Merge pull request #859 from dcSpark/nico/fix_errors_2

add openai reasoning support

shinkai-bin/shinkai-node/src/llm_provider/execution/job_execution_core.rs

@@ -1,4 +1,5 @@
 use crate::llm_provider::error::LLMProviderError;
+use crate::llm_provider::execution::chains::inference_chain_trait::InferenceChainResult;
 use crate::llm_provider::job_callback_manager::JobCallbackManager;
 use crate::llm_provider::job_manager::JobManager;
 use crate::llm_provider::llm_stopper::LLMStopper;
@@ -239,7 +240,7 @@ impl JobManager {
         let start = Instant::now();
 
         // Call the inference chain router to choose which chain to use, and call it
-        let inference_response = JobManager::inference_chain_router(
+        let (inference_response, inference_response_content) = match JobManager::inference_chain_router(
             db.clone(),
             llm_provider_found,
             full_job,
@@ -257,8 +258,21 @@ impl JobManager {
             // sqlite_logger.clone(),
             llm_stopper.clone(),
         )
-        .await?;
-        let inference_response_content = inference_response.response.clone();
+        .await
+        {
+            Ok(response) => (response.clone(), response.response),
+            Err(e) => {
+                let error_message = format!("{}", e);
+                // Create a minimal inference response with the error message
+                let error_response = InferenceChainResult {
+                    response: error_message.clone(),
+                    tps: None,
+                    answer_duration: None,
+                    tool_calls: None,
+                };
+                (error_response, error_message)
+            }
+        };
 
         let duration = start.elapsed();
         shinkai_log(

shinkai-bin/shinkai-node/src/llm_provider/execution/job_execution_helpers.rs

@@ -42,8 +42,8 @@ impl JobManager {
         let response = task_response?;
         shinkai_log(
             ShinkaiLogOption::JobExecution,
-            ShinkaiLogLevel::Debug,
-            format!("inference_llm_provider_markdown> response: {:?}", response).as_str(),
+            ShinkaiLogLevel::Info,
+            format!("inference_with_llm_provider> response: {:?}", response).as_str(),
         );
 
         response

shinkai-bin/shinkai-node/src/llm_provider/providers/openai.rs

@@ -6,7 +6,7 @@ use super::shared::openai_api::{openai_prepare_messages, MessageContent, OpenAIR
 use super::LLMService;
 use crate::llm_provider::execution::chains::inference_chain_trait::{FunctionCall, LLMInferenceResponse};
 use crate::llm_provider::llm_stopper::LLMStopper;
-use crate::managers::model_capabilities_manager::PromptResultEnum;
+use crate::managers::model_capabilities_manager::{ModelCapabilitiesManager, PromptResultEnum};
 use async_trait::async_trait;
 use futures::StreamExt;
 use reqwest::Client;
@@ -104,20 +104,32 @@ impl LLMService for OpenAI {
                     Err(e) => eprintln!("Failed to serialize tools_json: {:?}", e),
                 };
 
-                let mut payload = json!({
-                    "model": self.model_type,
-                    "messages": messages_json,
-                    "max_tokens": result.remaining_output_tokens,
-                    "stream": is_stream,
-                });
+                // Set up initial payload with appropriate token limit field based on model capabilities
+                let mut payload = if ModelCapabilitiesManager::has_reasoning_capabilities(&model) {
+                    json!({
+                        "model": self.model_type,
+                        "messages": messages_json,
+                        "max_completion_tokens": result.remaining_output_tokens,
+                        "stream": is_stream,
+                    })
+                } else {
+                    json!({
+                        "model": self.model_type,
+                        "messages": messages_json,
+                        "max_tokens": result.remaining_output_tokens,
+                        "stream": is_stream,
+                    })
+                };
 
                 // Conditionally add functions to the payload if tools_json is not empty
                 if !tools_json.is_empty() {
                     payload["functions"] = serde_json::Value::Array(tools_json.clone());
                 }
 
-                // Add options to payload
-                add_options_to_payload(&mut payload, config.as_ref());
+                // Only add options to payload for non-reasoning models
+                if !ModelCapabilitiesManager::has_reasoning_capabilities(&model) {
+                    add_options_to_payload(&mut payload, config.as_ref());
+                }
 
                 // Print payload as a pretty JSON string
                 match serde_json::to_string_pretty(&payload) {
@@ -248,6 +260,35 @@ pub async fn parse_openai_stream_chunk(
     inbox_name: Option<InboxName>,
     session_id: &str,
 ) -> Result<Option<String>, LLMProviderError> {
+    // If the buffer starts with '{', assume we might be receiving a JSON error.
+    if buffer.trim_start().starts_with('{') {
+        match serde_json::from_str::<JsonValue>(buffer) {
+            Ok(json_data) => {
+                // If it has an "error" field, record that and return immediately.
+                if let Some(error_obj) = json_data.get("error") {
+                    let code = error_obj
+                        .get("code")
+                        .and_then(|c| c.as_str())
+                        .unwrap_or("Unknown code")
+                        .to_string();
+                    let msg = error_obj
+                        .get("message")
+                        .and_then(|m| m.as_str())
+                        .unwrap_or("Unknown error");
+                    // Clear the buffer since we've consumed it
+                    buffer.clear();
+                    return Ok(Some(format!("{}: {}", code, msg)));
+                }
+                // Once parsed, clear the buffer since we've consumed it.
+                buffer.clear();
+            }
+            Err(_) => {
+                // It's not yet valid JSON (partial) - keep the buffer and wait for more data
+                return Ok(None);
+            }
+        }
+    }
+
     let mut error_message: Option<String> = None;
 
     loop {
@@ -571,7 +612,7 @@ pub async fn handle_streaming_response(
                 }
 
                 // Handle WebSocket updates for function calls
-                if let Some(ref manager) = ws_manager_trait {
+                if let Some(ref _manager) = ws_manager_trait {
                     if let Some(ref inbox_name) = inbox_name {
                         if let Some(last_function_call) = function_calls.last() {
                             send_tool_ws_update(&ws_manager_trait, Some(inbox_name.clone()), last_function_call)

shinkai-bin/shinkai-node/src/llm_provider/providers/shared/openai_api.rs

@@ -7,6 +7,7 @@ use serde::{Deserialize, Serialize};
 use serde_json::{self};
 use shinkai_message_primitives::schemas::llm_providers::serialized_llm_provider::LLMProviderInterface;
 use shinkai_message_primitives::schemas::prompts::Prompt;
+use shinkai_message_primitives::schemas::subprompts::{SubPrompt, SubPromptType};
 
 use super::shared_model_logic;
 
@@ -106,6 +107,17 @@ pub struct Usage {
 }
 
 pub fn openai_prepare_messages(model: &LLMProviderInterface, prompt: Prompt) -> Result<PromptResult, LLMProviderError> {
+    let mut prompt = prompt.clone();
+
+    // If this is a reasoning model, filter out system prompts before any processing
+    if ModelCapabilitiesManager::has_reasoning_capabilities(model) {
+        prompt.sub_prompts.retain(|sp| match sp {
+            SubPrompt::Content(SubPromptType::System, _, _) => false,
+            SubPrompt::Omni(SubPromptType::System, _, _, _) => false,
+            _ => true,
+        });
+    }
+
     let max_input_tokens = ModelCapabilitiesManager::get_max_input_tokens(model);
 
     // Generate the messages and filter out images
@@ -125,25 +137,35 @@ pub fn openai_prepare_messages(model: &LLMProviderInterface, prompt: Prompt) ->
         .into_iter()
         .partition(|message| message.role.is_some());
 
-
     // Convert both sets of messages to serde Value
     let messages_json = serde_json::to_value(messages_with_role)?;
 
     // Convert tools to serde Value with name transformation
     let tools_json = serde_json::to_value(
-        tools.clone().into_iter().map(|mut tool| {
-            if let Some(functions) = tool.functions.as_mut() {
-                for function in functions {
-                    // Replace any characters that aren't alphanumeric, underscore, or hyphen
-                    function.name = function.name
-                        .chars()
-                        .map(|c| if c.is_alphanumeric() || c == '_' || c == '-' { c } else { '_' })
-                        .collect::<String>()
-                        .to_lowercase();
+        tools
+            .clone()
+            .into_iter()
+            .map(|mut tool| {
+                if let Some(functions) = tool.functions.as_mut() {
+                    for function in functions {
+                        // Replace any characters that aren't alphanumeric, underscore, or hyphen
+                        function.name = function
+                            .name
+                            .chars()
+                            .map(|c| {
+                                if c.is_alphanumeric() || c == '_' || c == '-' {
+                                    c
+                                } else {
+                                    '_'
+                                }
+                            })
+                            .collect::<String>()
+                            .to_lowercase();
+                    }
                 }
-            }
-            tool
-        }).collect::<Vec<_>>()
+                tool
+            })
+            .collect::<Vec<_>>(),
     )?;
 
     // Convert messages_json and tools_json to Vec<serde_json::Value>
@@ -211,7 +233,10 @@ pub fn openai_prepare_messages(model: &LLMProviderInterface, prompt: Prompt) ->
     })
 }
 
-pub fn openai_prepare_messages_gemini(model: &LLMProviderInterface, prompt: Prompt) -> Result<PromptResult, LLMProviderError> {
+pub fn openai_prepare_messages_gemini(
+    model: &LLMProviderInterface,
+    prompt: Prompt,
+) -> Result<PromptResult, LLMProviderError> {
     let max_input_tokens = ModelCapabilitiesManager::get_max_input_tokens(model);
 
     // Generate the messages and filter out images
@@ -525,7 +550,10 @@ mod tests {
         let response: OpenAIResponse = serde_json::from_str(response_text).expect("Failed to deserialize");
 
         // Verify basic response fields
-        assert_eq!(response.id.clone().unwrap(), "chatcmpl-0cae310a-2b36-470a-9261-0f24d77b01bc");
+        assert_eq!(
+            response.id.clone().unwrap(),
+            "chatcmpl-0cae310a-2b36-470a-9261-0f24d77b01bc"
+        );
         assert_eq!(response.object, "chat.completion");
         assert_eq!(response.created, 1736736692);
         assert_eq!(response.system_fingerprint, Some("fp_9cb648b966".to_string()));
@@ -540,10 +568,10 @@ mod tests {
         let message = &choice.message;
         assert_eq!(message.role, "assistant");
         assert!(message.content.is_none());
-        
+
         let tool_calls = message.tool_calls.as_ref().expect("Should have tool_calls");
         assert_eq!(tool_calls.len(), 1);
-        
+
         let tool_call = &tool_calls[0];
         assert_eq!(tool_call.id, "call_sa3n");
         assert_eq!(tool_call.call_type, "function");
@@ -563,4 +591,47 @@ mod tests {
         let groq = response.groq.expect("Should have Groq info");
         assert_eq!(groq.id, "req_01jhes5nvkedsb8hcw0x912fa6");
     }
+
+    #[test]
+    fn test_system_prompt_filtering() {
+        // Create a prompt with both Content and Omni system prompts
+        let sub_prompts = vec![
+            SubPrompt::Content(
+                SubPromptType::System,
+                "System prompt that should be filtered".to_string(),
+                98,
+            ),
+            SubPrompt::Content(SubPromptType::User, "User message that should remain".to_string(), 100),
+            SubPrompt::Omni(
+                SubPromptType::UserLastMessage,
+                "Last user message that should remain".to_string(),
+                vec![],
+                100,
+            ),
+        ];
+
+        let mut prompt = Prompt::new();
+        prompt.add_sub_prompts(sub_prompts);
+
+        // Create a mock model with reasoning capabilities
+        let model = SerializedLLMProvider::mock_provider_with_reasoning().model;
+
+        // Process the prompt
+        let result = openai_prepare_messages(&model, prompt).expect("Failed to prepare messages");
+
+        // Extract the messages from the result
+        let messages = match &result.messages {
+            PromptResultEnum::Value(value) => value.as_array().unwrap(),
+            _ => panic!("Expected Value variant"),
+        };
+
+        // Verify that only non-system messages remain
+        assert_eq!(messages.len(), 2, "Should only have 2 messages after filtering");
+
+        // Check that the remaining messages are the user messages
+        for message in messages {
+            let role = message["role"].as_str().unwrap();
+            assert_eq!(role, "user", "All remaining messages should be user messages");
+        }
+    }
 }

shinkai-bin/shinkai-node/src/managers/model_capabilities_manager.rs

@@ -1,23 +1,16 @@
 use crate::llm_provider::{
-    error::LLMProviderError,
-    providers::shared::{openai_api::openai_prepare_messages, shared_model_logic::llama_prepare_messages},
+    error::LLMProviderError, providers::shared::{openai_api::openai_prepare_messages, shared_model_logic::llama_prepare_messages}
 };
 use shinkai_message_primitives::{
     schemas::{
-        llm_message::LlmMessage,
-        llm_providers::{
-            common_agent_llm_provider::ProviderOrAgent,
-            serialized_llm_provider::{LLMProviderInterface, SerializedLLMProvider},
-        },
-        prompts::Prompt,
-        shinkai_name::ShinkaiName,
-    },
-    shinkai_utils::utils::count_tokens_from_message_llama3,
+        llm_message::LlmMessage, llm_providers::{
+            common_agent_llm_provider::ProviderOrAgent, serialized_llm_provider::{LLMProviderInterface, SerializedLLMProvider}
+        }, prompts::Prompt, shinkai_name::ShinkaiName
+    }, shinkai_utils::utils::count_tokens_from_message_llama3
 };
 use shinkai_sqlite::SqliteManager;
 use std::{
-    fmt,
-    sync::{Arc, Weak},
+    fmt, sync::{Arc, Weak}
 };
 
 #[derive(Debug)]
@@ -726,6 +719,21 @@ impl ModelCapabilitiesManager {
             _ => false,
         }
     }
+
+    /// Returns whether the given model has reasoning capabilities
+    pub fn has_reasoning_capabilities(model: &LLMProviderInterface) -> bool {
+        match model {
+            LLMProviderInterface::OpenAI(openai) => {
+                openai.model_type.starts_with("o1")
+                    || openai.model_type.starts_with("o2")
+                    || openai.model_type.starts_with("o3")
+                    || openai.model_type.starts_with("o4")
+                    || openai.model_type.starts_with("o5")
+            }
+            LLMProviderInterface::Ollama(ollama) => ollama.model_type.starts_with("deepseek-r1"),
+            _ => false,
+        }
+    }
 }
 
 // TODO: add a tokenizer library only in the dev env and test that the