From 54d6f019819e8fe8315ae0406832cd4548cbfb7e Mon Sep 17 00:00:00 2001
From: ZhenYi <434836402@qq.com>
Date: Sun, 17 May 2026 17:32:29 +0800
Subject: [PATCH] fix(ai): cap max_output_tokens to 2048 hard limit

- Add resolve_room_max_tokens helper with hard cap at 2048
- Replace unwrap_or(4096) defaults across ai_service and process_ai
---
 libs/room/src/service/ai_service.rs | 16 +++++++++++++---
 libs/room/src/service/process_ai.rs | 20 ++++++++++++++++++--
 2 files changed, 31 insertions(+), 5 deletions(-)

diff --git a/libs/room/src/service/ai_service.rs b/libs/room/src/service/ai_service.rs
index 631b44b..77e589d 100644
--- a/libs/room/src/service/ai_service.rs
+++ b/libs/room/src/service/ai_service.rs
@@ -11,14 +11,24 @@ use sea_orm::{ColumnTrait, EntityTrait, QueryFilter};
 
 use crate::connection::RoomConnectionManager;
 use crate::error::RoomError;
-use crate::service::ai_streaming;
 use crate::service::ai_nonstreaming;
+use crate::service::ai_streaming;
 use crate::service::history;
 use crate::service::patterns::{mention_bracket_re, mention_tag_re};
 use agent::chat::{AiRequest, ChatService};
 use agent::react::ROOM_CONTEXT_PROMPT;
 use agent::tool::registry::ToolRegistry;
 
+const ROOM_DEFAULT_MAX_OUTPUT_TOKENS: i32 = 1024;
+const ROOM_MAX_OUTPUT_TOKENS_HARD_CAP: i32 = 2048;
+
+fn resolve_room_max_tokens(configured: Option<i64>) -> i32 {
+    configured
+        .and_then(|v| i32::try_from(v).ok())
+        .unwrap_or(ROOM_DEFAULT_MAX_OUTPUT_TOKENS)
+        .clamp(1, ROOM_MAX_OUTPUT_TOKENS_HARD_CAP)
+}
+
 /// Service responsible for AI message generation orchestration.
 /// Decides which execution path to use (streaming/nonstreaming, ReAct/chat)
 /// and dispatches accordingly.
@@ -228,7 +238,7 @@ impl RoomAiService {
             })
             .unwrap_or_else(|| "guest".into());
 
-        let max_tokens = ai_config.max_tokens.unwrap_or(4096) as i32;
+        let max_tokens = resolve_room_max_tokens(ai_config.max_tokens);
 
         let mut request = AiRequest {
             db: self.db.clone(),
@@ -304,7 +314,7 @@ impl RoomAiService {
 
                 // Send the billing error as a visible message in the room
                 let error_content = format!(
-                    "⚠️ Billing Error: Insufficient balance. Your project and personal account do not have enough funds to process this AI request. Please add credits to continue using AI features."
+                    "鈿狅笍 Billing Error: Insufficient balance. Your project and personal account do not have enough funds to process this AI request. Please add credits to continue using AI features."
                 );
                 let _ = super::ai_common::create_and_publish_ai_message(
                     &self.db,
diff --git a/libs/room/src/service/process_ai.rs b/libs/room/src/service/process_ai.rs
index 49a8c4b..45aad98 100644
--- a/libs/room/src/service/process_ai.rs
+++ b/libs/room/src/service/process_ai.rs
@@ -12,6 +12,16 @@ use agent::react::ROOM_CONTEXT_PROMPT;
 use agent::tool::registry::ToolRegistry;
 use models::projects::project_members;
 
+const ROOM_DEFAULT_MAX_OUTPUT_TOKENS: i32 = 1024;
+const ROOM_MAX_OUTPUT_TOKENS_HARD_CAP: i32 = 2048;
+
+fn resolve_room_max_tokens(configured: Option<i64>) -> i32 {
+    configured
+        .and_then(|v| i32::try_from(v).ok())
+        .unwrap_or(ROOM_DEFAULT_MAX_OUTPUT_TOKENS)
+        .clamp(1, ROOM_MAX_OUTPUT_TOKENS_HARD_CAP)
+}
+
 impl RoomService {
     pub async fn process_message_ai(
         &self,
@@ -127,7 +137,7 @@ impl RoomService {
             })
             .unwrap_or_else(|| "guest".into());
 
-        let max_tokens = ai_config.max_tokens.unwrap_or(4096) as i32;
+        let max_tokens = resolve_room_max_tokens(ai_config.max_tokens);
 
         let mut request = AiRequest {
             db: self.db.clone(),
@@ -164,7 +174,13 @@ impl RoomService {
             });
         request.history_cutoff_seq = cutoff_seq;
 
-        request.room_preamble = Some(build_room_preamble(&room, &project, &sender, &sender_role, &optimized_history));
+        request.room_preamble = Some(build_room_preamble(
+            &room,
+            &project,
+            &sender,
+            &sender_role,
+            &optimized_history,
+        ));
 
         let use_streaming = ai_config.stream;