stakpak_api/local/

mod.rs

// use crate::local::hooks::file_scratchpad_context::{
//     FileScratchpadContextHook, FileScratchpadContextHookOptions,
// };
use crate::local::hooks::inline_scratchpad_context::{
    InlineScratchpadContextHook, InlineScratchpadContextHookOptions,
};
use crate::{AgentProvider, ApiStreamError, GetMyAccountResponse};
use crate::{ListRuleBook, models::*};
use async_trait::async_trait;
use futures_util::Stream;
use libsql::{Builder, Connection};
use reqwest::Error as ReqwestError;
use reqwest::header::HeaderMap;
use rmcp::model::Content;
use stakpak_shared::hooks::{HookContext, HookRegistry, LifecycleEvent};
use stakpak_shared::models::integrations::anthropic::{AnthropicConfig, AnthropicModel};
use stakpak_shared::models::integrations::gemini::{GeminiConfig, GeminiModel};
use stakpak_shared::models::integrations::openai::{
    AgentModel, ChatCompletionChoice, ChatCompletionResponse, ChatCompletionStreamChoice,
    ChatCompletionStreamResponse, ChatMessage, FinishReason, MessageContent, OpenAIConfig,
    OpenAIModel, Role, Tool,
};
use stakpak_shared::models::integrations::search_service::*;
use stakpak_shared::models::llm::{
    GenerationDelta, LLMInput, LLMMessage, LLMMessageContent, LLMModel, LLMProviderConfig,
    LLMStreamInput, chat, chat_stream,
};
use stakpak_shared::tls_client::{TlsClientConfig, create_tls_client};
use std::pin::Pin;
use std::sync::Arc;
use tokio::sync::mpsc;
use uuid::Uuid;

mod context_managers;
mod db;
mod hooks;

#[cfg(test)]
mod tests;

#[derive(Clone, Debug)]
pub struct LocalClient {
    pub db: Connection,
    pub stakpak_base_url: Option<String>,
    pub anthropic_config: Option<AnthropicConfig>,
    pub openai_config: Option<OpenAIConfig>,
    pub gemini_config: Option<GeminiConfig>,
    pub model_options: ModelOptions,
    pub hook_registry: Option<Arc<HookRegistry<AgentState>>>,
    _search_services_orchestrator: Option<Arc<SearchServicesOrchestrator>>,
}

#[derive(Clone, Debug)]
pub struct ModelOptions {
    pub smart_model: Option<LLMModel>,
    pub eco_model: Option<LLMModel>,
    pub recovery_model: Option<LLMModel>,
}

#[derive(Clone, Debug)]
pub struct ModelSet {
    pub smart_model: LLMModel,
    pub eco_model: LLMModel,
    pub recovery_model: LLMModel,
    pub hook_registry: Option<Arc<HookRegistry<AgentState>>>,
    pub _search_services_orchestrator: Option<Arc<SearchServicesOrchestrator>>,
}

impl ModelSet {
    fn get_model(&self, agent_model: &AgentModel) -> LLMModel {
        match agent_model {
            AgentModel::Smart => self.smart_model.clone(),
            AgentModel::Eco => self.eco_model.clone(),
            AgentModel::Recovery => self.recovery_model.clone(),
        }
    }
}

impl From<ModelOptions> for ModelSet {
    fn from(value: ModelOptions) -> Self {
        let smart_model = value
            .smart_model
            .unwrap_or(LLMModel::Anthropic(AnthropicModel::Claude45Sonnet));
        let eco_model = value
            .eco_model
            .unwrap_or(LLMModel::Anthropic(AnthropicModel::Claude45Haiku));
        let recovery_model = value
            .recovery_model
            .unwrap_or(LLMModel::OpenAI(OpenAIModel::GPT5));

        Self {
            smart_model,
            eco_model,
            recovery_model,
            hook_registry: None,
            _search_services_orchestrator: None,
        }
    }
}

pub struct LocalClientConfig {
    pub stakpak_base_url: Option<String>,
    pub store_path: Option<String>,
    pub anthropic_config: Option<AnthropicConfig>,
    pub openai_config: Option<OpenAIConfig>,
    pub gemini_config: Option<GeminiConfig>,
    pub smart_model: Option<String>,
    pub eco_model: Option<String>,
    pub recovery_model: Option<String>,
    pub hook_registry: Option<HookRegistry<AgentState>>,
}

#[derive(Debug)]
enum StreamMessage {
    Delta(GenerationDelta),
    Ctx(Box<HookContext<AgentState>>),
}

const DEFAULT_STORE_PATH: &str = ".stakpak/data/local.db";
const TITLE_GENERATOR_PROMPT: &str = include_str!("./prompts/session_title_generator.v1.txt");

impl LocalClient {
    pub async fn new(config: LocalClientConfig) -> Result<Self, String> {
        let store_path = config
            .store_path
            .map(std::path::PathBuf::from)
            .unwrap_or_else(|| {
                std::env::home_dir()
                    .unwrap_or_default()
                    .join(DEFAULT_STORE_PATH)
            });

        if let Some(parent) = store_path.parent() {
            std::fs::create_dir_all(parent)
                .map_err(|e| format!("Failed to create database directory: {}", e))?;
        }

        let db = Builder::new_local(store_path.display().to_string())
            .build()
            .await
            .map_err(|e| e.to_string())?;

        let conn = db.connect().map_err(|e| e.to_string())?;

        // Initialize database schema
        db::init_schema(&conn).await?;

        let model_options = ModelOptions {
            smart_model: config.smart_model.map(LLMModel::from),
            eco_model: config.eco_model.map(LLMModel::from),
            recovery_model: config.recovery_model.map(LLMModel::from),
        };

        // Add hooks
        let mut hook_registry = config.hook_registry.unwrap_or_default();
        hook_registry.register(
            LifecycleEvent::BeforeInference,
            Box::new(InlineScratchpadContextHook::new(
                InlineScratchpadContextHookOptions {
                    model_options: model_options.clone(),
                    history_action_message_size_limit: Some(100),
                    history_action_message_keep_last_n: Some(1),
                    history_action_result_keep_last_n: Some(50),
                },
            )),
        );
        // hook_registry.register(
        //     LifecycleEvent::BeforeInference,
        //     Box::new(FileScratchpadContextHook::new(
        //         FileScratchpadContextHookOptions {
        //             history_action_message_size_limit: Some(100),
        //             history_action_message_keep_last_n: Some(1),
        //             history_action_result_keep_last_n: Some(50),
        //             scratchpad_path: None,
        //             todo_path: None,
        //             model_options: model_options.clone(),
        //             overwrite_if_different: Some(true),
        //         },
        //     )),
        // );

        Ok(Self {
            db: conn,
            stakpak_base_url: config.stakpak_base_url.map(|url| url + "/v1"),
            anthropic_config: config.anthropic_config,
            gemini_config: config.gemini_config,
            openai_config: config.openai_config,
            model_options,
            hook_registry: Some(Arc::new(hook_registry)),
            _search_services_orchestrator: Some(Arc::new(SearchServicesOrchestrator)),
        })
    }
}

#[async_trait]
impl AgentProvider for LocalClient {
    async fn get_my_account(&self) -> Result<GetMyAccountResponse, String> {
        Ok(GetMyAccountResponse {
            username: "local".to_string(),
            id: "local".to_string(),
            first_name: "local".to_string(),
            last_name: "local".to_string(),
        })
    }

    async fn list_rulebooks(&self) -> Result<Vec<ListRuleBook>, String> {
        if self.stakpak_base_url.is_none() {
            return Ok(vec![]);
        }

        let stakpak_base_url = self
            .stakpak_base_url
            .as_ref()
            .ok_or("Stakpak base URL not set")?;

        let url = format!("{}/rules", stakpak_base_url);

        let client = create_tls_client(
            TlsClientConfig::default().with_timeout(std::time::Duration::from_secs(300)),
        )?;

        let response = client
            .get(url)
            .send()
            .await
            .map_err(|e: ReqwestError| e.to_string())?;

        let value: serde_json::Value = response.json().await.map_err(|e| e.to_string())?;

        match serde_json::from_value::<ListRulebooksResponse>(value.clone()) {
            Ok(response) => Ok(response.results),
            Err(e) => {
                eprintln!("Failed to deserialize response: {}", e);
                eprintln!("Raw response: {}", value);
                Err("Failed to deserialize response:".into())
            }
        }
    }

    async fn get_rulebook_by_uri(&self, uri: &str) -> Result<RuleBook, String> {
        let stakpak_base_url = self
            .stakpak_base_url
            .as_ref()
            .ok_or("Stakpak base URL not set")?;

        let encoded_uri = urlencoding::encode(uri);

        let url = format!("{}/rules/{}", stakpak_base_url, encoded_uri);

        let client = create_tls_client(
            TlsClientConfig::default().with_timeout(std::time::Duration::from_secs(300)),
        )?;

        let response = client
            .get(&url)
            .send()
            .await
            .map_err(|e: ReqwestError| e.to_string())?;

        let value: serde_json::Value = response.json().await.map_err(|e| e.to_string())?;

        match serde_json::from_value::<RuleBook>(value.clone()) {
            Ok(response) => Ok(response),
            Err(e) => {
                eprintln!("Failed to deserialize response: {}", e);
                eprintln!("Raw response: {}", value);
                Err("Failed to deserialize response:".into())
            }
        }
    }

    async fn create_rulebook(
        &self,
        _uri: &str,
        _description: &str,
        _content: &str,
        _tags: Vec<String>,
        _visibility: Option<RuleBookVisibility>,
    ) -> Result<CreateRuleBookResponse, String> {
        // TODO: Implement create rulebook
        Err("Local provider does not support rulebooks yet".to_string())
    }

    async fn delete_rulebook(&self, _uri: &str) -> Result<(), String> {
        // TODO: Implement delete rulebook
        Err("Local provider does not support rulebooks yet".to_string())
    }

    async fn list_agent_sessions(&self) -> Result<Vec<AgentSession>, String> {
        db::list_sessions(&self.db).await
    }

    async fn get_agent_session(&self, session_id: Uuid) -> Result<AgentSession, String> {
        db::get_session(&self.db, session_id).await
    }

    async fn get_agent_session_stats(
        &self,
        _session_id: Uuid,
    ) -> Result<AgentSessionStats, String> {
        // TODO: Implement get agent session stats
        Ok(AgentSessionStats::default())
    }

    async fn get_agent_checkpoint(&self, checkpoint_id: Uuid) -> Result<RunAgentOutput, String> {
        db::get_checkpoint(&self.db, checkpoint_id).await
    }

    async fn get_agent_session_latest_checkpoint(
        &self,
        session_id: Uuid,
    ) -> Result<RunAgentOutput, String> {
        db::get_latest_checkpoint(&self.db, session_id).await
    }

    async fn chat_completion(
        &self,
        model: AgentModel,
        messages: Vec<ChatMessage>,
        tools: Option<Vec<Tool>>,
    ) -> Result<ChatCompletionResponse, String> {
        let mut ctx = HookContext::new(None, AgentState::new(model, messages, tools));

        if let Some(hook_registry) = &self.hook_registry {
            hook_registry
                .execute_hooks(&mut ctx, &LifecycleEvent::BeforeRequest)
                .await
                .map_err(|e| e.to_string())?
                .ok()?;
        }

        let current_checkpoint = self.initialize_session(&ctx.state.messages).await?;
        ctx.set_session_id(current_checkpoint.session.id);

        let new_message = self.run_agent_completion(&mut ctx, None).await?;
        ctx.state.append_new_message(new_message.clone());

        let result = self
            .update_session(&current_checkpoint, ctx.state.messages.clone())
            .await?;
        let checkpoint_created_at = result.checkpoint.created_at.timestamp() as u64;
        ctx.set_new_checkpoint_id(result.checkpoint.id);

        if let Some(hook_registry) = &self.hook_registry {
            hook_registry
                .execute_hooks(&mut ctx, &LifecycleEvent::AfterRequest)
                .await
                .map_err(|e| e.to_string())?
                .ok()?;
        }

        Ok(ChatCompletionResponse {
            id: ctx.new_checkpoint_id.unwrap().to_string(),
            object: "chat.completion".to_string(),
            created: checkpoint_created_at,
            model: ctx
                .state
                .llm_input
                .as_ref()
                .map(|llm_input| llm_input.model.clone().to_string())
                .unwrap_or_default(),
            choices: vec![ChatCompletionChoice {
                index: 0,
                message: ctx.state.messages.last().cloned().unwrap(),
                logprobs: None,
                finish_reason: FinishReason::Stop,
            }],
            usage: ctx
                .state
                .llm_output
                .as_ref()
                .map(|u| u.usage.clone())
                .unwrap_or_default(),
            system_fingerprint: None,
            metadata: None,
        })
    }

    async fn chat_completion_stream(
        &self,
        model: AgentModel,
        messages: Vec<ChatMessage>,
        tools: Option<Vec<Tool>>,
        _headers: Option<HeaderMap>,
    ) -> Result<
        (
            Pin<
                Box<dyn Stream<Item = Result<ChatCompletionStreamResponse, ApiStreamError>> + Send>,
            >,
            Option<String>,
        ),
        String,
    > {
        let mut ctx = HookContext::new(None, AgentState::new(model, messages, tools));

        if let Some(hook_registry) = &self.hook_registry {
            hook_registry
                .execute_hooks(&mut ctx, &LifecycleEvent::BeforeRequest)
                .await
                .map_err(|e| e.to_string())?
                .ok()?;
        }

        let current_checkpoint = self.initialize_session(&ctx.state.messages).await?;
        ctx.set_session_id(current_checkpoint.session.id);

        let (tx, mut rx) = mpsc::channel::<Result<StreamMessage, String>>(100);

        let _ = tx
            .send(Ok(StreamMessage::Delta(GenerationDelta::Content {
                content: format!(
                    "\n<checkpoint_id>{}</checkpoint_id>\n",
                    current_checkpoint.checkpoint.id
                ),
            })))
            .await;

        let client = self.clone();
        let self_clone = self.clone();
        let mut ctx_clone = ctx.clone();
        tokio::spawn(async move {
            let result = client
                .run_agent_completion(&mut ctx_clone, Some(tx.clone()))
                .await;

            match result {
                Err(e) => {
                    let _ = tx.send(Err(e)).await;
                }
                Ok(new_message) => {
                    ctx_clone.state.append_new_message(new_message.clone());
                    let _ = tx
                        .send(Ok(StreamMessage::Ctx(Box::new(ctx_clone.clone()))))
                        .await;

                    let output = self_clone
                        .update_session(&current_checkpoint, ctx_clone.state.messages.clone())
                        .await;

                    match output {
                        Err(e) => {
                            let _ = tx.send(Err(e)).await;
                        }
                        Ok(output) => {
                            ctx_clone.set_new_checkpoint_id(output.checkpoint.id);
                            let _ = tx.send(Ok(StreamMessage::Ctx(Box::new(ctx_clone)))).await;
                            let _ = tx
                                .send(Ok(StreamMessage::Delta(GenerationDelta::Content {
                                    content: format!(
                                        "\n<checkpoint_id>{}</checkpoint_id>\n",
                                        output.checkpoint.id
                                    ),
                                })))
                                .await;
                        }
                    }
                }
            }
        });

        let hook_registry = self.hook_registry.clone();
        let stream = async_stream::stream! {
            while let Some(delta_result) = rx.recv().await {
                match delta_result {
                    Ok(delta) => match delta {
                            StreamMessage::Ctx(updated_ctx) => {
                                ctx = *updated_ctx;
                            }
                            StreamMessage::Delta(delta) => {
                                yield Ok(ChatCompletionStreamResponse {
                                    id: ctx.request_id.to_string(),
                                    object: "chat.completion.chunk".to_string(),
                                    created: chrono::Utc::now().timestamp() as u64,
                                    model: ctx.state.llm_input.as_ref().map(|llm_input| llm_input.model.clone().to_string()).unwrap_or_default(),
                                    choices: vec![ChatCompletionStreamChoice {
                                        index: 0,
                                        delta: delta.into(),
                                        finish_reason: None,
                                    }],
                                    usage: ctx.state.llm_output.as_ref().map(|u| u.usage.clone()),
                                    metadata: None,
                                })
                            }
                        }
                    Err(e) => yield Err(ApiStreamError::Unknown(e)),
                }
            }

            if let Some(hook_registry) = hook_registry {
                hook_registry
                    .execute_hooks(&mut ctx, &LifecycleEvent::AfterRequest)
                    .await
                    .map_err(|e| e.to_string())?
                    .ok()?;
            }
        };

        Ok((Box::pin(stream), None))
    }

    async fn cancel_stream(&self, _request_id: String) -> Result<(), String> {
        Ok(())
    }

    async fn search_docs(&self, input: &SearchDocsRequest) -> Result<Vec<Content>, String> {
        let config = SearchServicesOrchestrator::start()
            .await
            .map_err(|e| e.to_string())?;

        // SECURITY TODO:
        // This uses plain-text, unauthenticated HTTP over localhost.
        // While acceptable for local development, this is an injection
        //
        // Mitigations to consider:
        // - Add mutual authentication (e.g., token)
        // - Validate the expected service identity

        let api_url = format!("http://localhost:{}", config.api_port);
        let search_client = SearchClient::new(api_url);

        let initial_query = if let Some(exclude) = &input.exclude_keywords {
            format!("{} -{}", input.keywords, exclude)
        } else {
            input.keywords.clone()
        };

        let llm_config = self.get_llm_config();
        let search_model = get_search_model(
            &llm_config,
            self.model_options.eco_model.clone(),
            self.model_options.smart_model.clone(),
        );

        let analysis = analyze_search_query(&llm_config, &search_model, &initial_query).await?;
        let required_documentation = analysis.required_documentation;
        let mut current_query = analysis.reformulated_query;
        let mut previous_queries = Vec::new();
        let mut final_valid_docs = Vec::new();
        let mut accumulated_needed_urls = Vec::new();

        const MAX_ITERATIONS: usize = 3;

        for _iteration in 0..MAX_ITERATIONS {
            previous_queries.push(current_query.clone());

            let search_results = search_client
                .search_and_scrape(current_query.clone(), None)
                .await
                .map_err(|e| e.to_string())?;

            if search_results.is_empty() {
                break;
            }

            let validation_result = validate_search_docs(
                &llm_config,
                &search_model,
                &search_results,
                &current_query,
                &required_documentation,
                &previous_queries,
                &accumulated_needed_urls,
            )
            .await?;

            for url in &validation_result.needed_urls {
                if !accumulated_needed_urls.contains(url) {
                    accumulated_needed_urls.push(url.clone());
                }
            }

            for doc in validation_result.valid_docs.into_iter() {
                let is_duplicate = final_valid_docs
                    .iter()
                    .any(|existing_doc: &ScrapedContent| existing_doc.url == doc.url);

                if !is_duplicate {
                    final_valid_docs.push(doc);
                }
            }

            if validation_result.is_satisfied {
                break;
            }

            if let Some(new_query) = validation_result.new_query {
                if new_query != current_query && !previous_queries.contains(&new_query) {
                    current_query = new_query;
                } else {
                    break;
                }
            } else {
                break;
            }
        }

        if final_valid_docs.is_empty() {
            return Ok(vec![Content::text("No results found".to_string())]);
        }

        let contents: Vec<Content> = final_valid_docs
            .into_iter()
            .map(|result| {
                Content::text(format!(
                    "Title: {}\nURL: {}\nContent: {}",
                    result.title.unwrap_or_default(),
                    result.url,
                    result.content.unwrap_or_default(),
                ))
            })
            .collect();

        Ok(contents)
    }

    async fn search_memory(&self, _input: &SearchMemoryRequest) -> Result<Vec<Content>, String> {
        // TODO: Implement search memory
        Ok(Vec::new())
    }

    async fn slack_read_messages(
        &self,
        _input: &SlackReadMessagesRequest,
    ) -> Result<Vec<Content>, String> {
        // TODO: Implement slack read messages
        Ok(Vec::new())
    }

    async fn slack_read_replies(
        &self,
        _input: &SlackReadRepliesRequest,
    ) -> Result<Vec<Content>, String> {
        // TODO: Implement slack read replies
        Ok(Vec::new())
    }

    async fn slack_send_message(
        &self,
        _input: &SlackSendMessageRequest,
    ) -> Result<Vec<Content>, String> {
        // TODO: Implement slack send message
        Ok(Vec::new())
    }

    async fn memorize_session(&self, _checkpoint_id: Uuid) -> Result<(), String> {
        // TODO: Implement memorize session
        Ok(())
    }
}

impl LocalClient {
    fn get_llm_config(&self) -> LLMProviderConfig {
        LLMProviderConfig {
            anthropic_config: self.anthropic_config.clone(),
            openai_config: self.openai_config.clone(),
            gemini_config: self.gemini_config.clone(),
        }
    }

    async fn run_agent_completion(
        &self,
        ctx: &mut HookContext<AgentState>,
        stream_channel_tx: Option<mpsc::Sender<Result<StreamMessage, String>>>,
    ) -> Result<ChatMessage, String> {
        if let Some(hook_registry) = &self.hook_registry {
            hook_registry
                .execute_hooks(ctx, &LifecycleEvent::BeforeInference)
                .await
                .map_err(|e| e.to_string())?
                .ok()?;
        }

        let input = if let Some(llm_input) = ctx.state.llm_input.clone() {
            llm_input
        } else {
            return Err(
                "Run agent completion: LLM input not found, make sure to register a context hook before inference"
                    .to_string(),
            );
        };

        let llm_config = self.get_llm_config();

        let (response_message, usage) = if let Some(tx) = stream_channel_tx {
            let (internal_tx, mut internal_rx) = mpsc::channel::<GenerationDelta>(100);
            let input = LLMStreamInput {
                model: input.model,
                messages: input.messages,
                max_tokens: input.max_tokens,
                tools: input.tools,
                stream_channel_tx: internal_tx,
            };

            let chat_future = async move {
                chat_stream(&llm_config, input)
                    .await
                    .map_err(|e| e.to_string())
            };

            let receive_future = async move {
                while let Some(delta) = internal_rx.recv().await {
                    if tx.send(Ok(StreamMessage::Delta(delta))).await.is_err() {
                        break;
                    }
                }
            };

            let (chat_result, _) = tokio::join!(chat_future, receive_future);
            let response = chat_result?;
            (response.choices[0].message.clone(), response.usage)
        } else {
            let response = chat(&llm_config, input).await.map_err(|e| e.to_string())?;
            (response.choices[0].message.clone(), response.usage)
        };

        ctx.state.set_llm_output(response_message, usage);

        if let Some(hook_registry) = &self.hook_registry {
            hook_registry
                .execute_hooks(ctx, &LifecycleEvent::AfterInference)
                .await
                .map_err(|e| e.to_string())?
                .ok()?;
        }

        let llm_output = ctx
            .state
            .llm_output
            .as_ref()
            .ok_or_else(|| "LLM output is missing from state".to_string())?;

        Ok(ChatMessage::from(llm_output))
    }

    async fn initialize_session(&self, messages: &[ChatMessage]) -> Result<RunAgentOutput, String> {
        // 1. Validate input
        if messages.is_empty() {
            return Err("At least one message is required".to_string());
        }

        // 2. Extract session/checkpoint ID or create new session
        let checkpoint_id = ChatMessage::last_server_message(messages).and_then(|message| {
            message
                .content
                .as_ref()
                .and_then(|content| content.extract_checkpoint_id())
        });

        let current_checkpoint = if let Some(checkpoint_id) = checkpoint_id {
            db::get_checkpoint(&self.db, checkpoint_id).await?
        } else {
            let title = self.generate_session_title(messages).await?;

            // Create new session
            let session_id = Uuid::new_v4();
            let now = chrono::Utc::now();
            let session = AgentSession {
                id: session_id,
                title,
                agent_id: AgentID::PabloV1,
                visibility: AgentSessionVisibility::Private,
                created_at: now,
                updated_at: now,
                checkpoints: vec![],
            };
            db::create_session(&self.db, &session).await?;

            // Create initial checkpoint (root)
            let checkpoint_id = Uuid::new_v4();
            let checkpoint = AgentCheckpointListItem {
                id: checkpoint_id,
                status: AgentStatus::Complete,
                execution_depth: 0,
                parent: None,
                created_at: now,
                updated_at: now,
            };
            let initial_state = AgentOutput::PabloV1 {
                messages: messages.to_vec(),
                node_states: serde_json::json!({}),
            };
            db::create_checkpoint(&self.db, session_id, &checkpoint, &initial_state).await?;

            db::get_checkpoint(&self.db, checkpoint_id).await?
        };

        Ok(current_checkpoint)
    }

    async fn update_session(
        &self,
        checkpoint_info: &RunAgentOutput,
        new_messages: Vec<ChatMessage>,
    ) -> Result<RunAgentOutput, String> {
        let now = chrono::Utc::now();
        let complete_checkpoint = AgentCheckpointListItem {
            id: Uuid::new_v4(),
            status: AgentStatus::Complete,
            execution_depth: checkpoint_info.checkpoint.execution_depth + 1,
            parent: Some(AgentParentCheckpoint {
                id: checkpoint_info.checkpoint.id,
            }),
            created_at: now,
            updated_at: now,
        };

        let mut new_state = checkpoint_info.output.clone();
        new_state.set_messages(new_messages);

        db::create_checkpoint(
            &self.db,
            checkpoint_info.session.id,
            &complete_checkpoint,
            &new_state,
        )
        .await?;

        Ok(RunAgentOutput {
            checkpoint: complete_checkpoint,
            session: checkpoint_info.session.clone(),
            output: new_state,
        })
    }

    async fn generate_session_title(&self, messages: &[ChatMessage]) -> Result<String, String> {
        let llm_config = self.get_llm_config();

        let llm_model = if let Some(eco_model) = &self.model_options.eco_model {
            eco_model.clone()
        } else if llm_config.openai_config.is_some() {
            LLMModel::OpenAI(OpenAIModel::GPT5Mini)
        } else if llm_config.anthropic_config.is_some() {
            LLMModel::Anthropic(AnthropicModel::Claude45Haiku)
        } else if llm_config.gemini_config.is_some() {
            LLMModel::Gemini(GeminiModel::Gemini25Flash)
        } else {
            return Err("No LLM config found".to_string());
        };

        let messages = vec![
            LLMMessage {
                role: "system".to_string(),
                content: LLMMessageContent::String(TITLE_GENERATOR_PROMPT.into()),
            },
            LLMMessage {
                role: "user".to_string(),
                content: LLMMessageContent::String(
                    messages
                        .iter()
                        .map(|msg| {
                            msg.content
                                .as_ref()
                                .unwrap_or(&MessageContent::String("".to_string()))
                                .to_string()
                        })
                        .collect(),
                ),
            },
        ];

        let input = LLMInput {
            model: llm_model,
            messages,
            max_tokens: 100,
            tools: None,
        };

        let response = chat(&llm_config, input).await.map_err(|e| e.to_string())?;

        Ok(response.choices[0].message.content.to_string())
    }
}

async fn analyze_search_query(
    llm_config: &LLMProviderConfig,
    model: &LLMModel,
    query: &str,
) -> Result<AnalysisResult, String> {
    let system_prompt = r#"You are an expert search query analyzer specializing in technical documentation retrieval.

## Your Task

Analyze the user's search query to:
1. Identify the specific types of documentation needed
2. Reformulate the query for optimal search engine results

## Guidelines for Required Documentation

Identify specific documentation types such as:
- API references and specifications
- Installation/setup guides
- Configuration documentation
- Tutorials and getting started guides
- Troubleshooting guides
- Architecture/design documents
- CLI/command references
- SDK/library documentation

## Guidelines for Query Reformulation

Create an optimized search query that:
- Uses specific technical terminology
- Includes relevant keywords (e.g., "documentation", "guide", "API")
- Removes ambiguous or filler words
- Targets authoritative sources when possible
- Is concise but comprehensive (5-10 words ideal)

## Response Format

Respond ONLY with valid XML in this exact structure:

<analysis>
  <required_documentation>
    <item>specific documentation type needed</item>
  </required_documentation>
  <reformulated_query>optimized search query string</reformulated_query>
</analysis>"#;

    let user_prompt = format!(
        r#"<user_query>{}</user_query>

Analyze this query and provide the required documentation types and an optimized search query."#,
        query
    );

    let response = chat(
        llm_config,
        LLMInput {
            model: model.clone(),
            messages: vec![
                LLMMessage {
                    role: Role::System.to_string(),
                    content: LLMMessageContent::String(system_prompt.to_string()),
                },
                LLMMessage {
                    role: Role::User.to_string(),
                    content: LLMMessageContent::String(user_prompt.to_string()),
                },
            ],
            max_tokens: 2000,
            tools: None,
        },
    )
    .await
    .map_err(|e| e.to_string())?;

    let content = response.choices[0].message.content.to_string();

    parse_analysis_xml(&content)
}

fn parse_analysis_xml(xml: &str) -> Result<AnalysisResult, String> {
    let extract_tag = |tag: &str| -> Option<String> {
        let start_tag = format!("<{}>", tag);
        let end_tag = format!("</{}>", tag);
        xml.find(&start_tag).and_then(|start| {
            let content_start = start + start_tag.len();
            xml[content_start..]
                .find(&end_tag)
                .map(|end| xml[content_start..content_start + end].trim().to_string())
        })
    };

    let extract_all_tags = |tag: &str| -> Vec<String> {
        let start_tag = format!("<{}>", tag);
        let end_tag = format!("</{}>", tag);
        let mut results = Vec::new();
        let mut search_start = 0;

        while let Some(start) = xml[search_start..].find(&start_tag) {
            let abs_start = search_start + start + start_tag.len();
            if let Some(end) = xml[abs_start..].find(&end_tag) {
                results.push(xml[abs_start..abs_start + end].trim().to_string());
                search_start = abs_start + end + end_tag.len();
            } else {
                break;
            }
        }
        results
    };

    let required_documentation = extract_all_tags("item");
    let reformulated_query =
        extract_tag("reformulated_query").ok_or("Failed to extract reformulated_query from XML")?;

    Ok(AnalysisResult {
        required_documentation,
        reformulated_query,
    })
}

async fn validate_search_docs(
    llm_config: &LLMProviderConfig,
    model: &LLMModel,
    docs: &[ScrapedContent],
    query: &str,
    required_documentation: &[String],
    previous_queries: &[String],
    accumulated_needed_urls: &[String],
) -> Result<ValidationResult, String> {
    let docs_preview = docs
        .iter()
        .enumerate()
        .take(10)
        .map(|(i, r)| {
            format!(
                "<doc index=\"{}\">\n  <title>{}</title>\n  <url>{}</url>\n</doc>",
                i + 1,
                r.title.clone().unwrap_or_else(|| "Untitled".to_string()),
                r.url
            )
        })
        .collect::<Vec<_>>()
        .join("\n");

    let required_docs_formatted = required_documentation
        .iter()
        .map(|d| format!("  <item>{}</item>", d))
        .collect::<Vec<_>>()
        .join("\n");

    let previous_queries_formatted = previous_queries
        .iter()
        .map(|q| format!("  <query>{}</query>", q))
        .collect::<Vec<_>>()
        .join("\n");

    let accumulated_urls_formatted = accumulated_needed_urls
        .iter()
        .map(|u| format!("  <url>{}</url>", u))
        .collect::<Vec<_>>()
        .join("\n");

    let system_prompt = r#"You are an expert search result validator. Your task is to evaluate whether search results adequately satisfy a documentation query.

## Evaluation Criteria

For each search result, assess:
1. **Relevance**: Does the document directly address the required documentation topics?
2. **Authority**: Is this an official source, documentation site, or authoritative reference?
3. **Completeness**: Does it provide comprehensive information, not just passing mentions?
4. **Freshness**: For technical docs, prefer current/maintained sources over outdated ones.

## Decision Guidelines

Mark results as SATISFIED when:
- All required documentation topics have at least one authoritative source
- The sources provide actionable, detailed information
- No critical gaps remain in coverage

Suggest a NEW QUERY when:
- Key topics are missing from results
- Results are too general or tangential
- A more specific query would yield better results
- Previous queries haven't addressed certain requirements

## Response Format

Respond ONLY with valid XML in this exact structure:

<validation>
  <is_satisfied>true or false</is_satisfied>
  <valid_docs>
    <doc><url>exact URL from results</url></doc>
  </valid_docs>
  <needed_urls>
    <url>specific URL pattern or domain still needed</url>
  </needed_urls>
  <new_query>refined search query if not satisfied, omit if satisfied</new_query>
  <reasoning>brief explanation of your assessment</reasoning>
</validation>"#;

    let user_prompt = format!(
        r#"<search_context>
  <original_query>{}</original_query>
  <required_documentation>
{}
  </required_documentation>
  <previous_queries>
{}
  </previous_queries>
  <accumulated_needed_urls>
{}
  </accumulated_needed_urls>
</search_context>

<current_results>
{}
</current_results>

Evaluate these search results against the requirements. Which documents are valid and relevant? Is the documentation requirement satisfied? If not, what specific query would help find missing information?"#,
        query,
        if required_docs_formatted.is_empty() {
            "    <item>None specified</item>".to_string()
        } else {
            required_docs_formatted
        },
        if previous_queries_formatted.is_empty() {
            "    <query>None</query>".to_string()
        } else {
            previous_queries_formatted
        },
        if accumulated_urls_formatted.is_empty() {
            "    <url>None</url>".to_string()
        } else {
            accumulated_urls_formatted
        },
        docs_preview
    );

    let response = chat(
        llm_config,
        LLMInput {
            model: model.clone(),
            messages: vec![
                LLMMessage {
                    role: Role::System.to_string(),
                    content: LLMMessageContent::String(system_prompt.to_string()),
                },
                LLMMessage {
                    role: Role::User.to_string(),
                    content: LLMMessageContent::String(user_prompt.to_string()),
                },
            ],
            max_tokens: 4000,
            tools: None,
        },
    )
    .await
    .map_err(|e| e.to_string())?;

    let content = response.choices[0].message.content.to_string();

    let validation = parse_validation_xml(&content, docs)?;

    Ok(validation)
}

fn parse_validation_xml(xml: &str, docs: &[ScrapedContent]) -> Result<ValidationResult, String> {
    let extract_tag = |tag: &str| -> Option<String> {
        let start_tag = format!("<{}>", tag);
        let end_tag = format!("</{}>", tag);
        xml.find(&start_tag).and_then(|start| {
            let content_start = start + start_tag.len();
            xml[content_start..]
                .find(&end_tag)
                .map(|end| xml[content_start..content_start + end].trim().to_string())
        })
    };

    let extract_all_tags = |tag: &str| -> Vec<String> {
        let start_tag = format!("<{}>", tag);
        let end_tag = format!("</{}>", tag);
        let mut results = Vec::new();
        let mut search_start = 0;

        while let Some(start) = xml[search_start..].find(&start_tag) {
            let abs_start = search_start + start + start_tag.len();
            if let Some(end) = xml[abs_start..].find(&end_tag) {
                results.push(xml[abs_start..abs_start + end].trim().to_string());
                search_start = abs_start + end + end_tag.len();
            } else {
                break;
            }
        }
        results
    };

    let is_satisfied = extract_tag("is_satisfied")
        .map(|s| s.to_lowercase() == "true")
        .unwrap_or(false);

    let valid_urls: Vec<String> = extract_all_tags("url")
        .into_iter()
        .filter(|url| docs.iter().any(|d| d.url == *url))
        .collect();

    let valid_docs: Vec<ScrapedContent> = valid_urls
        .iter()
        .filter_map(|url| docs.iter().find(|d| d.url == *url).cloned())
        .collect();

    let needed_urls: Vec<String> = extract_all_tags("url")
        .into_iter()
        .filter(|url| !docs.iter().any(|d| d.url == *url))
        .collect();

    let new_query = extract_tag("new_query").filter(|q| !q.is_empty() && q != "omit if satisfied");

    Ok(ValidationResult {
        is_satisfied,
        valid_docs,
        needed_urls,
        new_query,
    })
}

fn get_search_model(
    llm_config: &LLMProviderConfig,
    eco_model: Option<LLMModel>,
    smart_model: Option<LLMModel>,
) -> LLMModel {
    let base_model = eco_model.or(smart_model);

    match base_model {
        Some(LLMModel::OpenAI(_)) => LLMModel::OpenAI(OpenAIModel::O4Mini),
        Some(LLMModel::Anthropic(_)) => LLMModel::Anthropic(AnthropicModel::Claude45Haiku),
        Some(LLMModel::Gemini(_)) => LLMModel::Gemini(GeminiModel::Gemini3Flash),
        Some(LLMModel::Custom(model)) => LLMModel::Custom(model),
        None => {
            if llm_config.openai_config.is_some() {
                LLMModel::OpenAI(OpenAIModel::O4Mini)
            } else if llm_config.anthropic_config.is_some() {
                LLMModel::Anthropic(AnthropicModel::Claude45Haiku)
            } else if llm_config.gemini_config.is_some() {
                LLMModel::Gemini(GeminiModel::Gemini3Flash)
            } else {
                LLMModel::OpenAI(OpenAIModel::O4Mini)
            }
        }
    }
}