| 1 | use serde::Deserialize; |
| 2 | use thiserror::Error; |
| 3 | |
| 4 | use crate::{ |
| 5 | BranchSpec, BudgetSpec, CondSpec, ExpandSpec, GateSpec, LeafSpec, LoopUntilSpec, ModelPolicy, |
| 6 | PermissionSpec, PromotionPolicy, ReduceSpec, SequenceSpec, TeacherReviewSpec, WorkflowNode, |
| 7 | WorkflowSpec, validate_workflow_nodes, |
| 8 | }; |
| 9 | |
| 10 | pub type JavascriptWorkflowResult<T> = std::result::Result<T, JavascriptWorkflowError>; |
| 11 | |
| 12 | #[derive(Debug, Error)] |
| 13 | pub enum JavascriptWorkflowError { |
| 14 | #[error("workflow source contains unsupported construct `{construct}`")] |
| 15 | UnsupportedConstruct { construct: &'static str }, |
| 16 | #[error("workflow source did not call workflow({{...}})")] |
| 17 | MissingWorkflowCall, |
| 18 | #[error("workflow({{...}}) object could not be extracted: {0}")] |
| 19 | InvalidWorkflowObject(String), |
| 20 | #[error("invalid workflow JSON object: {0}")] |
| 21 | InvalidJson(serde_json::Error), |
| 22 | #[error("invalid workflow node: {0}")] |
| 23 | InvalidNode(String), |
| 24 | } |
| 25 | |
| 26 | pub fn compile_javascript_workflow( |
| 27 | identifier: &str, |
| 28 | source: &str, |
| 29 | ) -> JavascriptWorkflowResult<WorkflowSpec> { |
| 30 | compile_js_like_workflow(identifier, source) |
| 31 | } |
| 32 | |
| 33 | pub fn compile_typescript_workflow( |
| 34 | identifier: &str, |
| 35 | source: &str, |
| 36 | ) -> JavascriptWorkflowResult<WorkflowSpec> { |
| 37 | compile_js_like_workflow(identifier, source) |
| 38 | } |
| 39 | |
| 40 | fn compile_js_like_workflow( |
| 41 | _identifier: &str, |
| 42 | source: &str, |
| 43 | ) -> JavascriptWorkflowResult<WorkflowSpec> { |
| 44 | reject_unsupported_constructs(source)?; |
| 45 | let object = extract_workflow_object(source)?; |
| 46 | let authored = serde_json::from_str::<JsWorkflowSpec>(object) |
| 47 | .map_err(JavascriptWorkflowError::InvalidJson)?; |
| 48 | let mut workflow = authored.into_workflow(); |
| 49 | normalize_leaf_profiles(&mut workflow.nodes); |
| 50 | normalize_gate_roles(&mut workflow.gates); |
| 51 | if workflow.goal.trim().is_empty() { |
| 52 | return Err(JavascriptWorkflowError::InvalidNode( |
| 53 | "workflow goal cannot be empty".to_string(), |
| 54 | )); |
| 55 | } |
| 56 | validate_workflow_nodes(&workflow.nodes) |
| 57 | .map_err(|error| JavascriptWorkflowError::InvalidNode(error.to_string()))?; |
| 58 | Ok(workflow) |
| 59 | } |
| 60 | |
| 61 | // Role/profile names are case-insensitive roster keys; the IR stores the |
| 62 | // canonical lowercase form. Invalid tokens are left as-is so validation |
| 63 | // reports them. |
| 64 | fn normalize_leaf_profiles(nodes: &mut [WorkflowNode]) { |
| 65 | for node in nodes { |
| 66 | match node { |
| 67 | WorkflowNode::Leaf(spec) => { |
| 68 | if let Some(role) = spec.role.as_mut() { |
| 69 | *role = role.trim().to_lowercase(); |
| 70 | } |
| 71 | if let Some(profile) = spec.profile.as_mut() { |
| 72 | *profile = profile.trim().to_lowercase(); |
| 73 | } |
| 74 | } |
| 75 | WorkflowNode::BranchSet(spec) => normalize_leaf_profiles(&mut spec.children), |
| 76 | WorkflowNode::Sequence(spec) => normalize_leaf_profiles(&mut spec.children), |
| 77 | WorkflowNode::LoopUntil(spec) => normalize_leaf_profiles(&mut spec.children), |
| 78 | WorkflowNode::Cond(spec) => { |
| 79 | normalize_leaf_profiles(&mut spec.then_nodes); |
| 80 | normalize_leaf_profiles(&mut spec.else_nodes); |
| 81 | } |
| 82 | WorkflowNode::Expand(spec) => { |
| 83 | if let Some(template) = spec.template.as_deref_mut() { |
| 84 | normalize_leaf_profiles(std::slice::from_mut(template)); |
| 85 | } |
| 86 | } |
| 87 | WorkflowNode::Reduce(_) | WorkflowNode::TeacherReview(_) => {} |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | fn normalize_gate_roles(gates: &mut [GateSpec]) { |
| 93 | for gate in gates { |
| 94 | gate.role = gate.role.trim().to_lowercase(); |
| 95 | if let Some(blocks_role) = gate.blocks_role.as_mut() { |
| 96 | *blocks_role = blocks_role.trim().to_lowercase(); |
| 97 | } |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | fn reject_unsupported_constructs(source: &str) -> JavascriptWorkflowResult<()> { |
| 102 | for (needle, construct) in [ |
| 103 | ("import ", "import"), |
| 104 | ("import(", "dynamic import"), |
| 105 | ("require(", "require"), |
| 106 | ("fetch(", "fetch"), |
| 107 | ("XMLHttpRequest", "XMLHttpRequest"), |
| 108 | ("WebSocket", "WebSocket"), |
| 109 | ("process.", "process"), |
| 110 | ("Deno.", "Deno"), |
| 111 | ("Bun.", "Bun"), |
| 112 | ("child_process", "child_process"), |
| 113 | ("exec(", "exec"), |
| 114 | ("spawn(", "spawn"), |
| 115 | ("open(", "open"), |
| 116 | ("readFile", "readFile"), |
| 117 | ("writeFile", "writeFile"), |
| 118 | ("async ", "async"), |
| 119 | ("await ", "await"), |
| 120 | ("eval(", "eval"), |
| 121 | ("new Function", "Function"), |
| 122 | ] { |
| 123 | if source.contains(needle) { |
| 124 | return Err(JavascriptWorkflowError::UnsupportedConstruct { construct }); |
| 125 | } |
| 126 | } |
| 127 | Ok(()) |
| 128 | } |
| 129 | |
| 130 | fn extract_workflow_object(source: &str) -> JavascriptWorkflowResult<&str> { |
| 131 | let workflow_pos = source |
| 132 | .find("workflow") |
| 133 | .ok_or(JavascriptWorkflowError::MissingWorkflowCall)?; |
| 134 | let open_paren_rel = source[workflow_pos..] |
| 135 | .find('(') |
| 136 | .ok_or(JavascriptWorkflowError::MissingWorkflowCall)?; |
| 137 | let open_paren = workflow_pos + open_paren_rel; |
| 138 | let object_start = source[open_paren + 1..] |
| 139 | .char_indices() |
| 140 | .find_map(|(idx, ch)| { |
| 141 | if ch.is_whitespace() { |
| 142 | None |
| 143 | } else { |
| 144 | Some((open_paren + 1 + idx, ch)) |
| 145 | } |
| 146 | }) |
| 147 | .ok_or(JavascriptWorkflowError::MissingWorkflowCall)?; |
| 148 | if object_start.1 != '{' { |
| 149 | return Err(JavascriptWorkflowError::InvalidWorkflowObject( |
| 150 | "workflow(...) must receive a JSON-compatible object literal".to_string(), |
| 151 | )); |
| 152 | } |
| 153 | |
| 154 | let mut depth = 0usize; |
| 155 | let mut in_string: Option<char> = None; |
| 156 | let mut escape = false; |
| 157 | for (idx, ch) in source[object_start.0..].char_indices() { |
| 158 | let absolute = object_start.0 + idx; |
| 159 | if let Some(quote) = in_string { |
| 160 | if escape { |
| 161 | escape = false; |
| 162 | } else if ch == '\\' { |
| 163 | escape = true; |
| 164 | } else if ch == quote { |
| 165 | in_string = None; |
| 166 | } |
| 167 | continue; |
| 168 | } |
| 169 | |
| 170 | match ch { |
| 171 | '"' | '\'' | '`' => in_string = Some(ch), |
| 172 | '{' => depth += 1, |
| 173 | '}' => { |
| 174 | depth = depth.checked_sub(1).ok_or_else(|| { |
| 175 | JavascriptWorkflowError::InvalidWorkflowObject( |
| 176 | "unbalanced closing brace".to_string(), |
| 177 | ) |
| 178 | })?; |
| 179 | if depth == 0 { |
| 180 | return Ok(&source[object_start.0..=absolute]); |
| 181 | } |
| 182 | } |
| 183 | _ => {} |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | Err(JavascriptWorkflowError::InvalidWorkflowObject( |
| 188 | "missing closing brace for workflow object".to_string(), |
| 189 | )) |
| 190 | } |
| 191 | |
| 192 | #[derive(Debug, Deserialize)] |
| 193 | #[serde(deny_unknown_fields)] |
| 194 | struct JsWorkflowSpec { |
| 195 | #[serde(default)] |
| 196 | id: Option<String>, |
| 197 | goal: String, |
| 198 | #[serde(default)] |
| 199 | description: Option<String>, |
| 200 | #[serde(default)] |
| 201 | budget: BudgetSpec, |
| 202 | #[serde(default)] |
| 203 | permissions: PermissionSpec, |
| 204 | #[serde(default)] |
| 205 | model_policy: ModelPolicy, |
| 206 | #[serde(default)] |
| 207 | promotion_policy: PromotionPolicy, |
| 208 | #[serde(default)] |
| 209 | gates: Vec<GateSpec>, |
| 210 | #[serde(default)] |
| 211 | nodes: Vec<JsWorkflowNode>, |
| 212 | } |
| 213 | |
| 214 | impl JsWorkflowSpec { |
| 215 | fn into_workflow(self) -> WorkflowSpec { |
| 216 | WorkflowSpec { |
| 217 | id: self.id, |
| 218 | goal: self.goal, |
| 219 | description: self.description, |
| 220 | budget: self.budget, |
| 221 | permissions: self.permissions, |
| 222 | model_policy: self.model_policy, |
| 223 | promotion_policy: self.promotion_policy, |
| 224 | gates: self.gates, |
| 225 | nodes: self |
| 226 | .nodes |
| 227 | .into_iter() |
| 228 | .map(JsWorkflowNode::into_node) |
| 229 | .collect(), |
| 230 | } |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | #[derive(Debug, Deserialize)] |
| 235 | #[serde(untagged)] |
| 236 | enum JsWorkflowNode { |
| 237 | Raw(WorkflowNode), |
| 238 | Agent(JsAgentNode), |
| 239 | Branch(JsBranchNode), |
| 240 | Sequence(JsSequenceNode), |
| 241 | Reduce(JsReduceNode), |
| 242 | TeacherReview(JsTeacherReviewNode), |
| 243 | LoopUntil(JsLoopUntilNode), |
| 244 | Cond(JsCondNode), |
| 245 | Expand(JsExpandNode), |
| 246 | } |
| 247 | |
| 248 | impl JsWorkflowNode { |
| 249 | fn into_node(self) -> WorkflowNode { |
| 250 | match self { |
| 251 | Self::Raw(node) => node, |
| 252 | Self::Agent(node) => WorkflowNode::Leaf(node.agent), |
| 253 | Self::Branch(node) => WorkflowNode::BranchSet(node.branch.into_branch()), |
| 254 | Self::Sequence(node) => WorkflowNode::Sequence(node.sequence.into_sequence()), |
| 255 | Self::Reduce(node) => WorkflowNode::Reduce(node.reduce), |
| 256 | Self::TeacherReview(node) => WorkflowNode::TeacherReview(node.teacher_review), |
| 257 | Self::LoopUntil(node) => WorkflowNode::LoopUntil(node.loop_until.into_loop_until()), |
| 258 | Self::Cond(node) => WorkflowNode::Cond(node.cond.into_cond()), |
| 259 | Self::Expand(node) => WorkflowNode::Expand(node.expand.into_expand()), |
| 260 | } |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | #[derive(Debug, Deserialize)] |
| 265 | #[serde(deny_unknown_fields)] |
| 266 | struct JsAgentNode { |
| 267 | agent: LeafSpec, |
| 268 | } |
| 269 | |
| 270 | #[derive(Debug, Deserialize)] |
| 271 | #[serde(deny_unknown_fields)] |
| 272 | struct JsBranchNode { |
| 273 | branch: JsBranchSpec, |
| 274 | } |
| 275 | |
| 276 | #[derive(Debug, Deserialize)] |
| 277 | #[serde(deny_unknown_fields)] |
| 278 | struct JsBranchSpec { |
| 279 | id: String, |
| 280 | #[serde(default)] |
| 281 | description: Option<String>, |
| 282 | #[serde(default = "default_true")] |
| 283 | parallel: bool, |
| 284 | #[serde(default)] |
| 285 | budget: BudgetSpec, |
| 286 | #[serde(default)] |
| 287 | permissions: PermissionSpec, |
| 288 | #[serde(default)] |
| 289 | model_policy: ModelPolicy, |
| 290 | #[serde(default)] |
| 291 | children: Vec<JsWorkflowNode>, |
| 292 | } |
| 293 | |
| 294 | impl JsBranchSpec { |
| 295 | fn into_branch(self) -> BranchSpec { |
| 296 | BranchSpec { |
| 297 | id: self.id, |
| 298 | description: self.description, |
| 299 | parallel: self.parallel, |
| 300 | budget: self.budget, |
| 301 | permissions: self.permissions, |
| 302 | model_policy: self.model_policy, |
| 303 | children: self |
| 304 | .children |
| 305 | .into_iter() |
| 306 | .map(JsWorkflowNode::into_node) |
| 307 | .collect(), |
| 308 | } |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | #[derive(Debug, Deserialize)] |
| 313 | #[serde(deny_unknown_fields)] |
| 314 | struct JsSequenceNode { |
| 315 | sequence: JsSequenceSpec, |
| 316 | } |
| 317 | |
| 318 | #[derive(Debug, Deserialize)] |
| 319 | #[serde(deny_unknown_fields)] |
| 320 | struct JsSequenceSpec { |
| 321 | id: String, |
| 322 | #[serde(default)] |
| 323 | children: Vec<JsWorkflowNode>, |
| 324 | } |
| 325 | |
| 326 | impl JsSequenceSpec { |
| 327 | fn into_sequence(self) -> SequenceSpec { |
| 328 | SequenceSpec { |
| 329 | id: self.id, |
| 330 | children: self |
| 331 | .children |
| 332 | .into_iter() |
| 333 | .map(JsWorkflowNode::into_node) |
| 334 | .collect(), |
| 335 | } |
| 336 | } |
| 337 | } |
| 338 | |
| 339 | #[derive(Debug, Deserialize)] |
| 340 | #[serde(deny_unknown_fields)] |
| 341 | struct JsReduceNode { |
| 342 | reduce: ReduceSpec, |
| 343 | } |
| 344 | |
| 345 | #[derive(Debug, Deserialize)] |
| 346 | #[serde(deny_unknown_fields)] |
| 347 | struct JsTeacherReviewNode { |
| 348 | teacher_review: TeacherReviewSpec, |
| 349 | } |
| 350 | |
| 351 | #[derive(Debug, Deserialize)] |
| 352 | #[serde(deny_unknown_fields)] |
| 353 | struct JsLoopUntilNode { |
| 354 | loop_until: JsLoopUntilSpec, |
| 355 | } |
| 356 | |
| 357 | #[derive(Debug, Deserialize)] |
| 358 | #[serde(deny_unknown_fields)] |
| 359 | struct JsLoopUntilSpec { |
| 360 | id: String, |
| 361 | condition: String, |
| 362 | #[serde(default)] |
| 363 | max_iterations: Option<u32>, |
| 364 | #[serde(default)] |
| 365 | children: Vec<JsWorkflowNode>, |
| 366 | } |
| 367 | |
| 368 | impl JsLoopUntilSpec { |
| 369 | fn into_loop_until(self) -> LoopUntilSpec { |
| 370 | LoopUntilSpec { |
| 371 | id: self.id, |
| 372 | condition: self.condition, |
| 373 | max_iterations: self.max_iterations, |
| 374 | children: self |
| 375 | .children |
| 376 | .into_iter() |
| 377 | .map(JsWorkflowNode::into_node) |
| 378 | .collect(), |
| 379 | } |
| 380 | } |
| 381 | } |
| 382 | |
| 383 | #[derive(Debug, Deserialize)] |
| 384 | #[serde(deny_unknown_fields)] |
| 385 | struct JsCondNode { |
| 386 | cond: JsCondSpec, |
| 387 | } |
| 388 | |
| 389 | #[derive(Debug, Deserialize)] |
| 390 | #[serde(deny_unknown_fields)] |
| 391 | struct JsCondSpec { |
| 392 | id: String, |
| 393 | condition: String, |
| 394 | #[serde(default)] |
| 395 | then_nodes: Vec<JsWorkflowNode>, |
| 396 | #[serde(default)] |
| 397 | else_nodes: Vec<JsWorkflowNode>, |
| 398 | } |
| 399 | |
| 400 | impl JsCondSpec { |
| 401 | fn into_cond(self) -> CondSpec { |
| 402 | CondSpec { |
| 403 | id: self.id, |
| 404 | condition: self.condition, |
| 405 | then_nodes: self |
| 406 | .then_nodes |
| 407 | .into_iter() |
| 408 | .map(JsWorkflowNode::into_node) |
| 409 | .collect(), |
| 410 | else_nodes: self |
| 411 | .else_nodes |
| 412 | .into_iter() |
| 413 | .map(JsWorkflowNode::into_node) |
| 414 | .collect(), |
| 415 | } |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | #[derive(Debug, Deserialize)] |
| 420 | #[serde(deny_unknown_fields)] |
| 421 | struct JsExpandNode { |
| 422 | expand: JsExpandSpec, |
| 423 | } |
| 424 | |
| 425 | #[derive(Debug, Deserialize)] |
| 426 | #[serde(deny_unknown_fields)] |
| 427 | struct JsExpandSpec { |
| 428 | id: String, |
| 429 | source: String, |
| 430 | #[serde(default)] |
| 431 | max_children: Option<usize>, |
| 432 | #[serde(default)] |
| 433 | template: Option<Box<JsWorkflowNode>>, |
| 434 | } |
| 435 | |
| 436 | impl JsExpandSpec { |
| 437 | fn into_expand(self) -> ExpandSpec { |
| 438 | ExpandSpec { |
| 439 | id: self.id, |
| 440 | source: self.source, |
| 441 | max_children: self.max_children, |
| 442 | template: self.template.map(|node| Box::new(node.into_node())), |
| 443 | } |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | fn default_true() -> bool { |
| 448 | true |
| 449 | } |
| 450 | |
| 451 | #[cfg(test)] |
| 452 | mod tests { |
| 453 | use super::*; |
| 454 | use crate::{ |
| 455 | AgentType, GateKind, GateOn, GateOnFail, GateOutcome, GateState, LaneGateBoard, TaskMode, |
| 456 | }; |
| 457 | |
| 458 | #[test] |
| 459 | fn javascript_workflow_compiles_branch_reduce_to_ir() { |
| 460 | let source = r#" |
| 461 | export default workflow({ |
| 462 | "id": "js-audit", |
| 463 | "goal": "Audit a change with parallel agents", |
| 464 | "nodes": [ |
| 465 | { |
| 466 | "branch": { |
| 467 | "id": "parallel-audit", |
| 468 | "children": [ |
| 469 | { |
| 470 | "agent": { |
| 471 | "id": "docs-audit", |
| 472 | "prompt": "Inspect docs for missing updates", |
| 473 | "agent_type": "review", |
| 474 | "file_scope": ["docs"] |
| 475 | } |
| 476 | }, |
| 477 | { |
| 478 | "agent": { |
| 479 | "id": "tests-audit", |
| 480 | "prompt": "Inspect targeted tests", |
| 481 | "agent_type": "verifier", |
| 482 | "budget": { "max_steps": 4 } |
| 483 | } |
| 484 | } |
| 485 | ] |
| 486 | } |
| 487 | }, |
| 488 | { |
| 489 | "reduce": { |
| 490 | "id": "synthesize", |
| 491 | "inputs": ["docs-audit", "tests-audit"], |
| 492 | "prompt": "Merge the branch findings" |
| 493 | } |
| 494 | } |
| 495 | ] |
| 496 | }); |
| 497 | "#; |
| 498 | |
| 499 | let workflow = |
| 500 | compile_javascript_workflow("audit.workflow.js", source).expect("compile JS workflow"); |
| 501 | |
| 502 | assert_eq!(workflow.id.as_deref(), Some("js-audit")); |
| 503 | assert_eq!(workflow.nodes.len(), 2); |
| 504 | let WorkflowNode::BranchSet(branch) = &workflow.nodes[0] else { |
| 505 | panic!("first node should be a branch"); |
| 506 | }; |
| 507 | assert!(branch.parallel); |
| 508 | assert_eq!(branch.children.len(), 2); |
| 509 | let WorkflowNode::Leaf(leaf) = &branch.children[1] else { |
| 510 | panic!("second branch child should be a leaf"); |
| 511 | }; |
| 512 | assert_eq!(leaf.agent_type, AgentType::Verifier); |
| 513 | assert_eq!(leaf.budget.max_steps, Some(4)); |
| 514 | assert!(matches!(workflow.nodes[1], WorkflowNode::Reduce(_))); |
| 515 | } |
| 516 | |
| 517 | #[test] |
| 518 | fn typescript_workflow_allows_satisfies_suffix_without_executing_js() { |
| 519 | let source = r#" |
| 520 | export default workflow({ |
| 521 | "goal": "TS authored workflow", |
| 522 | "nodes": [ |
| 523 | { "agent": { "id": "scan", "prompt": "scan safely" } } |
| 524 | ] |
| 525 | } satisfies WorkflowSpec); |
| 526 | "#; |
| 527 | |
| 528 | let workflow = |
| 529 | compile_typescript_workflow("scan.workflow.ts", source).expect("compile TS workflow"); |
| 530 | |
| 531 | assert_eq!(workflow.goal, "TS authored workflow"); |
| 532 | assert_eq!(workflow.nodes.len(), 1); |
| 533 | } |
| 534 | |
| 535 | #[test] |
| 536 | fn javascript_workflow_accepts_and_normalizes_agent_profile() { |
| 537 | let source = r#" |
| 538 | workflow({ |
| 539 | "goal": "profile routing", |
| 540 | "nodes": [ |
| 541 | { "agent": { "id": "review", "prompt": "review the diff", "profile": " Reviewer " } }, |
| 542 | { "agent": { "id": "scan", "prompt": "scan safely" } } |
| 543 | ] |
| 544 | }); |
| 545 | "#; |
| 546 | |
| 547 | let workflow = compile_javascript_workflow("profile.workflow.js", source) |
| 548 | .expect("profile-carrying workflow should compile"); |
| 549 | |
| 550 | let WorkflowNode::Leaf(review) = &workflow.nodes[0] else { |
| 551 | panic!("first node should be a leaf"); |
| 552 | }; |
| 553 | assert_eq!(review.profile.as_deref(), Some("reviewer")); |
| 554 | let WorkflowNode::Leaf(scan) = &workflow.nodes[1] else { |
| 555 | panic!("second node should be a leaf"); |
| 556 | }; |
| 557 | assert_eq!(scan.profile, None); |
| 558 | } |
| 559 | |
| 560 | #[test] |
| 561 | fn javascript_workflow_accepts_and_normalizes_agent_role() { |
| 562 | let source = r#" |
| 563 | workflow({ |
| 564 | "goal": "role routing", |
| 565 | "nodes": [ |
| 566 | { "agent": { "id": "scout-issue", "prompt": "Investigate #4090. Read-only.", "role": " Scout " } }, |
| 567 | { "agent": { "id": "fix-it", "prompt": "Apply minimal fix.", "role": "implementer" } } |
| 568 | ] |
| 569 | }); |
| 570 | "#; |
| 571 | |
| 572 | let workflow = compile_javascript_workflow("role.workflow.js", source) |
| 573 | .expect("role-carrying workflow should compile"); |
| 574 | |
| 575 | let WorkflowNode::Leaf(scout) = &workflow.nodes[0] else { |
| 576 | panic!("first node should be a leaf"); |
| 577 | }; |
| 578 | assert_eq!(scout.role.as_deref(), Some("scout")); |
| 579 | assert_eq!(scout.profile, None); |
| 580 | // Provider/model are not required identity fields on role steps. |
| 581 | assert_eq!(scout.model_policy.provider, None); |
| 582 | assert_eq!(scout.model_policy.model, None); |
| 583 | |
| 584 | let WorkflowNode::Leaf(fix) = &workflow.nodes[1] else { |
| 585 | panic!("second node should be a leaf"); |
| 586 | }; |
| 587 | assert_eq!(fix.role.as_deref(), Some("implementer")); |
| 588 | } |
| 589 | |
| 590 | #[test] |
| 591 | fn javascript_workflow_accepts_gate_specs() { |
| 592 | let source = r#" |
| 593 | workflow({ |
| 594 | "goal": "role gates", |
| 595 | "gates": [ |
| 596 | { |
| 597 | "id": "scout-findings", |
| 598 | "role": " Scout ", |
| 599 | "on": "role_complete", |
| 600 | "gate": "approve", |
| 601 | "on_fail": "block", |
| 602 | "blocks_role": " Implementer ", |
| 603 | "artifact_kind": "findings" |
| 604 | } |
| 605 | ], |
| 606 | "nodes": [ |
| 607 | { "agent": { "id": "scout", "prompt": "Find risk.", "role": "scout" } }, |
| 608 | { "agent": { "id": "fix", "prompt": "Use findings.", "role": "implementer" } } |
| 609 | ] |
| 610 | }); |
| 611 | "#; |
| 612 | |
| 613 | let workflow = |
| 614 | compile_javascript_workflow("gates.workflow.js", source).expect("compile gates"); |
| 615 | |
| 616 | assert_eq!(workflow.gates.len(), 1); |
| 617 | let gate = &workflow.gates[0]; |
| 618 | assert_eq!(gate.id, "scout-findings"); |
| 619 | assert_eq!(gate.role, "scout"); |
| 620 | assert_eq!(gate.on, GateOn::RoleComplete); |
| 621 | assert_eq!(gate.gate, GateKind::Approve); |
| 622 | assert_eq!(gate.on_fail, GateOnFail::Block); |
| 623 | assert_eq!(gate.blocks_role.as_deref(), Some("implementer")); |
| 624 | assert_eq!(gate.artifact_kind.as_deref(), Some("findings")); |
| 625 | } |
| 626 | |
| 627 | #[test] |
| 628 | fn stopship_acceptance_fixture_is_read_only_and_gate_complete() { |
| 629 | let source = include_str!("../../../workflows/stopship.workflow.js"); |
| 630 | let workflow = compile_javascript_workflow("stopship.workflow.js", source) |
| 631 | .expect("compile stopship acceptance fixture"); |
| 632 | |
| 633 | assert_eq!(workflow.id.as_deref(), Some("stopship-release-acceptance")); |
| 634 | let WorkflowNode::Sequence(sequence) = &workflow.nodes[0] else { |
| 635 | panic!("acceptance fixture should begin with one ordered role chain"); |
| 636 | }; |
| 637 | let expected_children = [ |
| 638 | ("explore", 6, 480, 96_000), |
| 639 | ("implement", 4, 420, 72_000), |
| 640 | ("reviewer", 4, 420, 72_000), |
| 641 | ("test", 4, 420, 72_000), |
| 642 | ("release_lead", 3, 300, 48_000), |
| 643 | ]; |
| 644 | let mut aggregate_token_cap = 0_u64; |
| 645 | assert_eq!(sequence.children.len(), expected_children.len()); |
| 646 | for (node, (expected_role, max_steps, timeout_secs, max_tokens)) in |
| 647 | sequence.children.iter().zip(expected_children) |
| 648 | { |
| 649 | let WorkflowNode::Leaf(leaf) = node else { |
| 650 | panic!("acceptance role chain must contain only agent leaves"); |
| 651 | }; |
| 652 | assert_eq!(leaf.role.as_deref(), Some(expected_role)); |
| 653 | assert_eq!(leaf.mode, TaskMode::ReadOnly); |
| 654 | assert!(!leaf.permissions.allow_write); |
| 655 | // No explicit `allowed_tools` anywhere: the read-only lowering |
| 656 | // alone scopes the source-gathering explore role, and its |
| 657 | // catalog-visible surface (tool_search + deferred grep_files) is |
| 658 | // pinned from the TUI side (scout_surface_keeps_tool_search_ |
| 659 | // grep_files_activation_path). |
| 660 | assert!(leaf.permissions.allowed_tools.is_empty()); |
| 661 | assert_eq!( |
| 662 | leaf.permissions.deny_all_tools, |
| 663 | expected_role != "explore", |
| 664 | "only the source-gathering explore role should receive tools" |
| 665 | ); |
| 666 | assert!( |
| 667 | leaf.prompt.contains( |
| 668 | "first non-empty line of your response must be exactly APPROVE or exactly BLOCK" |
| 669 | ), |
| 670 | "{expected_role} must declare the host-readable verdict contract" |
| 671 | ); |
| 672 | assert!( |
| 673 | leaf.prompt |
| 674 | .contains("Do not put any words before that verdict") |
| 675 | && leaf.prompt.contains("Here is the verdict"), |
| 676 | "{expected_role} must reject verdict preambles that the host cannot parse" |
| 677 | ); |
| 678 | if expected_role == "explore" { |
| 679 | assert!( |
| 680 | leaf.prompt.contains("exactly one `tool_search` call") |
| 681 | && leaf.prompt.contains("exactly one `grep_files` call") |
| 682 | && leaf.prompt.contains( |
| 683 | "the content-search tool is deferred, so this activation is required before it can be called" |
| 684 | ), |
| 685 | "the scout must activate the deferred content-search tool before searching with it" |
| 686 | ); |
| 687 | assert!( |
| 688 | leaf.prompt.contains("Make no other tool calls") |
| 689 | && leaf.prompt.contains("nothing else"), |
| 690 | "the scout discovery must stay one bounded activation-plus-search round" |
| 691 | ); |
| 692 | assert_eq!( |
| 693 | leaf.file_scope |
| 694 | .iter() |
| 695 | .map(String::as_str) |
| 696 | .collect::<Vec<_>>(), |
| 697 | vec![ |
| 698 | "fleets/stopship.toml", |
| 699 | "crates/cli/src/lib.rs", |
| 700 | "crates/workflow/src/role_resolve.rs", |
| 701 | "crates/tui/src/tools/workflow/mod.rs", |
| 702 | "crates/lane/src/runtime.rs", |
| 703 | ], |
| 704 | "the scout grep must not include its own authored prompt" |
| 705 | ); |
| 706 | let repository_root = |
| 707 | std::path::Path::new(env!("CARGO_MANIFEST_DIR")).join("../.."); |
| 708 | for path in &leaf.file_scope { |
| 709 | assert!( |
| 710 | repository_root.join(path).is_file(), |
| 711 | "scout evidence path must exist: {path}" |
| 712 | ); |
| 713 | } |
| 714 | assert!( |
| 715 | leaf.prompt.contains( |
| 716 | "`include` set exactly to [`fleets/stopship.toml`, `crates/cli/src/lib.rs`, `crates/workflow/src/role_resolve.rs`, `crates/tui/src/tools/workflow/mod.rs`, `crates/lane/src/runtime.rs`]" |
| 717 | ) && leaf.prompt.contains("Matches outside that exact include list do not count"), |
| 718 | "the grep_files search must constrain the actual tool input, not only file scope metadata" |
| 719 | ); |
| 720 | assert!( |
| 721 | leaf.prompt.contains("if you can populate all seven") |
| 722 | && leaf |
| 723 | .prompt |
| 724 | .contains("never return BLOCK after citing all seven") |
| 725 | && leaf |
| 726 | .prompt |
| 727 | .contains("identify each missing owner as MISSING"), |
| 728 | "the scout verdict must follow its own complete evidence artifact" |
| 729 | ); |
| 730 | } else { |
| 731 | assert!( |
| 732 | leaf.prompt.contains("Tools are intentionally unavailable") |
| 733 | && leaf.prompt.contains("promoted handoff") |
| 734 | && leaf.prompt.contains("all seven owners") |
| 735 | && leaf.prompt.contains("one concise row per owner"), |
| 736 | "{expected_role} must consume promoted evidence without reopening discovery" |
| 737 | ); |
| 738 | } |
| 739 | assert_eq!( |
| 740 | leaf.file_scope |
| 741 | .iter() |
| 742 | .map(String::as_str) |
| 743 | .collect::<Vec<_>>(), |
| 744 | vec![ |
| 745 | "fleets/stopship.toml", |
| 746 | "crates/cli/src/lib.rs", |
| 747 | "crates/workflow/src/role_resolve.rs", |
| 748 | "crates/tui/src/tools/workflow/mod.rs", |
| 749 | "crates/lane/src/runtime.rs", |
| 750 | ], |
| 751 | "every acceptance role must carry the same promoted evidence boundary" |
| 752 | ); |
| 753 | assert_eq!(leaf.budget.max_steps, Some(max_steps), "{expected_role}"); |
| 754 | let response_budget = match max_steps { |
| 755 | 6 => "at most six model responses", |
| 756 | 4 => "at most four model responses", |
| 757 | 3 => "at most three model responses", |
| 758 | _ => unreachable!("unexpected stopship model-response budget"), |
| 759 | }; |
| 760 | assert!( |
| 761 | leaf.prompt.contains(response_budget) && leaf.prompt.contains("with no tool calls"), |
| 762 | "{expected_role} must reserve a response for its explicit verdict" |
| 763 | ); |
| 764 | assert_eq!( |
| 765 | leaf.budget.timeout_secs, |
| 766 | Some(timeout_secs), |
| 767 | "{expected_role}" |
| 768 | ); |
| 769 | assert_eq!(leaf.budget.max_tokens, Some(max_tokens), "{expected_role}"); |
| 770 | assert!( |
| 771 | max_tokens < u64::from(max_steps) * 24_000, |
| 772 | "{expected_role} verdict reserve must not raise its token ceiling" |
| 773 | ); |
| 774 | aggregate_token_cap = aggregate_token_cap.saturating_add(max_tokens); |
| 775 | assert!( |
| 776 | leaf.profile.is_none(), |
| 777 | "Fleet must resolve the declared role" |
| 778 | ); |
| 779 | } |
| 780 | assert_eq!( |
| 781 | aggregate_token_cap, 360_000, |
| 782 | "the fixture must stay globally bounded when no shared override is supplied" |
| 783 | ); |
| 784 | |
| 785 | let expected_gates = [ |
| 786 | ("explore", Some("implement"), "source_evidence"), |
| 787 | ("implement", Some("reviewer"), "verification_plan"), |
| 788 | ("reviewer", Some("test"), "review_report"), |
| 789 | ("test", Some("release_lead"), "verification_report"), |
| 790 | ("release_lead", None, "final_receipt"), |
| 791 | ]; |
| 792 | assert_eq!(workflow.gates.len(), expected_gates.len()); |
| 793 | for (gate, (role, blocked_role, artifact_kind)) in workflow.gates.iter().zip(expected_gates) |
| 794 | { |
| 795 | assert_eq!(gate.role, role); |
| 796 | assert_eq!(gate.on, GateOn::RoleComplete); |
| 797 | assert_eq!(gate.on_fail, GateOnFail::Block); |
| 798 | assert_eq!(gate.blocks_role.as_deref(), blocked_role); |
| 799 | assert_eq!(gate.max_retries, 0); |
| 800 | assert_eq!(gate.artifact_kind.as_deref(), Some(artifact_kind)); |
| 801 | assert!( |
| 802 | gate.require_explicit_verdict, |
| 803 | "{role} gate must fail closed when its verdict is missing or malformed" |
| 804 | ); |
| 805 | } |
| 806 | |
| 807 | let mut board = LaneGateBoard::new("lane-fixture-contract"); |
| 808 | board.install_gates(&workflow.gates); |
| 809 | assert_eq!( |
| 810 | board |
| 811 | .evaluate(&workflow.gates[0], GateOutcome::Pass) |
| 812 | .expect("successful role promotes its gate"), |
| 813 | GateState::Passed |
| 814 | ); |
| 815 | let failure = board |
| 816 | .evaluate( |
| 817 | &workflow.gates[3], |
| 818 | GateOutcome::Fail { |
| 819 | reason: "verifier receipt missing".to_string(), |
| 820 | }, |
| 821 | ) |
| 822 | .expect("failed verifier updates its gate"); |
| 823 | assert!(matches!(failure, GateState::Blocked { .. })); |
| 824 | assert!( |
| 825 | board |
| 826 | .role_is_blocked(&workflow.gates, "release_lead") |
| 827 | .is_some() |
| 828 | ); |
| 829 | } |
| 830 | |
| 831 | #[test] |
| 832 | fn javascript_workflow_rejects_invalid_agent_profiles() { |
| 833 | for bad in [r#""""#, r#""has space""#, r#""quote\"y""#, r#""a=b""#] { |
| 834 | let source = format!( |
| 835 | r#" |
| 836 | workflow({{ |
| 837 | "goal": "bad profile", |
| 838 | "nodes": [ |
| 839 | {{ "agent": {{ "id": "scan", "prompt": "scan safely", "profile": {bad} }} }} |
| 840 | ] |
| 841 | }}); |
| 842 | "# |
| 843 | ); |
| 844 | |
| 845 | let err = compile_javascript_workflow("bad-profile.workflow.js", &source) |
| 846 | .expect_err("invalid profile should be rejected"); |
| 847 | |
| 848 | assert!( |
| 849 | matches!(err, JavascriptWorkflowError::InvalidNode(_)), |
| 850 | "profile {bad} should fail as an invalid node, got {err:?}" |
| 851 | ); |
| 852 | assert!(err.to_string().contains("profile")); |
| 853 | } |
| 854 | } |
| 855 | |
| 856 | #[test] |
| 857 | fn javascript_workflow_rejects_runtime_effects() { |
| 858 | let source = r#" |
| 859 | import fs from "fs"; |
| 860 | workflow({ "goal": "bad", "nodes": [] }); |
| 861 | "#; |
| 862 | |
| 863 | let err = compile_javascript_workflow("bad.workflow.js", source) |
| 864 | .expect_err("imports must be rejected"); |
| 865 | |
| 866 | assert!(matches!( |
| 867 | err, |
| 868 | JavascriptWorkflowError::UnsupportedConstruct { |
| 869 | construct: "import" |
| 870 | } |
| 871 | )); |
| 872 | } |
| 873 | |
| 874 | #[test] |
| 875 | fn javascript_workflow_rejects_unknown_result_reference() { |
| 876 | let source = r#" |
| 877 | workflow({ |
| 878 | "goal": "bad dependency", |
| 879 | "nodes": [ |
| 880 | { |
| 881 | "agent": { |
| 882 | "id": "scan", |
| 883 | "prompt": "scan safely", |
| 884 | "depends_on_results": ["missing"] |
| 885 | } |
| 886 | } |
| 887 | ] |
| 888 | }); |
| 889 | "#; |
| 890 | |
| 891 | let err = compile_javascript_workflow("bad-reference.workflow.js", source) |
| 892 | .expect_err("validation must reject unknown result references"); |
| 893 | |
| 894 | assert!(matches!(err, JavascriptWorkflowError::InvalidNode(_))); |
| 895 | assert!(err.to_string().contains("missing")); |
| 896 | } |
| 897 | |
| 898 | #[test] |
| 899 | fn javascript_example_compiles_to_validated_ir() { |
| 900 | let source = include_str!("../../../workflows/issue_audit.workflow.js"); |
| 901 | let workflow = |
| 902 | compile_javascript_workflow("issue_audit.workflow.js", source).expect("compile"); |
| 903 | |
| 904 | workflow |
| 905 | .validate_for_fleet() |
| 906 | .expect("example lowers to IR that passes Fleet validation"); |
| 907 | assert_eq!(workflow.id.as_deref(), Some("issue-audit-js")); |
| 908 | assert_eq!(workflow.nodes.len(), 2, "{:#?}", workflow.nodes); |
| 909 | let WorkflowNode::BranchSet(branch) = &workflow.nodes[0] else { |
| 910 | panic!("first node should be the parallel audit branch"); |
| 911 | }; |
| 912 | let leaf_ids: Vec<&str> = branch |
| 913 | .children |
| 914 | .iter() |
| 915 | .map(|child| match child { |
| 916 | WorkflowNode::Leaf(leaf) => leaf.id.as_str(), |
| 917 | other => panic!("audit branch child should be an agent leaf: {other:?}"), |
| 918 | }) |
| 919 | .collect(); |
| 920 | assert_eq!(leaf_ids, ["code-audit", "test-audit", "docs-audit"]); |
| 921 | let WorkflowNode::Reduce(reduce) = &workflow.nodes[1] else { |
| 922 | panic!("second node should reduce the audit findings"); |
| 923 | }; |
| 924 | assert_eq!(reduce.id, "synthesize-release-risk"); |
| 925 | assert_eq!(reduce.inputs, leaf_ids); |
| 926 | } |
| 927 | } |
| 928 |