| 1 | use std::collections::HashMap; |
| 2 | use std::path::PathBuf; |
| 3 | use std::sync::Arc; |
| 4 | use std::time::Duration; |
| 5 | |
| 6 | use anyhow::Result; |
| 7 | use async_trait::async_trait; |
| 8 | use codewhale_protocol::{ToolKind, ToolOutput, ToolPayload}; |
| 9 | use serde::{Deserialize, Serialize}; |
| 10 | use serde_json::Value; |
| 11 | use tokio::sync::{OwnedRwLockReadGuard, OwnedRwLockWriteGuard, RwLock}; |
| 12 | |
| 13 | mod outcome; |
| 14 | mod prepared; |
| 15 | mod resources; |
| 16 | |
| 17 | pub use outcome::{ToolExecutionOutcome, ToolTerminalStatus}; |
| 18 | pub use prepared::PreparedToolCall; |
| 19 | pub use resources::{ResourceClaim, schedule_non_conflicting}; |
| 20 | |
| 21 | tokio::task_local! { |
| 22 | static TOOL_EXECUTION_LOCK_HELD: (); |
| 23 | } |
| 24 | |
| 25 | /// Capabilities that a tool may have or require. |
| 26 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)] |
| 27 | pub enum ToolCapability { |
| 28 | /// Tool only reads data, never modifies state. |
| 29 | ReadOnly, |
| 30 | /// Tool writes to the filesystem. |
| 31 | WritesFiles, |
| 32 | /// Tool executes arbitrary shell commands. |
| 33 | ExecutesCode, |
| 34 | /// Tool makes network requests. |
| 35 | Network, |
| 36 | /// Tool can be run in a sandbox. |
| 37 | Sandboxable, |
| 38 | /// Tool requires user approval before execution. |
| 39 | RequiresApproval, |
| 40 | } |
| 41 | |
| 42 | /// Approval requirement for a tool. |
| 43 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)] |
| 44 | pub enum ApprovalRequirement { |
| 45 | /// Never needs approval: safe read-only operations. |
| 46 | #[default] |
| 47 | Auto, |
| 48 | /// Suggest approval but allow user to skip. |
| 49 | Suggest, |
| 50 | /// Always require explicit user approval. |
| 51 | Required, |
| 52 | } |
| 53 | |
| 54 | /// Errors that can occur during tool execution. |
| 55 | #[derive(Debug, Clone, thiserror::Error)] |
| 56 | pub enum ToolError { |
| 57 | #[error("Failed to validate input: {message}")] |
| 58 | InvalidInput { message: String }, |
| 59 | #[error("Failed to validate input: missing required field '{field}'")] |
| 60 | MissingField { field: String }, |
| 61 | #[error("Failed to resolve path '{}': path escapes workspace", path.display())] |
| 62 | PathEscape { path: PathBuf }, |
| 63 | #[error("Failed to execute tool: {message}")] |
| 64 | ExecutionFailed { message: String }, |
| 65 | #[error("Failed to execute tool: operation timed out after {seconds}s")] |
| 66 | Timeout { seconds: u64 }, |
| 67 | #[error("Tool execution cancelled: {message}")] |
| 68 | Cancelled { message: String }, |
| 69 | #[error("Failed to locate tool: {message}")] |
| 70 | NotAvailable { message: String }, |
| 71 | #[error("Failed to authorize tool execution: {message}")] |
| 72 | PermissionDenied { message: String }, |
| 73 | } |
| 74 | |
| 75 | impl ToolError { |
| 76 | #[must_use] |
| 77 | pub fn invalid_input(msg: impl Into<String>) -> Self { |
| 78 | Self::InvalidInput { |
| 79 | message: msg.into(), |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | #[must_use] |
| 84 | pub fn missing_field(field: impl Into<String>) -> Self { |
| 85 | Self::MissingField { |
| 86 | field: field.into(), |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | #[must_use] |
| 91 | pub fn execution_failed(msg: impl Into<String>) -> Self { |
| 92 | Self::ExecutionFailed { |
| 93 | message: msg.into(), |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | #[must_use] |
| 98 | pub fn cancelled(msg: impl Into<String>) -> Self { |
| 99 | Self::Cancelled { |
| 100 | message: msg.into(), |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | #[must_use] |
| 105 | pub fn path_escape(path: impl Into<PathBuf>) -> Self { |
| 106 | Self::PathEscape { path: path.into() } |
| 107 | } |
| 108 | |
| 109 | #[must_use] |
| 110 | pub fn not_available(msg: impl Into<String>) -> Self { |
| 111 | Self::NotAvailable { |
| 112 | message: msg.into(), |
| 113 | } |
| 114 | } |
| 115 | |
| 116 | #[must_use] |
| 117 | pub fn permission_denied(msg: impl Into<String>) -> Self { |
| 118 | Self::PermissionDenied { |
| 119 | message: msg.into(), |
| 120 | } |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | /// Result of a tool execution. |
| 125 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 126 | pub struct ToolResult { |
| 127 | /// The output content, which may be JSON or plain text. |
| 128 | pub content: String, |
| 129 | /// Whether the execution was successful. |
| 130 | pub success: bool, |
| 131 | /// Optional structured metadata. |
| 132 | #[serde(skip_serializing_if = "Option::is_none")] |
| 133 | pub metadata: Option<Value>, |
| 134 | } |
| 135 | |
| 136 | impl ToolResult { |
| 137 | /// Create a successful result with content. |
| 138 | #[must_use] |
| 139 | pub fn success(content: impl Into<String>) -> Self { |
| 140 | Self { |
| 141 | content: content.into(), |
| 142 | success: true, |
| 143 | metadata: None, |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | /// Create an error result with message. |
| 148 | #[must_use] |
| 149 | pub fn error(message: impl Into<String>) -> Self { |
| 150 | Self { |
| 151 | content: message.into(), |
| 152 | success: false, |
| 153 | metadata: None, |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | /// Create a successful result from JSON. |
| 158 | pub fn json<T: Serialize>(value: &T) -> std::result::Result<Self, serde_json::Error> { |
| 159 | Ok(Self { |
| 160 | content: serde_json::to_string(value)?, |
| 161 | success: true, |
| 162 | metadata: None, |
| 163 | }) |
| 164 | } |
| 165 | |
| 166 | /// Add metadata to the result. |
| 167 | #[must_use] |
| 168 | pub fn with_metadata(mut self, metadata: Value) -> Self { |
| 169 | self.metadata = Some(metadata); |
| 170 | self |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | /// Name the JSON type of a value the way a tool schema would spell it. |
| 175 | #[must_use] |
| 176 | pub fn json_type_name(value: &Value) -> &'static str { |
| 177 | match value { |
| 178 | Value::Null => "null", |
| 179 | Value::Bool(_) => "boolean", |
| 180 | Value::Number(_) => "number", |
| 181 | Value::String(_) => "string", |
| 182 | Value::Array(_) => "array", |
| 183 | Value::Object(_) => "object", |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | /// Render a value for an error message, truncated so a huge payload cannot |
| 188 | /// swamp the transcript. |
| 189 | #[must_use] |
| 190 | pub fn value_preview(value: &Value) -> String { |
| 191 | let preview = value.to_string(); |
| 192 | if preview.chars().count() > 120 { |
| 193 | preview.chars().take(117).collect::<String>() + "..." |
| 194 | } else { |
| 195 | preview |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | /// The one error every type mismatch on a tool parameter produces. |
| 200 | /// |
| 201 | /// Names the parameter, the type that arrived, and the type the schema |
| 202 | /// declares, plus the offending value — everything the caller needs to fix |
| 203 | /// the call on the next turn without another round trip. |
| 204 | #[must_use] |
| 205 | pub fn type_mismatch(field: &str, value: &Value, expected: &str) -> ToolError { |
| 206 | ToolError::invalid_input(format!( |
| 207 | "field '{field}' must be {expected}; got {}. Received: {}", |
| 208 | json_type_name(value), |
| 209 | value_preview(value) |
| 210 | )) |
| 211 | } |
| 212 | |
| 213 | /// Whether a value counts as "the caller did not supply this field". |
| 214 | /// |
| 215 | /// JSON `null` is the wire spelling of absence, so an optional field set to |
| 216 | /// `null` takes its default rather than erroring. This is the *only* |
| 217 | /// tolerance in the optional extractors, and it is uniform across all of |
| 218 | /// them: `null` means no value, and no value is exactly what a default is |
| 219 | /// for. Every other type mismatch is an error. |
| 220 | fn is_absent(value: Option<&Value>) -> bool { |
| 221 | matches!(value, None | Some(Value::Null)) |
| 222 | } |
| 223 | |
| 224 | /// Helper to extract a required string field from JSON input. |
| 225 | pub fn required_str<'a>(input: &'a Value, field: &str) -> std::result::Result<&'a str, ToolError> { |
| 226 | if let Some(value) = input.get(field) { |
| 227 | if let Some(string_value) = value.as_str() { |
| 228 | return Ok(string_value); |
| 229 | } |
| 230 | |
| 231 | return Err(type_mismatch(field, value, "a string")); |
| 232 | } |
| 233 | |
| 234 | // When the field is missing, list the fields the caller *did* |
| 235 | // supply so the model can spot the mismatch without a retry. |
| 236 | let provided: Vec<&str> = input |
| 237 | .as_object() |
| 238 | .map(|obj| obj.keys().map(|k| k.as_str()).collect()) |
| 239 | .unwrap_or_default(); |
| 240 | if provided.is_empty() { |
| 241 | Err(ToolError::missing_field(field)) |
| 242 | } else { |
| 243 | let hint = format!( |
| 244 | "missing required field '{field}'. Input provided: {}", |
| 245 | provided.join(", ") |
| 246 | ); |
| 247 | Err(ToolError::invalid_input(hint)) |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | /// Helper to extract an optional string field from JSON input. |
| 252 | /// |
| 253 | /// A wrong type is an error, never a silent `None`. See [`type_mismatch`] |
| 254 | /// for why nothing is coerced. |
| 255 | pub fn optional_str<'a>( |
| 256 | input: &'a Value, |
| 257 | field: &str, |
| 258 | ) -> std::result::Result<Option<&'a str>, ToolError> { |
| 259 | let value = input.get(field); |
| 260 | if is_absent(value) { |
| 261 | return Ok(None); |
| 262 | } |
| 263 | let value = value.expect("is_absent covers the None case"); |
| 264 | value |
| 265 | .as_str() |
| 266 | .map(Some) |
| 267 | .ok_or_else(|| type_mismatch(field, value, "a string")) |
| 268 | } |
| 269 | |
| 270 | /// Helper to extract a required u64 field from JSON input. |
| 271 | /// |
| 272 | /// Absence (field missing or `null`) is a `missing_field` error; a value |
| 273 | /// that is present but not a u64 is a [`type_mismatch`] naming the field and |
| 274 | /// the expected type, so the caller fixes the field's type instead of |
| 275 | /// re-sending it as missing. |
| 276 | pub fn required_u64(input: &Value, field: &str) -> std::result::Result<u64, ToolError> { |
| 277 | let value = input.get(field); |
| 278 | if is_absent(value) { |
| 279 | return Err(ToolError::missing_field(field)); |
| 280 | } |
| 281 | let value = value.expect("is_absent covers the None case"); |
| 282 | value |
| 283 | .as_u64() |
| 284 | .ok_or_else(|| type_mismatch(field, value, "a non-negative integer")) |
| 285 | } |
| 286 | |
| 287 | /// Helper to extract an optional u64 field with default. |
| 288 | /// |
| 289 | /// A wrong type is an error, never a silent fall back to `default`. |
| 290 | pub fn optional_u64( |
| 291 | input: &Value, |
| 292 | field: &str, |
| 293 | default: u64, |
| 294 | ) -> std::result::Result<u64, ToolError> { |
| 295 | let value = input.get(field); |
| 296 | if is_absent(value) { |
| 297 | return Ok(default); |
| 298 | } |
| 299 | let value = value.expect("is_absent covers the None case"); |
| 300 | value |
| 301 | .as_u64() |
| 302 | .ok_or_else(|| type_mismatch(field, value, "a non-negative integer")) |
| 303 | } |
| 304 | |
| 305 | /// Helper to extract an optional bool field with default. |
| 306 | /// |
| 307 | /// A wrong type is an error, never a silent fall back to `default`. In |
| 308 | /// particular the string `"true"` is refused rather than coerced: the |
| 309 | /// default this used to fall back to is frequently the *opposite* of what |
| 310 | /// the caller asked for, and some of those defaults gate irreversible |
| 311 | /// actions. |
| 312 | pub fn optional_bool( |
| 313 | input: &Value, |
| 314 | field: &str, |
| 315 | default: bool, |
| 316 | ) -> std::result::Result<bool, ToolError> { |
| 317 | Ok(optional_bool_opt(input, field)?.unwrap_or(default)) |
| 318 | } |
| 319 | |
| 320 | /// Helper to extract an optional bool that has no default. |
| 321 | /// |
| 322 | /// `None` means the caller did not supply the field; a wrong type is an |
| 323 | /// error. Use this where "unset" is itself meaningful — an authority |
| 324 | /// declaration that is dropped instead of read is a restriction that |
| 325 | /// silently evaporates. |
| 326 | pub fn optional_bool_opt( |
| 327 | input: &Value, |
| 328 | field: &str, |
| 329 | ) -> std::result::Result<Option<bool>, ToolError> { |
| 330 | let value = input.get(field); |
| 331 | if is_absent(value) { |
| 332 | return Ok(None); |
| 333 | } |
| 334 | let value = value.expect("is_absent covers the None case"); |
| 335 | value |
| 336 | .as_bool() |
| 337 | .map(Some) |
| 338 | .ok_or_else(|| type_mismatch(field, value, "a boolean")) |
| 339 | } |
| 340 | |
| 341 | /// Descriptor that describes a tool available in the registry. |
| 342 | /// |
| 343 | /// Contains the tool's name, its JSON input/output schemas, and |
| 344 | /// execution constraints such as timeout and parallelism. |
| 345 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 346 | pub struct ToolDescriptor { |
| 347 | /// Unique name used to look up the tool. |
| 348 | pub name: String, |
| 349 | /// JSON Schema describing the tool's expected input parameters. |
| 350 | pub input_schema: Value, |
| 351 | /// JSON Schema describing the tool's output format. |
| 352 | pub output_schema: Value, |
| 353 | /// Whether multiple invocations of this tool may run concurrently. |
| 354 | pub supports_parallel_tool_calls: bool, |
| 355 | /// Optional per-call timeout in milliseconds; `None` means no timeout. |
| 356 | pub timeout_ms: Option<u64>, |
| 357 | } |
| 358 | |
| 359 | /// A [`ToolDescriptor`] together with its runtime configuration. |
| 360 | /// |
| 361 | /// Wraps a `ToolDescriptor` and exposes the parallelism flag directly so the |
| 362 | /// dispatcher can check it without digging into the inner spec. |
| 363 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 364 | pub struct ConfiguredToolDescriptor { |
| 365 | /// The underlying tool descriptor. |
| 366 | pub spec: ToolDescriptor, |
| 367 | /// Whether this tool supports concurrent invocations. |
| 368 | pub supports_parallel_tool_calls: bool, |
| 369 | } |
| 370 | |
| 371 | /// Identifies where a tool call originated from. |
| 372 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 373 | #[serde(rename_all = "snake_case")] |
| 374 | pub enum ToolCallSource { |
| 375 | /// Direct invocation from the model or user. |
| 376 | Direct, |
| 377 | /// Invocation through the JavaScript REPL environment. |
| 378 | JsRepl, |
| 379 | } |
| 380 | |
| 381 | /// A tool invocation request before it has been validated and dispatched. |
| 382 | /// |
| 383 | /// Contains the tool name, its input payload, and metadata about where the |
| 384 | /// call originated. |
| 385 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 386 | pub struct ToolCall { |
| 387 | /// Name of the tool to invoke. |
| 388 | pub name: String, |
| 389 | /// The input payload for the tool. |
| 390 | pub payload: ToolPayload, |
| 391 | /// Where this call originated (direct or REPL). |
| 392 | pub source: ToolCallSource, |
| 393 | /// Optional raw tool-call identifier from the upstream provider. |
| 394 | pub raw_tool_call_id: Option<String>, |
| 395 | } |
| 396 | |
| 397 | impl ToolCall { |
| 398 | /// Derive the execution subject for this call. |
| 399 | /// |
| 400 | /// For local shell payloads this returns the shell command and its |
| 401 | /// working directory; for all other payloads the tool name and the |
| 402 | /// provided `fallback_cwd` are returned instead. The third element |
| 403 | /// of the tuple is a human-readable kind label (`"shell"` or `"tool"`). |
| 404 | pub fn execution_subject(&self, fallback_cwd: &str) -> (String, String, &'static str) { |
| 405 | match &self.payload { |
| 406 | ToolPayload::LocalShell { params } => ( |
| 407 | params.command.clone(), |
| 408 | params |
| 409 | .cwd |
| 410 | .clone() |
| 411 | .unwrap_or_else(|| fallback_cwd.to_string()), |
| 412 | "shell", |
| 413 | ), |
| 414 | _ => (self.name.clone(), fallback_cwd.to_string(), "tool"), |
| 415 | } |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | /// A validated tool invocation ready to be handled. |
| 420 | /// |
| 421 | /// Created by the registry after a [`ToolCall`] passes validation, this |
| 422 | /// carries all the context a [`ToolHandler`] needs to execute the tool. |
| 423 | #[derive(Debug, Clone)] |
| 424 | pub struct ToolInvocation { |
| 425 | /// Unique identifier for this invocation (generated or from the provider). |
| 426 | pub call_id: String, |
| 427 | /// Name of the tool being invoked. |
| 428 | pub tool_name: String, |
| 429 | /// The input payload for the tool. |
| 430 | pub payload: ToolPayload, |
| 431 | /// Where this invocation originated. |
| 432 | pub source: ToolCallSource, |
| 433 | } |
| 434 | |
| 435 | /// Errors that can occur during tool dispatch and execution. |
| 436 | /// |
| 437 | /// Unlike [`ToolError`], which represents input validation failures within |
| 438 | /// a tool, `FunctionCallError` covers problems at the dispatch layer: the |
| 439 | /// tool was not found, its kind did not match, it was rejected because it |
| 440 | /// is mutating, it timed out, was cancelled, or its handler returned an |
| 441 | /// error. |
| 442 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 443 | pub enum FunctionCallError { |
| 444 | /// No tool with the given name is registered. |
| 445 | ToolNotFound { name: String }, |
| 446 | /// The payload kind does not match the handler's expected kind. |
| 447 | KindMismatch { expected: ToolKind, got: ToolKind }, |
| 448 | /// The tool is mutating but `allow_mutating` was `false`. |
| 449 | MutatingToolRejected { name: String }, |
| 450 | /// The tool execution exceeded its configured timeout. |
| 451 | TimedOut { name: String, timeout_ms: u64 }, |
| 452 | /// The tool execution was cancelled. |
| 453 | Cancelled { name: String }, |
| 454 | /// The tool handler returned an error. |
| 455 | ExecutionFailed { name: String, error: String }, |
| 456 | } |
| 457 | |
| 458 | /// Trait implemented by concrete tool handlers. |
| 459 | /// |
| 460 | /// Each registered tool is backed by a handler that reports its kind, |
| 461 | /// whether it is mutating, and performs the actual execution. |
| 462 | #[async_trait] |
| 463 | pub trait ToolHandler: Send + Sync { |
| 464 | /// The [`ToolKind`] this handler expects (e.g. `Function` or `Mcp`). |
| 465 | fn kind(&self) -> ToolKind; |
| 466 | |
| 467 | /// Returns `true` if `kind` matches this handler's expected kind. |
| 468 | /// |
| 469 | /// The default implementation compares against [`kind()`](ToolHandler::kind). |
| 470 | fn matches_kind(&self, kind: ToolKind) -> bool { |
| 471 | self.kind() == kind |
| 472 | } |
| 473 | |
| 474 | /// Whether this tool performs side-effects that require user approval. |
| 475 | /// |
| 476 | /// Defaults to `false` (read-only / safe). |
| 477 | fn is_mutating(&self) -> bool { |
| 478 | false |
| 479 | } |
| 480 | |
| 481 | /// Execute the tool with the given invocation context. |
| 482 | async fn handle( |
| 483 | &self, |
| 484 | invocation: ToolInvocation, |
| 485 | ) -> std::result::Result<ToolOutput, FunctionCallError>; |
| 486 | } |
| 487 | |
| 488 | /// Manages concurrent tool execution via a read/write lock. |
| 489 | /// |
| 490 | /// Parallel-safe tools acquire a read lock (allowing overlap), while |
| 491 | /// serial tools acquire a write lock (exclusive access). Reentrant calls |
| 492 | /// (e.g. a tool invoking another tool) skip locking to avoid deadlock. |
| 493 | #[derive(Debug)] |
| 494 | pub struct ToolCallRuntime { |
| 495 | execution_lock: Arc<RwLock<()>>, |
| 496 | } |
| 497 | |
| 498 | impl Default for ToolCallRuntime { |
| 499 | fn default() -> Self { |
| 500 | Self { |
| 501 | execution_lock: Arc::new(RwLock::new(())), |
| 502 | } |
| 503 | } |
| 504 | } |
| 505 | |
| 506 | #[derive(Debug)] |
| 507 | enum ToolExecutionGuard { |
| 508 | Parallel(#[allow(dead_code)] OwnedRwLockReadGuard<()>), |
| 509 | Serial(#[allow(dead_code)] OwnedRwLockWriteGuard<()>), |
| 510 | Reentrant, |
| 511 | } |
| 512 | |
| 513 | impl ToolCallRuntime { |
| 514 | async fn acquire(&self, supports_parallel: bool) -> ToolExecutionGuard { |
| 515 | if TOOL_EXECUTION_LOCK_HELD.try_with(|_| ()).is_ok() { |
| 516 | return ToolExecutionGuard::Reentrant; |
| 517 | } |
| 518 | |
| 519 | if supports_parallel { |
| 520 | ToolExecutionGuard::Parallel(self.execution_lock.clone().read_owned().await) |
| 521 | } else { |
| 522 | ToolExecutionGuard::Serial(self.execution_lock.clone().write_owned().await) |
| 523 | } |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | /// Central registry that maps tool names to their specs and handlers. |
| 528 | /// |
| 529 | /// Use [`register()`](ToolRegistry::register) to add tools, then |
| 530 | /// [`dispatch()`](ToolRegistry::dispatch) to invoke them. The registry |
| 531 | /// owns a [`ToolCallRuntime`] that manages concurrent execution. |
| 532 | #[derive(Default)] |
| 533 | pub struct ToolRegistry { |
| 534 | handlers: HashMap<String, Arc<dyn ToolHandler>>, |
| 535 | specs: HashMap<String, ConfiguredToolDescriptor>, |
| 536 | runtime: ToolCallRuntime, |
| 537 | } |
| 538 | |
| 539 | impl ToolRegistry { |
| 540 | /// Register a tool with its specification and handler. |
| 541 | /// |
| 542 | /// The tool's name is taken from `spec.name`. Returns an error if |
| 543 | /// registration fails (currently infallible, but the `Result` is |
| 544 | /// reserved for future validation). |
| 545 | pub fn register(&mut self, spec: ToolDescriptor, handler: Arc<dyn ToolHandler>) -> Result<()> { |
| 546 | let name = spec.name.clone(); |
| 547 | self.specs.insert( |
| 548 | name.clone(), |
| 549 | ConfiguredToolDescriptor { |
| 550 | supports_parallel_tool_calls: spec.supports_parallel_tool_calls, |
| 551 | spec, |
| 552 | }, |
| 553 | ); |
| 554 | self.handlers.insert(name, handler); |
| 555 | Ok(()) |
| 556 | } |
| 557 | |
| 558 | /// Return the configured specs for every registered tool. |
| 559 | pub fn list_specs(&self) -> Vec<ConfiguredToolDescriptor> { |
| 560 | self.specs.values().cloned().collect() |
| 561 | } |
| 562 | |
| 563 | /// Validate and execute a tool call. |
| 564 | /// |
| 565 | /// Looks up the tool by name, verifies the payload kind matches the |
| 566 | /// handler, enforces the `allow_mutating` guard, acquires the |
| 567 | /// appropriate execution lock, and forwards the call to the handler. |
| 568 | /// Returns a [`FunctionCallError`] if any validation step fails or |
| 569 | /// the handler returns an error. |
| 570 | pub async fn dispatch( |
| 571 | &self, |
| 572 | call: ToolCall, |
| 573 | allow_mutating: bool, |
| 574 | ) -> std::result::Result<ToolOutput, FunctionCallError> { |
| 575 | let handler = self.handlers.get(&call.name).cloned().ok_or_else(|| { |
| 576 | FunctionCallError::ToolNotFound { |
| 577 | name: call.name.clone(), |
| 578 | } |
| 579 | })?; |
| 580 | let configured = |
| 581 | self.specs |
| 582 | .get(&call.name) |
| 583 | .cloned() |
| 584 | .ok_or_else(|| FunctionCallError::ToolNotFound { |
| 585 | name: call.name.clone(), |
| 586 | })?; |
| 587 | |
| 588 | let payload_kind = tool_payload_kind(&call.payload); |
| 589 | let expected = handler.kind(); |
| 590 | if !handler.matches_kind(payload_kind) { |
| 591 | return Err(FunctionCallError::KindMismatch { |
| 592 | expected, |
| 593 | got: payload_kind, |
| 594 | }); |
| 595 | } |
| 596 | if handler.is_mutating() && !allow_mutating { |
| 597 | return Err(FunctionCallError::MutatingToolRejected { name: call.name }); |
| 598 | } |
| 599 | |
| 600 | let invocation = ToolInvocation { |
| 601 | call_id: call |
| 602 | .raw_tool_call_id |
| 603 | .clone() |
| 604 | .unwrap_or_else(|| format!("tool-call-{}", uuid::Uuid::new_v4())), |
| 605 | tool_name: call.name.clone(), |
| 606 | payload: call.payload, |
| 607 | source: call.source, |
| 608 | }; |
| 609 | |
| 610 | let _guard = self |
| 611 | .runtime |
| 612 | .acquire(configured.supports_parallel_tool_calls) |
| 613 | .await; |
| 614 | |
| 615 | TOOL_EXECUTION_LOCK_HELD |
| 616 | .scope( |
| 617 | (), |
| 618 | self.execute_with_timeout(handler, configured.spec.timeout_ms, invocation), |
| 619 | ) |
| 620 | .await |
| 621 | } |
| 622 | |
| 623 | async fn execute_with_timeout( |
| 624 | &self, |
| 625 | handler: Arc<dyn ToolHandler>, |
| 626 | timeout_ms: Option<u64>, |
| 627 | invocation: ToolInvocation, |
| 628 | ) -> std::result::Result<ToolOutput, FunctionCallError> { |
| 629 | if let Some(timeout_ms) = timeout_ms { |
| 630 | let name = invocation.tool_name.clone(); |
| 631 | match tokio::time::timeout( |
| 632 | Duration::from_millis(timeout_ms), |
| 633 | handler.handle(invocation), |
| 634 | ) |
| 635 | .await |
| 636 | { |
| 637 | Ok(result) => result, |
| 638 | Err(_) => Err(FunctionCallError::TimedOut { name, timeout_ms }), |
| 639 | } |
| 640 | } else { |
| 641 | handler.handle(invocation).await |
| 642 | } |
| 643 | } |
| 644 | } |
| 645 | |
| 646 | fn tool_payload_kind(payload: &ToolPayload) -> ToolKind { |
| 647 | match payload { |
| 648 | ToolPayload::Mcp { .. } => ToolKind::Mcp, |
| 649 | ToolPayload::Function { .. } |
| 650 | | ToolPayload::Custom { .. } |
| 651 | | ToolPayload::LocalShell { .. } => ToolKind::Function, |
| 652 | } |
| 653 | } |
| 654 | |
| 655 | #[cfg(test)] |
| 656 | mod tests { |
| 657 | use serde_json::json; |
| 658 | |
| 659 | use super::*; |
| 660 | |
| 661 | #[test] |
| 662 | fn tool_result_success_sets_plain_content() { |
| 663 | let content = "operation completed successfully"; |
| 664 | let result = ToolResult::success(content); |
| 665 | |
| 666 | assert!(result.success); |
| 667 | assert_eq!(result.content, content); |
| 668 | assert!(result.metadata.is_none()); |
| 669 | } |
| 670 | |
| 671 | #[test] |
| 672 | fn tool_result_json_round_trips_content() { |
| 673 | let result = ToolResult::json(&json!({"ok": true})).expect("json"); |
| 674 | assert!(result.success); |
| 675 | let content: serde_json::Value = |
| 676 | serde_json::from_str(&result.content).expect("content is valid json"); |
| 677 | assert_eq!(content, json!({"ok": true})); |
| 678 | } |
| 679 | |
| 680 | #[test] |
| 681 | fn helper_extractors_validate_shape() { |
| 682 | let input = json!({"name": "demo", "count": 7, "enabled": true}); |
| 683 | assert_eq!(required_str(&input, "name").expect("name"), "demo"); |
| 684 | assert_eq!(optional_str(&input, "name").unwrap(), Some("demo")); |
| 685 | assert_eq!(optional_str(&input, "missing").unwrap(), None); |
| 686 | assert_eq!(optional_str(&json!({"name": null}), "name").unwrap(), None); |
| 687 | assert_eq!(optional_u64(&input, "count", 0).unwrap(), 7); |
| 688 | assert!(optional_bool(&input, "enabled", false).unwrap()); |
| 689 | // "name" is present but a string: a type mismatch, not a missing |
| 690 | // field, so the caller fixes the type instead of re-sending the name. |
| 691 | let err = required_u64(&input, "name") |
| 692 | .expect_err("a present string is not a missing u64") |
| 693 | .to_string(); |
| 694 | assert!( |
| 695 | err.contains("field 'name' must be a non-negative integer"), |
| 696 | "{err}" |
| 697 | ); |
| 698 | } |
| 699 | |
| 700 | /// The rule, stated once: an optional parameter of the wrong JSON type is |
| 701 | /// an error that names the parameter, what arrived, and what was wanted. |
| 702 | /// `null` alone means "absent" and takes the default. |
| 703 | #[test] |
| 704 | fn optional_extractors_refuse_type_mismatches_instead_of_defaulting() { |
| 705 | // The shipping bug: a stringy "true" became the default `false`, |
| 706 | // which for `dry_run` is the opposite of what the caller asked and |
| 707 | // gates an irreversible action. |
| 708 | let err = optional_bool(&json!({"dry_run": "true"}), "dry_run", false) |
| 709 | .expect_err("a stringy bool must not become the default") |
| 710 | .to_string(); |
| 711 | assert!(err.contains("dry_run"), "{err}"); |
| 712 | assert!(err.contains("must be a boolean"), "{err}"); |
| 713 | assert!(err.contains("got string"), "{err}"); |
| 714 | assert!(err.contains("\"true\""), "{err}"); |
| 715 | |
| 716 | for bad in [json!("true"), json!(1), json!(0), json!([]), json!({})] { |
| 717 | assert!( |
| 718 | optional_bool(&json!({"flag": bad}), "flag", false).is_err(), |
| 719 | "optional_bool accepted {bad}" |
| 720 | ); |
| 721 | } |
| 722 | for bad in [json!("7"), json!(-1), json!(1.5), json!(true), json!([7])] { |
| 723 | assert!( |
| 724 | optional_u64(&json!({"n": bad}), "n", 42).is_err(), |
| 725 | "optional_u64 accepted {bad}" |
| 726 | ); |
| 727 | } |
| 728 | for bad in [json!(7), json!(true), json!(["a"]), json!({"a": 1})] { |
| 729 | assert!( |
| 730 | optional_str(&json!({"s": bad}), "s").is_err(), |
| 731 | "optional_str accepted {bad}" |
| 732 | ); |
| 733 | } |
| 734 | |
| 735 | // `null` is the wire spelling of absence, uniformly across all three. |
| 736 | assert!(optional_bool(&json!({"flag": null}), "flag", true).unwrap()); |
| 737 | assert_eq!(optional_u64(&json!({"n": null}), "n", 42).unwrap(), 42); |
| 738 | assert_eq!(optional_str(&json!({"s": null}), "s").unwrap(), None); |
| 739 | } |
| 740 | |
| 741 | #[test] |
| 742 | fn type_mismatch_truncates_a_huge_offending_value() { |
| 743 | let big = Value::String("x".repeat(500)); |
| 744 | let err = type_mismatch("body", &big, "a boolean").to_string(); |
| 745 | assert!(err.contains("body"), "{err}"); |
| 746 | assert!(err.ends_with("..."), "{err}"); |
| 747 | assert!(err.chars().count() < 250, "{err}"); |
| 748 | } |
| 749 | |
| 750 | #[test] |
| 751 | fn required_u64_distinguishes_missing_from_type_mismatch() { |
| 752 | // Absent (or null) is a missing-field error. |
| 753 | assert!(matches!( |
| 754 | required_u64(&json!({}), "count"), |
| 755 | Err(ToolError::MissingField { .. }) |
| 756 | )); |
| 757 | assert!(matches!( |
| 758 | required_u64(&json!({"count": null}), "count"), |
| 759 | Err(ToolError::MissingField { .. }) |
| 760 | )); |
| 761 | |
| 762 | // Present and valid values pass through, including the extremes. |
| 763 | assert_eq!(required_u64(&json!({"count": 42}), "count").unwrap(), 42); |
| 764 | assert_eq!( |
| 765 | required_u64(&json!({"count": u64::MAX}), "count").unwrap(), |
| 766 | u64::MAX |
| 767 | ); |
| 768 | |
| 769 | // Present but wrongly typed is a type mismatch naming the field and |
| 770 | // the expected type — never a missing-field misdirection. |
| 771 | for value in [json!(-1), json!(2.5), json!("42")] { |
| 772 | let err = required_u64(&json!({"count": value}), "count") |
| 773 | .expect_err("wrong type must not look missing") |
| 774 | .to_string(); |
| 775 | assert!( |
| 776 | err.contains("field 'count' must be a non-negative integer"), |
| 777 | "{err}" |
| 778 | ); |
| 779 | } |
| 780 | } |
| 781 | |
| 782 | #[test] |
| 783 | fn required_str_reports_provided_fields_on_missing_required_field() { |
| 784 | let input = json!({"path": "src/lib.rs", "content": "new body"}); |
| 785 | let err = required_str(&input, "replace").expect_err("replace is missing"); |
| 786 | let message = err.to_string(); |
| 787 | assert!(message.contains("missing required field 'replace'")); |
| 788 | assert!(message.contains("Input provided:")); |
| 789 | assert!(message.contains("path")); |
| 790 | assert!(message.contains("content")); |
| 791 | } |
| 792 | |
| 793 | #[test] |
| 794 | fn required_str_reports_wrong_type_when_field_exists() { |
| 795 | let input = json!({"replace": [{"path": "src/lib.rs", "content": "new body"}]}); |
| 796 | let err = required_str(&input, "replace").expect_err("replace has wrong type"); |
| 797 | let message = err.to_string(); |
| 798 | assert!(message.contains("field 'replace' must be a string")); |
| 799 | assert!(message.contains("got array")); |
| 800 | assert!(message.contains(r#""content":"new body""#)); |
| 801 | assert!(message.contains(r#""path":"src/lib.rs""#)); |
| 802 | } |
| 803 | |
| 804 | #[test] |
| 805 | fn tool_error_display_matches_legacy_text() { |
| 806 | let err = ToolError::missing_field("path"); |
| 807 | assert_eq!( |
| 808 | err.to_string(), |
| 809 | "Failed to validate input: missing required field 'path'" |
| 810 | ); |
| 811 | } |
| 812 | |
| 813 | #[test] |
| 814 | fn tool_error_missing_field_constructor() { |
| 815 | let err = ToolError::missing_field("my_field"); |
| 816 | assert!(matches!(err, ToolError::MissingField { field } if field == "my_field")); |
| 817 | } |
| 818 | |
| 819 | #[test] |
| 820 | fn tool_error_not_available_displays_reason() { |
| 821 | let err = ToolError::not_available("custom tool not found"); |
| 822 | |
| 823 | assert!(matches!(err, ToolError::NotAvailable { .. })); |
| 824 | assert_eq!( |
| 825 | err.to_string(), |
| 826 | "Failed to locate tool: custom tool not found" |
| 827 | ); |
| 828 | } |
| 829 | |
| 830 | #[test] |
| 831 | fn tool_error_permission_denied_displays_reason() { |
| 832 | let err = ToolError::permission_denied("unauthorized user"); |
| 833 | |
| 834 | assert!(matches!(err, ToolError::PermissionDenied { .. })); |
| 835 | assert_eq!( |
| 836 | err.to_string(), |
| 837 | "Failed to authorize tool execution: unauthorized user" |
| 838 | ); |
| 839 | } |
| 840 | |
| 841 | #[test] |
| 842 | fn tool_error_execution_failed_displays_reason() { |
| 843 | let err = ToolError::execution_failed("process crashed"); |
| 844 | |
| 845 | assert!( |
| 846 | matches!(err, ToolError::ExecutionFailed { ref message } if message == "process crashed") |
| 847 | ); |
| 848 | assert_eq!(err.to_string(), "Failed to execute tool: process crashed"); |
| 849 | } |
| 850 | |
| 851 | #[test] |
| 852 | fn tool_error_invalid_input_creates_correct_variant() { |
| 853 | let err = ToolError::invalid_input("test invalid message"); |
| 854 | match err { |
| 855 | ToolError::InvalidInput { message } => { |
| 856 | assert_eq!(message, "test invalid message"); |
| 857 | } |
| 858 | _ => panic!("Expected ToolError::InvalidInput, got {err:?}"), |
| 859 | } |
| 860 | } |
| 861 | |
| 862 | #[test] |
| 863 | fn tool_error_path_escape_display() { |
| 864 | let path = std::path::PathBuf::from("../outside"); |
| 865 | let err = ToolError::path_escape(path); |
| 866 | assert_eq!( |
| 867 | err.to_string(), |
| 868 | "Failed to resolve path '../outside': path escapes workspace" |
| 869 | ); |
| 870 | } |
| 871 | |
| 872 | #[test] |
| 873 | fn tool_call_execution_subject_uses_local_shell_command_and_cwd() { |
| 874 | let call = ToolCall { |
| 875 | name: "shell".to_string(), |
| 876 | payload: ToolPayload::LocalShell { |
| 877 | params: codewhale_protocol::LocalShellParams { |
| 878 | command: "ls -l".to_string(), |
| 879 | cwd: Some("/custom/dir".to_string()), |
| 880 | timeout_ms: None, |
| 881 | }, |
| 882 | }, |
| 883 | source: ToolCallSource::Direct, |
| 884 | raw_tool_call_id: None, |
| 885 | }; |
| 886 | |
| 887 | assert_eq!( |
| 888 | call.execution_subject("/fallback/dir"), |
| 889 | ("ls -l".to_string(), "/custom/dir".to_string(), "shell") |
| 890 | ); |
| 891 | } |
| 892 | |
| 893 | #[test] |
| 894 | fn tool_call_execution_subject_falls_back_for_shell_without_cwd() { |
| 895 | let call = ToolCall { |
| 896 | name: "shell".to_string(), |
| 897 | payload: ToolPayload::LocalShell { |
| 898 | params: codewhale_protocol::LocalShellParams { |
| 899 | command: "echo hello".to_string(), |
| 900 | cwd: None, |
| 901 | timeout_ms: None, |
| 902 | }, |
| 903 | }, |
| 904 | source: ToolCallSource::Direct, |
| 905 | raw_tool_call_id: None, |
| 906 | }; |
| 907 | |
| 908 | assert_eq!( |
| 909 | call.execution_subject("/fallback/dir"), |
| 910 | ( |
| 911 | "echo hello".to_string(), |
| 912 | "/fallback/dir".to_string(), |
| 913 | "shell" |
| 914 | ) |
| 915 | ); |
| 916 | } |
| 917 | |
| 918 | #[test] |
| 919 | fn tool_call_execution_subject_uses_tool_name_for_non_shell_payloads() { |
| 920 | let call = ToolCall { |
| 921 | name: "my_tool".to_string(), |
| 922 | payload: ToolPayload::Function { |
| 923 | arguments: "{}".to_string(), |
| 924 | }, |
| 925 | source: ToolCallSource::Direct, |
| 926 | raw_tool_call_id: None, |
| 927 | }; |
| 928 | |
| 929 | assert_eq!( |
| 930 | call.execution_subject("/fallback/dir"), |
| 931 | ("my_tool".to_string(), "/fallback/dir".to_string(), "tool") |
| 932 | ); |
| 933 | } |
| 934 | } |
| 935 |