| 1 | //! Exercise native plan lowering and the real provider-binding spawn boundary. |
| 2 | |
| 3 | use super::*; |
| 4 | use crate::client::CodewhaleClient; |
| 5 | use crate::config::{ApiProvider, Config}; |
| 6 | use crate::fleet::exact::{ExactFleetWorkflow, StaticFleetRouter}; |
| 7 | use crate::fleet::members::add_fleet_model; |
| 8 | use crate::fleet::store::{FleetFile, FleetScope, save_fleet, set_selected}; |
| 9 | use crate::tools::subagent::new_shared_subagent_manager; |
| 10 | use codewhale_workflow::{FleetDocument, QualifiedFleetId}; |
| 11 | use std::sync::atomic::AtomicUsize; |
| 12 | use std::time::Duration; |
| 13 | |
| 14 | const TARGET_MODEL: &str = "openai/gpt-4.1"; |
| 15 | const TARGET_SELECTOR: &str = "openrouter/openai/gpt-4.1"; |
| 16 | |
| 17 | struct RouteFixture { |
| 18 | config: Config, |
| 19 | runtime: SubAgentRuntime, |
| 20 | parent_calls: Arc<AtomicUsize>, |
| 21 | target_calls: Arc<AtomicUsize>, |
| 22 | target_bodies: Arc<Mutex<Vec<Value>>>, |
| 23 | } |
| 24 | |
| 25 | impl RouteFixture { |
| 26 | async fn new(workspace: &Path) -> Self { |
| 27 | // WorkflowVm dispatches on its own thread. Keep saved routes in the |
| 28 | // explicit workspace: the test-only personal-home resolver correctly |
| 29 | // refuses another thread's environment guard. |
| 30 | let fleet = FleetFile::new("Workflow fixture".into(), None).expect("empty saved Fleet"); |
| 31 | save_fleet(&fleet, FleetScope::Workspace, workspace).expect("save fixture Fleet"); |
| 32 | set_selected(&fleet.name, FleetScope::Workspace, workspace).expect("select fixture Fleet"); |
| 33 | let (parent, parent_calls, _) = tests::fake_chat_client_capturing("parent route").await; |
| 34 | let (target, target_calls, target_bodies) = |
| 35 | tests::fake_chat_client_capturing("frozen route result").await; |
| 36 | let mut config = Config { |
| 37 | api_key: Some("fixture-key".into()), |
| 38 | base_url: Some(parent.base_url().to_string()), |
| 39 | ..Default::default() |
| 40 | }; |
| 41 | config.set_provider_api_key_override(ApiProvider::Openrouter, Some("fixture-key".into())); |
| 42 | config.set_provider_base_url_override( |
| 43 | ApiProvider::Openrouter, |
| 44 | Some(target.base_url().to_string()), |
| 45 | ); |
| 46 | let client = CodewhaleClient::new(&config).expect("parent fixture client"); |
| 47 | let manager = new_shared_subagent_manager(workspace.to_path_buf(), 4); |
| 48 | let runtime = SubAgentRuntime::new( |
| 49 | client, |
| 50 | "deepseek-v4-pro".into(), |
| 51 | ToolContext::new(workspace.to_path_buf()).with_state_namespace("session-test"), |
| 52 | true, |
| 53 | None, |
| 54 | manager, |
| 55 | ) |
| 56 | .with_api_config(config.clone()); |
| 57 | Self { |
| 58 | config, |
| 59 | runtime, |
| 60 | parent_calls, |
| 61 | target_calls, |
| 62 | target_bodies, |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | fn driver(&self, fleet: WorkflowFleetBinding) -> Arc<SubAgentWorkflowDriver> { |
| 67 | let workspace = self.runtime.context.workspace.clone(); |
| 68 | let state = WorkflowWorkspaceState::open(&workspace); |
| 69 | let run_id = format!("shortlist-{}", Uuid::new_v4()); |
| 70 | state.runs.lock().expect("runs").insert( |
| 71 | run_id.clone(), |
| 72 | WorkflowRunRecord::new( |
| 73 | run_id.clone(), |
| 74 | Some("session-test".into()), |
| 75 | None, |
| 76 | None, |
| 77 | None, |
| 78 | ), |
| 79 | ); |
| 80 | SubAgentWorkflowDriver::new( |
| 81 | run_id, |
| 82 | "session-test".into(), |
| 83 | self.runtime.manager.clone(), |
| 84 | self.runtime.clone(), |
| 85 | state, |
| 86 | None, |
| 87 | fleet, |
| 88 | Vec::new(), |
| 89 | workspace, |
| 90 | ) |
| 91 | } |
| 92 | |
| 93 | fn assert_target_request(&self) { |
| 94 | assert_eq!( |
| 95 | self.parent_calls.load(Ordering::SeqCst), |
| 96 | 0, |
| 97 | "parent route must not receive the child" |
| 98 | ); |
| 99 | assert!( |
| 100 | self.target_calls.load(Ordering::SeqCst) > 0, |
| 101 | "the child must actually reach the selected provider" |
| 102 | ); |
| 103 | let bodies = self.target_bodies.lock().expect("request bodies"); |
| 104 | assert!(!bodies.is_empty()); |
| 105 | for body in bodies.iter() { |
| 106 | assert_eq!( |
| 107 | body["model"], TARGET_MODEL, |
| 108 | "provider wire model must match the saved route" |
| 109 | ); |
| 110 | } |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | /// The real VM has its own OS thread. Enroll it in this fixture's sealed test |
| 115 | /// environment before provider construction reads Config, and retain enrollment |
| 116 | /// while the real child tasks run on that same reactor. Otherwise Config's test |
| 117 | /// reader blocks on the environment lock held by the test awaiting the VM. |
| 118 | struct ScopedWorkflowDriver { |
| 119 | inner: Arc<SubAgentWorkflowDriver>, |
| 120 | ticket: crate::test_support::EnvScopeTicket, |
| 121 | membership: Mutex<Option<crate::test_support::EnvScopeMembership>>, |
| 122 | } |
| 123 | |
| 124 | #[async_trait] |
| 125 | impl WorkflowDriver for ScopedWorkflowDriver { |
| 126 | async fn spawn_task(&self, request: TaskRequest) -> Result<SpawnedTask, DriverError> { |
| 127 | { |
| 128 | let mut membership = self.membership.lock().expect("VM environment membership"); |
| 129 | if membership.is_none() { |
| 130 | *membership = Some( |
| 131 | crate::test_support::join_env_scope(Some(self.ticket)) |
| 132 | .expect("the originating test still owns its environment"), |
| 133 | ); |
| 134 | } |
| 135 | } |
| 136 | self.inner.spawn_task(request).await |
| 137 | } |
| 138 | |
| 139 | fn cancel_all(&self) { |
| 140 | self.inner.cancel_all(); |
| 141 | } |
| 142 | |
| 143 | fn budget(&self) -> BudgetSnapshot { |
| 144 | self.inner.budget() |
| 145 | } |
| 146 | |
| 147 | fn progress(&self, event: ProgressEvent) { |
| 148 | self.inner.progress(event); |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | async fn run_script(source: &str, driver: Arc<SubAgentWorkflowDriver>) -> Result<Value, String> { |
| 153 | let scoped = Arc::new(ScopedWorkflowDriver { |
| 154 | inner: driver.clone(), |
| 155 | ticket: crate::test_support::env_scope_ticket().expect("fixture owns test environment"), |
| 156 | membership: Mutex::new(None), |
| 157 | }); |
| 158 | let result = tokio::time::timeout( |
| 159 | Duration::from_secs(10), |
| 160 | WorkflowVm::new().run_script(source, json!({}), scoped), |
| 161 | ) |
| 162 | .await; |
| 163 | driver.cancel_all(); |
| 164 | result |
| 165 | .expect("workflow fixture must settle within ten seconds") |
| 166 | .map_err(|error| error.to_string()) |
| 167 | } |
| 168 | |
| 169 | fn native_script(child: Value) -> String { |
| 170 | let spec = structured_plan_to_workflow_spec(&json!({ |
| 171 | "goal": "review the route fixture", "risk": "read_only", "children": [child], |
| 172 | })) |
| 173 | .expect("valid native plan"); |
| 174 | lower_declarative_workflow_to_imperative_js(&spec).expect("lower native plan") |
| 175 | } |
| 176 | |
| 177 | fn select_role(workspace: &Path, model: &str, provider: &str) { |
| 178 | let fleet = FleetFile::parse(&format!( |
| 179 | "schema = 'fleet'\nschema_revision = 2\nname = 'Selected roster'\n\ |
| 180 | [[members]]\nid = 'auditor'\nrole = 'reviewer'\nprovider = '{provider}'\n\ |
| 181 | model = '{model}'\nreasoning = 'off'\n" |
| 182 | )) |
| 183 | .expect("saved roster"); |
| 184 | save_fleet(&fleet, FleetScope::Workspace, workspace).expect("save selected roster"); |
| 185 | set_selected(&fleet.name, FleetScope::Workspace, workspace).expect("select roster"); |
| 186 | } |
| 187 | |
| 188 | fn exact_document(reasoning: &str) -> FleetDocument { |
| 189 | let router = if reasoning == "auto" { |
| 190 | "reasoning_router = 'fixture-router'\n" |
| 191 | } else { |
| 192 | "" |
| 193 | }; |
| 194 | FleetDocument::parse(&format!( |
| 195 | "name = 'frozen-audit'\nschema = 'exact'\n{router}\ |
| 196 | [[members]]\nid = 'auditor'\nrole = 'reviewer'\nprovider = 'openrouter'\n\ |
| 197 | model = '{TARGET_MODEL}'\nreasoning = '{reasoning}'\n" |
| 198 | )) |
| 199 | .expect("exact Fleet document") |
| 200 | } |
| 201 | |
| 202 | #[tokio::test] |
| 203 | #[allow(clippy::await_holding_lock)] |
| 204 | async fn native_shortlisted_model_reaches_its_configured_provider_request() { |
| 205 | let _retry = tests::workflow_test_retry_guard(); |
| 206 | let _env = crate::test_support::lock_test_env(); |
| 207 | let root = tempfile::tempdir().unwrap(); |
| 208 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 209 | let fixture = RouteFixture::new(root.path()).await; |
| 210 | add_fleet_model(root.path(), "openrouter", TARGET_MODEL, &[]).unwrap(); |
| 211 | let source = native_script(json!({ |
| 212 | "prompt": "read-only route check", "type": "reviewer", "model": TARGET_SELECTOR, |
| 213 | })); |
| 214 | run_script(&source, fixture.driver(WorkflowFleetBinding::None)) |
| 215 | .await |
| 216 | .expect("shortlisted native child runs"); |
| 217 | fixture.assert_target_request(); |
| 218 | } |
| 219 | |
| 220 | #[tokio::test] |
| 221 | #[allow(clippy::await_holding_lock)] |
| 222 | async fn native_non_shortlisted_model_is_rejected_without_a_provider_call() { |
| 223 | let _retry = tests::workflow_test_retry_guard(); |
| 224 | let _env = crate::test_support::lock_test_env(); |
| 225 | let root = tempfile::tempdir().unwrap(); |
| 226 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 227 | let fixture = RouteFixture::new(root.path()).await; |
| 228 | add_fleet_model(root.path(), "openrouter", TARGET_MODEL, &[]).unwrap(); |
| 229 | let source = native_script(json!({ |
| 230 | "prompt": "read-only route check", "type": "reviewer", "model": "openrouter/outside-fixture", |
| 231 | })); |
| 232 | let error = run_script(&source, fixture.driver(WorkflowFleetBinding::None)) |
| 233 | .await |
| 234 | .expect_err("closed shortlist"); |
| 235 | assert!(error.contains("outside the selected Fleet"), "{error}"); |
| 236 | assert_eq!(fixture.parent_calls.load(Ordering::SeqCst), 0); |
| 237 | assert_eq!(fixture.target_calls.load(Ordering::SeqCst), 0); |
| 238 | } |
| 239 | |
| 240 | #[tokio::test] |
| 241 | #[allow(clippy::await_holding_lock)] |
| 242 | async fn native_role_without_model_uses_the_saved_provider_and_thinking() { |
| 243 | let _retry = tests::workflow_test_retry_guard(); |
| 244 | let _env = crate::test_support::lock_test_env(); |
| 245 | let root = tempfile::tempdir().unwrap(); |
| 246 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 247 | let fixture = RouteFixture::new(root.path()).await; |
| 248 | select_role(root.path(), TARGET_MODEL, "openrouter"); |
| 249 | let source = native_script(json!({"prompt": "read-only route check", "profile": "auditor"})); |
| 250 | let driver = fixture.driver(WorkflowFleetBinding::None); |
| 251 | run_script(&source, driver.clone()) |
| 252 | .await |
| 253 | .expect("saved role runs"); |
| 254 | fixture.assert_target_request(); |
| 255 | let runs = driver.state.runs.lock().expect("runs"); |
| 256 | let run = runs.get(&driver.run_id).expect("run record"); |
| 257 | let started = run |
| 258 | .events |
| 259 | .iter() |
| 260 | .find(|event| event.event_type() == "task_started") |
| 261 | .expect("task receipt"); |
| 262 | let WorkflowUiEventKind::TaskStarted(started) = &started.kind else { |
| 263 | unreachable!() |
| 264 | }; |
| 265 | assert_eq!(started.resolved_profile.as_deref(), Some("auditor")); |
| 266 | assert_eq!(started.resolved_provider, "openrouter"); |
| 267 | assert_eq!(started.resolved_model, TARGET_MODEL); |
| 268 | assert_eq!(started.effective_reasoning.as_deref(), Some("off")); |
| 269 | } |
| 270 | |
| 271 | #[tokio::test] |
| 272 | #[allow(clippy::await_holding_lock)] |
| 273 | async fn exact_fleet_frozen_route_survives_a_conflicting_selected_roster() { |
| 274 | let _retry = tests::workflow_test_retry_guard(); |
| 275 | let _env = crate::test_support::lock_test_env(); |
| 276 | let root = tempfile::tempdir().unwrap(); |
| 277 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 278 | let fixture = RouteFixture::new(root.path()).await; |
| 279 | // Capture first. Later selected-Pod edits must affect neither this member |
| 280 | // nor the preflighted provider endpoint, even when the member id collides. |
| 281 | let operation = ExactFleetWorkflow::capture( |
| 282 | &exact_document("off"), |
| 283 | QualifiedFleetId { |
| 284 | name: "frozen-audit".into(), |
| 285 | origin: "workspace".into(), |
| 286 | }, |
| 287 | "2026-09-13T00:00:00Z", |
| 288 | Some(&fixture.config), |
| 289 | &[], |
| 290 | ) |
| 291 | .expect("preflight exact route against loopback provider"); |
| 292 | select_role(root.path(), "deepseek-v4-pro", "deepseek"); |
| 293 | let source = |
| 294 | native_script(json!({"prompt": "read-only frozen route check", "profile": "auditor"})); |
| 295 | run_script( |
| 296 | &source, |
| 297 | fixture.driver(WorkflowFleetBinding::Exact(Arc::new(operation))), |
| 298 | ) |
| 299 | .await |
| 300 | .expect("frozen exact child runs independently of selected Fleet"); |
| 301 | fixture.assert_target_request(); |
| 302 | } |
| 303 | |
| 304 | #[tokio::test] |
| 305 | #[allow(clippy::await_holding_lock)] |
| 306 | async fn exact_fleet_model_override_is_rejected_before_router_or_provider() { |
| 307 | let _retry = tests::workflow_test_retry_guard(); |
| 308 | let _env = crate::test_support::lock_test_env(); |
| 309 | let root = tempfile::tempdir().unwrap(); |
| 310 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 311 | let fixture = RouteFixture::new(root.path()).await; |
| 312 | let router = StaticFleetRouter::new(r#"{"reasoning":"high"}"#); |
| 313 | let operation = ExactFleetWorkflow::for_tests( |
| 314 | &exact_document("auto"), |
| 315 | QualifiedFleetId { |
| 316 | name: "frozen-audit".into(), |
| 317 | origin: "workspace".into(), |
| 318 | }, |
| 319 | Some(router.clone()), |
| 320 | ); |
| 321 | let error = run_script( |
| 322 | "return await task({description:'read-only check', profile:'auditor', writeAuthority:'read_only', model:'openrouter/outside-fixture'});", |
| 323 | fixture.driver(WorkflowFleetBinding::Exact(Arc::new(operation))), |
| 324 | ).await.expect_err("exact Fleet refuses task route overrides"); |
| 325 | assert!( |
| 326 | error.contains("task option `model` is not allowed"), |
| 327 | "{error}" |
| 328 | ); |
| 329 | assert!( |
| 330 | router.seen.lock().unwrap().is_empty(), |
| 331 | "rejection must precede reasoning spend" |
| 332 | ); |
| 333 | assert_eq!(fixture.parent_calls.load(Ordering::SeqCst), 0); |
| 334 | assert_eq!(fixture.target_calls.load(Ordering::SeqCst), 0); |
| 335 | } |
| 336 | |
| 337 | #[tokio::test] |
| 338 | #[allow(clippy::await_holding_lock)] |
| 339 | async fn exact_fleet_changed_provider_endpoint_is_rejected_before_dispatch() { |
| 340 | let _retry = tests::workflow_test_retry_guard(); |
| 341 | let _env = crate::test_support::lock_test_env(); |
| 342 | let root = tempfile::tempdir().unwrap(); |
| 343 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 344 | let mut fixture = RouteFixture::new(root.path()).await; |
| 345 | let operation = ExactFleetWorkflow::capture( |
| 346 | &exact_document("off"), |
| 347 | QualifiedFleetId { |
| 348 | name: "frozen-audit".into(), |
| 349 | origin: "workspace".into(), |
| 350 | }, |
| 351 | "2026-09-13T00:00:00Z", |
| 352 | Some(&fixture.config), |
| 353 | &[], |
| 354 | ) |
| 355 | .expect("capture exact route"); |
| 356 | let mut changed = fixture.config.clone(); |
| 357 | changed.set_provider_base_url_override( |
| 358 | ApiProvider::Openrouter, |
| 359 | Some(fixture.runtime.client.base_url().to_string()), |
| 360 | ); |
| 361 | fixture.runtime.api_config = Some(Arc::new(changed)); |
| 362 | let error = run_script( |
| 363 | "return await task({description:'read-only check', profile:'auditor', writeAuthority:'read_only'});", |
| 364 | fixture.driver(WorkflowFleetBinding::Exact(Arc::new(operation))), |
| 365 | ) |
| 366 | .await |
| 367 | .expect_err("changed endpoint cannot inherit the frozen receipt"); |
| 368 | assert!(error.contains("exact Fleet route changed"), "{error}"); |
| 369 | assert_eq!(fixture.parent_calls.load(Ordering::SeqCst), 0); |
| 370 | assert_eq!(fixture.target_calls.load(Ordering::SeqCst), 0); |
| 371 | } |
| 372 | |
| 373 | #[tokio::test] |
| 374 | #[allow(clippy::await_holding_lock)] |
| 375 | async fn native_exact_fleet_builder_keeps_the_plan_read_only_ceiling() { |
| 376 | let _retry = tests::workflow_test_retry_guard(); |
| 377 | let _env = crate::test_support::lock_test_env(); |
| 378 | let root = tempfile::tempdir().unwrap(); |
| 379 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 380 | let fixture = RouteFixture::new(root.path()).await; |
| 381 | let document = FleetDocument::parse(&format!( |
| 382 | "name = 'frozen-builder'\nschema = 'exact'\n[[members]]\n\ |
| 383 | id = 'builder-one'\nrole = 'builder'\nprovider = 'openrouter'\n\ |
| 384 | model = '{TARGET_MODEL}'\nreasoning = 'off'\n" |
| 385 | )) |
| 386 | .unwrap(); |
| 387 | let operation = ExactFleetWorkflow::capture( |
| 388 | &document, |
| 389 | QualifiedFleetId { |
| 390 | name: "frozen-builder".into(), |
| 391 | origin: "workspace".into(), |
| 392 | }, |
| 393 | "2026-09-13T00:00:00Z", |
| 394 | Some(&fixture.config), |
| 395 | &[], |
| 396 | ) |
| 397 | .expect("capture exact builder"); |
| 398 | let source = |
| 399 | native_script(json!({"prompt": "inspect without changing files", "role": "builder"})); |
| 400 | run_script( |
| 401 | &source, |
| 402 | fixture.driver(WorkflowFleetBinding::Exact(Arc::new(operation))), |
| 403 | ) |
| 404 | .await |
| 405 | .expect("a native plan may narrow an exact builder"); |
| 406 | fixture.assert_target_request(); |
| 407 | let manager = fixture.runtime.manager.read().await; |
| 408 | let records = manager.list_worker_records(); |
| 409 | assert_eq!(records.len(), 1); |
| 410 | let profile = &records[0].spec.runtime_profile; |
| 411 | assert!( |
| 412 | !profile.permissions.write, |
| 413 | "the authored read_only mode must remain executable policy" |
| 414 | ); |
| 415 | assert_eq!(profile.shell, crate::worker_profile::ShellPolicy::None); |
| 416 | assert_eq!( |
| 417 | profile.tools, |
| 418 | crate::worker_profile::ToolScope::Explicit(vec!["File".into()]) |
| 419 | ); |
| 420 | assert_eq!( |
| 421 | records[0].spec.child_route.as_ref().unwrap().provider_id, |
| 422 | "openrouter" |
| 423 | ); |
| 424 | } |
| 425 | |
| 426 | #[tokio::test] |
| 427 | #[allow(clippy::await_holding_lock)] |
| 428 | async fn exact_fleet_task_cannot_widen_a_read_only_role_before_router_or_provider() { |
| 429 | let _retry = tests::workflow_test_retry_guard(); |
| 430 | let _env = crate::test_support::lock_test_env(); |
| 431 | let root = tempfile::tempdir().unwrap(); |
| 432 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 433 | let fixture = RouteFixture::new(root.path()).await; |
| 434 | let router = StaticFleetRouter::new(r#"{"reasoning":"high"}"#); |
| 435 | let operation = ExactFleetWorkflow::for_tests( |
| 436 | &exact_document("auto"), |
| 437 | QualifiedFleetId { |
| 438 | name: "frozen-audit".into(), |
| 439 | origin: "workspace".into(), |
| 440 | }, |
| 441 | Some(router.clone()), |
| 442 | ); |
| 443 | let error = run_script( |
| 444 | "return await task({description:'attempt widening', profile:'auditor', writeAuthority:'workspace_write', writeRoots:['src']});", |
| 445 | fixture.driver(WorkflowFleetBinding::Exact(Arc::new(operation))), |
| 446 | ).await.expect_err("task cannot widen the Runtime role"); |
| 447 | assert!(error.contains("cannot request write authority"), "{error}"); |
| 448 | assert!(router.seen.lock().unwrap().is_empty()); |
| 449 | assert_eq!(fixture.parent_calls.load(Ordering::SeqCst), 0); |
| 450 | assert_eq!(fixture.target_calls.load(Ordering::SeqCst), 0); |
| 451 | } |
| 452 |