||
Err(_) => return Err("the request ended without an answer".into()),
|
||||
}
|
||||
};
|
||||
let v: Value = serde_json::from_slice(&bytes).unwrap_or(Value::Null);
|
||||
if status != 200 {
|
||||
return Err(http_error(status, &v));
|
||||
}
|
||||
Ok(v)
|
||||
Ok((status, serde_json::from_slice(&bytes).unwrap_or(Value::Null)))
|
||||
}
|
||||
}
|
||||
|
||||
@@ -776,4 +805,63 @@ mod tests {
|
||||
let ctx = StartContext { cwd: std::env::temp_dir(), bridge: None, tools: None, wake: Box::new(|| {}) };
|
||||
assert!(ApiBackend::start(cfg, ctx).is_err());
|
||||
}
|
||||
|
||||
/// A worker with one function tool whose endpoint answers each POST with the next scripted
|
||||
/// (status, body); the requests it received are returned too.
|
||||
fn scripted_worker(answers: Vec<(u16, Value)>) -> (Worker, Arc<std::sync::Mutex<Vec<Value>>>) {
|
||||
let answers = std::sync::Mutex::new(answers.into_iter().rev().collect::<Vec<_>>());
|
||||
let seen = Arc::new(std::sync::Mutex::new(Vec::<Value>::new()));
|
||||
let log = seen.clone();
|
||||
let http: HttpFn = Arc::new(move |r: &HttpRequest| {
|
||||
log.lock().unwrap().push(serde_json::from_slice(r.body.as_deref().unwrap()).unwrap());
|
||||
let (status, v) = answers.lock().unwrap().pop().expect("no request beyond the script");
|
||||
Ok((status, v.to_string().into_bytes()))
|
||||
});
|
||||
let cfg = ApiConfig { name: "Test".into(), base_url: "http://localhost:1/v1".into(), model: "m".into(), key: Some("sk-secret".into()), http };
|
||||
let (events, _rx) = channel();
|
||||
let w = Worker {
|
||||
cfg,
|
||||
mcp: None,
|
||||
next: 0,
|
||||
tools: vec![json!({"type": "function", "function": {"name": "measure"}})],
|
||||
history: vec![json!({"role": "user", "content": "hi"})],
|
||||
events,
|
||||
wake: Arc::new(|| {}),
|
||||
cancel: Arc::new(AtomicBool::new(false)),
|
||||
no_reasoning: Default::default(),
|
||||
};
|
||||
(w, seen)
|
||||
}
|
||||
|
||||
fn refusal() -> (u16, Value) {
|
||||
(400, json!({"error": {"message": "Function tools with reasoning_effort are not supported for gpt-6.1-sol in /v1/chat/completions."}}))
|
||||
}
|
||||
|
||||
/// Newer OpenAI models answer 400 to function tools without `reasoning_effort: none`: the
|
||||
/// request is replayed once with it, and the next ones carry it at once.
|
||||
#[test]
|
||||
fn a_refusal_of_tools_with_reasoning_effort_is_replayed_once_with_none() {
|
||||
let ok = json!({"choices": [{"message": {"role": "assistant", "content": "Done."}}]});
|
||||
let (w, seen) = scripted_worker(vec![refusal(), (200, ok.clone()), (200, ok.clone())]);
|
||||
assert_eq!(w.request().unwrap(), ok);
|
||||
assert_eq!(w.request().unwrap(), ok);
|
||||
let sent = seen.lock().unwrap();
|
||||
assert_eq!(sent.len(), 3);
|
||||
assert!(sent[0].get("reasoning_effort").is_none());
|
||||
assert_eq!(sent[1]["reasoning_effort"], "none");
|
||||
assert_eq!(sent[2]["reasoning_effort"], "none");
|
||||
}
|
||||
|
||||
/// No loop: a second refusal is the error, after exactly two requests; the message shows the
|
||||
/// endpoint's words but never the key; an unrelated 400 is not replayed at all.
|
||||
#[test]
|
||||
fn the_replay_happens_once_and_only_for_that_refusal() {
|
||||
let (w, seen) = scripted_worker(vec![refusal(), refusal()]);
|
||||
let e = w.request().unwrap_err();
|
||||
assert!(e.starts_with("HTTP 400") && e.contains("reasoning_effort") && !e.contains("sk-secret"), "{e}");
|
||||
assert_eq!(seen.lock().unwrap().len(), 2);
|
||||
let (w, seen) = scripted_worker(vec![(400, json!({"error": {"message": "context length exceeded"}}))]);
|
||||
assert!(w.request().unwrap_err().contains("context length"));
|
||||
assert_eq!(seen.lock().unwrap().len(), 1);
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user