gigi-cli 1.0.0

use anyhow::{Context, Result};
use async_trait::async_trait;
use reqwest::Client;
use serde_json;

use super::cloud::{build_openai_messages, build_openai_tools, parse_openai_response, OpenAIResponse};
use super::provider::ModelProvider;
use super::types::*;

// =============================================================================
// Local Model Provider
//
// Supports Ollama, LM Studio, llama.cpp server, or any custom endpoint
// that exposes an OpenAI-compatible chat completions API.
// =============================================================================

/// The type of local model runtime being used.
#[derive(Debug, Clone)]
pub enum LocalRuntime {
    Ollama,
    LmStudio,
    LlamaCpp,
    Custom,
}

impl LocalRuntime {
    /// The chat completions endpoint path for this runtime.
    fn completions_path(&self) -> &str {
        match self {
            // Ollama exposes OpenAI-compatible endpoint
            Self::Ollama => "/v1/chat/completions",
            // LM Studio uses standard OpenAI path
            Self::LmStudio => "/v1/chat/completions",
            // llama.cpp server uses /v1/chat/completions when in OAI mode
            Self::LlamaCpp => "/v1/chat/completions",
            // Custom — assume same path
            Self::Custom => "/v1/chat/completions",
        }
    }

    fn display_name(&self) -> &str {
        match self {
            Self::Ollama => "ollama",
            Self::LmStudio => "lm_studio",
            Self::LlamaCpp => "llama_cpp",
            Self::Custom => "custom_local",
        }
    }
}

pub struct LocalModelProvider {
    client: Client,
    base_url: String,
    model: String,
    runtime: LocalRuntime,
    supports_tool_calls: bool,
}

impl LocalModelProvider {
    /// Create a new local model provider.
    ///
    /// # Arguments
    /// * `base_url` - The base URL of the local server (e.g., "http://localhost:11434" for Ollama)
    /// * `model` - The model name (e.g., "llama3", "codellama")
    /// * `runtime` - Which local runtime is being used
    /// * `supports_tool_calls` - Whether this model/runtime supports OpenAI-style tool calls.
    ///   Most local models don't, so default to false. When false, the agent loop
    ///   will attempt to parse tool invocations from the model's text output.
    pub fn new(
        base_url: String,
        model: String,
        runtime: LocalRuntime,
        supports_tool_calls: bool,
    ) -> Self {
        Self {
            client: Client::new(),
            base_url,
            model,
            runtime,
            supports_tool_calls,
        }
    }

    /// Convenience constructor for Ollama with defaults.
    pub fn ollama(model: Option<String>) -> Self {
        Self::new(
            "http://localhost:11434".to_string(),
            model.unwrap_or_else(|| "llama3".to_string()),
            LocalRuntime::Ollama,
            false,
        )
    }

    /// Convenience constructor for LM Studio with defaults.
    pub fn lm_studio(model: Option<String>) -> Self {
        Self::new(
            "http://localhost:1234".to_string(),
            model.unwrap_or_else(|| "local-model".to_string()),
            LocalRuntime::LmStudio,
            false,
        )
    }

    /// Convenience constructor for llama.cpp server with defaults.
    pub fn llama_cpp(model: Option<String>) -> Self {
        Self::new(
            "http://localhost:8080".to_string(),
            model.unwrap_or_else(|| "local-model".to_string()),
            LocalRuntime::LlamaCpp,
            false,
        )
    }
}

#[async_trait]
impl ModelProvider for LocalModelProvider {
    async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse> {
        let url = format!("{}{}", self.base_url, self.runtime.completions_path());

        let messages = if self.supports_tool_calls {
            build_openai_messages(&request)
        } else {
            build_text_only_messages(&request)
        };

        let mut body = serde_json::json!({
            "model": self.model,
            "messages": messages,
            "max_tokens": request.max_tokens,
            "stream": false,
        });

        // Only send tools if the runtime actually supports them
        if self.supports_tool_calls {
            let tools = build_openai_tools(&request);
            if !tools.is_empty() {
                body["tools"] = serde_json::json!(tools);
            }
        }

        let response = self
            .client
            .post(&url)
            .header("Content-Type", "application/json")
            .json(&body)
            .send()
            .await
            .with_context(|| {
                format!(
                    "Failed to connect to local model at {}. Is {} running?",
                    self.base_url,
                    self.runtime.display_name()
                )
            })?;

        if !response.status().is_success() {
            let status = response.status();
            let error_text = response.text().await.unwrap_or_default();
            anyhow::bail!(
                "Local model error ({}) from {}: {}",
                status,
                self.runtime.display_name(),
                error_text
            );
        }

        let api_response: OpenAIResponse = response
            .json()
            .await
            .context("Failed to parse response from local model")?;

        parse_openai_response(api_response)
    }

    fn name(&self) -> &str {
        self.runtime.display_name()
    }

    fn model_id(&self) -> &str {
        &self.model
    }

    fn supports_tools(&self) -> bool {
        self.supports_tool_calls
    }
}

fn build_text_only_messages(request: &CompletionRequest) -> Vec<serde_json::Value> {
    let mut messages = Vec::new();

    // Add system message if not empty
    if !request.system.is_empty() {
        messages.push(serde_json::json!({
            "role": "system",
            "content": request.system,
        }));
    }

    for msg in &request.messages {
        let role = match msg.role {
            crate::query::types::Role::User => "user",
            crate::query::types::Role::Assistant => "assistant",
            crate::query::types::Role::System => "system",
        };

        let mut content = String::new();

        for block in &msg.content {
            match block {
                ContentBlock::Text { text } => {
                    content.push_str(text);
                }
                ContentBlock::ToolUse { id: _, name, input } => {
                    content.push_str(&format!(
                        "\n<tool_call>\n{{\n  \"name\": \"{}\",\n  \"input\": {}\n}}\n</tool_call>\n",
                        name, serde_json::to_string_pretty(input).unwrap_or_default()
                    ));
                }
                ContentBlock::ToolResult { tool_use_id: _, content: result, is_error: _ } => {
                    content.push_str(&format!(
                        "\n<tool_result>\n{}\n</tool_result>\n",
                        result
                    ));
                }
            }
        }

        messages.push(serde_json::json!({
            "role": role,
            "content": content,
        }));
    }

    messages
}