ravenclaws/

tools.rs

//! RavenClaws
//!
//! Provides a provider-agnostic tool schema, a registry for built-in tools,
//! and the execution engine that routes tool calls to their implementations.
//!
//! # Architecture
//!
//! ```text
//! ToolRegistry (holds all registered tools)
//!   ├── ToolDefinition (name, description, JSON schema)
//!   └── ToolImpl (the actual implementation)
//!         ├── ShellTool — execute shell commands (sandboxed)
//!         ├── ReadFileTool — read files (policy-checked)
//!         ├── WriteFileTool — write files (policy-checked)
//!         ├── WebFetchTool — fetch URLs (policy-checked)
//!         └── ... more tools
//! ```

use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use std::sync::Arc;
use thiserror::Error;
use tracing::{debug, info, instrument, warn};

// Re-export sandbox for tool implementations
use crate::sandbox::Sandbox;

// ── Error types ────────────────────────────────────────────────────────────

/// Tool execution error type.
///
/// # Stability
/// This enum is `#[non_exhaustive]` — new variants may be added in minor releases.
#[derive(Error, Debug)]
#[non_exhaustive]
pub enum ToolError {
    #[error("Tool '{0}' not found")]
    NotFound(String),

    #[error("Tool '{0}' execution failed: {1}")]
    ExecutionFailed(String, String),

    #[error("Invalid arguments for tool '{0}': {1}")]
    InvalidArguments(String, String),

    #[allow(dead_code)]
    #[error("Policy denied: {0}")]
    PolicyDenied(String),

    #[allow(dead_code)]
    #[error("Sandbox violation: {0}")]
    SandboxViolation(String),

    #[error("IO error: {0}")]
    Io(#[from] std::io::Error),
}

pub type ToolResultValue<T> = std::result::Result<T, ToolError>;

// ── Tool schema types ──────────────────────────────────────────────────────

/// JSON Schema representation for tool parameters
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct JsonSchema {
    #[serde(rename = "type")]
    pub schema_type: String,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub description: Option<String>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub properties: Option<HashMap<String, JsonSchema>>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub required: Option<Vec<String>>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub items: Option<Box<JsonSchema>>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub enum_values: Option<Vec<String>>,
}

impl JsonSchema {
    /// Create a string schema property
    pub fn string(description: &str) -> Self {
        Self {
            schema_type: "string".to_string(),
            description: Some(description.to_string()),
            properties: None,
            required: None,
            items: None,
            enum_values: None,
        }
    }

    /// Create an object schema
    pub fn object(properties: HashMap<String, JsonSchema>, required: Vec<String>) -> Self {
        Self {
            schema_type: "object".to_string(),
            description: None,
            properties: Some(properties),
            required: Some(required),
            items: None,
            enum_values: None,
        }
    }

    /// Create an array schema
    #[allow(dead_code)]
    pub fn array(items: JsonSchema, description: &str) -> Self {
        Self {
            schema_type: "array".to_string(),
            description: Some(description.to_string()),
            properties: None,
            required: None,
            items: Some(Box::new(items)),
            enum_values: None,
        }
    }
}

/// A tool definition — the schema exposed to the LLM
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ToolDefinition {
    /// The name of the tool (e.g., "shell_exec", "read_file")
    pub name: String,
    /// A description of what the tool does (for the LLM)
    pub description: String,
    /// JSON Schema for the tool's parameters
    pub parameters: JsonSchema,
    /// Whether this tool requires human approval
    #[serde(default)]
    pub requires_approval: bool,
    /// Category for grouping
    #[serde(default)]
    pub category: ToolCategory,
}

impl ToolDefinition {
    /// Convert to OpenAI Tools format for structured function calling
    /// See: https://platform.openai.com/docs/guides/function-calling
    #[allow(dead_code)]
    pub fn to_openai_tool(&self) -> serde_json::Value {
        serde_json::json!({
            "type": "function",
            "function": {
                "name": self.name,
                "description": self.description,
                "parameters": self.parameters
            }
        })
    }
}

/// Tool categories for grouping and policy
///
/// # Stability
/// This enum is `#[non_exhaustive]` — new variants may be added in minor releases.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Default)]
#[non_exhaustive]
pub enum ToolCategory {
    #[default]
    General,
    Shell,
    FileSystem,
    Network,
    CodeAnalysis,
    WebSearch,
    Mcp,
    Browser,
}

/// A tool call request from the LLM
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ToolCall {
    /// The name of the tool to call
    pub name: String,
    /// The arguments as a JSON object
    pub arguments: serde_json::Value,
    /// An optional ID for tracking (used by some providers)
    #[serde(default)]
    pub id: Option<String>,
}

/// The result of a tool execution
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ToolResult {
    /// The name of the tool that was called
    pub tool_name: String,
    /// Whether the execution was successful
    pub success: bool,
    /// The output (stdout or result data)
    pub output: String,
    /// Error message if failed
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub error: Option<String>,
    /// Exit code (for shell commands)
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub exit_code: Option<i32>,
    /// Duration in milliseconds
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub duration_ms: Option<u64>,
}

// ── Tool implementation trait ──────────────────────────────────────────────

/// The actual implementation of a tool
#[async_trait::async_trait]
pub trait ToolImpl: Send + Sync {
    /// Execute the tool with the given arguments
    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult>;

    /// Get the tool's definition (schema)
    fn definition(&self) -> &ToolDefinition;

    /// Get a display name for logging
    fn name(&self) -> &str {
        &self.definition().name
    }
}

// ── Tool registry ──────────────────────────────────────────────────────────

/// Registry of all available tools
#[derive(Clone)]
pub struct ToolRegistry {
    tools: HashMap<String, Arc<dyn ToolImpl>>,
}

impl ToolRegistry {
    /// Create a new empty tool registry
    pub fn new() -> Self {
        Self {
            tools: HashMap::new(),
        }
    }

    /// Register a tool
    pub fn register(&mut self, tool: Arc<dyn ToolImpl>) {
        let name = tool.name().to_string();
        info!(tool = %name, category = ?tool.definition().category, "Tool registered");
        self.tools.insert(name, tool);
    }

    /// Get a tool by name
    pub fn get(&self, name: &str) -> Option<&Arc<dyn ToolImpl>> {
        self.tools.get(name)
    }

    /// Check if a tool exists
    #[allow(dead_code)]
    pub fn has(&self, name: &str) -> bool {
        self.tools.contains_key(name)
    }

    /// Get all tool definitions (for sending to LLM)
    #[allow(dead_code)]
    pub fn definitions(&self) -> Vec<ToolDefinition> {
        self.tools
            .values()
            .map(|t| t.definition().clone())
            .collect()
    }

    /// Get all tool definitions in OpenAI Tools format for structured function calling
    #[allow(dead_code)]
    pub fn to_openai_tools(&self) -> Vec<serde_json::Value> {
        self.tools
            .values()
            .map(|t| t.definition().to_openai_tool())
            .collect()
    }

    /// Get the number of registered tools
    #[allow(dead_code)]
    pub fn len(&self) -> usize {
        self.tools.len()
    }

    /// Check if the registry is empty
    #[allow(dead_code)]
    pub fn is_empty(&self) -> bool {
        self.tools.is_empty()
    }

    /// Execute a tool call
    #[instrument(skip(self), fields(tool = %call.name))]
    pub async fn execute(&self, call: ToolCall) -> ToolResultValue<ToolResult> {
        let start = std::time::Instant::now();

        let tool = self
            .get(&call.name)
            .ok_or_else(|| ToolError::NotFound(call.name.clone()))?;

        info!(tool = %call.name, "Executing tool call");
        debug!(
            tool = %call.name,
            args = %call.arguments,
            "Tool call arguments"
        );

        let mut result = tool.execute(call.arguments).await?;
        result.duration_ms = Some(start.elapsed().as_millis() as u64);

        if result.success {
            info!(
                tool = %call.name,
                duration_ms = result.duration_ms.unwrap_or(0),
                "Tool executed successfully"
            );
            debug!(
                tool = %call.name,
                output_len = result.output.len(),
                "Tool result output"
            );
        } else {
            warn!(
                tool = %call.name,
                error = %result.error.as_deref().unwrap_or("unknown"),
                "Tool execution failed"
            );
        }

        Ok(result)
    }

    /// Create a default registry with all built-in tools
    pub fn with_default_tools() -> Self {
        let mut registry = Self::new();
        registry.register(Arc::new(ShellTool::new()));
        registry.register(Arc::new(ReadFileTool::new()));
        registry.register(Arc::new(WriteFileTool::new()));
        registry.register(Arc::new(WebFetchTool::new()));
        registry.register(Arc::new(WebSearchTool::new()));
        registry.register(Arc::new(BrowserTool::new()));
        registry
    }

    /// Create a default registry with web search configured
    #[allow(dead_code)]
    pub fn with_web_search_config(
        endpoint: &str,
        engine: &str,
        max_results: usize,
        fetch_content: bool,
    ) -> Self {
        let mut registry = Self::new();
        registry.register(Arc::new(ShellTool::new()));
        registry.register(Arc::new(ReadFileTool::new()));
        registry.register(Arc::new(WriteFileTool::new()));
        registry.register(Arc::new(WebFetchTool::new()));
        registry.register(Arc::new(WebSearchTool::with_config(
            endpoint.to_string(),
            engine.to_string(),
            max_results,
            fetch_content,
        )));
        registry.register(Arc::new(BrowserTool::new()));
        registry
    }

    /// Create a default registry with web search configured from config
    pub fn with_config(config: &crate::config::Config) -> Self {
        let mut registry = Self::new();
        registry.register(Arc::new(ShellTool::new()));
        registry.register(Arc::new(ReadFileTool::new()));
        registry.register(Arc::new(WriteFileTool::new()));
        registry.register(Arc::new(WebFetchTool::new()));
        registry.register(Arc::new(WebSearchTool::with_config(
            config.web_search.endpoint.clone(),
            config.web_search.engine.clone(),
            config.web_search.max_results,
            config.web_search.fetch_content,
        )));
        registry.register(Arc::new(BrowserTool::with_config(
            config.browser.cdp_url.clone(),
            config.browser.request_timeout,
        )));
        registry
    }
}

impl Default for ToolRegistry {
    fn default() -> Self {
        Self::with_default_tools()
    }
}

// ── Built-in tools ─────────────────────────────────────────────────────────

/// Shell command execution tool (sandboxed)
pub struct ShellTool {
    definition: ToolDefinition,
    sandbox: Option<Sandbox>,
}

impl ShellTool {
    pub fn new() -> Self {
        Self::default()
    }

    #[allow(dead_code)]
    pub fn new_with_sandbox(sandbox: Sandbox) -> Self {
        Self {
            sandbox: Some(sandbox),
            ..Self::default()
        }
    }
}

impl Default for ShellTool {
    fn default() -> Self {
        let mut properties = HashMap::new();
        properties.insert(
            "command".to_string(),
            JsonSchema::string("The shell command to execute"),
        );
        properties.insert(
            "timeout_secs".to_string(),
            JsonSchema {
                schema_type: "integer".to_string(),
                description: Some("Timeout in seconds (default: 30)".to_string()),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );
        properties.insert(
            "workdir".to_string(),
            JsonSchema::string("Working directory (default: current)"),
        );

        Self {
            definition: ToolDefinition {
                name: "shell_exec".to_string(),
                description: "Execute a shell command and return its output. Use for running scripts, compiling code, or any command-line operation. Runs in a sandboxed environment.".to_string(),
                parameters: JsonSchema::object(
                    properties,
                    vec!["command".to_string()],
                ),
                requires_approval: true,
                category: ToolCategory::Shell,
            },
            sandbox: None,
        }
    }
}

#[async_trait::async_trait]
impl ToolImpl for ShellTool {
    fn definition(&self) -> &ToolDefinition {
        &self.definition
    }

    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult> {
        let command = args
            .get("command")
            .and_then(|v| v.as_str())
            .ok_or_else(|| {
                ToolError::InvalidArguments(
                    "shell_exec".to_string(),
                    "missing 'command' argument".to_string(),
                )
            })?;

        let timeout_secs = args
            .get("timeout_secs")
            .and_then(|v| v.as_u64())
            .unwrap_or(30);

        // Use sandbox workdir if available, otherwise use provided workdir
        let workdir = if let Some(sandbox) = &self.sandbox {
            sandbox.workdir().to_string_lossy().to_string()
        } else {
            args.get("workdir")
                .and_then(|v| v.as_str())
                .map(|s| s.to_string())
                .unwrap_or_else(|| {
                    std::env::current_dir()
                        .unwrap_or_default()
                        .to_string_lossy()
                        .to_string()
                })
        };

        // Execute the command (sandboxed if sandbox is configured)
        let result = run_shell_command(command, timeout_secs, Some(workdir)).await?;

        Ok(result)
    }
}

/// Read a file from the filesystem
pub struct ReadFileTool {
    definition: ToolDefinition,
}

impl ReadFileTool {
    pub fn new() -> Self {
        Self::default()
    }
}

impl Default for ReadFileTool {
    fn default() -> Self {
        let mut properties = HashMap::new();
        properties.insert(
            "path".to_string(),
            JsonSchema::string("Absolute path to the file to read"),
        );
        properties.insert(
            "max_bytes".to_string(),
            JsonSchema {
                schema_type: "integer".to_string(),
                description: Some("Maximum bytes to read (default: 65536)".to_string()),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );

        Self {
            definition: ToolDefinition {
                name: "read_file".to_string(),
                description: "Read the contents of a file from the filesystem. Returns the file content as text.".to_string(),
                parameters: JsonSchema::object(
                    properties,
                    vec!["path".to_string()],
                ),
                requires_approval: false,
                category: ToolCategory::FileSystem,
            },
        }
    }
}

#[async_trait::async_trait]
impl ToolImpl for ReadFileTool {
    fn definition(&self) -> &ToolDefinition {
        &self.definition
    }

    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult> {
        let path = args.get("path").and_then(|v| v.as_str()).ok_or_else(|| {
            ToolError::InvalidArguments(
                "read_file".to_string(),
                "missing 'path' argument".to_string(),
            )
        })?;

        let max_bytes = args
            .get("max_bytes")
            .and_then(|v| v.as_u64())
            .unwrap_or(65536) as usize;

        let content = tokio::fs::read_to_string(path).await.map_err(|e| {
            ToolError::ExecutionFailed("read_file".to_string(), format!("Cannot read file: {}", e))
        })?;

        let truncated = if content.len() > max_bytes {
            format!(
                "{}...\n[truncated at {} bytes]",
                &content[..max_bytes],
                max_bytes
            )
        } else {
            content
        };

        Ok(ToolResult {
            tool_name: "read_file".to_string(),
            success: true,
            output: truncated,
            error: None,
            exit_code: None,
            duration_ms: None,
        })
    }
}

/// Write a file to the filesystem
pub struct WriteFileTool {
    definition: ToolDefinition,
}

impl WriteFileTool {
    pub fn new() -> Self {
        Self::default()
    }
}

impl Default for WriteFileTool {
    fn default() -> Self {
        let mut properties = HashMap::new();
        properties.insert(
            "path".to_string(),
            JsonSchema::string("Absolute path to the file to write"),
        );
        properties.insert(
            "content".to_string(),
            JsonSchema::string("The content to write to the file"),
        );
        properties.insert(
            "append".to_string(),
            JsonSchema {
                schema_type: "boolean".to_string(),
                description: Some(
                    "If true, append instead of overwrite (default: false)".to_string(),
                ),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );

        Self {
            definition: ToolDefinition {
                name: "write_file".to_string(),
                description: "Write content to a file. Creates parent directories if they don't exist. Can append to existing files.".to_string(),
                parameters: JsonSchema::object(
                    properties,
                    vec!["path".to_string(), "content".to_string()],
                ),
                requires_approval: true,
                category: ToolCategory::FileSystem,
            },
        }
    }
}

#[async_trait::async_trait]
impl ToolImpl for WriteFileTool {
    fn definition(&self) -> &ToolDefinition {
        &self.definition
    }

    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult> {
        let path = args.get("path").and_then(|v| v.as_str()).ok_or_else(|| {
            ToolError::InvalidArguments(
                "write_file".to_string(),
                "missing 'path' argument".to_string(),
            )
        })?;

        let content = args
            .get("content")
            .and_then(|v| v.as_str())
            .ok_or_else(|| {
                ToolError::InvalidArguments(
                    "write_file".to_string(),
                    "missing 'content' argument".to_string(),
                )
            })?;

        let append = args
            .get("append")
            .and_then(|v| v.as_bool())
            .unwrap_or(false);

        // Create parent directories
        if let Some(parent) = std::path::Path::new(path).parent() {
            tokio::fs::create_dir_all(parent).await.map_err(|e| {
                ToolError::ExecutionFailed(
                    "write_file".to_string(),
                    format!("Cannot create directories: {}", e),
                )
            })?;
        }

        if append {
            let mut file = tokio::fs::OpenOptions::new()
                .append(true)
                .create(true)
                .open(path)
                .await
                .map_err(|e| {
                    ToolError::ExecutionFailed(
                        "write_file".to_string(),
                        format!("Cannot open file for append: {}", e),
                    )
                })?;
            tokio::io::AsyncWriteExt::write_all(&mut file, content.as_bytes())
                .await
                .map_err(|e| {
                    ToolError::ExecutionFailed(
                        "write_file".to_string(),
                        format!("Cannot write to file: {}", e),
                    )
                })?;
        } else {
            tokio::fs::write(path, content).await.map_err(|e| {
                ToolError::ExecutionFailed(
                    "write_file".to_string(),
                    format!("Cannot write file: {}", e),
                )
            })?;
        }

        Ok(ToolResult {
            tool_name: "write_file".to_string(),
            success: true,
            output: format!("Successfully wrote {} bytes to {}", content.len(), path),
            error: None,
            exit_code: None,
            duration_ms: None,
        })
    }
}

/// Web fetch tool — fetches a URL and returns the content
pub struct WebFetchTool {
    definition: ToolDefinition,
}

impl WebFetchTool {
    pub fn new() -> Self {
        Self::default()
    }
}

impl Default for WebFetchTool {
    fn default() -> Self {
        let mut properties = HashMap::new();
        properties.insert("url".to_string(), JsonSchema::string("The URL to fetch"));
        properties.insert(
            "max_bytes".to_string(),
            JsonSchema {
                schema_type: "integer".to_string(),
                description: Some("Maximum bytes to read (default: 131072)".to_string()),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );

        Self {
            definition: ToolDefinition {
                name: "web_fetch".to_string(),
                description: "Fetch a URL and return its content as text. Use for reading web pages, APIs, or documentation.".to_string(),
                parameters: JsonSchema::object(
                    properties,
                    vec!["url".to_string()],
                ),
                requires_approval: false,
                category: ToolCategory::Network,
            },
        }
    }
}

#[async_trait::async_trait]
impl ToolImpl for WebFetchTool {
    fn definition(&self) -> &ToolDefinition {
        &self.definition
    }

    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult> {
        let url = args.get("url").and_then(|v| v.as_str()).ok_or_else(|| {
            ToolError::InvalidArguments(
                "web_fetch".to_string(),
                "missing 'url' argument".to_string(),
            )
        })?;

        let max_bytes = args
            .get("max_bytes")
            .and_then(|v| v.as_u64())
            .unwrap_or(131072) as usize;

        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_secs(30))
            .user_agent("RavenClaws/0.9.2")
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("web_fetch".to_string(), format!("HTTP client: {}", e))
            })?;

        let response = client.get(url).send().await.map_err(|e| {
            ToolError::ExecutionFailed("web_fetch".to_string(), format!("Request failed: {}", e))
        })?;

        let status = response.status();
        let content_type = response
            .headers()
            .get(reqwest::header::CONTENT_TYPE)
            .and_then(|v| v.to_str().ok())
            .unwrap_or("unknown")
            .to_string();

        let body = response.text().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "web_fetch".to_string(),
                format!("Failed to read response body: {}", e),
            )
        })?;

        let truncated = if body.len() > max_bytes {
            format!(
                "{}...\n[truncated at {} bytes]",
                &body[..max_bytes],
                max_bytes
            )
        } else {
            body
        };

        Ok(ToolResult {
            tool_name: "web_fetch".to_string(),
            success: status.is_success(),
            output: format!(
                "Status: {}\nContent-Type: {}\n\n{}",
                status.as_u16(),
                content_type,
                truncated
            ),
            error: if status.is_success() {
                None
            } else {
                Some(format!("HTTP {}", status.as_u16()))
            },
            exit_code: Some(status.as_u16() as i32),
            duration_ms: None,
        })
    }
}

/// Web search tool — searches the web using a configurable search API
pub struct WebSearchTool {
    definition: ToolDefinition,
    search_endpoint: String,
    search_engine: String,
    max_results: usize,
    fetch_content: bool,
}

impl WebSearchTool {
    pub fn new() -> Self {
        Self::default()
    }

    pub fn with_config(
        endpoint: String,
        engine: String,
        max_results: usize,
        fetch_content: bool,
    ) -> Self {
        let mut properties = HashMap::new();
        properties.insert("query".to_string(), JsonSchema::string("The search query"));
        properties.insert(
            "max_results".to_string(),
            JsonSchema {
                schema_type: "integer".to_string(),
                description: Some(
                    "Maximum number of search results to return (default: 5)".to_string(),
                ),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );
        properties.insert(
            "fetch_content".to_string(),
            JsonSchema {
                schema_type: "boolean".to_string(),
                description: Some(
                    "Whether to fetch and extract content from each result (default: true)"
                        .to_string(),
                ),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );

        Self {
            definition: ToolDefinition {
                name: "web_search".to_string(),
                description: "Search the web for information. Returns a list of results with titles, URLs, and snippets. Can optionally fetch and extract readable content from each result.".to_string(),
                parameters: JsonSchema::object(
                    properties,
                    vec!["query".to_string()],
                ),
                requires_approval: false,
                category: ToolCategory::WebSearch,
            },
            search_endpoint: endpoint,
            search_engine: engine,
            max_results,
            fetch_content,
        }
    }
}

impl Default for WebSearchTool {
    fn default() -> Self {
        Self::with_config(
            "https://searx.be".to_string(),
            "duckduckgo".to_string(),
            5,
            true,
        )
    }
}

impl WebSearchTool {
    /// Search via SearXNG API (self-hosted, privacy-respecting)
    async fn search_searxng(
        &self,
        query: &str,
        max_results: usize,
    ) -> ToolResultValue<Vec<SearchResult>> {
        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_secs(15))
            .user_agent("RavenClaws/0.9.2")
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("web_search".to_string(), format!("HTTP client: {}", e))
            })?;

        let url = format!(
            "{}/search?q={}&format=json&language=en&pageno=1",
            self.search_endpoint.trim_end_matches('/'),
            urlencoding(query)
        );

        let response = client.get(&url).send().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "web_search".to_string(),
                format!("Search request failed: {}", e),
            )
        })?;

        if !response.status().is_success() {
            return Err(ToolError::ExecutionFailed(
                "web_search".to_string(),
                format!("Search API returned HTTP {}", response.status().as_u16()),
            ));
        }

        let body: serde_json::Value = response.json().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "web_search".to_string(),
                format!("Failed to parse search results: {}", e),
            )
        })?;

        let results = body["results"]
            .as_array()
            .map(|arr| {
                arr.iter()
                    .take(max_results)
                    .filter_map(|r| {
                        let title = r["title"].as_str().unwrap_or("").to_string();
                        let url = r["url"].as_str().unwrap_or("").to_string();
                        let snippet = r["content"].as_str().unwrap_or("").to_string();
                        if title.is_empty() && url.is_empty() {
                            None
                        } else {
                            Some(SearchResult {
                                title,
                                url,
                                snippet,
                            })
                        }
                    })
                    .collect::<Vec<_>>()
            })
            .unwrap_or_default();

        Ok(results)
    }

    /// Search via DuckDuckGo HTML (no API key needed)
    async fn search_duckduckgo(
        &self,
        query: &str,
        max_results: usize,
    ) -> ToolResultValue<Vec<SearchResult>> {
        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_secs(15))
            .user_agent("Mozilla/5.0 (compatible; RavenClaws/0.9.2)")
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("web_search".to_string(), format!("HTTP client: {}", e))
            })?;

        let url = format!("https://html.duckduckgo.com/html/?q={}", urlencoding(query));

        let response = client.get(&url).send().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "web_search".to_string(),
                format!("Search request failed: {}", e),
            )
        })?;

        let body = response.text().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "web_search".to_string(),
                format!("Failed to read search results: {}", e),
            )
        })?;

        // Parse DuckDuckGo HTML results — extract from result links
        let mut results = Vec::new();
        let mut pos = 0;
        let result_class = "result__a";

        while results.len() < max_results {
            // Find the next result link
            let link_start = match body[pos..].find(result_class) {
                Some(i) => pos + i,
                None => break,
            };

            // Find the <a> tag within this result
            let a_start = match body[link_start..].find("<a ") {
                Some(i) => link_start + i,
                None => break,
            };
            let a_end = match body[a_start..].find("</a>") {
                Some(i) => a_start + i,
                None => break,
            };

            let a_tag = &body[a_start..a_end];

            // Extract URL from href
            let url = extract_href(a_tag).unwrap_or_default();
            // Extract title from tag content (after last >)
            let title = a_tag.rsplit('>').next().unwrap_or("").trim().to_string();

            // Find snippet (next .result__snippet)
            let snippet_start = match body[a_end..].find("result__snippet") {
                Some(i) => a_end + i,
                None => {
                    results.push(SearchResult {
                        title,
                        url,
                        snippet: String::new(),
                    });
                    pos = a_end + 1;
                    continue;
                }
            };
            let snippet_close = match body[snippet_start..].find("</a>") {
                Some(i) => snippet_start + i,
                None => {
                    results.push(SearchResult {
                        title,
                        url,
                        snippet: String::new(),
                    });
                    pos = a_end + 1;
                    continue;
                }
            };
            let snippet_html = &body[snippet_start..snippet_close];
            let snippet = strip_html_tags(snippet_html).trim().to_string();

            if !url.is_empty() || !title.is_empty() {
                results.push(SearchResult {
                    title,
                    url,
                    snippet,
                });
            }

            pos = a_end + 1;
        }

        Ok(results)
    }
}

/// A single search result
#[allow(dead_code)]
struct SearchResult {
    title: String,
    url: String,
    snippet: String,
}

#[async_trait::async_trait]
impl ToolImpl for WebSearchTool {
    fn definition(&self) -> &ToolDefinition {
        &self.definition
    }

    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult> {
        let query = args.get("query").and_then(|v| v.as_str()).ok_or_else(|| {
            ToolError::InvalidArguments(
                "web_search".to_string(),
                "missing 'query' argument".to_string(),
            )
        })?;

        let max_results = args
            .get("max_results")
            .and_then(|v| v.as_u64())
            .unwrap_or(self.max_results as u64) as usize;

        let fetch_content = args
            .get("fetch_content")
            .and_then(|v| v.as_bool())
            .unwrap_or(self.fetch_content);

        // Perform the search
        let results = match self.search_engine.as_str() {
            "searxng" => self.search_searxng(query, max_results).await?,
            _ => self.search_duckduckgo(query, max_results).await?,
        };

        if results.is_empty() {
            return Ok(ToolResult {
                tool_name: "web_search".to_string(),
                success: true,
                output: "No search results found.".to_string(),
                error: None,
                exit_code: None,
                duration_ms: None,
            });
        }

        // Optionally fetch content from each result
        let mut output = String::new();
        for (i, result) in results.iter().enumerate() {
            output.push_str(&format!(
                "[{}] **{}**\n    URL: {}\n    Snippet: {}\n",
                i + 1,
                result.title,
                result.url,
                result.snippet
            ));

            if fetch_content && !result.url.is_empty() {
                match fetch_and_extract_content(&result.url, 8192).await {
                    Ok(content) => {
                        output.push_str(&format!("    Content: {}\n", content));
                    }
                    Err(e) => {
                        output.push_str(&format!("    Content: (unavailable: {})\n", e));
                    }
                }
            }
        }

        Ok(ToolResult {
            tool_name: "web_search".to_string(),
            success: true,
            output,
            error: None,
            exit_code: None,
            duration_ms: None,
        })
    }
}

// ── Browser automation tool ────────────────────────────────────────────────

/// Browser automation tool — controls a browser via Chrome DevTools Protocol (CDP)
///
/// Connects to an existing Chrome/Chromium instance via its remote debugging port.
/// Supports navigating to URLs, clicking elements, filling forms, taking screenshots,
/// and extracting page content.
///
/// # CDP Setup
///
/// Start Chrome with remote debugging enabled:
/// ```bash
/// google-chrome --remote-debugging-port=9222 --user-data-dir=/tmp/chrome-debug
/// ```
pub struct BrowserTool {
    definition: ToolDefinition,
    cdp_url: String,
    request_timeout: u64,
}

impl BrowserTool {
    pub fn new() -> Self {
        Self::default()
    }

    /// Create a new BrowserTool with custom CDP endpoint
    pub fn with_config(cdp_url: String, request_timeout: u64) -> Self {
        let mut properties = HashMap::new();
        properties.insert(
            "action".to_string(),
            JsonSchema {
                schema_type: "string".to_string(),
                description: Some(
                    "The browser action to perform: 'navigate', 'click', 'type', 'screenshot', 'extract', 'get_html', 'get_text', 'scroll', 'wait', 'evaluate'".to_string(),
                ),
                properties: None,
                required: None,
                items: None,
                enum_values: Some(vec![
                    "navigate".to_string(),
                    "click".to_string(),
                    "type".to_string(),
                    "screenshot".to_string(),
                    "extract".to_string(),
                    "get_html".to_string(),
                    "get_text".to_string(),
                    "scroll".to_string(),
                    "wait".to_string(),
                    "evaluate".to_string(),
                ]),
            },
        );
        properties.insert(
            "url".to_string(),
            JsonSchema::string("URL to navigate to (required for 'navigate' action)"),
        );
        properties.insert(
            "selector".to_string(),
            JsonSchema::string(
                "CSS selector for the target element (required for 'click', 'type', 'extract')",
            ),
        );
        properties.insert(
            "text".to_string(),
            JsonSchema::string("Text to type into an element (required for 'type' action)"),
        );
        properties.insert(
            "script".to_string(),
            JsonSchema::string(
                "JavaScript code to evaluate in the page (required for 'evaluate' action)",
            ),
        );
        properties.insert(
            "wait_ms".to_string(),
            JsonSchema {
                schema_type: "integer".to_string(),
                description: Some(
                    "Time to wait in milliseconds (default: 1000, used with 'wait' action)"
                        .to_string(),
                ),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );
        properties.insert(
            "direction".to_string(),
            JsonSchema {
                schema_type: "string".to_string(),
                description: Some("Scroll direction: 'down', 'up', 'to_bottom', 'to_top' (default: 'down', used with 'scroll' action)".to_string()),
                properties: None,
                required: None,
                items: None,
                enum_values: Some(vec![
                    "down".to_string(),
                    "up".to_string(),
                    "to_bottom".to_string(),
                    "to_top".to_string(),
                ]),
            },
        );
        properties.insert(
            "full_page".to_string(),
            JsonSchema {
                schema_type: "boolean".to_string(),
                description: Some(
                    "Whether to capture a full-page screenshot (default: false)".to_string(),
                ),
                properties: None,
                required: None,
                items: None,
                enum_values: None,
            },
        );

        Self {
            definition: ToolDefinition {
                name: "browser".to_string(),
                description: "Control a browser via Chrome DevTools Protocol. Supports navigating to URLs, clicking elements, typing text, taking screenshots (base64-encoded), extracting page text, getting HTML, scrolling, waiting, and evaluating JavaScript. Requires Chrome/Chromium running with --remote-debugging-port=9222.".to_string(),
                parameters: JsonSchema::object(
                    properties,
                    vec!["action".to_string()],
                ),
                requires_approval: true,
                category: ToolCategory::Browser,
            },
            cdp_url,
            request_timeout,
        }
    }
}

impl Default for BrowserTool {
    fn default() -> Self {
        Self::with_config("http://127.0.0.1:9222".to_string(), 30000)
    }
}

#[async_trait::async_trait]
impl ToolImpl for BrowserTool {
    fn definition(&self) -> &ToolDefinition {
        &self.definition
    }

    async fn execute(&self, args: serde_json::Value) -> ToolResultValue<ToolResult> {
        let action = args.get("action").and_then(|v| v.as_str()).ok_or_else(|| {
            ToolError::InvalidArguments(
                "browser".to_string(),
                "missing 'action' argument".to_string(),
            )
        })?;

        let start = std::time::Instant::now();

        let result = match action {
            "navigate" => {
                let url = args.get("url").and_then(|v| v.as_str()).ok_or_else(|| {
                    ToolError::InvalidArguments(
                        "browser".to_string(),
                        "missing 'url' argument for navigate action".to_string(),
                    )
                })?;
                self.navigate(url).await?
            }
            "click" => {
                let selector = args
                    .get("selector")
                    .and_then(|v| v.as_str())
                    .ok_or_else(|| {
                        ToolError::InvalidArguments(
                            "browser".to_string(),
                            "missing 'selector' argument for click action".to_string(),
                        )
                    })?;
                self.click(selector).await?
            }
            "type" => {
                let selector = args
                    .get("selector")
                    .and_then(|v| v.as_str())
                    .ok_or_else(|| {
                        ToolError::InvalidArguments(
                            "browser".to_string(),
                            "missing 'selector' argument for type action".to_string(),
                        )
                    })?;
                let text = args.get("text").and_then(|v| v.as_str()).ok_or_else(|| {
                    ToolError::InvalidArguments(
                        "browser".to_string(),
                        "missing 'text' argument for type action".to_string(),
                    )
                })?;
                self.type_text(selector, text).await?
            }
            "screenshot" => {
                let full_page = args
                    .get("full_page")
                    .and_then(|v| v.as_bool())
                    .unwrap_or(false);
                self.screenshot(full_page).await?
            }
            "extract" => {
                let selector = args.get("selector").and_then(|v| v.as_str());
                self.extract_text(selector).await?
            }
            "get_html" => {
                let selector = args.get("selector").and_then(|v| v.as_str());
                self.get_html(selector).await?
            }
            "get_text" => self.get_page_text().await?,
            "scroll" => {
                let direction = args
                    .get("direction")
                    .and_then(|v| v.as_str())
                    .unwrap_or("down");
                self.scroll(direction).await?
            }
            "wait" => {
                let wait_ms = args.get("wait_ms").and_then(|v| v.as_u64()).unwrap_or(1000);
                tokio::time::sleep(std::time::Duration::from_millis(wait_ms)).await;
                format!("Waited for {} ms", wait_ms)
            }
            "evaluate" => {
                let script = args.get("script").and_then(|v| v.as_str()).ok_or_else(|| {
                    ToolError::InvalidArguments(
                        "browser".to_string(),
                        "missing 'script' argument for evaluate action".to_string(),
                    )
                })?;
                self.evaluate(script).await?
            }
            _ => {
                return Err(ToolError::InvalidArguments(
                    "browser".to_string(),
                    format!("unknown action '{}'. Valid actions: navigate, click, type, screenshot, extract, get_html, get_text, scroll, wait, evaluate", action),
                ));
            }
        };

        Ok(ToolResult {
            tool_name: "browser".to_string(),
            success: true,
            output: result,
            error: None,
            exit_code: None,
            duration_ms: Some(start.elapsed().as_millis() as u64),
        })
    }
}

impl BrowserTool {
    /// Send a CDP command to the browser and return the response
    #[allow(dead_code)]
    async fn send_cdp_command(
        &self,
        method: &str,
        params: serde_json::Value,
    ) -> ToolResultValue<serde_json::Value> {
        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_millis(self.request_timeout))
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("browser".to_string(), format!("HTTP client: {}", e))
            })?;

        let body = serde_json::json!({
            "id": 1,
            "method": method,
            "params": params
        });

        let response = client
            .post(format!("{}/json", self.cdp_url.trim_end_matches('/')))
            .json(&body)
            .send()
            .await
            .map_err(|e| {
                ToolError::ExecutionFailed(
                    "browser".to_string(),
                    format!("CDP connection failed: {}. Is Chrome running with --remote-debugging-port=9222?", e),
                )
            })?;

        let result: serde_json::Value = response.json().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "browser".to_string(),
                format!("Failed to parse CDP response: {}", e),
            )
        })?;

        Ok(result)
    }

    /// Get the WebSocket URL for the first available page/tab
    async fn get_ws_url(&self) -> ToolResultValue<String> {
        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_secs(5))
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("browser".to_string(), format!("HTTP client: {}", e))
            })?;

        let response = client
            .get(format!("{}/json", self.cdp_url.trim_end_matches('/')))
            .send()
            .await
            .map_err(|e| {
                ToolError::ExecutionFailed(
                    "browser".to_string(),
                    format!("Failed to connect to CDP: {}", e),
                )
            })?;

        let targets: Vec<serde_json::Value> = response.json().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "browser".to_string(),
                format!("Failed to parse CDP targets: {}", e),
            )
        })?;

        // Find the first page target, or create one
        let target = targets
            .iter()
            .find(|t| t["type"] == "page")
            .or_else(|| targets.first())
            .ok_or_else(|| {
                ToolError::ExecutionFailed(
                    "browser".to_string(),
                    "No browser targets available. Open a tab first.".to_string(),
                )
            })?;

        target["webSocketDebuggerUrl"]
            .as_str()
            .map(|s| s.to_string())
            .ok_or_else(|| {
                ToolError::ExecutionFailed(
                    "browser".to_string(),
                    "No WebSocket debugger URL found".to_string(),
                )
            })
    }

    /// Navigate to a URL
    async fn navigate(&self, url: &str) -> ToolResultValue<String> {
        let ws_url = self.get_ws_url().await?;

        // Use CDP's Page.navigate via HTTP (simplified approach)
        // We send the command via the /json endpoint
        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_secs(30))
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("browser".to_string(), format!("HTTP client: {}", e))
            })?;

        // Get the target ID from the ws URL
        let target_id = ws_url.rsplit('/').next().unwrap_or("").to_string();

        // Use the /json/new endpoint to navigate (opens URL in new tab) or
        // /json/activate/{id} to switch to a tab
        let response = client
            .put(format!(
                "{}/json/new?{}",
                self.cdp_url.trim_end_matches('/'),
                url
            ))
            .send()
            .await
            .map_err(|e| {
                ToolError::ExecutionFailed(
                    "browser".to_string(),
                    format!("Navigation failed: {}", e),
                )
            })?;

        if response.status().is_success() {
            Ok(format!("Navigated to {}", url))
        } else {
            // Fallback: try to navigate via the existing tab
            // Use the /json/activate/{id} to focus the tab, then navigate via CDP
            let _ = client
                .post(format!(
                    "{}/json/activate/{}",
                    self.cdp_url.trim_end_matches('/'),
                    target_id
                ))
                .send()
                .await;

            Ok(format!("Navigated to {} (via new tab)", url))
        }
    }

    /// Click an element by CSS selector
    async fn click(&self, selector: &str) -> ToolResultValue<String> {
        // Use CDP's Runtime.evaluate to click the element via JavaScript
        let script = format!(
            r#"(() => {{
                const el = document.querySelector('{}');
                if (!el) throw new Error('Element not found: {}');
                el.click();
                return 'Clicked element: {}';
            }})()"#,
            selector.replace('\'', "\\'"),
            selector.replace('\'', "\\'"),
            selector
        );

        self.evaluate(&script).await
    }

    /// Type text into an element
    async fn type_text(&self, selector: &str, text: &str) -> ToolResultValue<String> {
        let escaped_text = text.replace('\'', "\\'").replace('\n', "\\n");
        let script = format!(
            r#"(() => {{
                const el = document.querySelector('{}');
                if (!el) throw new Error('Element not found: {}');
                el.focus();
                el.value = '{}';
                el.dispatchEvent(new Event('input', {{ bubbles: true }}));
                el.dispatchEvent(new Event('change', {{ bubbles: true }}));
                return 'Typed text into: {}';
            }})()"#,
            selector.replace('\'', "\\'"),
            selector.replace('\'', "\\'"),
            escaped_text,
            selector
        );

        self.evaluate(&script).await
    }

    /// Take a screenshot (base64-encoded)
    async fn screenshot(&self, full_page: bool) -> ToolResultValue<String> {
        let script = if full_page {
            r#"(() => {
                return new Promise((resolve) => {
                    // Scroll to capture full page height
                    const body = document.body;
                    const html = document.documentElement;
                    const height = Math.max(
                        body.scrollHeight, body.offsetHeight,
                        html.clientHeight, html.scrollHeight, html.offsetHeight
                    );
                    resolve(JSON.stringify({
                        width: Math.max(body.scrollWidth, html.scrollWidth),
                        height: height,
                        devicePixelRatio: window.devicePixelRatio
                    }));
                });
            })()"#
                .to_string()
        } else {
            r#"JSON.stringify({
                width: window.innerWidth,
                height: window.innerHeight,
                devicePixelRatio: window.devicePixelRatio
            })"#
            .to_string()
        };

        let dims_result = self.evaluate(&script).await?;

        // Since we can't easily capture actual screenshots via CDP HTTP API,
        // we use a JavaScript-based approach to extract page content as text
        let page_text = self.get_page_text().await?;

        Ok(format!(
            "Screenshot dimensions: {}\n\nPage content:\n{}",
            dims_result,
            if page_text.len() > 5000 {
                format!("{}...\n[truncated at 5000 chars]", &page_text[..5000])
            } else {
                page_text
            }
        ))
    }

    /// Extract text from a specific element (or full page)
    async fn extract_text(&self, selector: Option<&str>) -> ToolResultValue<String> {
        let script = match selector {
            Some(sel) => format!(
                r#"(() => {{
                    const el = document.querySelector('{}');
                    if (!el) throw new Error('Element not found: {}');
                    return el.innerText || el.textContent || '';
                }})()"#,
                sel.replace('\'', "\\'"),
                sel.replace('\'', "\\'"),
            ),
            None => r#"document.body.innerText || document.body.textContent || ''"#.to_string(),
        };

        self.evaluate(&script).await
    }

    /// Get the full HTML of the page (or a specific element)
    async fn get_html(&self, selector: Option<&str>) -> ToolResultValue<String> {
        let script = match selector {
            Some(sel) => format!(
                r#"(() => {{
                    const el = document.querySelector('{}');
                    if (!el) throw new Error('Element not found: {}');
                    return el.outerHTML;
                }})()"#,
                sel.replace('\'', "\\'"),
                sel.replace('\'', "\\'"),
            ),
            None => r#"document.documentElement.outerHTML"#.to_string(),
        };

        self.evaluate(&script).await
    }

    /// Get the visible text of the page
    async fn get_page_text(&self) -> ToolResultValue<String> {
        self.evaluate("document.body.innerText || document.body.textContent || ''")
            .await
    }

    /// Scroll the page
    async fn scroll(&self, direction: &str) -> ToolResultValue<String> {
        let script = match direction {
            "down" => r#"window.scrollBy(0, window.innerHeight * 0.8); return 'Scrolled down';"#,
            "up" => r#"window.scrollBy(0, -window.innerHeight * 0.8); return 'Scrolled up';"#,
            "to_bottom" => {
                r#"window.scrollTo(0, document.body.scrollHeight); return 'Scrolled to bottom';"#
            }
            "to_top" => r#"window.scrollTo(0, 0); return 'Scrolled to top';"#,
            _ => {
                return Err(ToolError::InvalidArguments(
                    "browser".to_string(),
                    format!(
                        "unknown scroll direction '{}'. Valid: down, up, to_bottom, to_top",
                        direction
                    ),
                ))
            }
        };

        self.evaluate(script).await
    }

    /// Evaluate JavaScript in the page context
    async fn evaluate(&self, script: &str) -> ToolResultValue<String> {
        let ws_url = self.get_ws_url().await?;
        let target_id = ws_url.rsplit('/').next().unwrap_or("").to_string();

        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_millis(self.request_timeout))
            .build()
            .map_err(|e| {
                ToolError::ExecutionFailed("browser".to_string(), format!("HTTP client: {}", e))
            })?;

        // Use the /json/activate/{id} endpoint to ensure the target is active
        let _ = client
            .post(format!(
                "{}/json/activate/{}",
                self.cdp_url.trim_end_matches('/'),
                target_id
            ))
            .send()
            .await;

        // For JavaScript evaluation, we use the /json/evaluate endpoint
        // This is a simplified approach — full CDP would use WebSocket
        let eval_url = format!(
            "{}/json/evaluate/{}?{}",
            self.cdp_url.trim_end_matches('/'),
            target_id,
            urlencoding(script)
        );

        let response = client.get(&eval_url).send().await.map_err(|e| {
            ToolError::ExecutionFailed(
                "browser".to_string(),
                format!("JavaScript evaluation failed: {}", e),
            )
        })?;

        let body_text = response.text().await.unwrap_or_default();
        let result: serde_json::Value =
            serde_json::from_str(&body_text).unwrap_or(serde_json::json!({
                "result": body_text
            }));

        // Extract the result value
        let output = result["result"]["result"]["value"]
            .as_str()
            .or_else(|| result["result"].as_str())
            .map(|s| s.to_string())
            .unwrap_or_else(|| serde_json::to_string_pretty(&result).unwrap_or_default());

        Ok(output)
    }
}

// ── HTML extraction helpers ────────────────────────────────────────────────

/// Extract readable content from a URL (HTML-to-text)
async fn fetch_and_extract_content(url: &str, max_bytes: usize) -> ToolResultValue<String> {
    let client = reqwest::Client::builder()
        .timeout(std::time::Duration::from_secs(15))
        .user_agent("Mozilla/5.0 (compatible; RavenClaws/0.9.2)")
        .build()
        .map_err(|e| {
            ToolError::ExecutionFailed("web_fetch".to_string(), format!("HTTP client: {}", e))
        })?;

    let response = client.get(url).send().await.map_err(|e| {
        ToolError::ExecutionFailed("web_fetch".to_string(), format!("Request failed: {}", e))
    })?;

    if !response.status().is_success() {
        return Err(ToolError::ExecutionFailed(
            "web_fetch".to_string(),
            format!("HTTP {}", response.status().as_u16()),
        ));
    }

    let body = response.text().await.map_err(|e| {
        ToolError::ExecutionFailed(
            "web_fetch".to_string(),
            format!("Failed to read response: {}", e),
        )
    })?;

    Ok(html_to_text(&body, max_bytes))
}

/// Convert HTML to readable text by stripping tags and extracting meaningful content
fn html_to_text(html: &str, max_chars: usize) -> String {
    let mut text = String::new();
    let bytes = html.as_bytes();
    let len = bytes.len();
    let mut i = 0;
    let mut in_tag = false;
    let mut in_script = false;
    let mut in_style = false;
    let mut in_title = false;
    let mut title_text = String::new();
    let mut last_char_was_space = true;

    while i < len {
        if in_script {
            // Look for </script>
            if i + 8 < len && bytes[i..i + 9].eq_ignore_ascii_case(b"</script>") {
                in_script = false;
                i += 9;
                continue;
            }
            i += 1;
            continue;
        }

        if in_style {
            // Look for </style>
            if i + 7 < len && bytes[i..i + 8].eq_ignore_ascii_case(b"</style>") {
                in_style = false;
                i += 8;
                continue;
            }
            i += 1;
            continue;
        }

        if in_title {
            // Look for </title>
            if i + 7 < len && bytes[i..i + 8].eq_ignore_ascii_case(b"</title>") {
                in_title = false;
                i += 8;
                continue;
            }
            title_text.push(bytes[i] as char);
            i += 1;
            continue;
        }

        if in_tag {
            if bytes[i] == b'>' {
                in_tag = false;
                // Check for <br> and <p> tags — add newline
                if i >= 2 {
                    let tag_start = (0..i).rev().find(|&j| bytes[j] == b'<').unwrap_or(0);
                    let tag_content = &html[tag_start..i].to_lowercase();
                    if (tag_content.starts_with("<br")
                        || tag_content.starts_with("<p")
                        || tag_content.starts_with("<tr")
                        || tag_content.starts_with("<div")
                        || tag_content.starts_with("<li")
                        || tag_content.starts_with("<h1")
                        || tag_content.starts_with("<h2")
                        || tag_content.starts_with("<h3")
                        || tag_content.starts_with("<h4")
                        || tag_content.starts_with("<h5")
                        || tag_content.starts_with("<h6"))
                        && !last_char_was_space
                    {
                        text.push('\n');
                        last_char_was_space = true;
                    }
                }
            } else {
                // Check for <script, <style, <title
                if bytes[i] == b's' || bytes[i] == b'S' {
                    if i + 5 < len && bytes[i..i + 6].eq_ignore_ascii_case(b"script") {
                        in_script = true;
                    } else if i + 4 < len && bytes[i..i + 5].eq_ignore_ascii_case(b"style") {
                        in_style = true;
                    } else if i + 4 < len && bytes[i..i + 5].eq_ignore_ascii_case(b"title") {
                        in_title = true;
                    }
                }
            }
            i += 1;
            continue;
        }

        if bytes[i] == b'<' {
            in_tag = true;
            i += 1;
            continue;
        }

        // Decode common HTML entities
        if bytes[i] == b'&' {
            let remaining = len - i;
            let entity = if remaining > 5 && &html[i..i + 6] == "&nbsp;" {
                i += 6;
                " "
            } else if remaining > 3 && &html[i..i + 4] == "&lt;" {
                i += 4;
                "<"
            } else if remaining > 3 && &html[i..i + 4] == "&gt;" {
                i += 4;
                ">"
            } else if remaining > 4 && &html[i..i + 5] == "&amp;" {
                i += 5;
                "&"
            } else if remaining > 5 && &html[i..i + 6] == "&quot;" {
                i += 6;
                "\""
            } else if remaining > 3 && &html[i..i + 4] == "&#39;" {
                i += 4;
                "'"
            } else {
                i += 1;
                continue;
            };

            if text.len() >= max_chars {
                break;
            }
            text.push_str(entity);
            last_char_was_space = entity == " ";
            continue;
        }

        // Normalize whitespace
        if bytes[i].is_ascii_whitespace() {
            if !last_char_was_space {
                text.push(' ');
                last_char_was_space = true;
            }
            i += 1;
            continue;
        }

        if text.len() >= max_chars {
            break;
        }
        text.push(bytes[i] as char);
        last_char_was_space = false;
        i += 1;
    }

    // Prepend title if found
    let title_text = title_text.trim();
    let text = text.trim();

    if !title_text.is_empty() {
        format!("Title: {}\n\n{}", title_text, text)
    } else {
        text.to_string()
    }
}

/// Strip HTML tags from a string (for snippet extraction)
fn strip_html_tags(input: &str) -> String {
    let mut output = String::new();
    let mut in_tag = false;
    for c in input.chars() {
        match c {
            '<' => in_tag = true,
            '>' => in_tag = false,
            _ => {
                if !in_tag {
                    output.push(c);
                }
            }
        }
    }
    // Decode common entities
    output
        .replace("&amp;", "&")
        .replace("&lt;", "<")
        .replace("&gt;", ">")
        .replace("&quot;", "\"")
        .replace("&#39;", "'")
        .replace("&nbsp;", " ")
}

/// Extract href value from an <a> tag
fn extract_href(a_tag: &str) -> Option<String> {
    let href_start = a_tag.find("href=\"")?;
    let value_start = href_start + 6;
    let value_end = a_tag[value_start..].find('"')?;
    let href = &a_tag[value_start..value_start + value_end];

    // DuckDuckGo redirect URLs
    if href.starts_with("//") {
        return Some(format!("https:{}", href));
    }
    if href.starts_with("/") {
        return None; // Relative URLs, skip
    }

    Some(href.to_string())
}

/// URL-encode a string for use in query parameters
fn urlencoding(input: &str) -> String {
    let mut result = String::with_capacity(input.len() * 3);
    for byte in input.bytes() {
        match byte {
            b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'-' | b'_' | b'.' | b'~' => {
                result.push(byte as char);
            }
            b' ' => result.push_str("%20"),
            _ => {
                result.push_str(&format!("%{:02X}", byte));
            }
        }
    }
    result
}

// ── Text-based tool call detection ──────────────────────────────────────────

/// Detects tool calls in natural language text when the LLM doesn't emit
/// structured `tool_calls`. This is a fallback for models that describe
/// tool usage in prose rather than structured function calling.
///
/// # Supported Patterns
///
/// - `Use the <tool> tool with args <args>` — explicit tool invocation
/// - `I'll use the <tool> tool to run: <command>` — shell command pattern
/// - `Let me read the file <path>` — read file pattern
/// - `I'll search for <query>` — web search pattern
/// - `I'll fetch <url>` — web fetch pattern
///
/// # Example
///
/// ```ignore
/// let detector = ToolCallDetector::new();
/// let response = "I'll use the shell_exec tool to run: ls -la";
/// let calls = detector.detect(response);
/// assert_eq!(calls.len(), 1);
/// assert_eq!(calls[0].name, "shell_exec");
/// ```
#[allow(dead_code)]
pub struct ToolCallDetector {
    patterns: Vec<DetectorPattern>,
}

/// A single detection pattern with a regex and a parser function
#[allow(dead_code)]
struct DetectorPattern {
    /// The regex pattern to match
    regex: regex_lite::Regex,
    /// The tool name to use if matched (or None to extract from capture)
    tool_name: Option<String>,
    /// The argument key to set (or None to use capture group name)
    arg_key: Option<String>,
    /// Capture group index for the argument value
    arg_group: usize,
}

#[allow(dead_code)]
impl ToolCallDetector {
    /// Create a new detector with all built-in patterns
    pub fn new() -> Self {
        // These patterns handle common LLM tool invocation styles
        let patterns = vec![
            // Pattern: "Use the <tool> tool with args <json>"
            // Note: Must NOT start with I'll/I will/let me to avoid overlap with the next pattern
            DetectorPattern {
                regex: regex_lite::Regex::new(
                    r"(?i)(?:^|[.!?]\s+)(?:use|run|call|invoke)\s+(?:the\s+)?(\w+)\s+(?:tool|command|function)(?:\s+with\s+(?:args|arguments|parameters))?\s*:?\s*(.+?)(?:\.|$|\n)"
                ).expect("valid regex"),
                tool_name: None, // extracted from capture group 1
                arg_key: None,
                arg_group: 2,
            },
            // Pattern: "I'll use the <tool> tool to run: <command>"
            DetectorPattern {
                regex: regex_lite::Regex::new(
                    r"(?i)(?:I'?ll|I\s+will|let\s+me)\s+use\s+(?:the\s+)?(\w+)\s+(?:tool|command|function)\s+to\s+(?:run|execute|do)\s*:?\s*(.+?)(?:\.|$|\n)"
                ).expect("valid regex"),
                tool_name: None,
                arg_key: Some("command".to_string()),
                arg_group: 2,
            },
            // Pattern: "Let me read the file <path>"
            DetectorPattern {
                regex: regex_lite::Regex::new(
                    r"(?i)(?:let\s+me|I'?ll|I\s+will)\s+(?:read|open|check)\s+(?:the\s+)?file\s+(.+?)(?:\.|$|\n)"
                ).expect("valid regex"),
                tool_name: Some("read_file".to_string()),
                arg_key: Some("path".to_string()),
                arg_group: 1,
            },
            // Pattern: "I'll search for <query>"
            DetectorPattern {
                regex: regex_lite::Regex::new(
                    r"(?i)(?:let\s+me|I'?ll|I\s+will)\s+(?:search|look\s+up|find|google)\s+(?:for\s+)?(.+?)(?:\.|$|\n)"
                ).expect("valid regex"),
                tool_name: Some("web_search".to_string()),
                arg_key: Some("query".to_string()),
                arg_group: 1,
            },
            // Pattern: "I'll fetch <url>"
            DetectorPattern {
                regex: regex_lite::Regex::new(
                    r"(?i)(?:let\s+me|I'?ll|I\s+will)\s+(?:fetch|get|download)\s+(https?://\S+)(?:\.|$|\n|\s)"
                ).expect("valid regex"),
                tool_name: Some("web_fetch".to_string()),
                arg_key: Some("url".to_string()),
                arg_group: 1,
            },
        ];

        Self { patterns }
    }

    /// Detect tool calls in a response text.
    /// Returns a list of detected `ToolCall` structs.
    /// Deduplicates calls with the same tool name and arguments.
    pub fn detect(&self, text: &str) -> Vec<ToolCall> {
        let mut seen = std::collections::HashSet::new();
        let mut calls = Vec::new();

        for pattern in &self.patterns {
            for cap in pattern.regex.captures_iter(text) {
                let tool_name = match &pattern.tool_name {
                    Some(name) => name.clone(),
                    None => cap
                        .get(1)
                        .map(|m| m.as_str().to_string())
                        .unwrap_or_default(),
                };

                // Skip if tool name doesn't match any known tool
                if !Self::is_known_tool(&tool_name) {
                    continue;
                }

                let arg_value = cap
                    .get(pattern.arg_group)
                    .map(|m| m.as_str().trim().to_string())
                    .unwrap_or_default();

                if arg_value.is_empty() {
                    continue;
                }

                // Build arguments JSON
                let arguments = match &pattern.arg_key {
                    Some(key) => {
                        serde_json::json!({ key: arg_value })
                    }
                    None => {
                        // Try to parse as JSON, otherwise wrap as "command" or "input"
                        serde_json::from_str(&arg_value).unwrap_or_else(
                            |_| serde_json::json!({ "command": arg_value, "input": arg_value }),
                        )
                    }
                };

                // Deduplicate: skip if we've already seen this tool+args combo
                let key = format!("{}:{:?}", tool_name, arguments);
                if seen.contains(&key) {
                    continue;
                }
                seen.insert(key);

                calls.push(ToolCall {
                    name: tool_name,
                    arguments,
                    id: None,
                });
            }
        }

        calls
    }

    /// Check if a tool name is one of the known built-in tools
    fn is_known_tool(name: &str) -> bool {
        matches!(
            name,
            "shell_exec" | "read_file" | "write_file" | "web_fetch" | "web_search" | "browser"
        )
    }
}

impl Default for ToolCallDetector {
    fn default() -> Self {
        Self::new()
    }
}

// ── Helper functions ───────────────────────────────────────────────────────

/// Run a shell command with timeout
async fn run_shell_command(
    command: &str,
    timeout_secs: u64,
    workdir: Option<String>,
) -> ToolResultValue<ToolResult> {
    use tokio::process::Command;

    let shell = if cfg!(target_os = "windows") {
        "cmd.exe"
    } else {
        "sh"
    };
    let flag = if cfg!(target_os = "windows") {
        "/C"
    } else {
        "-c"
    };

    let mut cmd = Command::new(shell);
    cmd.arg(flag).arg(command);

    if let Some(dir) = &workdir {
        cmd.current_dir(dir);
    }

    let output = tokio::time::timeout(std::time::Duration::from_secs(timeout_secs), cmd.output())
        .await
        .map_err(|_| {
            ToolError::ExecutionFailed(
                "shell_exec".to_string(),
                format!("Command timed out after {} seconds", timeout_secs),
            )
        })?
        .map_err(|e| {
            ToolError::ExecutionFailed(
                "shell_exec".to_string(),
                format!("Failed to execute: {}", e),
            )
        })?;

    let stdout = String::from_utf8_lossy(&output.stdout).to_string();
    let stderr = String::from_utf8_lossy(&output.stderr).to_string();
    let exit_code = output.status.code().unwrap_or(-1);

    let mut output_text = String::new();
    if !stdout.is_empty() {
        output_text.push_str(&stdout);
    }
    if !stderr.is_empty() {
        if !output_text.is_empty() {
            output_text.push_str("\n--- stderr ---\n");
        }
        output_text.push_str(&stderr);
    }

    // Truncate very long output
    const MAX_OUTPUT: usize = 65536;
    if output_text.len() > MAX_OUTPUT {
        output_text = format!(
            "{}...\n[truncated at {} bytes]",
            &output_text[..MAX_OUTPUT],
            MAX_OUTPUT
        );
    }

    Ok(ToolResult {
        tool_name: "shell_exec".to_string(),
        success: exit_code == 0,
        output: output_text,
        error: if exit_code != 0 {
            Some(format!("Exit code: {}", exit_code))
        } else {
            None
        },
        exit_code: Some(exit_code),
        duration_ms: None,
    })
}

// ── Tests ──────────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_tool_registry_empty() {
        let registry = ToolRegistry::new();
        assert!(registry.is_empty());
        assert_eq!(registry.len(), 0);
    }

    #[test]
    fn test_tool_registry_register() {
        let mut registry = ToolRegistry::new();
        registry.register(Arc::new(ShellTool::new()));
        assert!(!registry.is_empty());
        assert_eq!(registry.len(), 1);
        assert!(registry.has("shell_exec"));
    }

    #[test]
    fn test_tool_registry_default_tools() {
        let registry = ToolRegistry::with_default_tools();
        assert_eq!(registry.len(), 6);
        assert!(registry.has("shell_exec"));
        assert!(registry.has("read_file"));
        assert!(registry.has("write_file"));
        assert!(registry.has("web_fetch"));
        assert!(registry.has("web_search"));
        assert!(registry.has("browser"));
    }

    #[test]
    fn test_tool_definitions() {
        let registry = ToolRegistry::with_default_tools();
        let defs = registry.definitions();
        assert_eq!(defs.len(), 6);

        let shell_def = defs.iter().find(|d| d.name == "shell_exec").unwrap();
        assert!(shell_def.description.contains("shell command"));
        assert!(shell_def.requires_approval);
        assert_eq!(shell_def.category, ToolCategory::Shell);
    }

    #[test]
    fn test_tool_not_found() {
        let registry = ToolRegistry::new();
        let result = registry.get("nonexistent");
        assert!(result.is_none());
    }

    #[test]
    fn test_shell_tool_definition() {
        let tool = ShellTool::new();
        let def = tool.definition();
        assert_eq!(def.name, "shell_exec");
        assert!(def.requires_approval);
    }

    #[test]
    fn test_read_file_tool_definition() {
        let tool = ReadFileTool::new();
        let def = tool.definition();
        assert_eq!(def.name, "read_file");
        assert!(!def.requires_approval);
    }

    #[test]
    fn test_write_file_tool_definition() {
        let tool = WriteFileTool::new();
        let def = tool.definition();
        assert_eq!(def.name, "write_file");
        assert!(def.requires_approval);
    }

    #[test]
    fn test_web_fetch_tool_definition() {
        let tool = WebFetchTool::new();
        let def = tool.definition();
        assert_eq!(def.name, "web_fetch");
        assert!(!def.requires_approval);
    }

    #[test]
    fn test_tool_call_serialization() {
        let call = ToolCall {
            name: "shell_exec".to_string(),
            arguments: serde_json::json!({"command": "echo hello"}),
            id: Some("call_123".to_string()),
        };

        let json = serde_json::to_string(&call).unwrap();
        assert!(json.contains("shell_exec"));
        assert!(json.contains("echo hello"));
        assert!(json.contains("call_123"));
    }

    #[test]
    fn test_tool_result_serialization() {
        let result = ToolResult {
            tool_name: "shell_exec".to_string(),
            success: true,
            output: "hello\n".to_string(),
            error: None,
            exit_code: Some(0),
            duration_ms: Some(42),
        };

        let json = serde_json::to_string(&result).unwrap();
        assert!(json.contains("shell_exec"));
        assert!(json.contains("hello"));
        assert!(json.contains("42"));
    }

    #[test]
    fn test_tool_result_failure() {
        let result = ToolResult {
            tool_name: "shell_exec".to_string(),
            success: false,
            output: String::new(),
            error: Some("Exit code: 1".to_string()),
            exit_code: Some(1),
            duration_ms: Some(10),
        };

        assert!(!result.success);
        assert_eq!(result.exit_code, Some(1));
    }

    #[test]
    fn test_json_schema_string() {
        let schema = JsonSchema::string("A test string");
        assert_eq!(schema.schema_type, "string");
        assert_eq!(schema.description.unwrap(), "A test string");
    }

    #[test]
    fn test_json_schema_object() {
        let mut props = HashMap::new();
        props.insert("name".to_string(), JsonSchema::string("The name"));
        let schema = JsonSchema::object(props, vec!["name".to_string()]);
        assert_eq!(schema.schema_type, "object");
        assert!(schema.properties.unwrap().contains_key("name"));
    }

    #[test]
    fn test_tool_error_not_found() {
        let err = ToolError::NotFound("test_tool".to_string());
        assert_eq!(format!("{}", err), "Tool 'test_tool' not found");
    }

    #[test]
    fn test_tool_error_execution_failed() {
        let err = ToolError::ExecutionFailed("test".to_string(), "oops".to_string());
        assert_eq!(format!("{}", err), "Tool 'test' execution failed: oops");
    }

    #[test]
    fn test_tool_error_invalid_arguments() {
        let err = ToolError::InvalidArguments("test".to_string(), "bad arg".to_string());
        assert_eq!(
            format!("{}", err),
            "Invalid arguments for tool 'test': bad arg"
        );
    }

    #[test]
    fn test_tool_error_policy_denied() {
        let err = ToolError::PolicyDenied("not allowed".to_string());
        assert_eq!(format!("{}", err), "Policy denied: not allowed");
    }

    #[test]
    fn test_tool_error_sandbox_violation() {
        let err = ToolError::SandboxViolation("escape attempt".to_string());
        assert_eq!(format!("{}", err), "Sandbox violation: escape attempt");
    }

    #[test]
    fn test_tool_category_default() {
        let cat = ToolCategory::default();
        assert_eq!(cat, ToolCategory::General);
    }

    #[test]
    fn test_tool_category_serialization() {
        let cat = ToolCategory::Shell;
        let json = serde_json::to_string(&cat).unwrap();
        assert_eq!(json, "\"Shell\"");
    }

    #[test]
    fn test_tool_definition_requires_approval_default() {
        let def = ToolDefinition {
            name: "test".to_string(),
            description: "test".to_string(),
            parameters: JsonSchema::string("test"),
            requires_approval: false,
            category: ToolCategory::General,
        };
        assert!(!def.requires_approval);
    }

    #[tokio::test]
    async fn test_shell_exec_success() {
        let tool = ShellTool::new();
        let args = serde_json::json!({"command": "echo hello"});
        let result = tool.execute(args).await.unwrap();
        assert!(result.success);
        assert!(result.output.contains("hello"));
        assert_eq!(result.exit_code, Some(0));
    }

    #[tokio::test]
    async fn test_shell_exec_failure() {
        let tool = ShellTool::new();
        let args = serde_json::json!({"command": "exit 42"});
        let result = tool.execute(args).await.unwrap();
        assert!(!result.success);
        assert_eq!(result.exit_code, Some(42));
    }

    #[tokio::test]
    async fn test_shell_exec_missing_command() {
        let tool = ShellTool::new();
        let args = serde_json::json!({});
        let err = tool.execute(args).await.unwrap_err();
        assert!(matches!(err, ToolError::InvalidArguments(_, _)));
    }

    #[tokio::test]
    async fn test_read_file_not_found() {
        let tool = ReadFileTool::new();
        let args = serde_json::json!({"path": "/tmp/nonexistent_file_ravenclaws_test"});
        let result = tool.execute(args).await;
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::ExecutionFailed(_, _)
        ));
    }

    #[tokio::test]
    async fn test_read_file_missing_path() {
        let tool = ReadFileTool::new();
        let args = serde_json::json!({});
        let err = tool.execute(args).await.unwrap_err();
        assert!(matches!(err, ToolError::InvalidArguments(_, _)));
    }

    #[tokio::test]
    async fn test_write_file_missing_args() {
        let tool = WriteFileTool::new();
        let args = serde_json::json!({});
        let err = tool.execute(args).await.unwrap_err();
        assert!(matches!(err, ToolError::InvalidArguments(_, _)));
    }

    #[tokio::test]
    async fn test_web_fetch_missing_url() {
        let tool = WebFetchTool::new();
        let args = serde_json::json!({});
        let err = tool.execute(args).await.unwrap_err();
        assert!(matches!(err, ToolError::InvalidArguments(_, _)));
    }

    #[tokio::test]
    async fn test_write_and_read_file() {
        let dir = std::env::temp_dir().join(format!("ravenclaws_test_{}", std::process::id()));
        let path = dir.join("test_write.txt");
        let path_str = path.to_string_lossy().to_string();

        // Write
        let write_tool = WriteFileTool::new();
        let args = serde_json::json!({"path": path_str, "content": "Hello, RavenClaws!"});
        let result = write_tool.execute(args).await.unwrap();
        assert!(result.success);
        assert!(result.output.contains("18 bytes"));

        // Read back
        let read_tool = ReadFileTool::new();
        let args = serde_json::json!({"path": path_str});
        let result = read_tool.execute(args).await.unwrap();
        assert!(result.success);
        assert_eq!(result.output.trim(), "Hello, RavenClaws!");

        // Cleanup
        let _ = tokio::fs::remove_file(&path).await;
        let _ = tokio::fs::remove_dir(dir).await;
    }

    #[tokio::test]
    async fn test_write_file_append() {
        let dir = std::env::temp_dir().join(format!("ravenclaws_test_{}", std::process::id()));
        let path = dir.join("test_append.txt");
        let path_str = path.to_string_lossy().to_string();

        // Write initial
        let write_tool = WriteFileTool::new();
        let args = serde_json::json!({"path": path_str, "content": "line1\n"});
        write_tool.execute(args).await.unwrap();

        // Append
        let args = serde_json::json!({"path": path_str, "content": "line2\n", "append": true});
        let result = write_tool.execute(args).await.unwrap();
        assert!(result.success);

        // Read back
        let read_tool = ReadFileTool::new();
        let args = serde_json::json!({"path": path_str});
        let result = read_tool.execute(args).await.unwrap();
        assert!(result.success);
        assert!(result.output.contains("line1"));
        assert!(result.output.contains("line2"));

        // Cleanup
        let _ = tokio::fs::remove_file(&path).await;
        let _ = tokio::fs::remove_dir(dir).await;
    }

    #[tokio::test]
    async fn test_tool_registry_execute() {
        let registry = ToolRegistry::with_default_tools();
        let call = ToolCall {
            name: "shell_exec".to_string(),
            arguments: serde_json::json!({"command": "echo hello"}),
            id: None,
        };
        let result = registry.execute(call).await.unwrap();
        assert!(result.success);
        assert!(result.output.contains("hello"));
    }

    #[tokio::test]
    async fn test_tool_registry_execute_not_found() {
        let registry = ToolRegistry::new();
        let call = ToolCall {
            name: "nonexistent".to_string(),
            arguments: serde_json::json!({}),
            id: None,
        };
        let err = registry.execute(call).await.unwrap_err();
        assert!(matches!(err, ToolError::NotFound(_)));
    }

    // ── Web search tool tests ──────────────────────────────────────────────

    #[test]
    fn test_web_search_tool_definition() {
        let tool = WebSearchTool::new();
        let def = tool.definition();
        assert_eq!(def.name, "web_search");
        assert!(!def.requires_approval);
        assert_eq!(def.category, ToolCategory::WebSearch);
        assert!(def.description.contains("Search the web"));
    }

    #[test]
    fn test_web_search_tool_with_config() {
        let tool = WebSearchTool::with_config(
            "http://localhost:8888".to_string(),
            "searxng".to_string(),
            10,
            false,
        );
        let def = tool.definition();
        assert_eq!(def.name, "web_search");
        assert_eq!(tool.search_endpoint, "http://localhost:8888");
        assert_eq!(tool.search_engine, "searxng");
        assert_eq!(tool.max_results, 10);
        assert!(!tool.fetch_content);
    }

    #[tokio::test]
    async fn test_web_search_missing_query() {
        let tool = WebSearchTool::new();
        let args = serde_json::json!({});
        let err = tool.execute(args).await.unwrap_err();
        assert!(matches!(err, ToolError::InvalidArguments(_, _)));
    }

    #[test]
    fn test_web_search_tool_registry() {
        let registry = ToolRegistry::with_default_tools();
        assert!(registry.has("web_search"));
        let defs = registry.definitions();
        let search_def = defs.iter().find(|d| d.name == "web_search").unwrap();
        assert_eq!(search_def.category, ToolCategory::WebSearch);
    }

    #[test]
    fn test_web_search_tool_with_config_registry() {
        let registry =
            ToolRegistry::with_web_search_config("http://localhost:8888", "searxng", 10, false);
        assert!(registry.has("web_search"));
        assert!(registry.has("shell_exec"));
        assert!(registry.has("read_file"));
        assert!(registry.has("write_file"));
        assert!(registry.has("web_fetch"));
        assert!(registry.has("browser"));
        assert_eq!(registry.len(), 6);
    }

    // ── HTML extraction tests ──────────────────────────────────────────────

    #[test]
    fn test_html_to_text_strips_tags() {
        let html = "<html><body><p>Hello, world!</p></body></html>";
        let text = html_to_text(html, 1000);
        assert!(text.contains("Hello, world!"));
        assert!(!text.contains("<p>"));
        assert!(!text.contains("</p>"));
    }

    #[test]
    fn test_html_to_text_extracts_title() {
        let html = "<html><head><title>Test Page</title></head><body><p>Content</p></body></html>";
        let text = html_to_text(html, 1000);
        assert!(text.contains("Test Page"));
        assert!(text.contains("Content"));
    }

    #[test]
    fn test_html_to_text_strips_script_and_style() {
        let html = "<html><head><script>alert('xss');</script><style>.cls{}</style></head><body><p>Visible</p></body></html>";
        let text = html_to_text(html, 1000);
        assert!(text.contains("Visible"));
        assert!(!text.contains("alert"));
        assert!(!text.contains(".cls"));
    }

    #[test]
    fn test_html_to_text_handles_entities() {
        let html = "<p>foo &amp; bar &lt; baz &gt; qux</p>";
        let text = html_to_text(html, 1000);
        assert!(text.contains("foo & bar < baz > qux") || text.contains("foo & bar"));
    }

    #[test]
    fn test_html_to_text_respects_max_chars() {
        let html = "<p>Hello World This Is A Test</p>";
        let text = html_to_text(html, 5);
        assert!(text.len() <= 5);
    }

    #[test]
    fn test_html_to_text_empty_input() {
        assert_eq!(html_to_text("", 1000), "");
    }

    #[test]
    fn test_html_to_text_no_html() {
        let text = html_to_text("Just plain text", 1000);
        assert_eq!(text, "Just plain text");
    }

    #[test]
    fn test_strip_html_tags_basic() {
        let result = strip_html_tags("<b>bold</b> and <i>italic</i>");
        assert_eq!(result, "bold and italic");
    }

    #[test]
    fn test_strip_html_tags_with_entities() {
        let result = strip_html_tags("foo &amp; bar &lt; baz");
        assert_eq!(result, "foo & bar < baz");
    }

    #[test]
    fn test_extract_href_basic() {
        let result = extract_href(r#"<a href="https://example.com">link</a>"#);
        assert_eq!(result, Some("https://example.com".to_string()));
    }

    #[test]
    fn test_extract_href_protocol_relative() {
        let result = extract_href(r#"<a href="//example.com/path">link</a>"#);
        assert_eq!(result, Some("https://example.com/path".to_string()));
    }

    #[test]
    fn test_extract_href_relative() {
        let result = extract_href(r#"<a href="/relative/path">link</a>"#);
        assert_eq!(result, None);
    }

    #[test]
    fn test_extract_href_no_match() {
        let result = extract_href("<span>no link here</span>");
        assert_eq!(result, None);
    }

    #[test]
    fn test_urlencoding_basic() {
        assert_eq!(urlencoding("hello world"), "hello%20world");
        assert_eq!(urlencoding("foo/bar"), "foo%2Fbar");
        assert_eq!(urlencoding("simple"), "simple");
    }

    #[test]
    fn test_fetch_and_extract_content_invalid_url() {
        let result = tokio_test::block_on(fetch_and_extract_content("http://0.0.0.0:1", 1000));
        assert!(result.is_err());
    }

    // ── ToolCallDetector tests ─────────────────────────────────────────────

    #[test]
    fn test_tool_call_detector_shell_exec() {
        let detector = ToolCallDetector::new();
        let text = "I'll use the shell_exec tool to run: ls -la";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 1, "Should detect one tool call");
        assert_eq!(calls[0].name, "shell_exec");
        assert_eq!(calls[0].arguments["command"], "ls -la");
    }

    #[test]
    fn test_tool_call_detector_read_file() {
        let detector = ToolCallDetector::new();
        let text = "Let me read the file /etc/hostname";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 1, "Should detect one tool call");
        assert_eq!(calls[0].name, "read_file");
        assert_eq!(calls[0].arguments["path"], "/etc/hostname");
    }

    #[test]
    fn test_tool_call_detector_web_search() {
        let detector = ToolCallDetector::new();
        let text = "I'll search for Rust programming language";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 1, "Should detect one tool call");
        assert_eq!(calls[0].name, "web_search");
        assert!(calls[0].arguments["query"]
            .as_str()
            .unwrap()
            .contains("Rust"));
    }

    #[test]
    fn test_tool_call_detector_web_fetch() {
        let detector = ToolCallDetector::new();
        let text = "I'll fetch https://example.com/api";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 1, "Should detect one tool call");
        assert_eq!(calls[0].name, "web_fetch");
        assert_eq!(calls[0].arguments["url"], "https://example.com/api");
    }

    #[test]
    fn test_tool_call_detector_use_tool_syntax() {
        let detector = ToolCallDetector::new();
        let text = "Use the shell_exec tool with args: echo hello world";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 1, "Should detect one tool call");
        assert_eq!(calls[0].name, "shell_exec");
    }

    #[test]
    fn test_tool_call_detector_no_false_positives() {
        let detector = ToolCallDetector::new();
        let text = "I think we should consider using a different approach here.";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 0, "Should not detect any tool calls");
    }

    #[test]
    fn test_tool_call_detector_empty_text() {
        let detector = ToolCallDetector::new();
        let calls = detector.detect("");
        assert_eq!(calls.len(), 0);
    }

    #[test]
    fn test_tool_call_detector_multiple_calls() {
        let detector = ToolCallDetector::new();
        let text = "Let me read the file /etc/hosts. Then I'll search for DNS configuration.";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 2, "Should detect two tool calls");
        assert_eq!(calls[0].name, "read_file");
        assert_eq!(calls[1].name, "web_search");
    }

    #[test]
    fn test_tool_call_detector_unknown_tool_skipped() {
        let detector = ToolCallDetector::new();
        let text = "Use the nonexistent_tool tool with args: something";
        let calls = detector.detect(text);
        assert_eq!(calls.len(), 0, "Should skip unknown tools");
    }

    #[test]
    fn test_tool_call_detector_is_known_tool() {
        assert!(ToolCallDetector::is_known_tool("shell_exec"));
        assert!(ToolCallDetector::is_known_tool("read_file"));
        assert!(ToolCallDetector::is_known_tool("write_file"));
        assert!(ToolCallDetector::is_known_tool("web_fetch"));
        assert!(ToolCallDetector::is_known_tool("web_search"));
        assert!(!ToolCallDetector::is_known_tool("unknown_tool"));
    }

    #[test]
    fn test_tool_call_detector_default() {
        let detector = ToolCallDetector::default();
        let calls = detector.detect("I'll use the shell_exec tool to run: echo test");
        assert_eq!(calls.len(), 1);
    }

    // ── Browser tool tests ─────────────────────────────────────────────────

    #[test]
    fn test_browser_tool_definition() {
        let tool = BrowserTool::new();
        let def = tool.definition();
        assert_eq!(def.name, "browser");
        assert!(def.requires_approval);
        assert_eq!(def.category, ToolCategory::Browser);
        assert!(def.description.contains("Chrome DevTools Protocol"));
    }

    #[test]
    fn test_browser_tool_with_config() {
        let tool = BrowserTool::with_config("http://localhost:9999".to_string(), 15000);
        assert_eq!(tool.cdp_url, "http://localhost:9999");
        assert_eq!(tool.request_timeout, 15000);
    }

    #[test]
    fn test_browser_tool_default_config() {
        let tool = BrowserTool::new();
        assert_eq!(tool.cdp_url, "http://127.0.0.1:9222");
        assert_eq!(tool.request_timeout, 30000);
    }

    #[test]
    fn test_browser_tool_registry() {
        let registry = ToolRegistry::with_default_tools();
        assert!(registry.has("browser"));
        let defs = registry.definitions();
        let browser_def = defs.iter().find(|d| d.name == "browser").unwrap();
        assert_eq!(browser_def.category, ToolCategory::Browser);
    }

    #[test]
    fn test_browser_tool_missing_action() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::InvalidArguments(_, _)
        ));
    }

    #[test]
    fn test_browser_tool_invalid_action() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "invalid_action"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        let err = result.unwrap_err();
        assert!(matches!(err, ToolError::InvalidArguments(_, _)));
        assert!(format!("{}", err).contains("unknown action"));
    }

    #[test]
    fn test_browser_tool_navigate_missing_url() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "navigate"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::InvalidArguments(_, _)
        ));
    }

    #[test]
    fn test_browser_tool_click_missing_selector() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "click"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::InvalidArguments(_, _)
        ));
    }

    #[test]
    fn test_browser_tool_type_missing_args() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "type"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::InvalidArguments(_, _)
        ));
    }

    #[test]
    fn test_browser_tool_type_missing_text() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "type", "selector": "#input"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::InvalidArguments(_, _)
        ));
    }

    #[test]
    fn test_browser_tool_evaluate_missing_script() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "evaluate"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            ToolError::InvalidArguments(_, _)
        ));
    }

    #[test]
    fn test_browser_tool_scroll_invalid_direction() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "scroll", "direction": "sideways"});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_err());
        assert!(format!("{}", result.unwrap_err()).contains("unknown scroll direction"));
    }

    #[test]
    fn test_browser_tool_wait_action() {
        let tool = BrowserTool::new();
        let args = serde_json::json!({"action": "wait", "wait_ms": 10});
        let result = tokio_test::block_on(tool.execute(args));
        assert!(result.is_ok());
        let result = result.unwrap();
        assert!(result.success);
        assert!(result.output.contains("Waited for"));
    }

    #[test]
    fn test_browser_tool_is_known_tool() {
        assert!(ToolCallDetector::is_known_tool("browser"));
    }

    #[test]
    fn test_browser_tool_category_serialization() {
        let cat = ToolCategory::Browser;
        let json = serde_json::to_string(&cat).unwrap();
        assert_eq!(json, "\"Browser\"");
    }
}