fix(inference): honour max_tokens / max_completion_tokens in local server

src/openhuman/inference/http/http_tests.rs

@@ -140,3 +140,39 @@ fn strip_temperature_suffix_only_removes_numeric_suffixes() {
     assert_eq!(strip_temperature_suffix("llama3.1:8b@1"), "llama3.1:8b");
     assert_eq!(strip_temperature_suffix("gpt@beta"), "gpt@beta");
 }
+
+/// Asserts that `ChatCompletionRequest` parses `max_completion_tokens` and respects
+/// fallback precedence over `max_tokens` (Refs #5498).
+#[test]
+fn test_chat_completion_request_deserializes_max_completion_tokens() {
+    use crate::openhuman::inference::http::types::ChatCompletionRequest;
+
+    let json_both = serde_json::json!({
+        "model": "gpt-5",
+        "messages": [{ "role": "user", "content": "hello" }],
+        "max_tokens": 100,
+        "max_completion_tokens": 200
+    });
+    let parsed: ChatCompletionRequest = serde_json::from_value(json_both).unwrap();
+    assert_eq!(parsed.max_tokens, Some(100));
+    assert_eq!(parsed.max_completion_tokens, Some(200));
+    assert_eq!(
+        parsed.max_completion_tokens.or(parsed.max_tokens),
+        Some(200)
+    );
+
+    let json_legacy = serde_json::json!({
+        "model": "gpt-4o",
+        "messages": [{ "role": "user", "content": "hello" }],
+        "max_tokens": 150
+    });
+    let parsed_legacy: ChatCompletionRequest = serde_json::from_value(json_legacy).unwrap();
+    assert_eq!(parsed_legacy.max_tokens, Some(150));
+    assert_eq!(parsed_legacy.max_completion_tokens, None);
+    assert_eq!(
+        parsed_legacy
+            .max_completion_tokens
+            .or(parsed_legacy.max_tokens),
+        Some(150)
+    );
+}

src/openhuman/inference/http/server.rs

@@ -145,9 +145,12 @@ async fn chat_completions_handler(
     let completion_id = format!("chatcmpl-{}", uuid::Uuid::new_v4());
     let created = chrono::Utc::now().timestamp();
     let model_name = req.model.clone();
-    let model_request = ModelRequest::new(messages)
+    let mut model_request = ModelRequest::new(messages)
         .with_model(model_id.clone())
         .with_temperature(temperature);
+    if let Some(tokens) = req.max_completion_tokens.or(req.max_tokens) {
+        model_request = model_request.with_max_tokens(tokens);
+    }
 
     if req.stream {
         let model_stream = match chat_model.stream(&(), model_request).await {

src/openhuman/inference/http/types.rs

@@ -4,16 +4,25 @@ use serde::{Deserialize, Serialize};
 
 // ── Chat Completions ──────────────────────────────────────────────────────────
 
+/// Request payload for OpenAI-compatible chat completions (`POST /v1/chat/completions`).
 #[derive(Debug, Deserialize)]
 pub struct ChatCompletionRequest {
+    /// Identifier of the model to query.
     pub model: String,
+    /// List of input messages in conversation order.
     pub messages: Vec<ChatCompletionMessage>,
+    /// Whether to stream back partial progress via SSE chunks.
     #[serde(default)]
     pub stream: bool,
+    /// Sampling temperature between 0.0 and 2.0.
     #[serde(default)]
     pub temperature: Option<f64>,
+    /// Legacy maximum tokens limit (superseded by `max_completion_tokens`).
     #[serde(default)]
     pub max_tokens: Option<u32>,
+    /// Maximum number of output tokens for newer OpenAI / reasoning models (#5498).
+    #[serde(default)]
+    pub max_completion_tokens: Option<u32>,
     /// Optional tool definitions (ignored if the provider doesn't support them).
     #[serde(default)]
     pub tools: Option<serde_json::Value>,