| 1 | //! The sandboxed QuickJS VM that executes Workflow scripts. |
| 2 | //! |
| 3 | //! Threading model (design §2.2): `rquickjs` contexts and every `'js` value |
| 4 | //! are `!Send`, so each run gets a dedicated OS thread with its own |
| 5 | //! current-thread tokio reactor. Host functions do no heavy work inline — |
| 6 | //! only `Send` data (JSON strings, [`TaskRequest`]s, oneshot replies) crosses |
| 7 | //! to the driver; conversion back into JS values happens on the VM thread |
| 8 | //! after the await resolves. |
| 9 | //! |
| 10 | //! Sandbox: the context registers only standard ECMAScript intrinsics plus |
| 11 | //! the Workflow globals (`task`, `parallel`, `pipeline`, `log`, `phase`, |
| 12 | //! `budget`, `args`). There is no module loader, no fs/net/process access, |
| 13 | //! and `Date`/`Math.random` are overridden to throw so recorded runs stay |
| 14 | //! deterministic for replay. |
| 15 | |
| 16 | use std::cell::Cell; |
| 17 | use std::env; |
| 18 | use std::rc::Rc; |
| 19 | use std::sync::atomic::{AtomicBool, Ordering}; |
| 20 | use std::sync::{Arc, OnceLock}; |
| 21 | |
| 22 | use rquickjs::function::{Async, Func}; |
| 23 | use rquickjs::{AsyncContext, AsyncRuntime, CatchResultExt, CaughtError, Ctx, Promise, Value}; |
| 24 | use serde::Deserialize; |
| 25 | use tokio::sync::{OwnedSemaphorePermit, Semaphore, oneshot, watch}; |
| 26 | |
| 27 | use crate::driver::{ProgressEvent, TaskCompletion, TaskRequest, WorkflowDriver}; |
| 28 | use crate::error::{TaskError, TaskErrorKind, WorkflowJsError}; |
| 29 | use crate::schema::{ |
| 30 | ReplyDecodeError, SCHEMA_REPAIR_MAX_ATTEMPTS, carried_raw, compile_schema, decode_reply, |
| 31 | repair_prompt, |
| 32 | }; |
| 33 | use crate::{ |
| 34 | CODEMODE_MAX_TOOL_CALLS, PARALLEL_MAX_ITEMS, ToolCallRequest, ToolInvoker, |
| 35 | WORKFLOW_LIFETIME_CAP, normalize_profile, |
| 36 | }; |
| 37 | |
| 38 | const DEFAULT_VM_MEMORY_LIMIT_BYTES: usize = 32 * 1024 * 1024; |
| 39 | const MIN_VM_MEMORY_LIMIT_BYTES: usize = 4 * 1024 * 1024; |
| 40 | const MAX_VM_MEMORY_LIMIT_BYTES: usize = 512 * 1024 * 1024; |
| 41 | const DEFAULT_VM_STACK_BYTES: usize = 1024 * 1024; |
| 42 | const MIN_VM_STACK_BYTES: usize = 128 * 1024; |
| 43 | const MAX_VM_STACK_BYTES: usize = 8 * 1024 * 1024; |
| 44 | const DEFAULT_VM_THREAD_STACK_BYTES: usize = 2 * 1024 * 1024; |
| 45 | const MIN_VM_THREAD_STACK_BYTES: usize = 512 * 1024; |
| 46 | const MAX_VM_THREAD_STACK_BYTES: usize = 16 * 1024 * 1024; |
| 47 | const DEFAULT_MAX_CONCURRENT_VMS: usize = 4; |
| 48 | const MAX_CONCURRENT_VMS: usize = 256; |
| 49 | |
| 50 | const VM_MEMORY_LIMIT_MB_ENV: &str = "CODEWHALE_WORKFLOW_JS_MEMORY_LIMIT_MB"; |
| 51 | const VM_STACK_KB_ENV: &str = "CODEWHALE_WORKFLOW_JS_STACK_KB"; |
| 52 | const VM_THREAD_STACK_KB_ENV: &str = "CODEWHALE_WORKFLOW_JS_THREAD_STACK_KB"; |
| 53 | const VM_MAX_CONCURRENT_ENV: &str = "CODEWHALE_WORKFLOW_JS_MAX_CONCURRENT"; |
| 54 | |
| 55 | /// Resource limits applied to the QuickJS runtime before any script runs. |
| 56 | /// |
| 57 | /// There is deliberately no wall-clock timeout here: cancellation (dropping |
| 58 | /// the run future, or the driver's cancel cascade) is the deadline mechanism. |
| 59 | #[derive(Debug, Clone, Copy)] |
| 60 | pub struct VmLimits { |
| 61 | /// QuickJS heap ceiling in bytes (default 32 MiB). |
| 62 | pub memory_limit_bytes: usize, |
| 63 | /// Maximum interpreter stack in bytes (default 1 MiB). |
| 64 | pub max_stack_bytes: usize, |
| 65 | } |
| 66 | |
| 67 | impl Default for VmLimits { |
| 68 | fn default() -> Self { |
| 69 | Self::from_env() |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | impl VmLimits { |
| 74 | pub fn from_env() -> Self { |
| 75 | Self { |
| 76 | memory_limit_bytes: env_usize_bytes( |
| 77 | VM_MEMORY_LIMIT_MB_ENV, |
| 78 | 1024 * 1024, |
| 79 | MIN_VM_MEMORY_LIMIT_BYTES, |
| 80 | MAX_VM_MEMORY_LIMIT_BYTES, |
| 81 | DEFAULT_VM_MEMORY_LIMIT_BYTES, |
| 82 | ), |
| 83 | max_stack_bytes: env_usize_bytes( |
| 84 | VM_STACK_KB_ENV, |
| 85 | 1024, |
| 86 | MIN_VM_STACK_BYTES, |
| 87 | MAX_VM_STACK_BYTES, |
| 88 | DEFAULT_VM_STACK_BYTES, |
| 89 | ), |
| 90 | } |
| 91 | } |
| 92 | } |
| 93 | |
| 94 | fn env_usize_bytes(name: &str, unit: usize, min: usize, max: usize, default: usize) -> usize { |
| 95 | env::var(name) |
| 96 | .ok() |
| 97 | .and_then(|raw| raw.parse::<usize>().ok()) |
| 98 | .and_then(|value| value.checked_mul(unit)) |
| 99 | .map(|bytes| bytes.clamp(min, max)) |
| 100 | .unwrap_or(default) |
| 101 | } |
| 102 | |
| 103 | fn max_concurrent_vms() -> usize { |
| 104 | env::var(VM_MAX_CONCURRENT_ENV) |
| 105 | .ok() |
| 106 | .and_then(|raw| raw.parse::<usize>().ok()) |
| 107 | .map(|value| value.clamp(1, MAX_CONCURRENT_VMS)) |
| 108 | .unwrap_or(DEFAULT_MAX_CONCURRENT_VMS) |
| 109 | } |
| 110 | |
| 111 | fn vm_thread_stack_bytes() -> usize { |
| 112 | env_usize_bytes( |
| 113 | VM_THREAD_STACK_KB_ENV, |
| 114 | 1024, |
| 115 | MIN_VM_THREAD_STACK_BYTES, |
| 116 | MAX_VM_THREAD_STACK_BYTES, |
| 117 | DEFAULT_VM_THREAD_STACK_BYTES, |
| 118 | ) |
| 119 | } |
| 120 | |
| 121 | fn vm_admission() -> &'static Arc<Semaphore> { |
| 122 | static ADMISSION: OnceLock<Arc<Semaphore>> = OnceLock::new(); |
| 123 | ADMISSION.get_or_init(|| Arc::new(Semaphore::new(max_concurrent_vms()))) |
| 124 | } |
| 125 | |
| 126 | /// Executes Workflow scripts, one isolated QuickJS runtime per run. |
| 127 | /// |
| 128 | /// Every [`WorkflowVm::run_script`] call spins up a fresh interpreter on a |
| 129 | /// dedicated thread, so runs share nothing (globals, heap, interned atoms) |
| 130 | /// and a wedged script can never stall a sibling run. |
| 131 | #[derive(Debug, Clone, Default)] |
| 132 | pub struct WorkflowVm { |
| 133 | limits: VmLimits, |
| 134 | } |
| 135 | |
| 136 | impl WorkflowVm { |
| 137 | /// A VM with the default [`VmLimits`]. |
| 138 | pub fn new() -> Self { |
| 139 | Self::default() |
| 140 | } |
| 141 | |
| 142 | /// A VM with explicit resource limits. |
| 143 | pub fn with_limits(limits: VmLimits) -> Self { |
| 144 | Self { limits } |
| 145 | } |
| 146 | |
| 147 | /// Run one Workflow script to completion. |
| 148 | /// |
| 149 | /// * `source` is the script body; it is wrapped in an async function, so |
| 150 | /// top-level `await` and `return` both work. The returned value is the |
| 151 | /// script's `return` value, JSON-encoded (`undefined` becomes `null`). |
| 152 | /// * `args` is exposed verbatim to the script as the `args` global. |
| 153 | /// * `driver` executes `task()` spawns and receives progress events. A |
| 154 | /// driver instance is scoped to exactly one run: `cancel_all` is always |
| 155 | /// invoked at run teardown (success, script error, or cancellation), so |
| 156 | /// stray children never outlive the script that spawned them. |
| 157 | /// |
| 158 | /// Cancellation cascade (design §9): dropping the returned future cancels |
| 159 | /// the run — the interrupt handler aborts executing JS, pending `task()` |
| 160 | /// awaits resolve to errors, and `driver.cancel_all()` is invoked |
| 161 | /// immediately from the dropping thread. |
| 162 | pub async fn run_script( |
| 163 | &self, |
| 164 | source: &str, |
| 165 | args: serde_json::Value, |
| 166 | driver: Arc<dyn WorkflowDriver>, |
| 167 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 168 | self.run_script_with_cancel(source, args, driver, WorkflowRunCancel::new()) |
| 169 | .await |
| 170 | } |
| 171 | |
| 172 | /// Like [`Self::run_script`], but accepts an external cancel handle so the |
| 173 | /// host can interrupt the VM without dropping the run future. |
| 174 | pub async fn run_script_with_cancel( |
| 175 | &self, |
| 176 | source: &str, |
| 177 | args: serde_json::Value, |
| 178 | driver: Arc<dyn WorkflowDriver>, |
| 179 | cancel: WorkflowRunCancel, |
| 180 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 181 | self.run_inner(source, args, driver, None, cancel).await |
| 182 | } |
| 183 | |
| 184 | /// Run a code-mode script: the same VM, plus the `tools.call()` host |
| 185 | /// binding backed by `invoker`. Workflow runs (no invoker) never see the |
| 186 | /// binding, so the Workflow sandbox keeps its documented host surface. |
| 187 | pub async fn run_tools_script( |
| 188 | &self, |
| 189 | source: &str, |
| 190 | args: serde_json::Value, |
| 191 | driver: Arc<dyn WorkflowDriver>, |
| 192 | invoker: Arc<dyn ToolInvoker>, |
| 193 | cancel: WorkflowRunCancel, |
| 194 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 195 | self.run_inner(source, args, driver, Some(invoker), cancel) |
| 196 | .await |
| 197 | } |
| 198 | |
| 199 | async fn run_inner( |
| 200 | &self, |
| 201 | source: &str, |
| 202 | args: serde_json::Value, |
| 203 | driver: Arc<dyn WorkflowDriver>, |
| 204 | invoker: Option<Arc<dyn ToolInvoker>>, |
| 205 | cancel: WorkflowRunCancel, |
| 206 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 207 | let args_json = serde_json::to_string(&args) |
| 208 | .map_err(|err| WorkflowJsError::InvalidArgs(err.to_string()))?; |
| 209 | let cancel = cancel.0; |
| 210 | let (result_tx, result_rx) = oneshot::channel(); |
| 211 | let mut guard = RunGuard { |
| 212 | cancel: cancel.clone(), |
| 213 | driver: driver.clone(), |
| 214 | armed: true, |
| 215 | }; |
| 216 | |
| 217 | let permit = vm_admission() |
| 218 | .clone() |
| 219 | .acquire_owned() |
| 220 | .await |
| 221 | .map_err(|_| WorkflowJsError::VmInit("VM admission gate closed".to_string()))?; |
| 222 | let limits = self.limits; |
| 223 | let source = source.to_string(); |
| 224 | let thread_driver = driver.clone(); |
| 225 | let thread_cancel = cancel.clone(); |
| 226 | let thread_invoker = invoker.clone(); |
| 227 | let spawned = std::thread::Builder::new() |
| 228 | .name("workflow-js-vm".to_string()) |
| 229 | .stack_size(vm_thread_stack_bytes()) |
| 230 | .spawn(move || { |
| 231 | let _permit: OwnedSemaphorePermit = permit; |
| 232 | let outcome = vm_thread_main( |
| 233 | source, |
| 234 | args_json, |
| 235 | thread_driver.clone(), |
| 236 | thread_cancel, |
| 237 | thread_invoker, |
| 238 | limits, |
| 239 | ); |
| 240 | // Run teardown: this driver is scoped to one run, so any task |
| 241 | // still in flight is unreachable now — cancel the cascade. |
| 242 | thread_driver.cancel_all(); |
| 243 | let _ = result_tx.send(outcome); |
| 244 | }); |
| 245 | if let Err(err) = spawned { |
| 246 | guard.armed = false; |
| 247 | return Err(WorkflowJsError::VmInit(format!( |
| 248 | "failed to spawn VM thread: {err}" |
| 249 | ))); |
| 250 | } |
| 251 | |
| 252 | match result_rx.await { |
| 253 | Ok(outcome) => { |
| 254 | // The VM thread has already torn down and cancelled children. |
| 255 | guard.armed = false; |
| 256 | outcome |
| 257 | } |
| 258 | // VM thread panicked before reporting; leave the guard armed so |
| 259 | // its drop (right now, at return) cancels outstanding tasks. |
| 260 | Err(_) => Err(WorkflowJsError::VmTerminated( |
| 261 | "VM thread exited without reporting a result".to_string(), |
| 262 | )), |
| 263 | } |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | /// Cooperative cancel signal shared by the run future (guard side) and the VM |
| 268 | /// thread. The atomic flag feeds the QuickJS interrupt handler (sync, called |
| 269 | /// mid-bytecode); the watch channel wakes host futures parked on driver |
| 270 | /// completions. |
| 271 | #[derive(Clone)] |
| 272 | pub struct WorkflowRunCancel(CancelHandle); |
| 273 | |
| 274 | impl WorkflowRunCancel { |
| 275 | #[must_use] |
| 276 | pub fn new() -> Self { |
| 277 | Self(CancelHandle::new()) |
| 278 | } |
| 279 | |
| 280 | pub fn cancel(&self) { |
| 281 | self.0.cancel(); |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | impl Default for WorkflowRunCancel { |
| 286 | fn default() -> Self { |
| 287 | Self::new() |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | #[derive(Clone)] |
| 292 | struct CancelHandle { |
| 293 | flag: Arc<AtomicBool>, |
| 294 | tx: Arc<watch::Sender<bool>>, |
| 295 | } |
| 296 | |
| 297 | impl CancelHandle { |
| 298 | fn new() -> Self { |
| 299 | let (tx, _rx) = watch::channel(false); |
| 300 | Self { |
| 301 | flag: Arc::new(AtomicBool::new(false)), |
| 302 | tx: Arc::new(tx), |
| 303 | } |
| 304 | } |
| 305 | |
| 306 | fn cancel(&self) { |
| 307 | self.flag.store(true, Ordering::SeqCst); |
| 308 | self.tx.send_replace(true); |
| 309 | } |
| 310 | |
| 311 | fn is_cancelled(&self) -> bool { |
| 312 | self.flag.load(Ordering::SeqCst) |
| 313 | } |
| 314 | |
| 315 | async fn cancelled(&self) { |
| 316 | let mut rx = self.tx.subscribe(); |
| 317 | let _ = rx.wait_for(|cancelled| *cancelled).await; |
| 318 | } |
| 319 | |
| 320 | fn flag_arc(&self) -> Arc<AtomicBool> { |
| 321 | self.flag.clone() |
| 322 | } |
| 323 | } |
| 324 | |
| 325 | /// Fires the cancel cascade if the caller drops the run future before the VM |
| 326 | /// reports a result. |
| 327 | struct RunGuard { |
| 328 | cancel: CancelHandle, |
| 329 | driver: Arc<dyn WorkflowDriver>, |
| 330 | armed: bool, |
| 331 | } |
| 332 | |
| 333 | impl Drop for RunGuard { |
| 334 | fn drop(&mut self) { |
| 335 | if self.armed { |
| 336 | self.cancel.cancel(); |
| 337 | self.driver.cancel_all(); |
| 338 | } |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | fn vm_thread_main( |
| 343 | source: String, |
| 344 | args_json: String, |
| 345 | driver: Arc<dyn WorkflowDriver>, |
| 346 | cancel: CancelHandle, |
| 347 | invoker: Option<Arc<dyn ToolInvoker>>, |
| 348 | limits: VmLimits, |
| 349 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 350 | let reactor = tokio::runtime::Builder::new_current_thread() |
| 351 | .enable_all() |
| 352 | .build() |
| 353 | .map_err(|err| WorkflowJsError::VmInit(format!("failed to build VM reactor: {err}")))?; |
| 354 | reactor.block_on(run_in_vm( |
| 355 | source, args_json, driver, cancel, invoker, limits, |
| 356 | )) |
| 357 | } |
| 358 | |
| 359 | async fn run_in_vm( |
| 360 | source: String, |
| 361 | args_json: String, |
| 362 | driver: Arc<dyn WorkflowDriver>, |
| 363 | cancel: CancelHandle, |
| 364 | invoker: Option<Arc<dyn ToolInvoker>>, |
| 365 | limits: VmLimits, |
| 366 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 367 | let runtime = AsyncRuntime::new().map_err(|err| WorkflowJsError::VmInit(err.to_string()))?; |
| 368 | runtime.set_memory_limit(limits.memory_limit_bytes).await; |
| 369 | runtime.set_max_stack_size(limits.max_stack_bytes).await; |
| 370 | let interrupt_flag = cancel.flag_arc(); |
| 371 | runtime |
| 372 | .set_interrupt_handler(Some(Box::new(move || { |
| 373 | interrupt_flag.load(Ordering::Acquire) |
| 374 | }))) |
| 375 | .await; |
| 376 | let context = AsyncContext::full(&runtime) |
| 377 | .await |
| 378 | .map_err(|err| WorkflowJsError::VmInit(err.to_string()))?; |
| 379 | |
| 380 | let result = context |
| 381 | .async_with(async |ctx| run_in_ctx(ctx, source, args_json, driver, invoker, cancel).await) |
| 382 | .await; |
| 383 | drop(context); |
| 384 | runtime.run_gc().await; |
| 385 | result |
| 386 | } |
| 387 | |
| 388 | async fn run_in_ctx( |
| 389 | ctx: Ctx<'_>, |
| 390 | source: String, |
| 391 | args_json: String, |
| 392 | driver: Arc<dyn WorkflowDriver>, |
| 393 | invoker: Option<Arc<dyn ToolInvoker>>, |
| 394 | cancel: CancelHandle, |
| 395 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 396 | install_host(&ctx, driver, invoker.clone(), cancel.clone(), &args_json)?; |
| 397 | ctx.eval::<(), _>(prelude()) |
| 398 | .catch(&ctx) |
| 399 | .map_err(|err| WorkflowJsError::VmInit(format!("prelude failed: {err}")))?; |
| 400 | if invoker.is_some() { |
| 401 | ctx.eval::<(), _>(CODEMODE_PRELUDE) |
| 402 | .catch(&ctx) |
| 403 | .map_err(|err| WorkflowJsError::VmInit(format!("codemode prelude failed: {err}")))?; |
| 404 | } |
| 405 | |
| 406 | let desugared = desugar_export_default(&source); |
| 407 | let wrapped = format!("(async () => {{\n{desugared}\n}})()"); |
| 408 | let promise = ctx |
| 409 | .eval::<Promise, _>(wrapped) |
| 410 | .catch(&ctx) |
| 411 | .map_err(|err| script_error(&cancel, err))?; |
| 412 | let value = promise |
| 413 | .into_future::<Value>() |
| 414 | .await |
| 415 | .catch(&ctx) |
| 416 | .map_err(|err| script_error(&cancel, err))?; |
| 417 | js_value_to_json(&ctx, value) |
| 418 | } |
| 419 | |
| 420 | /// Rewrite the documented module-style authoring shape |
| 421 | /// (`export default async function (args) { ... }`) into the script form the |
| 422 | /// VM actually evals. Sources are wrapped in an async IIFE, where the |
| 423 | /// module-only `export` keyword is a syntax error, so without this every |
| 424 | /// imperative `export default` workflow (including the #4131 dogfood |
| 425 | /// fixtures) failed to parse. The default export is captured, invoked with |
| 426 | /// the `args` global when it is a function, and its result becomes the run |
| 427 | /// result; a non-function default export is returned as-is. |
| 428 | fn desugar_export_default(source: &str) -> String { |
| 429 | const EXPORT_DEFAULT: &str = "export default"; |
| 430 | let Some(offset) = line_leading_export_default(source) else { |
| 431 | return source.to_string(); |
| 432 | }; |
| 433 | let mut out = source.to_string(); |
| 434 | out.replace_range( |
| 435 | offset..offset + EXPORT_DEFAULT.len(), |
| 436 | "globalThis.__workflow_default =", |
| 437 | ); |
| 438 | out.push('\n'); |
| 439 | out.push_str( |
| 440 | ";{\n const __wf_default = globalThis.__workflow_default;\n delete globalThis.__workflow_default;\n if (typeof __wf_default === \"function\") {\n return await __wf_default(args);\n }\n if (__wf_default !== undefined) {\n return __wf_default;\n }\n}\n", |
| 441 | ); |
| 442 | out |
| 443 | } |
| 444 | |
| 445 | /// Return the byte offset of a line-leading `export default` token that is |
| 446 | /// actual JavaScript syntax, not text inside a string, template literal, or |
| 447 | /// comment. This intentionally recognizes only the documented authoring shape |
| 448 | /// instead of attempting to implement a general JavaScript module parser. |
| 449 | fn line_leading_export_default(source: &str) -> Option<usize> { |
| 450 | const EXPORT_DEFAULT: &[u8] = b"export default"; |
| 451 | let bytes = source.as_bytes(); |
| 452 | let mut idx = 0usize; |
| 453 | let mut quote = None; |
| 454 | let mut escaped = false; |
| 455 | let mut line_comment = false; |
| 456 | let mut block_comment = false; |
| 457 | let mut line_has_only_whitespace = true; |
| 458 | |
| 459 | while idx < bytes.len() { |
| 460 | let byte = bytes[idx]; |
| 461 | |
| 462 | if line_comment { |
| 463 | if byte == b'\n' { |
| 464 | line_comment = false; |
| 465 | line_has_only_whitespace = true; |
| 466 | } |
| 467 | idx += 1; |
| 468 | continue; |
| 469 | } |
| 470 | |
| 471 | if block_comment { |
| 472 | if byte == b'*' && bytes.get(idx + 1) == Some(&b'/') { |
| 473 | block_comment = false; |
| 474 | line_has_only_whitespace = false; |
| 475 | idx += 2; |
| 476 | continue; |
| 477 | } |
| 478 | if byte == b'\n' { |
| 479 | line_has_only_whitespace = true; |
| 480 | } else if !byte.is_ascii_whitespace() { |
| 481 | line_has_only_whitespace = false; |
| 482 | } |
| 483 | idx += 1; |
| 484 | continue; |
| 485 | } |
| 486 | |
| 487 | if let Some(active_quote) = quote { |
| 488 | if byte == b'\n' { |
| 489 | line_has_only_whitespace = true; |
| 490 | escaped = false; |
| 491 | } else { |
| 492 | if !byte.is_ascii_whitespace() { |
| 493 | line_has_only_whitespace = false; |
| 494 | } |
| 495 | if escaped { |
| 496 | escaped = false; |
| 497 | } else if byte == b'\\' { |
| 498 | escaped = true; |
| 499 | } else if byte == active_quote { |
| 500 | quote = None; |
| 501 | } |
| 502 | } |
| 503 | idx += 1; |
| 504 | continue; |
| 505 | } |
| 506 | |
| 507 | if byte == b'\n' { |
| 508 | line_has_only_whitespace = true; |
| 509 | idx += 1; |
| 510 | continue; |
| 511 | } |
| 512 | if line_has_only_whitespace && byte.is_ascii_whitespace() { |
| 513 | idx += 1; |
| 514 | continue; |
| 515 | } |
| 516 | if line_has_only_whitespace && bytes[idx..].starts_with(EXPORT_DEFAULT) { |
| 517 | return Some(idx); |
| 518 | } |
| 519 | |
| 520 | line_has_only_whitespace = false; |
| 521 | if byte == b'/' && bytes.get(idx + 1) == Some(&b'/') { |
| 522 | line_comment = true; |
| 523 | idx += 2; |
| 524 | } else if byte == b'/' && bytes.get(idx + 1) == Some(&b'*') { |
| 525 | block_comment = true; |
| 526 | idx += 2; |
| 527 | } else { |
| 528 | if matches!(byte, b'\'' | b'"' | b'`') { |
| 529 | quote = Some(byte); |
| 530 | } |
| 531 | idx += 1; |
| 532 | } |
| 533 | } |
| 534 | |
| 535 | None |
| 536 | } |
| 537 | |
| 538 | fn script_error(cancel: &CancelHandle, err: CaughtError<'_>) -> WorkflowJsError { |
| 539 | if cancel.is_cancelled() { |
| 540 | WorkflowJsError::Cancelled |
| 541 | } else { |
| 542 | WorkflowJsError::Script(err.to_string()) |
| 543 | } |
| 544 | } |
| 545 | |
| 546 | fn js_value_to_json<'js>( |
| 547 | ctx: &Ctx<'js>, |
| 548 | value: Value<'js>, |
| 549 | ) -> Result<serde_json::Value, WorkflowJsError> { |
| 550 | if value.is_undefined() { |
| 551 | return Ok(serde_json::Value::Null); |
| 552 | } |
| 553 | let text = ctx |
| 554 | .json_stringify(value) |
| 555 | .map_err(|err| WorkflowJsError::ResultEncoding(err.to_string()))?; |
| 556 | match text { |
| 557 | None => Ok(serde_json::Value::Null), |
| 558 | Some(text) => { |
| 559 | let text = text |
| 560 | .to_string() |
| 561 | .map_err(|err| WorkflowJsError::ResultEncoding(err.to_string()))?; |
| 562 | serde_json::from_str(&text) |
| 563 | .map_err(|err| WorkflowJsError::ResultEncoding(err.to_string())) |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | |
| 568 | /// Prelude fragment for code-mode runs only: captures `__codemode_call` into |
| 569 | /// the frozen `globalThis.tools` surface, then deletes the raw binding. |
| 570 | /// Workflow runs never eval this, so their documented host surface is |
| 571 | /// unchanged — `tools` simply does not exist there. |
| 572 | const CODEMODE_PRELUDE: &str = r#" |
| 573 | (() => { |
| 574 | const hostCall = __codemode_call; |
| 575 | globalThis.tools = Object.freeze({ |
| 576 | call: async (tool, input) => { |
| 577 | const envelope = JSON.parse( |
| 578 | await hostCall(JSON.stringify({ tool, input: input === undefined ? {} : input })) |
| 579 | ); |
| 580 | if (envelope.error !== undefined) { |
| 581 | const err = new Error(envelope.error); |
| 582 | err.kind = envelope.error_kind; |
| 583 | throw err; |
| 584 | } |
| 585 | return envelope.result; |
| 586 | }, |
| 587 | }); |
| 588 | try { delete globalThis.__codemode_call; } catch (_) { /* frozen shape */ } |
| 589 | })(); |
| 590 | "#; |
| 591 | |
| 592 | /// The `tools.call()` host call. Infallible at the binding level, like |
| 593 | /// `task_host`: outcomes return through the `{result}` / |
| 594 | /// `{error, error_kind}` envelope so the prelude rethrows typed errors. |
| 595 | /// Gate refusals arrive as admission errors (nothing ran); a nested tool |
| 596 | /// that ran and failed arrives as an agent error (work failed). |
| 597 | async fn tools_call_host( |
| 598 | call_json: String, |
| 599 | invoker: Arc<dyn ToolInvoker>, |
| 600 | cancel: CancelHandle, |
| 601 | invoked: Rc<Cell<u64>>, |
| 602 | ) -> String { |
| 603 | let outcome = tools_call_host_inner(call_json, invoker, cancel, invoked).await; |
| 604 | let envelope = match outcome { |
| 605 | Ok(result) => serde_json::json!({ "result": result }), |
| 606 | Err(TaskError { kind, message }) => { |
| 607 | serde_json::json!({ "error": message, "error_kind": kind.as_str() }) |
| 608 | } |
| 609 | }; |
| 610 | envelope.to_string() |
| 611 | } |
| 612 | |
| 613 | async fn tools_call_host_inner( |
| 614 | call_json: String, |
| 615 | invoker: Arc<dyn ToolInvoker>, |
| 616 | cancel: CancelHandle, |
| 617 | invoked: Rc<Cell<u64>>, |
| 618 | ) -> Result<serde_json::Value, TaskError> { |
| 619 | let admission = |message: String| TaskError::new(TaskErrorKind::Admission, message); |
| 620 | let request: ToolCallRequest = serde_json::from_str(&call_json).map_err(|err| { |
| 621 | admission(format!( |
| 622 | "tools.call(): expected {{\"tool\", \"input\"}} JSON: {err}" |
| 623 | )) |
| 624 | })?; |
| 625 | if request.tool.trim().is_empty() { |
| 626 | return Err(admission( |
| 627 | "tools.call(): `tool` must be a non-empty string".to_string(), |
| 628 | )); |
| 629 | } |
| 630 | if !request.input.is_object() { |
| 631 | return Err(admission( |
| 632 | "tools.call(): `input` must be a JSON object".to_string(), |
| 633 | )); |
| 634 | } |
| 635 | if invoked.get() >= CODEMODE_MAX_TOOL_CALLS { |
| 636 | return Err(admission(format!( |
| 637 | "tools.call(): per-run tool-call cap ({CODEMODE_MAX_TOOL_CALLS}) reached" |
| 638 | ))); |
| 639 | } |
| 640 | if cancel.is_cancelled() { |
| 641 | return Err(TaskError::new( |
| 642 | TaskErrorKind::Cancelled, |
| 643 | "tools.call(): run cancelled".to_string(), |
| 644 | )); |
| 645 | } |
| 646 | invoked.set(invoked.get() + 1); |
| 647 | let response = tokio::select! { |
| 648 | _ = cancel.cancelled() => { |
| 649 | return Err(TaskError::new( |
| 650 | TaskErrorKind::Cancelled, |
| 651 | "tools.call(): run cancelled".to_string(), |
| 652 | )); |
| 653 | } |
| 654 | response = invoker.invoke(request) => response.map_err(|err| { |
| 655 | TaskError::new( |
| 656 | TaskErrorKind::from(&err), |
| 657 | format!("tools.call(): {err}"), |
| 658 | ) |
| 659 | })?, |
| 660 | }; |
| 661 | if response.ok { |
| 662 | Ok(response.result) |
| 663 | } else { |
| 664 | let message = response |
| 665 | .result |
| 666 | .as_str() |
| 667 | .unwrap_or("tools.call(): tool failed without a message") |
| 668 | .to_string(); |
| 669 | Err(TaskError::new(TaskErrorKind::Agent, message)) |
| 670 | } |
| 671 | } |
| 672 | |
| 673 | fn install_host( |
| 674 | ctx: &Ctx<'_>, |
| 675 | driver: Arc<dyn WorkflowDriver>, |
| 676 | invoker: Option<Arc<dyn ToolInvoker>>, |
| 677 | cancel: CancelHandle, |
| 678 | args_json: &str, |
| 679 | ) -> Result<(), WorkflowJsError> { |
| 680 | let globals = ctx.globals(); |
| 681 | |
| 682 | let args_value: Value = ctx |
| 683 | .json_parse(args_json) |
| 684 | .map_err(|err| WorkflowJsError::InvalidArgs(err.to_string()))?; |
| 685 | globals.set("args", args_value).map_err(init_err)?; |
| 686 | |
| 687 | // Per-run lifetime counter (design §4.3): counts spawn *attempts*, and the |
| 688 | // check + increment happen with no await in between so a parallel burst |
| 689 | // cannot slip past the cap on the single-threaded VM. |
| 690 | let spawned = Rc::new(Cell::new(0u64)); |
| 691 | |
| 692 | let task_driver = driver.clone(); |
| 693 | let task_cancel = cancel.clone(); |
| 694 | globals |
| 695 | .set( |
| 696 | "__workflow_task", |
| 697 | Func::from(Async(move |opts_json: String| { |
| 698 | let driver = task_driver.clone(); |
| 699 | let cancel = task_cancel.clone(); |
| 700 | let spawned = spawned.clone(); |
| 701 | async move { task_host(opts_json, driver, cancel, spawned).await } |
| 702 | })), |
| 703 | ) |
| 704 | .map_err(init_err)?; |
| 705 | |
| 706 | // Code-mode runs only: per-run tool-call counter. Same single-threaded |
| 707 | // check+increment discipline as `spawned` above — no await between the |
| 708 | // cap check and the increment, so a burst cannot slip past it. |
| 709 | if let Some(invoker) = invoker { |
| 710 | let invoked = Rc::new(Cell::new(0u64)); |
| 711 | let call_invoker = invoker.clone(); |
| 712 | let call_cancel = cancel.clone(); |
| 713 | globals |
| 714 | .set( |
| 715 | "__codemode_call", |
| 716 | Func::from(Async(move |call_json: String| { |
| 717 | let invoker = call_invoker.clone(); |
| 718 | let cancel = call_cancel.clone(); |
| 719 | let invoked = invoked.clone(); |
| 720 | async move { tools_call_host(call_json, invoker, cancel, invoked).await } |
| 721 | })), |
| 722 | ) |
| 723 | .map_err(init_err)?; |
| 724 | } |
| 725 | |
| 726 | let log_driver = driver.clone(); |
| 727 | globals |
| 728 | .set( |
| 729 | "__workflow_log", |
| 730 | Func::from(move |message: String| { |
| 731 | log_driver.progress(ProgressEvent::Log { message }); |
| 732 | }), |
| 733 | ) |
| 734 | .map_err(init_err)?; |
| 735 | |
| 736 | // Structured twin of the prelude's "every slot failed" breadcrumb (R9): |
| 737 | // a dead fan-out of script-thrown thunks leaves no task record behind, |
| 738 | // so the host needs a typed event — not a log line — to keep the run's |
| 739 | // terminal status honest. |
| 740 | let fanout_driver = driver.clone(); |
| 741 | globals |
| 742 | .set( |
| 743 | "__workflow_every_slot_failed", |
| 744 | Func::from(move |construct: String, failed: u32, total: u32| { |
| 745 | fanout_driver.progress(ProgressEvent::FanoutAllSlotsFailed { |
| 746 | construct, |
| 747 | failed, |
| 748 | total, |
| 749 | }); |
| 750 | }), |
| 751 | ) |
| 752 | .map_err(init_err)?; |
| 753 | |
| 754 | // Structured twin of the per-slot "dropped a failed slot as null" |
| 755 | // breadcrumb (R9): a PARTIALLY failed fan-out still resolves and leaves |
| 756 | // no task record for the dropped slot, so without this event the ledger |
| 757 | // cannot see the loss and records a clean Completed. |
| 758 | let fanout_driver = driver.clone(); |
| 759 | globals |
| 760 | .set( |
| 761 | "__workflow_slot_dropped", |
| 762 | Func::from(move |construct: String, kind: String, slot: u32| { |
| 763 | fanout_driver.progress(ProgressEvent::FanoutSlotDropped { |
| 764 | construct, |
| 765 | kind, |
| 766 | slot, |
| 767 | }); |
| 768 | }), |
| 769 | ) |
| 770 | .map_err(init_err)?; |
| 771 | |
| 772 | let phase_driver = driver.clone(); |
| 773 | globals |
| 774 | .set( |
| 775 | "__workflow_phase", |
| 776 | Func::from(move |title: String| { |
| 777 | phase_driver.progress(ProgressEvent::Phase { title }); |
| 778 | }), |
| 779 | ) |
| 780 | .map_err(init_err)?; |
| 781 | |
| 782 | // Budget reads are live driver snapshots (design §5.2). NaN encodes |
| 783 | // "no ceiling" for `total`; the prelude maps it to `null`. |
| 784 | let total_driver = driver.clone(); |
| 785 | globals |
| 786 | .set( |
| 787 | "__workflow_budget_total", |
| 788 | Func::from(move || -> f64 { |
| 789 | match total_driver.budget().total { |
| 790 | Some(total) => total as f64, |
| 791 | None => f64::NAN, |
| 792 | } |
| 793 | }), |
| 794 | ) |
| 795 | .map_err(init_err)?; |
| 796 | |
| 797 | let spent_driver = driver.clone(); |
| 798 | globals |
| 799 | .set( |
| 800 | "__workflow_budget_spent", |
| 801 | Func::from(move || -> f64 { spent_driver.budget().spent as f64 }), |
| 802 | ) |
| 803 | .map_err(init_err)?; |
| 804 | |
| 805 | globals |
| 806 | .set( |
| 807 | "__workflow_budget_remaining", |
| 808 | Func::from(move || -> f64 { |
| 809 | match driver.budget().remaining() { |
| 810 | Some(remaining) => remaining as f64, |
| 811 | None => f64::INFINITY, |
| 812 | } |
| 813 | }), |
| 814 | ) |
| 815 | .map_err(init_err)?; |
| 816 | |
| 817 | Ok(()) |
| 818 | } |
| 819 | |
| 820 | fn init_err(err: rquickjs::Error) -> WorkflowJsError { |
| 821 | WorkflowJsError::VmInit(err.to_string()) |
| 822 | } |
| 823 | |
| 824 | /// The `task()` host call. Everything that can go wrong is reported through |
| 825 | /// the JSON envelope (`{"error": ..., "error_kind": ...}`) so the prelude |
| 826 | /// re-throws it as a real JS `Error` with a script-side stack and a typed |
| 827 | /// [`TaskErrorKind`] on `.kind` (R9). The kind is assigned here, where the |
| 828 | /// failure actually happened — never re-derived from the message text. |
| 829 | async fn task_host( |
| 830 | opts_json: String, |
| 831 | driver: Arc<dyn WorkflowDriver>, |
| 832 | cancel: CancelHandle, |
| 833 | spawned: Rc<Cell<u64>>, |
| 834 | ) -> String { |
| 835 | let outcome = task_host_inner(opts_json, driver, cancel, spawned).await; |
| 836 | let envelope = match outcome { |
| 837 | Ok(value) => serde_json::json!({ "value": value }), |
| 838 | Err(TaskError { kind, message }) => { |
| 839 | serde_json::json!({ "error": message, "error_kind": kind.as_str() }) |
| 840 | } |
| 841 | }; |
| 842 | envelope.to_string() |
| 843 | } |
| 844 | |
| 845 | /// Best-effort `label`/`phase` from raw `task()` options, for rejection |
| 846 | /// receipts when the options never survived parsing. |
| 847 | fn task_identity_hint(opts_json: &str) -> (Option<String>, Option<String>) { |
| 848 | let value: serde_json::Value = |
| 849 | serde_json::from_str(opts_json).unwrap_or(serde_json::Value::Null); |
| 850 | let pluck = |key: &str| { |
| 851 | value |
| 852 | .get(key) |
| 853 | .and_then(serde_json::Value::as_str) |
| 854 | .map(str::trim) |
| 855 | .filter(|text| !text.is_empty()) |
| 856 | .map(str::to_string) |
| 857 | }; |
| 858 | (pluck("label"), pluck("phase")) |
| 859 | } |
| 860 | |
| 861 | /// Record a pre-spawn `task()` rejection on the host ledger, then hand the |
| 862 | /// message back for the JS throw. Rejections that never reach `spawn_task` |
| 863 | /// would otherwise be invisible to the run record (#5035's surviving gap). |
| 864 | fn reject_task(driver: &Arc<dyn WorkflowDriver>, opts_json: &str, message: String) -> String { |
| 865 | let (label, phase) = task_identity_hint(opts_json); |
| 866 | driver.progress(ProgressEvent::TaskRejected { |
| 867 | label, |
| 868 | phase, |
| 869 | message: message.clone(), |
| 870 | }); |
| 871 | message |
| 872 | } |
| 873 | |
| 874 | /// Emit the terminal schema-failure receipt for a `task()` whose reply failed |
| 875 | /// `responseSchema` with no repair left to try, and hand the message back for |
| 876 | /// the JS throw. `note` (when present) names why a repair was skipped, so the |
| 877 | /// operator can tell "repair refused to run" from "repair also failed". |
| 878 | fn fail_schema( |
| 879 | driver: &Arc<dyn WorkflowDriver>, |
| 880 | task_id: String, |
| 881 | attempt: u32, |
| 882 | error: &ReplyDecodeError, |
| 883 | raw: String, |
| 884 | raw_truncated: bool, |
| 885 | note: Option<String>, |
| 886 | ) -> String { |
| 887 | let message = match note { |
| 888 | Some(note) => format!("{} (repair skipped: {note})", error.message()), |
| 889 | None => error.message().to_string(), |
| 890 | }; |
| 891 | driver.progress(ProgressEvent::TaskSchemaValidationFailed { |
| 892 | task_id, |
| 893 | kind: error.kind().to_string(), |
| 894 | attempt, |
| 895 | message: message.clone(), |
| 896 | raw, |
| 897 | raw_truncated, |
| 898 | }); |
| 899 | message |
| 900 | } |
| 901 | |
| 902 | /// Build the repair request for `next_attempt` (#5583): the same child |
| 903 | /// identity, budget fields, and schema as the original, with a repair prompt |
| 904 | /// carrying the original task, the schema, the failed reply, and why it |
| 905 | /// failed, plus the wall clock the first attempt did not spend. |
| 906 | fn repair_request( |
| 907 | original: &TaskRequest, |
| 908 | next_attempt: u32, |
| 909 | error: &ReplyDecodeError, |
| 910 | failed_raw: &str, |
| 911 | wall_time_secs: Option<u64>, |
| 912 | ) -> TaskRequest { |
| 913 | let mut request = original.clone(); |
| 914 | let schema = original |
| 915 | .response_schema |
| 916 | .as_ref() |
| 917 | .expect("repair only runs when responseSchema is set"); |
| 918 | // The bracket prefix identifies the repair at a glance on progress |
| 919 | // surfaces and lets tests script the repair reply by rule order. |
| 920 | request.description = format!( |
| 921 | "[schema repair {next_attempt}] {}", |
| 922 | repair_prompt(&original.description, schema, failed_raw, error) |
| 923 | ); |
| 924 | request.wall_time_secs = wall_time_secs; |
| 925 | // Label and phase stay inherited so progress surfaces group the repair |
| 926 | // with its task. |
| 927 | request |
| 928 | } |
| 929 | |
| 930 | /// The one cancellation error: the run's deadline fired. Fatal in every |
| 931 | /// `parallel()` / `pipeline()` mode — a cancelled run must never resolve into |
| 932 | /// a slot value. |
| 933 | fn cancelled_task() -> TaskError { |
| 934 | TaskError::new(TaskErrorKind::Cancelled, "task(): run cancelled") |
| 935 | } |
| 936 | |
| 937 | /// A terminal `responseSchema` failure, already receipted by [`fail_schema`]. |
| 938 | fn schema_task(message: String) -> TaskError { |
| 939 | TaskError::new(TaskErrorKind::Schema, message) |
| 940 | } |
| 941 | |
| 942 | async fn task_host_inner( |
| 943 | opts_json: String, |
| 944 | driver: Arc<dyn WorkflowDriver>, |
| 945 | cancel: CancelHandle, |
| 946 | spawned: Rc<Cell<u64>>, |
| 947 | ) -> Result<serde_json::Value, TaskError> { |
| 948 | let admission = |message: String| TaskError::new(TaskErrorKind::Admission, message); |
| 949 | let request = parse_task_options(&opts_json) |
| 950 | .map_err(|message| admission(reject_task(&driver, &opts_json, message)))?; |
| 951 | // Compile the schema before spawning so a malformed one fails fast |
| 952 | // instead of burning a subagent. |
| 953 | let validator = request |
| 954 | .response_schema |
| 955 | .as_ref() |
| 956 | .map(compile_schema) |
| 957 | .transpose() |
| 958 | .map_err(|message| admission(reject_task(&driver, &opts_json, message)))?; |
| 959 | |
| 960 | // Lifetime backstop (design §4.3) — checked and bumped before any await. |
| 961 | if spawned.get() >= WORKFLOW_LIFETIME_CAP { |
| 962 | return Err(admission(reject_task( |
| 963 | &driver, |
| 964 | &opts_json, |
| 965 | format!( |
| 966 | "task(): Workflow lifetime agent cap ({WORKFLOW_LIFETIME_CAP}) reached for this run" |
| 967 | ), |
| 968 | ))); |
| 969 | } |
| 970 | // Fast-fail budget gate. The authoritative reservation lives in the |
| 971 | // driver (design §5.3); this only stops obviously-doomed spawns early. |
| 972 | let snapshot = driver.budget(); |
| 973 | if snapshot.exhausted() { |
| 974 | return Err(TaskError::new( |
| 975 | TaskErrorKind::Budget, |
| 976 | reject_task( |
| 977 | &driver, |
| 978 | &opts_json, |
| 979 | format!( |
| 980 | "task(): budget exhausted ({} of {} tokens spent)", |
| 981 | snapshot.spent, |
| 982 | snapshot.total.unwrap_or(0) |
| 983 | ), |
| 984 | ), |
| 985 | )); |
| 986 | } |
| 987 | if cancel.is_cancelled() { |
| 988 | return Err(cancelled_task()); |
| 989 | } |
| 990 | |
| 991 | // Bounded schema repair (#5583): after a failed `responseSchema` decode, |
| 992 | // re-ask the same route before throwing. `None` is the default single |
| 993 | // repair; `Some(0)` disables it. The first attempt is attempt 1, so the |
| 994 | // task is schema-terminal once `attempt` reaches this ceiling. |
| 995 | let last_attempt = 1 + request.schema_repair_attempts.unwrap_or(1); |
| 996 | // The wall clock is shared across attempts: a repair inherits the time |
| 997 | // the first attempt did not spend, not a fresh budget. |
| 998 | let started = std::time::Instant::now(); |
| 999 | let mut wall_time_secs_left = request.wall_time_secs; |
| 1000 | let mut current = request.clone(); |
| 1001 | let mut attempt: u32 = 0; |
| 1002 | loop { |
| 1003 | attempt += 1; |
| 1004 | spawned.set(spawned.get() + 1); |
| 1005 | let spawned_task = driver |
| 1006 | .spawn_task(current.clone()) |
| 1007 | .await |
| 1008 | .map_err(|err| TaskError::new(TaskErrorKind::from(&err), err.to_string()))?; |
| 1009 | let task_id = spawned_task.task_id; |
| 1010 | let completion_rx = spawned_task.completion; |
| 1011 | let completion = tokio::select! { |
| 1012 | _ = cancel.cancelled() => return Err(cancelled_task()), |
| 1013 | completion = completion_rx => completion.map_err(|_| { |
| 1014 | TaskError::new( |
| 1015 | TaskErrorKind::Driver, |
| 1016 | "task(): driver dropped the completion channel", |
| 1017 | ) |
| 1018 | })?, |
| 1019 | }; |
| 1020 | |
| 1021 | let text = match completion { |
| 1022 | TaskCompletion::Completed { text } => text, |
| 1023 | TaskCompletion::Failed { message } => { |
| 1024 | return Err(TaskError::new( |
| 1025 | TaskErrorKind::Agent, |
| 1026 | format!("task(): subagent failed: {message}"), |
| 1027 | )); |
| 1028 | } |
| 1029 | TaskCompletion::Cancelled => { |
| 1030 | return Err(TaskError::new( |
| 1031 | TaskErrorKind::Cancelled, |
| 1032 | "task(): subagent cancelled", |
| 1033 | )); |
| 1034 | } |
| 1035 | TaskCompletion::BudgetExhausted { message } => { |
| 1036 | return Err(TaskError::new( |
| 1037 | TaskErrorKind::Budget, |
| 1038 | format!("task(): budget exhausted: {message}"), |
| 1039 | )); |
| 1040 | } |
| 1041 | }; |
| 1042 | // Without a schema the raw text is the contract; with one, the decode |
| 1043 | // decides — and a failure may still be repaired. |
| 1044 | let Some(validator) = validator.as_ref() else { |
| 1045 | return Ok(serde_json::Value::String(text)); |
| 1046 | }; |
| 1047 | let error = match decode_reply(&text, validator) { |
| 1048 | Ok(value) => return Ok(value), |
| 1049 | Err(error) => error, |
| 1050 | }; |
| 1051 | let (raw, raw_truncated) = carried_raw(&text); |
| 1052 | if attempt >= last_attempt { |
| 1053 | return Err(schema_task(fail_schema( |
| 1054 | &driver, |
| 1055 | task_id, |
| 1056 | attempt, |
| 1057 | &error, |
| 1058 | raw, |
| 1059 | raw_truncated, |
| 1060 | None, |
| 1061 | ))); |
| 1062 | } |
| 1063 | // The attempt failed but a repair remains: record it as a receipt |
| 1064 | // (visible even when the repair succeeds), then re-run the admission |
| 1065 | // gates — a repair is a real child, not a free retry. |
| 1066 | driver.progress(ProgressEvent::TaskSchemaRepairAttempted { |
| 1067 | task_id: task_id.clone(), |
| 1068 | kind: error.kind().to_string(), |
| 1069 | attempt, |
| 1070 | message: error.message().to_string(), |
| 1071 | raw: raw.clone(), |
| 1072 | raw_truncated, |
| 1073 | }); |
| 1074 | if spawned.get() >= WORKFLOW_LIFETIME_CAP { |
| 1075 | return Err(schema_task(fail_schema( |
| 1076 | &driver, |
| 1077 | task_id, |
| 1078 | attempt, |
| 1079 | &error, |
| 1080 | raw, |
| 1081 | raw_truncated, |
| 1082 | Some(format!( |
| 1083 | "workflow lifetime agent cap ({WORKFLOW_LIFETIME_CAP}) reached" |
| 1084 | )), |
| 1085 | ))); |
| 1086 | } |
| 1087 | let snapshot = driver.budget(); |
| 1088 | if snapshot.exhausted() { |
| 1089 | return Err(schema_task(fail_schema( |
| 1090 | &driver, |
| 1091 | task_id, |
| 1092 | attempt, |
| 1093 | &error, |
| 1094 | raw, |
| 1095 | raw_truncated, |
| 1096 | Some("budget exhausted".to_string()), |
| 1097 | ))); |
| 1098 | } |
| 1099 | if cancel.is_cancelled() { |
| 1100 | return Err(cancelled_task()); |
| 1101 | } |
| 1102 | if let Some(wall) = wall_time_secs_left { |
| 1103 | let remaining = wall.saturating_sub(started.elapsed().as_secs()); |
| 1104 | if remaining == 0 { |
| 1105 | return Err(schema_task(fail_schema( |
| 1106 | &driver, |
| 1107 | task_id, |
| 1108 | attempt, |
| 1109 | &error, |
| 1110 | raw, |
| 1111 | raw_truncated, |
| 1112 | Some("no wall-time left from wallTimeSecs".to_string()), |
| 1113 | ))); |
| 1114 | } |
| 1115 | wall_time_secs_left = Some(remaining); |
| 1116 | } |
| 1117 | current = repair_request(&request, attempt + 1, &error, &raw, wall_time_secs_left); |
| 1118 | } |
| 1119 | } |
| 1120 | |
| 1121 | /// JS-facing option names for `task()` (design §3.3). Unknown fields are |
| 1122 | /// rejected so a typo (`responseschema`) fails loudly instead of being |
| 1123 | /// silently dropped. Every multi-word field also accepts its snake_case |
| 1124 | /// spelling, and the `agent` tool's `workspace_policy` name is accepted as an |
| 1125 | /// alias for worktree isolation — the two spawn surfaces are written by the |
| 1126 | /// same authors (often models), so a schema that runs on one must not be an |
| 1127 | /// unknown-field error on the other. |
| 1128 | #[derive(Debug, Deserialize)] |
| 1129 | #[serde(rename_all = "camelCase", deny_unknown_fields)] |
| 1130 | struct TaskOptions { |
| 1131 | #[serde(alias = "title")] |
| 1132 | description: Option<String>, |
| 1133 | prompt: Option<String>, |
| 1134 | #[serde(alias = "type", alias = "subagent_type")] |
| 1135 | subagent_type: Option<String>, |
| 1136 | /// Fleet role name (#4177). Preferred step identity field. |
| 1137 | role: Option<String>, |
| 1138 | profile: Option<String>, |
| 1139 | model: Option<String>, |
| 1140 | #[serde(alias = "model_strength")] |
| 1141 | model_strength: Option<String>, |
| 1142 | thinking: Option<String>, |
| 1143 | cwd: Option<String>, |
| 1144 | #[serde(default)] |
| 1145 | worktree: bool, |
| 1146 | /// `agent`-tool alias for worktree isolation: "shared" | "worktree". |
| 1147 | #[serde(default, alias = "workspace_policy")] |
| 1148 | workspace_policy: Option<String>, |
| 1149 | #[serde(alias = "write_authority")] |
| 1150 | write_authority: Option<String>, |
| 1151 | #[serde(default, alias = "write_roots")] |
| 1152 | write_roots: Vec<String>, |
| 1153 | #[serde(default, alias = "exact_files")] |
| 1154 | exact_files: Vec<String>, |
| 1155 | #[serde(default, alias = "coordination_contracts")] |
| 1156 | coordination_contracts: Vec<String>, |
| 1157 | #[serde(default)] |
| 1158 | dependencies: Vec<String>, |
| 1159 | #[serde(default)] |
| 1160 | acceptance: Vec<String>, |
| 1161 | #[serde(alias = "allowed_tools")] |
| 1162 | allowed_tools: Option<Vec<String>>, |
| 1163 | #[serde(alias = "max_depth")] |
| 1164 | max_depth: Option<u32>, |
| 1165 | #[serde(alias = "token_budget")] |
| 1166 | token_budget: Option<u64>, |
| 1167 | #[serde(alias = "max_steps")] |
| 1168 | max_steps: Option<u32>, |
| 1169 | #[serde(alias = "wall_time_secs")] |
| 1170 | wall_time_secs: Option<u64>, |
| 1171 | #[serde(alias = "response_schema")] |
| 1172 | response_schema: Option<serde_json::Value>, |
| 1173 | /// Bounded `responseSchema` repair attempts after a failed decode |
| 1174 | /// (#5583): re-ask the same route with the schema and the failed reply. |
| 1175 | /// Defaults to one; `0` disables; capped at |
| 1176 | /// [`SCHEMA_REPAIR_MAX_ATTEMPTS`] so repair stays a bounded recovery. |
| 1177 | #[serde(default, alias = "schema_repair_attempts")] |
| 1178 | schema_repair_attempts: Option<u32>, |
| 1179 | label: Option<String>, |
| 1180 | phase: Option<String>, |
| 1181 | } |
| 1182 | |
| 1183 | fn parse_task_options(opts_json: &str) -> Result<TaskRequest, String> { |
| 1184 | let mut options: TaskOptions = |
| 1185 | serde_json::from_str(opts_json).map_err(|err| format!("task(): invalid options: {err}"))?; |
| 1186 | if let Some(policy) = options.workspace_policy.take() { |
| 1187 | match policy.trim().to_ascii_lowercase().as_str() { |
| 1188 | "worktree" => options.worktree = true, |
| 1189 | "shared" => { |
| 1190 | if options.worktree { |
| 1191 | return Err( |
| 1192 | "task(): workspacePolicy 'shared' conflicts with worktree: true" |
| 1193 | .to_string(), |
| 1194 | ); |
| 1195 | } |
| 1196 | } |
| 1197 | other => { |
| 1198 | return Err(format!( |
| 1199 | "task(): workspacePolicy must be shared or worktree; got {other:?}" |
| 1200 | )); |
| 1201 | } |
| 1202 | } |
| 1203 | } |
| 1204 | let description = options |
| 1205 | .prompt |
| 1206 | .or(options.description) |
| 1207 | .filter(|description| !description.trim().is_empty()) |
| 1208 | .ok_or_else(|| "task(): 'description' (or 'prompt') is required".to_string())?; |
| 1209 | let role = options |
| 1210 | .role |
| 1211 | .as_deref() |
| 1212 | .map(normalize_profile) |
| 1213 | .transpose() |
| 1214 | .map_err(|err| format!("task(): role: {err}"))?; |
| 1215 | let profile = options |
| 1216 | .profile |
| 1217 | .as_deref() |
| 1218 | .map(normalize_profile) |
| 1219 | .transpose() |
| 1220 | .map_err(|err| format!("task(): {err}"))?; |
| 1221 | options.write_roots = normalize_task_paths("writeRoots", options.write_roots, 32)?; |
| 1222 | options.exact_files = normalize_task_paths("exactFiles", options.exact_files, 32)?; |
| 1223 | let cwd = options |
| 1224 | .cwd |
| 1225 | .take() |
| 1226 | .map(|value| normalize_task_paths("cwd", vec![value], 1)) |
| 1227 | .transpose()? |
| 1228 | .and_then(|mut paths| paths.pop()); |
| 1229 | options.coordination_contracts = |
| 1230 | normalize_task_string_list("coordinationContracts", options.coordination_contracts, 16)?; |
| 1231 | options.dependencies = normalize_task_string_list("dependencies", options.dependencies, 8)?; |
| 1232 | options.acceptance = normalize_task_string_list("acceptance", options.acceptance, 8)?; |
| 1233 | let write_authority = options |
| 1234 | .write_authority |
| 1235 | .as_deref() |
| 1236 | .map(|value| value.trim().to_ascii_lowercase()) |
| 1237 | .map(|value| match value.as_str() { |
| 1238 | "read_only" | "workspace_write" | "worktree_write" => Ok(value), |
| 1239 | _ => Err(format!( |
| 1240 | "task(): writeAuthority must be read_only, workspace_write, or worktree_write; got {value:?}" |
| 1241 | )), |
| 1242 | }) |
| 1243 | .transpose()?; |
| 1244 | if write_authority.as_deref() == Some("worktree_write") && !options.worktree { |
| 1245 | return Err("task(): writeAuthority worktree_write requires worktree: true".to_string()); |
| 1246 | } |
| 1247 | let role_kind = role.as_deref().and_then(task_role_kind); |
| 1248 | let type_kind = options.subagent_type.as_deref().and_then(task_role_kind); |
| 1249 | if let (Some(role_kind), Some(type_kind)) = (role_kind, type_kind) |
| 1250 | && role_kind != type_kind |
| 1251 | { |
| 1252 | return Err("task(): role and subagentType declare contradictory authorities".to_string()); |
| 1253 | } |
| 1254 | let declared_kind = role_kind.or(type_kind); |
| 1255 | if matches!(declared_kind, Some(TaskRoleKind::ReadOnly)) |
| 1256 | && write_authority |
| 1257 | .as_deref() |
| 1258 | .is_some_and(|authority| authority != "read_only") |
| 1259 | { |
| 1260 | return Err("task(): read-only roles cannot declare write-capable authority".to_string()); |
| 1261 | } |
| 1262 | if write_authority |
| 1263 | .as_deref() |
| 1264 | .is_some_and(|authority| authority != "read_only") |
| 1265 | && options.write_roots.is_empty() |
| 1266 | && options.exact_files.is_empty() |
| 1267 | && options.coordination_contracts.is_empty() |
| 1268 | { |
| 1269 | return Err( |
| 1270 | "task(): write-capable authority requires writeRoots, exactFiles, or coordinationContracts" |
| 1271 | .to_string(), |
| 1272 | ); |
| 1273 | } |
| 1274 | let explicit_write_identity = declared_kind == Some(TaskRoleKind::Implementer) |
| 1275 | || (declared_kind == Some(TaskRoleKind::General) |
| 1276 | && (role.is_some() || options.subagent_type.is_some())) |
| 1277 | || (profile.is_some() && declared_kind.is_none()); |
| 1278 | if explicit_write_identity |
| 1279 | && write_authority.as_deref() != Some("read_only") |
| 1280 | && options.write_roots.is_empty() |
| 1281 | && options.exact_files.is_empty() |
| 1282 | && options.coordination_contracts.is_empty() |
| 1283 | { |
| 1284 | return Err( |
| 1285 | "task(): explicit write-capable identities require writeRoots, exactFiles, or coordinationContracts" |
| 1286 | .to_string(), |
| 1287 | ); |
| 1288 | } |
| 1289 | if let Some(attempts) = options.schema_repair_attempts |
| 1290 | && attempts > SCHEMA_REPAIR_MAX_ATTEMPTS |
| 1291 | { |
| 1292 | return Err(format!( |
| 1293 | "task(): schemaRepairAttempts is bounded to {SCHEMA_REPAIR_MAX_ATTEMPTS}; \ |
| 1294 | repair is a bounded recovery, not a retry loop" |
| 1295 | )); |
| 1296 | } |
| 1297 | Ok(TaskRequest { |
| 1298 | description, |
| 1299 | subagent_type: options.subagent_type, |
| 1300 | role, |
| 1301 | profile, |
| 1302 | model: options.model, |
| 1303 | model_strength: options.model_strength, |
| 1304 | thinking: options.thinking, |
| 1305 | cwd, |
| 1306 | worktree: options.worktree, |
| 1307 | write_authority, |
| 1308 | write_roots: options.write_roots, |
| 1309 | exact_files: options.exact_files, |
| 1310 | coordination_contracts: options.coordination_contracts, |
| 1311 | dependencies: options.dependencies, |
| 1312 | acceptance: options.acceptance, |
| 1313 | allowed_tools: options.allowed_tools, |
| 1314 | // Host-imposed only: a script cannot set (or clear) a deny list. |
| 1315 | disallowed_tools: Vec::new(), |
| 1316 | max_depth: options.max_depth, |
| 1317 | token_budget: options.token_budget, |
| 1318 | max_steps: options.max_steps, |
| 1319 | wall_time_secs: options.wall_time_secs, |
| 1320 | response_schema: options.response_schema, |
| 1321 | schema_repair_attempts: options.schema_repair_attempts, |
| 1322 | label: options.label, |
| 1323 | phase: options.phase, |
| 1324 | }) |
| 1325 | } |
| 1326 | |
| 1327 | fn normalize_task_string_list( |
| 1328 | field: &str, |
| 1329 | values: Vec<String>, |
| 1330 | limit: usize, |
| 1331 | ) -> Result<Vec<String>, String> { |
| 1332 | if values.len() > limit { |
| 1333 | return Err(format!("task(): {field} accepts at most {limit} entries")); |
| 1334 | } |
| 1335 | let mut normalized = Vec::new(); |
| 1336 | for value in values { |
| 1337 | let value = value.trim(); |
| 1338 | if value.is_empty() || value.chars().count() > 512 { |
| 1339 | return Err(format!( |
| 1340 | "task(): {field} entries must be 1..=512 characters" |
| 1341 | )); |
| 1342 | } |
| 1343 | if !normalized.iter().any(|existing| existing == value) { |
| 1344 | normalized.push(value.to_string()); |
| 1345 | } |
| 1346 | } |
| 1347 | Ok(normalized) |
| 1348 | } |
| 1349 | |
| 1350 | fn normalize_task_paths( |
| 1351 | field: &str, |
| 1352 | values: Vec<String>, |
| 1353 | limit: usize, |
| 1354 | ) -> Result<Vec<String>, String> { |
| 1355 | if values.len() > limit { |
| 1356 | return Err(format!("task(): {field} accepts at most {limit} entries")); |
| 1357 | } |
| 1358 | let mut normalized = Vec::new(); |
| 1359 | for raw in values { |
| 1360 | let raw = raw.trim().replace('\\', "/"); |
| 1361 | let windows_drive = raw.as_bytes().get(1) == Some(&b':') |
| 1362 | && raw.as_bytes().first().is_some_and(u8::is_ascii_alphabetic); |
| 1363 | if raw.is_empty() |
| 1364 | || raw.chars().count() > 512 |
| 1365 | || raw.starts_with('/') |
| 1366 | || raw.starts_with("//") |
| 1367 | || windows_drive |
| 1368 | || raw.chars().any(|ch| matches!(ch, '\0' | '\r' | '\n')) |
| 1369 | { |
| 1370 | return Err(format!( |
| 1371 | "task(): {field} entries must be bounded repo-relative paths" |
| 1372 | )); |
| 1373 | } |
| 1374 | let mut segments = Vec::new(); |
| 1375 | for segment in raw.split('/') { |
| 1376 | match segment { |
| 1377 | "" | "." => {} |
| 1378 | ".." => { |
| 1379 | return Err(format!( |
| 1380 | "task(): {field} paths cannot contain parent traversal" |
| 1381 | )); |
| 1382 | } |
| 1383 | value => segments.push(value), |
| 1384 | } |
| 1385 | } |
| 1386 | let path = if segments.is_empty() { |
| 1387 | ".".to_string() |
| 1388 | } else { |
| 1389 | segments.join("/") |
| 1390 | }; |
| 1391 | if !normalized.contains(&path) { |
| 1392 | normalized.push(path); |
| 1393 | } |
| 1394 | } |
| 1395 | Ok(normalized) |
| 1396 | } |
| 1397 | |
| 1398 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 1399 | enum TaskRoleKind { |
| 1400 | ReadOnly, |
| 1401 | General, |
| 1402 | Implementer, |
| 1403 | } |
| 1404 | |
| 1405 | fn task_role_kind(value: &str) -> Option<TaskRoleKind> { |
| 1406 | match value.trim().to_ascii_lowercase().as_str() { |
| 1407 | "explore" | "explorer" | "scout" | "plan" | "planner" | "review" | "reviewer" |
| 1408 | | "verify" | "verifier" => Some(TaskRoleKind::ReadOnly), |
| 1409 | "general" | "worker" => Some(TaskRoleKind::General), |
| 1410 | "implement" | "implementer" | "builder" => Some(TaskRoleKind::Implementer), |
| 1411 | _ => None, |
| 1412 | } |
| 1413 | } |
| 1414 | |
| 1415 | /// The JS prelude injected before every script: determinism bans, the |
| 1416 | /// `task`/`parallel`/`pipeline`/`log`/`phase` stdlib (design §7), and the |
| 1417 | /// `budget` global. |
| 1418 | fn prelude() -> String { |
| 1419 | PRELUDE_TEMPLATE.replace("__MAX_ITEMS__", &PARALLEL_MAX_ITEMS.to_string()) |
| 1420 | } |
| 1421 | |
| 1422 | const PRELUDE_TEMPLATE: &str = r#""use strict"; |
| 1423 | (() => { |
| 1424 | const banned = (name) => () => { |
| 1425 | throw new Error(name + " is unavailable in Workflow scripts: runs must be deterministic for record/replay"); |
| 1426 | }; |
| 1427 | const BannedDate = function Date() { |
| 1428 | throw new Error("new Date()/Date() is unavailable in Workflow scripts: runs must be deterministic for record/replay"); |
| 1429 | }; |
| 1430 | BannedDate.now = banned("Date.now()"); |
| 1431 | BannedDate.parse = banned("Date.parse()"); |
| 1432 | BannedDate.UTC = banned("Date.UTC()"); |
| 1433 | globalThis.Date = BannedDate; |
| 1434 | Math.random = banned("Math.random()"); |
| 1435 | |
| 1436 | // Capture temporary host bindings into this closure, then strip them from |
| 1437 | // globalThis so scripts only see the documented Workflow surface (#4129). |
| 1438 | const hostTask = __workflow_task; |
| 1439 | const hostLog = __workflow_log; |
| 1440 | const hostEverySlotFailed = __workflow_every_slot_failed; |
| 1441 | const hostSlotDropped = __workflow_slot_dropped; |
| 1442 | const hostPhase = __workflow_phase; |
| 1443 | const hostBudgetTotal = __workflow_budget_total; |
| 1444 | const hostBudgetSpent = __workflow_budget_spent; |
| 1445 | const hostBudgetRemaining = __workflow_budget_remaining; |
| 1446 | |
| 1447 | const MAX_ITEMS = __MAX_ITEMS__; |
| 1448 | const taskErrorText = (err) => String(err && err.message !== undefined ? err.message : err); |
| 1449 | |
| 1450 | // Typed slot errors (R9). Every error thrown by task() carries a |
| 1451 | // host-assigned `kind` copied off the task envelope; anything else that |
| 1452 | // reaches a slot was thrown by the script itself and reports as "script". |
| 1453 | // |
| 1454 | // The kind is read from the error object and never guessed from message |
| 1455 | // text. A substring classifier let a child's own words ("...budget |
| 1456 | // exhausted...", "...responseSchema...") forge a fatal classification and |
| 1457 | // abort a healthy run, and it could not tell a genuine subagent failure |
| 1458 | // apart from a plain `throw new Error(...)` in a stage. |
| 1459 | const HOST_KINDS = ["admission", "budget", "cancelled", "agent", "schema", "driver"]; |
| 1460 | const SCRIPT_KIND = "script"; |
| 1461 | const taskErrorKind = (err) => |
| 1462 | err !== null && typeof err === "object" && HOST_KINDS.indexOf(err.kind) !== -1 |
| 1463 | ? err.kind |
| 1464 | : SCRIPT_KIND; |
| 1465 | // Fatal kinds are never absorbed into a slot value: cancellation is the |
| 1466 | // run's own deadline, and a schema breach means the contract the caller |
| 1467 | // explicitly asked for was not met. `mode: "partial"` opts out for schema |
| 1468 | // (and only schema) by keeping it as a structured slot value instead. |
| 1469 | const isFatalTaskError = (err) => { |
| 1470 | const kind = taskErrorKind(err); |
| 1471 | return kind === "cancelled" || kind === "schema"; |
| 1472 | }; |
| 1473 | |
| 1474 | // Stamp the resolved kind onto an error that is about to be rethrown, so a |
| 1475 | // script's own `catch (err) { err.kind }` reads the same vocabulary the |
| 1476 | // slot classifier used. Host errors already carry theirs; this only names |
| 1477 | // the script throws, which would otherwise surface as `undefined`. |
| 1478 | const stampKind = (err, kind) => { |
| 1479 | if (err !== null && typeof err === "object" && err.kind === undefined) { |
| 1480 | try { |
| 1481 | err.kind = kind; |
| 1482 | } catch (_) { |
| 1483 | // A frozen error keeps whatever it has; the log line still names it. |
| 1484 | } |
| 1485 | } |
| 1486 | return err; |
| 1487 | }; |
| 1488 | |
| 1489 | const SLOT_MODES = ["settled", "fail-fast", "partial"]; |
| 1490 | // `settled` is the default and is exactly today's behavior: a non-fatal |
| 1491 | // slot failure resolves to `null` so an author need not handle every error. |
| 1492 | // An unrecognized mode throws rather than silently falling back — a typo |
| 1493 | // like `mode: "failfast"` used to read as `settled` and quietly keep |
| 1494 | // dropping slots the author believed were now fatal. |
| 1495 | const slotMode = (fn, opts) => { |
| 1496 | if (opts === null || typeof opts !== "object" || opts.mode === undefined) { |
| 1497 | return "settled"; |
| 1498 | } |
| 1499 | if (SLOT_MODES.indexOf(opts.mode) === -1) { |
| 1500 | throw new Error( |
| 1501 | fn + "(): unknown mode " + JSON.stringify(opts.mode) + |
| 1502 | "; expected one of " + SLOT_MODES.join(", ") |
| 1503 | ); |
| 1504 | } |
| 1505 | return opts.mode; |
| 1506 | }; |
| 1507 | |
| 1508 | // The failure ledger for one fan-out, attached to the resolved array as a |
| 1509 | // non-enumerable `errors` property. Non-enumerable and non-index, so the |
| 1510 | // array's contents, length, and JSON encoding are byte-identical to before: |
| 1511 | // `results.filter(Boolean)` still works, and a script that wants to know |
| 1512 | // WHY a slot is null can now ask instead of guessing. |
| 1513 | const attachSlotErrors = (results, errors) => { |
| 1514 | errors.sort((a, b) => a.index - b.index); |
| 1515 | Object.defineProperty(results, "errors", { |
| 1516 | value: Object.freeze(errors.map((entry) => Object.freeze(entry))), |
| 1517 | enumerable: false, |
| 1518 | configurable: false, |
| 1519 | writable: false, |
| 1520 | }); |
| 1521 | return results; |
| 1522 | }; |
| 1523 | |
| 1524 | globalThis.task = async (opts) => { |
| 1525 | if (opts === null || typeof opts !== "object") { |
| 1526 | throw new TypeError("task(): expected an options object"); |
| 1527 | } |
| 1528 | const envelope = JSON.parse(await hostTask(JSON.stringify(opts))); |
| 1529 | if (envelope.error !== undefined) { |
| 1530 | const err = new Error(envelope.error); |
| 1531 | // The host always names the kind; a missing one means an envelope this |
| 1532 | // prelude did not produce, which is not a typed task failure. |
| 1533 | err.kind = HOST_KINDS.indexOf(envelope.error_kind) !== -1 |
| 1534 | ? envelope.error_kind |
| 1535 | : SCRIPT_KIND; |
| 1536 | throw err; |
| 1537 | } |
| 1538 | return envelope.value; |
| 1539 | }; |
| 1540 | |
| 1541 | globalThis.parallel = (thunks, opts) => { |
| 1542 | if (!Array.isArray(thunks)) { |
| 1543 | throw new TypeError("parallel(): expected an array of thunks"); |
| 1544 | } |
| 1545 | if (thunks.length > MAX_ITEMS) { |
| 1546 | throw new Error("parallel(): max " + MAX_ITEMS + " items per call"); |
| 1547 | } |
| 1548 | const mode = slotMode("parallel", opts); |
| 1549 | const failFast = mode === "fail-fast"; |
| 1550 | const partial = mode === "partial"; |
| 1551 | const errors = []; |
| 1552 | // Returns the slot value, or throws to reject the whole fan-out. |
| 1553 | const onSlotError = (index, err) => { |
| 1554 | const kind = taskErrorKind(err); |
| 1555 | const message = taskErrorText(err); |
| 1556 | stampKind(err, kind); |
| 1557 | // Cancellation is the run's deadline in every mode, partial included. |
| 1558 | if (kind === "cancelled") throw err; |
| 1559 | if (partial) { |
| 1560 | // Opt-in partial mode: every non-cancellation slot failure becomes a |
| 1561 | // structured value the script can branch on. It never masquerades as |
| 1562 | // a success -- `__taskError` is the whole point of the shape. |
| 1563 | errors.push({ index: index, kind: kind, message: message }); |
| 1564 | hostLog( |
| 1565 | "parallel(): partial mode kept a failed slot as __taskError (kind=" + |
| 1566 | kind + ", slot " + index + "): " + message |
| 1567 | ); |
| 1568 | return { __taskError: { index: index, kind: kind, message: message } }; |
| 1569 | } |
| 1570 | if (isFatalTaskError(err)) throw err; |
| 1571 | if (failFast) { |
| 1572 | hostLog( |
| 1573 | "parallel(): fail-fast slot error (kind=" + kind + ", slot " + index + "): " + message |
| 1574 | ); |
| 1575 | throw err; |
| 1576 | } |
| 1577 | errors.push({ index: index, kind: kind, message: message }); |
| 1578 | hostLog( |
| 1579 | "parallel(): dropped a failed slot as null (kind=" + kind + ", slot " + index + "): " + |
| 1580 | message |
| 1581 | ); |
| 1582 | hostSlotDropped("parallel", kind, index); |
| 1583 | return null; |
| 1584 | }; |
| 1585 | const slots = thunks.map((thunk, index) => { |
| 1586 | try { |
| 1587 | return Promise.resolve(typeof thunk === "function" ? thunk() : thunk) |
| 1588 | .catch((err) => onSlotError(index, err)); |
| 1589 | } catch (err) { |
| 1590 | try { |
| 1591 | return onSlotError(index, err); |
| 1592 | } catch (rethrown) { |
| 1593 | return Promise.reject(rethrown); |
| 1594 | } |
| 1595 | } |
| 1596 | }); |
| 1597 | return Promise.all(slots).then((results) => { |
| 1598 | // A fan-out where nothing survived is a dead fan-out, not resilience. |
| 1599 | // The default stays ergonomic (the array still resolves) but the run |
| 1600 | // log says so in one line an operator can grep for, and the structured |
| 1601 | // event below lets the host status classifier refuse to call such a |
| 1602 | // run a plain success. |
| 1603 | if (results.length > 0 && errors.length === results.length) { |
| 1604 | hostLog( |
| 1605 | "parallel(): every slot failed (" + errors.length + " of " + results.length + |
| 1606 | "); no work survived this fan-out" |
| 1607 | ); |
| 1608 | hostEverySlotFailed("parallel", errors.length, results.length); |
| 1609 | } |
| 1610 | return attachSlotErrors(results, errors); |
| 1611 | }); |
| 1612 | }; |
| 1613 | |
| 1614 | globalThis.pipeline = (items, ...stages) => { |
| 1615 | if (!Array.isArray(items)) { |
| 1616 | throw new TypeError("pipeline(): expected an array of items"); |
| 1617 | } |
| 1618 | if (items.length > MAX_ITEMS) { |
| 1619 | throw new Error("pipeline(): max " + MAX_ITEMS + " items per call"); |
| 1620 | } |
| 1621 | // Options overload: pipeline(items, { stages: [...], mode: "fail-fast" }). |
| 1622 | let mode = "settled"; |
| 1623 | if ( |
| 1624 | stages.length === 1 && |
| 1625 | stages[0] !== null && |
| 1626 | typeof stages[0] === "object" && |
| 1627 | Array.isArray(stages[0].stages) |
| 1628 | ) { |
| 1629 | mode = slotMode("pipeline", stages[0]); |
| 1630 | stages = stages[0].stages; |
| 1631 | } |
| 1632 | const failFast = mode === "fail-fast"; |
| 1633 | const partial = mode === "partial"; |
| 1634 | const errors = []; |
| 1635 | return Promise.all(items.map(async (item, index) => { |
| 1636 | let value = item; |
| 1637 | for (const stage of stages) { |
| 1638 | try { |
| 1639 | value = await stage(value, item, index); |
| 1640 | } catch (err) { |
| 1641 | const kind = taskErrorKind(err); |
| 1642 | const message = taskErrorText(err); |
| 1643 | stampKind(err, kind); |
| 1644 | if (kind === "cancelled") throw err; |
| 1645 | if (partial) { |
| 1646 | errors.push({ index: index, kind: kind, message: message }); |
| 1647 | hostLog( |
| 1648 | "pipeline(): partial mode kept item " + index + |
| 1649 | " as __taskError (kind=" + kind + "): " + message |
| 1650 | ); |
| 1651 | return { __taskError: { index: index, kind: kind, message: message } }; |
| 1652 | } |
| 1653 | if (isFatalTaskError(err)) throw err; |
| 1654 | if (failFast) { |
| 1655 | hostLog( |
| 1656 | "pipeline(): fail-fast stage error on item " + index + |
| 1657 | " (kind=" + kind + "): " + message |
| 1658 | ); |
| 1659 | throw err; |
| 1660 | } |
| 1661 | errors.push({ index: index, kind: kind, message: message }); |
| 1662 | hostLog( |
| 1663 | "pipeline(): dropped item " + index + " as null (kind=" + kind + "): " + message |
| 1664 | ); |
| 1665 | hostSlotDropped("pipeline", kind, index); |
| 1666 | return null; |
| 1667 | } |
| 1668 | } |
| 1669 | return value; |
| 1670 | })).then((results) => { |
| 1671 | if (results.length > 0 && errors.length === results.length) { |
| 1672 | hostLog( |
| 1673 | "pipeline(): every item failed (" + errors.length + " of " + results.length + |
| 1674 | "); no work survived this pipeline" |
| 1675 | ); |
| 1676 | hostEverySlotFailed("pipeline", errors.length, results.length); |
| 1677 | } |
| 1678 | return attachSlotErrors(results, errors); |
| 1679 | }); |
| 1680 | }; |
| 1681 | |
| 1682 | globalThis.log = (message) => { |
| 1683 | hostLog(typeof message === "string" ? message : (JSON.stringify(message) ?? String(message))); |
| 1684 | }; |
| 1685 | globalThis.phase = (title) => { |
| 1686 | hostPhase(String(title)); |
| 1687 | }; |
| 1688 | |
| 1689 | const total = hostBudgetTotal(); |
| 1690 | globalThis.budget = Object.freeze({ |
| 1691 | total: Number.isNaN(total) ? null : total, |
| 1692 | spent: () => hostBudgetSpent(), |
| 1693 | remaining: () => hostBudgetRemaining(), |
| 1694 | }); |
| 1695 | |
| 1696 | for (const name of [ |
| 1697 | "__workflow_task", |
| 1698 | "__workflow_log", |
| 1699 | "__workflow_every_slot_failed", |
| 1700 | "__workflow_phase", |
| 1701 | "__workflow_budget_total", |
| 1702 | "__workflow_budget_spent", |
| 1703 | "__workflow_budget_remaining", |
| 1704 | ]) { |
| 1705 | try { |
| 1706 | delete globalThis[name]; |
| 1707 | } catch (_) { |
| 1708 | // Non-configurable bindings stay; the inventory test will fail closed. |
| 1709 | } |
| 1710 | } |
| 1711 | })(); |
| 1712 | "#; |
| 1713 |