zclaw_openfang/crates/zclaw-runtime/src/driver/openai.rs

//! OpenAI-compatible driver implementation

use async_trait::async_trait;
use async_stream::stream;
use futures::{Stream, StreamExt};
use secrecy::{ExposeSecret, SecretString};
use reqwest::Client;
use serde::{Deserialize, Serialize};
use std::pin::Pin;
use zclaw_types::{Result, ZclawError};

use super::{CompletionRequest, CompletionResponse, ContentBlock, LlmDriver, StopReason};
use crate::stream::StreamChunk;

/// OpenAI-compatible driver
pub struct OpenAiDriver {
    client: Client,
    api_key: SecretString,
    base_url: String,
}

impl OpenAiDriver {
    pub fn new(api_key: SecretString) -> Self {
        Self {
            client: Client::builder()
                .user_agent(crate::USER_AGENT)
                .http1_only()
                .timeout(std::time::Duration::from_secs(120))  // 2 minute timeout
                .connect_timeout(std::time::Duration::from_secs(30))  // 30 second connect timeout
                .build()
                .unwrap_or_else(|_| Client::new()),
            api_key,
            base_url: "https://api.openai.com/v1".to_string(),
        }
    }

    pub fn with_base_url(api_key: SecretString, base_url: String) -> Self {
        Self {
            client: Client::builder()
                .user_agent(crate::USER_AGENT)
                .http1_only()
                .timeout(std::time::Duration::from_secs(120))  // 2 minute timeout
                .connect_timeout(std::time::Duration::from_secs(30))  // 30 second connect timeout
                .build()
                .unwrap_or_else(|_| Client::new()),
            api_key,
            base_url,
        }
    }
}

#[async_trait]
impl LlmDriver for OpenAiDriver {
    fn provider(&self) -> &str {
        "openai"
    }

    fn is_configured(&self) -> bool {
        !self.api_key.expose_secret().is_empty()
    }

    async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse> {
        let api_request = self.build_api_request(&request);

        // Debug: log the request details
        let url = format!("{}/chat/completions", self.base_url);
        let request_body = serde_json::to_string(&api_request).unwrap_or_default();
        tracing::debug!(target: "openai_driver", "Sending request to: {}", url);
        tracing::trace!(target: "openai_driver", "Request body: {}", request_body);

        let response = self.client
            .post(&url)
            .header("Authorization", format!("Bearer {}", self.api_key.expose_secret()))
            .header("Accept", "*/*")
            .json(&api_request)
            .send()
            .await
            .map_err(|e| ZclawError::LlmError(format!("HTTP request failed: {}", e)))?;

        if !response.status().is_success() {
            let status = response.status();
            let body = response.text().await.unwrap_or_default();
            tracing::warn!(target: "openai_driver", "API error {}: {}", status, body);
            return Err(ZclawError::LlmError(format!("API error {}: {}", status, body)));
        }

        tracing::debug!(target: "openai_driver", "Response status: {}", response.status());

        let api_response: OpenAiResponse = response
            .json()
            .await
            .map_err(|e| ZclawError::LlmError(format!("Failed to parse response: {}", e)))?;

        Ok(self.convert_response(api_response, request.model))
    }

    fn stream(
        &self,
        request: CompletionRequest,
    ) -> Pin<Box<dyn Stream<Item = Result<StreamChunk>> + Send + '_>> {
        // Check if we should use non-streaming mode for tool calls
        // Some providers don't support streaming with tools:
        // - Alibaba DashScope: "tools暂时无法与stream=True同时使用"
        // - Zhipu GLM: May have similar limitations
        let has_tools = !request.tools.is_empty();
        let needs_non_streaming = self.base_url.contains("dashscope") ||
                          self.base_url.contains("aliyuncs") ||
                          self.base_url.contains("bigmodel.cn");

        tracing::debug!(target: "openai_driver", "stream config: base_url={}, has_tools={}, needs_non_streaming={}",
                  self.base_url, has_tools, needs_non_streaming);

        if has_tools && needs_non_streaming {
            tracing::info!(target: "openai_driver", "Provider detected that may not support streaming with tools, using non-streaming mode. URL: {}", self.base_url);
            // Use non-streaming mode and convert to stream
            return self.stream_from_complete(request);
        }

        let mut stream_request = self.build_api_request(&request);
        stream_request.stream = true;

        // Debug: log the request details
        let url = format!("{}/chat/completions", self.base_url);
        let request_body = serde_json::to_string(&stream_request).unwrap_or_default();
        tracing::debug!("[OpenAiDriver:stream] Sending request to: {}", url);
        tracing::debug!("[OpenAiDriver:stream] Request body length: {} bytes", request_body.len());
        tracing::trace!("[OpenAiDriver:stream] Request body: {}", request_body);

        let base_url = self.base_url.clone();
        let api_key = self.api_key.expose_secret().to_string();

        Box::pin(stream! {
            tracing::debug!("[OpenAI:stream] POST to {}/chat/completions", base_url);
            tracing::debug!("[OpenAI:stream] Request model={}, stream={}", stream_request.model, stream_request.stream);
            let response = match self.client
                .post(format!("{}/chat/completions", base_url))
                .header("Authorization", format!("Bearer {}", api_key))
                .header("Content-Type", "application/json")
                .timeout(std::time::Duration::from_secs(120))  // 2 minute timeout
                .json(&stream_request)
                .send()
                .await
            {
                Ok(r) => {
                    tracing::debug!("[OpenAI:stream] Response status: {}, content-type: {:?}", r.status(), r.headers().get("content-type"));
                    r
                },
                Err(e) => {
                    tracing::debug!("[OpenAI:stream] HTTP request FAILED: {:?}", e);
                    yield Err(ZclawError::LlmError(format!("HTTP request failed: {}", e)));
                    return;
                }
            };

            if !response.status().is_success() {
                let status = response.status();
                let body = response.text().await.unwrap_or_default();
                tracing::debug!("[OpenAI:stream] API error {}: {}", status, &body[..body.len().min(500)]);
                yield Err(ZclawError::LlmError(format!("API error {}: {}", status, body)));
                return;
            }

            let mut byte_stream = response.bytes_stream();
            let mut accumulated_tool_calls: std::collections::HashMap<String, (String, String)> = std::collections::HashMap::new();
            let mut current_tool_id: Option<String> = None;
            let mut sse_event_count: usize = 0;
            let mut raw_bytes_total: usize = 0;

            while let Some(chunk_result) = byte_stream.next().await {
                let chunk = match chunk_result {
                    Ok(c) => c,
                    Err(e) => {
                        tracing::debug!("[OpenAI:stream] Byte stream error: {:?}", e);
                        yield Err(ZclawError::LlmError(format!("Stream error: {}", e)));
                        continue;
                    }
                };

                raw_bytes_total += chunk.len();
                let text = String::from_utf8_lossy(&chunk);
                // Log first 500 bytes of raw data for debugging SSE format
                if raw_bytes_total <= 600 {
                    tracing::debug!("[OpenAI:stream] RAW chunk ({} bytes): {:?}", text.len(), &text[..text.len().min(500)]);
                }
                for line in text.lines() {
                    let trimmed = line.trim();
                    if trimmed.is_empty() || trimmed.starts_with(':') {
                        continue; // Skip empty lines and SSE comments
                    }
                    // Handle both "data: " (standard) and "data:" (no space)
                    let data = if let Some(d) = trimmed.strip_prefix("data: ") {
                        Some(d)
                    } else if let Some(d) = trimmed.strip_prefix("data:") {
                        Some(d.trim_start())
                    } else {
                        None
                    };
                    if let Some(data) = data {
                        sse_event_count += 1;
                        if sse_event_count <= 3 || data == "[DONE]" {
                            tracing::debug!("[OpenAI:stream] SSE #{}: {}", sse_event_count, &data[..data.len().min(300)]);
                        }
                        if data == "[DONE]" {
                            tracing::debug!("[OpenAI:stream] Received [DONE], total SSE events: {}, raw bytes: {}", sse_event_count, raw_bytes_total);

                            // Emit ToolUseEnd for all accumulated tool calls (skip invalid ones with empty name)
                            for (id, (name, args)) in &accumulated_tool_calls {
                                // Skip tool calls with empty name - they are invalid
                                if name.is_empty() {
                                    tracing::warn!("[OpenAI] Skipping invalid tool call with empty name: id={}", id);
                                    continue;
                                }
                                tracing::debug!("[OpenAI] Emitting ToolUseEnd: id={}, name={}, args={}", id, name, args);
                                // Ensure parsed args is always a valid JSON object
                                let parsed_args: serde_json::Value = if args.is_empty() {
                                    serde_json::json!({})
                                } else {
                                    serde_json::from_str(args).unwrap_or_else(|e| {
                                        tracing::warn!("[OpenAI] Failed to parse tool args '{}': {}, using empty object", args, e);
                                        serde_json::json!({})
                                    })
                                };
                                yield Ok(StreamChunk::ToolUseEnd {
                                    id: id.clone(),
                                    input: parsed_args,
                                });
                            }

                            yield Ok(StreamChunk::Complete {
                                input_tokens: 0,
                                output_tokens: 0,
                                stop_reason: "end_turn".to_string(),
                            });
                            continue;
                        }

                        match serde_json::from_str::<OpenAiStreamResponse>(data) {
                            Ok(resp) => {
                                if let Some(choice) = resp.choices.first() {
                                    let delta = &choice.delta;

                                    // Handle text content
                                    if let Some(content) = &delta.content {
                                        if !content.is_empty() {
                                            tracing::debug!("[OpenAI:stream] TextDelta: {} chars", content.len());
                                            yield Ok(StreamChunk::TextDelta { delta: content.clone() });
                                        }
                                    }

                                    // Handle reasoning_content (Kimi, Qwen, DeepSeek, GLM thinking)
                                    if let Some(reasoning) = &delta.reasoning_content {
                                        if !reasoning.is_empty() {
                                            tracing::debug!("[OpenAI:stream] ThinkingDelta (reasoning_content): {} chars", reasoning.len());
                                            yield Ok(StreamChunk::ThinkingDelta { delta: reasoning.clone() });
                                        }
                                    }

                                    // Handle tool calls
                                    if let Some(tool_calls) = &delta.tool_calls {
                                        tracing::trace!("[OpenAI] Received tool_calls delta: {:?}", tool_calls);
                                        for tc in tool_calls {
                                            // Tool call start - has id and name
                                            if let Some(id) = &tc.id {
                                                // Get function name if available
                                                let name = tc.function.as_ref()
                                                    .and_then(|f| f.name.clone())
                                                    .unwrap_or_default();

                                                // Only emit ToolUseStart if we have a valid tool name
                                                if !name.is_empty() {
                                                    tracing::debug!("[OpenAI] ToolUseStart: id={}, name={}", id, name);
                                                    current_tool_id = Some(id.clone());
                                                    accumulated_tool_calls.insert(id.clone(), (name.clone(), String::new()));
                                                    yield Ok(StreamChunk::ToolUseStart {
                                                        id: id.clone(),
                                                        name,
                                                    });
                                                } else {
                                                    tracing::debug!("[OpenAI] Tool call with empty name, waiting for name delta: id={}", id);
                                                    // Still track the tool call but don't emit yet
                                                    current_tool_id = Some(id.clone());
                                                    accumulated_tool_calls.insert(id.clone(), (String::new(), String::new()));
                                                }
                                            }

                                            // Tool call delta - has arguments
                                            if let Some(function) = &tc.function {
                                                tracing::trace!("[OpenAI] Function delta: name={:?}, arguments={:?}", function.name, function.arguments);
                                                if let Some(args) = &function.arguments {
                                                    tracing::debug!("[OpenAI] ToolUseDelta: args={}", args);
                                                    // Try to find the tool by id or use current
                                                    let tool_id = tc.id.as_ref()
                                                        .or(current_tool_id.as_ref())
                                                        .cloned()
                                                        .unwrap_or_default();

                                                    yield Ok(StreamChunk::ToolUseDelta {
                                                        id: tool_id.clone(),
                                                        delta: args.clone(),
                                                    });

                                                    // Accumulate arguments
                                                    if let Some(entry) = accumulated_tool_calls.get_mut(&tool_id) {
                                                        tracing::debug!("[OpenAI] Accumulating args for tool {}: '{}' -> '{}'", tool_id, args, entry.1);
                                                        entry.1.push_str(args);
                                                    } else {
                                                        tracing::warn!("[OpenAI] No entry found for tool_id '{}' to accumulate args", tool_id);
                                                    }
                                                }
                                            }
                                        }
                                    }
                                }
                            }
                            Err(e) => {
                                tracing::warn!("[OpenAI] Failed to parse SSE: {}, data: {}", e, data);
                            }
                        }
                    }
                }
            }
            tracing::debug!("[OpenAI:stream] Byte stream ended. Total: {} SSE events, {} raw bytes", sse_event_count, raw_bytes_total);
        })
    }
}

impl OpenAiDriver {
    /// Check if this is a Coding Plan endpoint (requires coding context)
    fn is_coding_plan_endpoint(&self) -> bool {
        self.base_url.contains("coding.dashscope") ||
        self.base_url.contains("coding/paas") ||
        self.base_url.contains("api.kimi.com/coding")
    }

    fn build_api_request(&self, request: &CompletionRequest) -> OpenAiRequest {
        // For Coding Plan endpoints, auto-add a coding assistant system prompt if not provided
        let system_prompt = if request.system.is_none() && self.is_coding_plan_endpoint() {
            Some("你是一个专业的编程助手，可以帮助用户解决编程问题、写代码、调试等。".to_string())
        } else {
            request.system.clone()
        };

        // Build messages with tool result truncation to prevent payload overflow.
        // Most LLM APIs have a 2-4MB HTTP payload limit.
        const MAX_TOOL_RESULT_BYTES: usize = 32_768; // 32KB per tool result
        const MAX_PAYLOAD_BYTES: usize = 1_800_000; // 1.8MB (under 2MB API limit)

        let mut messages: Vec<OpenAiMessage> = Vec::new();
        let mut pending_tool_calls: Option<Vec<OpenAiToolCall>> = None;
        let mut pending_content: Option<String> = None;
        let mut pending_reasoning: Option<String> = None;

        let flush_pending = |tc: &mut Option<Vec<OpenAiToolCall>>,
                             c: &mut Option<String>,
                             r: &mut Option<String>,
                             out: &mut Vec<OpenAiMessage>| {
            let calls = tc.take();
            let content = c.take();
            let reasoning = r.take();

            if let Some(calls) = calls {
                if !calls.is_empty() {
                    // Merge assistant content + reasoning into the tool call message.
                    // IMPORTANT: Some APIs (Kimi, Qwen) require `content` to be non-empty
                    // even when tool_calls is set. Use a meaningful placeholder if content is empty.
                    let content_value = content.filter(|s| !s.trim().is_empty())
                        .unwrap_or_else(|| "正在调用工具...".to_string());
                    out.push(OpenAiMessage {
                        role: "assistant".to_string(),
                        content: Some(content_value),
                        reasoning_content: reasoning.filter(|s| !s.is_empty()),
                        tool_calls: Some(calls),
                        tool_call_id: None,
                    });
                    return;
                }
            }
            // No tool calls — emit a plain assistant message.
            // Ensure content is always Some() and non-empty to satisfy API requirements.
            if content.is_some() || reasoning.is_some() {
                let content_value = content.filter(|s| !s.trim().is_empty())
                    .unwrap_or_else(|| "正在思考...".to_string());
                out.push(OpenAiMessage {
                    role: "assistant".to_string(),
                    content: Some(content_value),
                    reasoning_content: reasoning.filter(|s| !s.is_empty()),
                    tool_calls: None,
                    tool_call_id: None,
                });
            }
        };

        for msg in &request.messages {
            match msg {
                zclaw_types::Message::User { content } => {
                    flush_pending(&mut pending_tool_calls, &mut pending_content, &mut pending_reasoning, &mut messages);
                    messages.push(OpenAiMessage {
                        role: "user".to_string(),
                        content: Some(content.clone()),
                        tool_calls: None,
                        tool_call_id: None,
                        reasoning_content: None,
                    });
                }
                zclaw_types::Message::Assistant { content, thinking } => {
                    flush_pending(&mut pending_tool_calls, &mut pending_content, &mut pending_reasoning, &mut messages);
                    // Don't push immediately — wait to see if next messages are ToolUse
                    pending_content = Some(content.clone());
                    pending_reasoning = thinking.clone();
                }
                zclaw_types::Message::System { content } => {
                    flush_pending(&mut pending_tool_calls, &mut pending_content, &mut pending_reasoning, &mut messages);
                    messages.push(OpenAiMessage {
                        role: "system".to_string(),
                        content: Some(content.clone()),
                        tool_calls: None,
                        tool_call_id: None,
                        reasoning_content: None,
                    });
                }
                zclaw_types::Message::ToolUse { id, tool, input } => {
                    // Accumulate tool calls — they'll be merged with the pending assistant message
                    let args = if input.is_null() {
                        "{}".to_string()
                    } else {
                        serde_json::to_string(input).unwrap_or_else(|_| "{}".to_string())
                    };
                    pending_tool_calls
                        .get_or_insert_with(Vec::new)
                        .push(OpenAiToolCall {
                            id: id.clone(),
                            r#type: "function".to_string(),
                            function: FunctionCall {
                                name: tool.to_string(),
                                arguments: args,
                            },
                        });
                }
                zclaw_types::Message::ToolResult { tool_call_id, output, is_error, .. } => {
                    flush_pending(&mut pending_tool_calls, &mut pending_content, &mut pending_reasoning, &mut messages);
                    let content_str = if *is_error {
                        format!("Error: {}", output)
                    } else {
                        output.to_string()
                    };
                    // Truncate oversized tool results to prevent payload overflow
                    let truncated = if content_str.len() > MAX_TOOL_RESULT_BYTES {
                        let mut s = String::from(&content_str[..MAX_TOOL_RESULT_BYTES]);
                        s.push_str("\n\n... [内容已截断，原文过大]");
                        s
                    } else {
                        content_str
                    };
                    messages.push(OpenAiMessage {
                        role: "tool".to_string(),
                        content: Some(truncated),
                        tool_calls: None,
                        tool_call_id: Some(tool_call_id.clone()),
                        reasoning_content: None,
                    });
                }
            }
        }
        // Flush any remaining accumulated assistant content and/or tool calls
        flush_pending(&mut pending_tool_calls, &mut pending_content, &mut pending_reasoning, &mut messages);

        // Add system prompt if provided
        let mut messages = messages;
        if let Some(system) = &system_prompt {
            messages.insert(0, OpenAiMessage {
                role: "system".to_string(),
                content: Some(system.clone()),
                tool_calls: None,
                tool_call_id: None,
                reasoning_content: None,
            });
        }

        let tools: Vec<OpenAiTool> = request.tools
            .iter()
            .map(|t| OpenAiTool {
                r#type: "function".to_string(),
                function: FunctionDef {
                    name: t.name.clone(),
                    description: t.description.clone(),
                    parameters: t.input_schema.clone(),
                },
            })
            .collect();

        let api_request = OpenAiRequest {
            model: request.model.clone(),  // Use model ID directly without any transformation
            messages,
            max_tokens: request.max_tokens,
            temperature: request.temperature,
            stop: if request.stop.is_empty() { None } else { Some(request.stop.clone()) },
            stream: request.stream,
            tools: if tools.is_empty() { None } else { Some(tools) },
            reasoning_effort: request.reasoning_effort.clone(),
        };

        // Pre-send payload size validation
        if let Ok(serialized) = serde_json::to_string(&api_request) {
            if serialized.len() > MAX_PAYLOAD_BYTES {
                tracing::warn!(
                    target: "openai_driver",
                    "Request payload too large: {} bytes (limit: {}), truncating messages",
                    serialized.len(),
                    MAX_PAYLOAD_BYTES
                );
                return Self::truncate_messages_to_fit(api_request, MAX_PAYLOAD_BYTES);
            }
            tracing::debug!(
                target: "openai_driver",
                "Request payload size: {} bytes (limit: {})",
                serialized.len(),
                MAX_PAYLOAD_BYTES
            );
        }

        api_request
    }

    /// Emergency truncation: drop oldest non-system messages until payload fits
    fn truncate_messages_to_fit(mut request: OpenAiRequest, _max_bytes: usize) -> OpenAiRequest {
        // Keep system message (if any) and last 4 non-system messages
        let has_system = request.messages.first()
            .map(|m| m.role == "system")
            .unwrap_or(false);

        let non_system: Vec<OpenAiMessage> = request.messages.into_iter()
            .filter(|m| m.role != "system")
            .collect();

        // Keep last N messages and truncate any remaining large tool results
        let keep_count = 4.min(non_system.len());
        let start = non_system.len() - keep_count;
        let kept: Vec<OpenAiMessage> = non_system.into_iter()
            .skip(start)
            .map(|mut msg| {
                // Additional per-message truncation for tool results
                if msg.role == "tool" {
                    if let Some(ref content) = msg.content {
                        if content.len() > 16_384 {
                            let mut s = String::from(&content[..16_384]);
                            s.push_str("\n\n... [上下文压缩截断]");
                            msg.content = Some(s);
                        }
                    }
                }
                msg
            })
            .collect();

        let mut messages = Vec::new();
        if has_system {
            messages.push(OpenAiMessage {
                role: "system".to_string(),
                content: Some("You are a helpful AI assistant. (注意：对话历史已被压缩以适应上下文大小限制)".to_string()),
                tool_calls: None,
                tool_call_id: None,
                reasoning_content: None,
            });
        }
        messages.extend(kept);

        request.messages = messages;
        request
    }

    fn convert_response(&self, api_response: OpenAiResponse, model: String) -> CompletionResponse {
        let choice = api_response.choices.first();

        tracing::debug!("[OpenAiDriver:convert_response] Processing response: {} choices, first choice: {:?}", api_response.choices.len(), choice.map(|c| format!("content={:?}, tool_calls={:?}, finish_reason={:?}", c.message.content, c.message.tool_calls.as_ref().map(|tc| tc.len()), c.finish_reason)));

        let (content, stop_reason) = match choice {
            Some(c) => {
                // Priority: tool_calls > non-empty content > empty content
                // This is important because some providers return empty content with tool_calls
                let has_tool_calls = c.message.tool_calls.as_ref().map(|tc| !tc.is_empty()).unwrap_or(false);
                let has_content = c.message.content.as_ref().map(|t| !t.is_empty()).unwrap_or(false);
                let has_reasoning = c.message.reasoning_content.as_ref().map(|t| !t.is_empty()).unwrap_or(false);

                let blocks = if has_tool_calls {
                    // Tool calls take priority — safe to unwrap after has_tool_calls check
                    let tool_calls = c.message.tool_calls.as_ref().cloned().unwrap_or_default();
                    tracing::debug!("[OpenAiDriver:convert_response] Using tool_calls: {} calls", tool_calls.len());
                    tool_calls.iter().map(|tc| ContentBlock::ToolUse {
                        id: tc.id.clone(),
                        name: tc.function.name.clone(),
                        input: serde_json::from_str(&tc.function.arguments).unwrap_or(serde_json::Value::Null),
                    }).collect()
                } else if has_content {
                    // Non-empty content — safe to unwrap after has_content check
                    let text = c.message.content.as_deref().unwrap_or("");
                    tracing::debug!("[OpenAiDriver:convert_response] Using text content: {} chars", text.len());
                    vec![ContentBlock::Text { text: text.to_string() }]
                } else if has_reasoning {
                    // Content empty but reasoning_content present (Kimi, Qwen, DeepSeek)
                    let reasoning = c.message.reasoning_content.as_deref().unwrap_or("");
                    tracing::debug!("[OpenAiDriver:convert_response] Using reasoning_content: {} chars", reasoning.len());
                    vec![ContentBlock::Text { text: reasoning.to_string() }]
                } else {
                    // No content or tool_calls
                    tracing::debug!("[OpenAiDriver:convert_response] No content or tool_calls, using empty text");
                    vec![ContentBlock::Text { text: String::new() }]
                };

                let stop = match c.finish_reason.as_deref() {
                    Some("stop") => StopReason::EndTurn,
                    Some("length") => StopReason::MaxTokens,
                    Some("tool_calls") => StopReason::ToolUse,
                    _ => StopReason::EndTurn,
                };

                (blocks, stop)
            }
            None => {
                tracing::debug!("[OpenAiDriver:convert_response] No choices in response");
                (vec![ContentBlock::Text { text: String::new() }], StopReason::EndTurn)
            }
        };

        let (input_tokens, output_tokens) = api_response.usage
            .map(|u| (u.prompt_tokens, u.completion_tokens))
            .unwrap_or((0, 0));

        CompletionResponse {
            content,
            model,
            input_tokens,
            output_tokens,
            stop_reason,
        }
    }

    /// Convert a non-streaming completion to a stream for providers that don't support streaming with tools
    fn stream_from_complete(&self, request: CompletionRequest) -> Pin<Box<dyn Stream<Item = Result<StreamChunk>> + Send + '_>> {
        // Build non-streaming request
        let mut complete_request = self.build_api_request(&request);
        complete_request.stream = false;

        // Capture values before entering the stream
        let base_url = self.base_url.clone();
        let api_key = self.api_key.expose_secret().to_string();
        let model = request.model.clone();

        tracing::debug!(target: "openai_driver", "stream_from_complete: Starting non-streaming request to: {}/chat/completions", base_url);

        Box::pin(stream! {
            let url = format!("{}/chat/completions", base_url);
            tracing::debug!(target: "openai_driver", "stream_from_complete: Sending non-streaming request to: {}", url);

            let response = match self.client
                .post(&url)
                .header("Authorization", format!("Bearer {}", api_key))
                .header("Content-Type", "application/json")
                .timeout(std::time::Duration::from_secs(120))
                .json(&complete_request)
                .send()
                .await
            {
                Ok(r) => r,
                Err(e) => {
                    yield Err(ZclawError::LlmError(format!("HTTP request failed: {}", e)));
                    return;
                }
            };

            if !response.status().is_success() {
                let status = response.status();
                let body = response.text().await.unwrap_or_default();
                yield Err(ZclawError::LlmError(format!("API error {}: {}", status, body)));
                return;
            }

            let api_response: OpenAiResponse = match response.json().await {
                Ok(r) => r,
                Err(e) => {
                    tracing::warn!(target: "openai_driver", "stream_from_complete: Failed to parse response: {}", e);
                    yield Err(ZclawError::LlmError(format!("Failed to parse response: {}", e)));
                    return;
                }
            };

            tracing::debug!(target: "openai_driver", "stream_from_complete: Got response with {} choices", api_response.choices.len());
            if let Some(choice) = api_response.choices.first() {
                tracing::debug!(target: "openai_driver", "stream_from_complete: First choice: content={:?}, tool_calls={:?}, finish_reason={:?}",
                    choice.message.content.as_ref().map(|c| {
                        if c.len() > 100 {
                            // 使用 floor_char_boundary 确保不在多字节字符中间截断
                            let end = c.floor_char_boundary(100);
                            &c[..end]
                        } else {
                            c.as_str()
                        }
                    }),
                    choice.message.tool_calls.as_ref().map(|tc| tc.len()),
                    choice.finish_reason);
            }

            // Convert response to stream chunks
            let completion = self.convert_response(api_response, model.clone());
            tracing::debug!(target: "openai_driver", "stream_from_complete: Converted to {} content blocks, stop_reason: {:?}", completion.content.len(), completion.stop_reason);

            // Emit content blocks as stream chunks
            for block in &completion.content {
                tracing::debug!(target: "openai_driver", "stream_from_complete: Emitting block: {:?}", block);
                match block {
                    ContentBlock::Text { text } => {
                        if !text.is_empty() {
                            tracing::debug!(target: "openai_driver", "stream_from_complete: Emitting TextDelta: {} chars", text.len());
                            yield Ok(StreamChunk::TextDelta { delta: text.clone() });
                        }
                    }
                    ContentBlock::Thinking { thinking } => {
                        yield Ok(StreamChunk::ThinkingDelta { delta: thinking.clone() });
                    }
                    ContentBlock::ToolUse { id, name, input } => {
                        tracing::debug!(target: "openai_driver", "stream_from_complete: Emitting ToolUse: id={}, name={}", id, name);
                        // Emit tool use start
                        yield Ok(StreamChunk::ToolUseStart {
                            id: id.clone(),
                            name: name.clone(),
                        });
                        // Emit tool use delta with arguments
                        if !input.is_null() {
                            let args_str = serde_json::to_string(input).unwrap_or_default();
                            yield Ok(StreamChunk::ToolUseDelta {
                                id: id.clone(),
                                delta: args_str,
                            });
                        }
                        // Emit tool use end
                        yield Ok(StreamChunk::ToolUseEnd {
                            id: id.clone(),
                            input: input.clone(),
                        });
                    }
                    ContentBlock::ToolResult { .. } => {
                        // ToolResult is only used in request messages, never in responses
                    }
                }
            }

            // Emit completion
            yield Ok(StreamChunk::Complete {
                input_tokens: completion.input_tokens,
                output_tokens: completion.output_tokens,
                stop_reason: match completion.stop_reason {
                    StopReason::EndTurn => "end_turn",
                    StopReason::MaxTokens => "max_tokens",
                    StopReason::ToolUse => "tool_use",
                    StopReason::StopSequence => "stop",
                    StopReason::Error => "error",
                }.to_string(),
            });
        })
    }
}

// OpenAI API types

#[derive(Serialize)]
struct OpenAiRequest {
    model: String,
    messages: Vec<OpenAiMessage>,
    #[serde(skip_serializing_if = "Option::is_none")]
    max_tokens: Option<u32>,
    #[serde(skip_serializing_if = "Option::is_none")]
    temperature: Option<f32>,
    #[serde(skip_serializing_if = "Option::is_none")]
    stop: Option<Vec<String>>,
    #[serde(default)]
    stream: bool,
    #[serde(skip_serializing_if = "Option::is_none")]
    tools: Option<Vec<OpenAiTool>>,
    #[serde(skip_serializing_if = "Option::is_none")]
    reasoning_effort: Option<String>,
}

#[derive(Serialize)]
struct OpenAiMessage {
    role: String,
    #[serde(skip_serializing_if = "Option::is_none")]
    content: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tool_calls: Option<Vec<OpenAiToolCall>>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tool_call_id: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
    reasoning_content: Option<String>,
}

#[derive(Serialize)]
struct OpenAiToolCall {
    id: String,
    r#type: String,
    function: FunctionCall,
}

impl Default for OpenAiToolCall {
    fn default() -> Self {
        Self {
            id: String::new(),
            r#type: "function".to_string(),
            function: FunctionCall {
                name: String::new(),
                arguments: String::new(),
            },
        }
    }
}

#[derive(Serialize)]
struct FunctionCall {
    name: String,
    arguments: String,
}

#[derive(Serialize)]
struct OpenAiTool {
    r#type: String,
    function: FunctionDef,
}

#[derive(Serialize)]
struct FunctionDef {
    name: String,
    description: String,
    parameters: serde_json::Value,
}

#[derive(Deserialize, Default)]
struct OpenAiResponse {
    #[serde(default)]
    choices: Vec<OpenAiChoice>,
    #[serde(default)]
    usage: Option<OpenAiUsage>,
}

#[derive(Deserialize, Default, Clone)]
struct OpenAiChoice {
    #[serde(default)]
    message: OpenAiResponseMessage,
    #[serde(default)]
    finish_reason: Option<String>,
}

#[derive(Deserialize, Default, Clone)]
struct OpenAiResponseMessage {
    #[serde(default)]
    content: Option<String>,
    #[serde(default)]
    reasoning_content: Option<String>,
    #[serde(default)]
    tool_calls: Option<Vec<OpenAiToolCallResponse>>,
}

#[derive(Deserialize, Default, Clone)]
struct OpenAiToolCallResponse {
    #[serde(default)]
    id: String,
    #[serde(default)]
    function: FunctionCallResponse,
}

#[derive(Deserialize, Default, Clone)]
struct FunctionCallResponse {
    #[serde(default)]
    name: String,
    #[serde(default)]
    arguments: String,
}

#[derive(Deserialize, Default)]
struct OpenAiUsage {
    #[serde(default)]
    prompt_tokens: u32,
    #[serde(default)]
    completion_tokens: u32,
}

// OpenAI Streaming types

#[derive(Debug, Deserialize)]
struct OpenAiStreamResponse {
    #[serde(default)]
    choices: Vec<OpenAiStreamChoice>,
}

#[derive(Debug, Deserialize)]
struct OpenAiStreamChoice {
    #[serde(default)]
    delta: OpenAiDelta,
    #[serde(default)]
    #[allow(dead_code)] // Used for deserialization, not accessed
    finish_reason: Option<String>,
}

#[derive(Debug, Deserialize, Default)]
struct OpenAiDelta {
    #[serde(default)]
    content: Option<String>,
    #[serde(default)]
    reasoning_content: Option<String>,
    #[serde(default)]
    tool_calls: Option<Vec<OpenAiToolCallDelta>>,
}

#[derive(Debug, Deserialize)]
struct OpenAiToolCallDelta {
    #[serde(default)]
    id: Option<String>,
    #[serde(default)]
    function: Option<OpenAiFunctionDelta>,
}

#[derive(Debug, Deserialize)]
struct OpenAiFunctionDelta {
    #[serde(default)]
    name: Option<String>,
    #[serde(default)]
    arguments: Option<String>,
}