| 1 | //! Consumer regressions for operator-visible routes (#5915/#5955). |
| 2 | use super::*; |
| 3 | |
| 4 | #[tokio::test] |
| 5 | async fn roster_matches_actual_start_receipts_and_refreshes_live_role_defaults() { |
| 6 | let _env = crate::test_support::lock_test_env(); |
| 7 | let root = tempdir().unwrap(); |
| 8 | let (client, calls, _, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 9 | let config = crate::config::Config { |
| 10 | subagents: Some(crate::config::SubagentsConfig { |
| 11 | worker_model: Some("deepseek-v4-flash".into()), |
| 12 | ..Default::default() |
| 13 | }), |
| 14 | ..Default::default() |
| 15 | } |
| 16 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 17 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 8); |
| 18 | let context = ToolContext::new(root.path()).with_state_namespace("roster-route-consumer"); |
| 19 | let mut runtime = SubAgentRuntime::new( |
| 20 | client, |
| 21 | "deepseek-v4-pro".into(), |
| 22 | context.clone(), |
| 23 | false, |
| 24 | None, |
| 25 | manager.clone(), |
| 26 | ) |
| 27 | .with_api_config(config); |
| 28 | runtime |
| 29 | .role_models |
| 30 | .insert("general".into(), "deepseek-v4-pro".into()); |
| 31 | let tool = AgentTool::new(manager.clone(), runtime); |
| 32 | let query = tool |
| 33 | .execute(json!({"action":"roster"}), &context) |
| 34 | .await |
| 35 | .unwrap(); |
| 36 | let roster: Value = serde_json::from_str(&query.content).unwrap(); |
| 37 | assert_eq!( |
| 38 | calls.load(Ordering::SeqCst), |
| 39 | 0, |
| 40 | "discovery must not send inference" |
| 41 | ); |
| 42 | let rows = roster["members"].as_array().unwrap(); |
| 43 | assert_eq!(rows.len(), 8); |
| 44 | assert_eq!( |
| 45 | rows[0]["route"]["model"], "deepseek-v4-flash", |
| 46 | "live config supersedes launch default" |
| 47 | ); |
| 48 | for row in rows { |
| 49 | assert!(row["route"].is_object(), "route missing: {row}"); |
| 50 | assert_eq!(row["route"]["reachability"], "unverified"); |
| 51 | let role = row["role"].as_str().unwrap(); |
| 52 | let mut request = json!({"action":"start", "type":role, "prompt":"Say done."}); |
| 53 | if role == "custom" { |
| 54 | request["allowed_tools"] = json!(["Read"]); |
| 55 | } |
| 56 | let started = tool.execute(request, &context).await.unwrap(); |
| 57 | let metadata = started.metadata.as_ref().unwrap(); |
| 58 | let receipt = &metadata["child_route"]; |
| 59 | for (discovery, dispatch) in [ |
| 60 | ("provider", "provider_id"), |
| 61 | ("model", "model_id"), |
| 62 | ("reasoning_effort", "effective_reasoning"), |
| 63 | ("source", "route_source"), |
| 64 | ] { |
| 65 | assert_eq!( |
| 66 | row["route"][discovery], receipt[dispatch], |
| 67 | "{role}: {discovery}" |
| 68 | ); |
| 69 | } |
| 70 | manager |
| 71 | .write() |
| 72 | .await |
| 73 | .cancel_agent(metadata["agent_id"].as_str().unwrap()) |
| 74 | .unwrap(); |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | #[tokio::test] |
| 79 | async fn roster_preserves_unknown_and_non_metered_costs_and_invalid_role_errors() { |
| 80 | let _env = crate::test_support::lock_test_env(); |
| 81 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 82 | crate::provider_lake::clear_live_snapshot(); |
| 83 | for (provider, model, vendor, expected_cost, reason) in [ |
| 84 | ("deepseek", "deepseek-v4-flash", None, "paid", None), |
| 85 | ( |
| 86 | "openrouter", |
| 87 | "qwen/qwen3.7-plus", |
| 88 | Some("cerebras"), |
| 89 | "unknown", |
| 90 | Some("routing_dependent_price"), |
| 91 | ), |
| 92 | ( |
| 93 | "ollama", |
| 94 | "fixture-local-model", |
| 95 | None, |
| 96 | "not_money_metered", |
| 97 | Some("not_money_metered"), |
| 98 | ), |
| 99 | ] { |
| 100 | let root = tempdir().unwrap(); |
| 101 | let mut config = crate::config::Config { |
| 102 | provider: Some(provider.into()), |
| 103 | ..Default::default() |
| 104 | }; |
| 105 | let selected = config |
| 106 | .provider_config_for_mut( |
| 107 | &config.test_identity_for_kind(ProviderKind::parse(provider).unwrap()), |
| 108 | ) |
| 109 | .unwrap(); |
| 110 | selected.api_key = Some("roster-private-fixture-key".into()); |
| 111 | selected.model = Some(model.into()); |
| 112 | selected.vendor = vendor.map(str::to_string); |
| 113 | let client = CodewhaleClient::new(&config).unwrap(); |
| 114 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 1); |
| 115 | let runtime = SubAgentRuntime::new( |
| 116 | client, |
| 117 | model.into(), |
| 118 | ToolContext::new(root.path()), |
| 119 | false, |
| 120 | None, |
| 121 | manager, |
| 122 | ) |
| 123 | .with_api_config(config); |
| 124 | let row = |
| 125 | resolved_role_roster_entry(&runtime, &spawn_roster(&runtime), &FleetRole::Worker).await; |
| 126 | assert_eq!( |
| 127 | row["route"]["cost_class"], expected_cost, |
| 128 | "{provider}: {row}" |
| 129 | ); |
| 130 | assert_eq!(row["route"]["unpriced_reason"], json!(reason)); |
| 131 | assert!(!row.to_string().contains("roster-private-fixture-key")); |
| 132 | } |
| 133 | let mut runtime = stub_runtime(); |
| 134 | runtime |
| 135 | .role_models |
| 136 | .insert("general".into(), "invalid\nmodel".into()); |
| 137 | let row = |
| 138 | resolved_role_roster_entry(&runtime, &spawn_roster(&runtime), &FleetRole::Worker).await; |
| 139 | assert!(row["route"].is_null()); |
| 140 | assert!(row["route_error"].as_str().is_some()); |
| 141 | let other = |
| 142 | resolved_role_roster_entry(&runtime, &spawn_roster(&runtime), &FleetRole::Reviewer).await; |
| 143 | assert!( |
| 144 | other["route"].is_object(), |
| 145 | "one bad role must not hide other routes: {other}" |
| 146 | ); |
| 147 | } |
| 148 | |
| 149 | #[tokio::test] |
| 150 | async fn advertised_task_route_overrides_reach_start_and_foreign_models_fail_before_admission() { |
| 151 | let _env = crate::test_support::lock_test_env(); |
| 152 | let root = tempdir().unwrap(); |
| 153 | let (client, _, _, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 154 | let config = crate::config::Config { |
| 155 | ..Default::default() |
| 156 | } |
| 157 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 158 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 2); |
| 159 | let context = ToolContext::new(root.path()).with_state_namespace("explicit-task-route"); |
| 160 | let runtime = SubAgentRuntime::new( |
| 161 | client, |
| 162 | "deepseek-v4-flash".into(), |
| 163 | context.clone(), |
| 164 | false, |
| 165 | None, |
| 166 | manager.clone(), |
| 167 | ) |
| 168 | .with_api_config(config); |
| 169 | let tool = AgentTool::new(manager.clone(), runtime); |
| 170 | let schema = tool.input_schema(); |
| 171 | for field in ["model", "model_strength", "thinking"] { |
| 172 | assert!(schema["properties"].get(field).is_some()); |
| 173 | } |
| 174 | let started = tool.execute(json!({"action":"start", "type":"explore", "prompt":"Say done.", "model":"deepseek-v4-pro", "model_strength":"faster", "thinking":"high"}), &context).await.unwrap(); |
| 175 | let metadata = started.metadata.as_ref().unwrap(); |
| 176 | let receipt = &metadata["child_route"]; |
| 177 | assert_eq!(receipt["model_id"], "deepseek-v4-pro"); |
| 178 | assert_eq!(receipt["route_source"], "task.model"); |
| 179 | assert_eq!(receipt["effective_reasoning"], "high"); |
| 180 | manager |
| 181 | .write() |
| 182 | .await |
| 183 | .cancel_agent(metadata["agent_id"].as_str().unwrap()) |
| 184 | .unwrap(); |
| 185 | let error = tool.execute(json!({"action":"start", "type":"explore", "prompt":"Say done.", "model":"claude-fable-5"}), &context).await.unwrap_err(); |
| 186 | assert!(error.to_string().contains("provider"), "{error}"); |
| 187 | } |
| 188 | |
| 189 | struct ProjectProfilesGuard(bool); |
| 190 | impl ProjectProfilesGuard { |
| 191 | fn enabled() -> Self { |
| 192 | let previous = crate::fleet::roster::project_agent_profiles_enabled(); |
| 193 | crate::fleet::roster::set_project_agent_profiles_enabled(true); |
| 194 | Self(previous) |
| 195 | } |
| 196 | } |
| 197 | impl Drop for ProjectProfilesGuard { |
| 198 | fn drop(&mut self) { |
| 199 | crate::fleet::roster::set_project_agent_profiles_enabled(self.0); |
| 200 | } |
| 201 | } |
| 202 | |
| 203 | #[tokio::test] |
| 204 | async fn saved_profile_discovery_and_actual_start_share_current_instructions_route_and_trust() { |
| 205 | let _env = crate::test_support::lock_test_env(); |
| 206 | let root = tempdir().unwrap(); |
| 207 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 208 | let _project = ProjectProfilesGuard::enabled(); |
| 209 | let profile_dir = root.path().join(".codewhale/agents"); |
| 210 | std::fs::create_dir_all(&profile_dir).unwrap(); |
| 211 | let profile = profile_dir.join("bug-hunter.toml"); |
| 212 | let (client, calls, bodies, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 213 | let config = crate::config::Config { |
| 214 | subagents: Some(crate::config::SubagentsConfig { |
| 215 | explorer_model: Some("invalid\nrole-default".into()), |
| 216 | ..Default::default() |
| 217 | }), |
| 218 | ..Default::default() |
| 219 | } |
| 220 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 221 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 2); |
| 222 | let context = ToolContext::new(root.path()).with_state_namespace("saved-profile-consumer"); |
| 223 | let runtime = SubAgentRuntime::new( |
| 224 | client, |
| 225 | "deepseek-v4-flash".into(), |
| 226 | context.clone(), |
| 227 | false, |
| 228 | None, |
| 229 | manager.clone(), |
| 230 | ) |
| 231 | .with_api_config(config); |
| 232 | let tool = AgentTool::new(manager.clone(), runtime); |
| 233 | for (model, instruction) in [ |
| 234 | ("deepseek-v4-pro", "Inspect only changed parser branches."), |
| 235 | ("deepseek-v4-flash", "Inspect the new queue consumer."), |
| 236 | ] { |
| 237 | std::fs::write(&profile, format!("id = \"bug-hunter\"\nbase_role = \"scout\"\nmodel = \"{model}\"\nreasoning_effort = \"high\"\npersona = \"{instruction}\"\n")).unwrap(); |
| 238 | let before = calls.load(Ordering::SeqCst); |
| 239 | let discovered = tool |
| 240 | .execute(json!({"action":"roster"}), &context) |
| 241 | .await |
| 242 | .unwrap(); |
| 243 | assert_eq!( |
| 244 | calls.load(Ordering::SeqCst), |
| 245 | before, |
| 246 | "roster must never infer" |
| 247 | ); |
| 248 | let roster: Value = serde_json::from_str(&discovered.content).unwrap(); |
| 249 | let row = roster["profiles"] |
| 250 | .as_array() |
| 251 | .unwrap() |
| 252 | .iter() |
| 253 | .find(|row| row["member_id"] == "bug-hunter") |
| 254 | .unwrap(); |
| 255 | assert_eq!(row["route"]["model"], model); |
| 256 | assert_eq!(row["route"]["reasoning_effort"], "high"); |
| 257 | let started = tool |
| 258 | .execute( |
| 259 | json!({"profile":"bug-hunter", "prompt":"Inspect the assigned slice."}), |
| 260 | &context, |
| 261 | ) |
| 262 | .await |
| 263 | .unwrap(); |
| 264 | let meta = started.metadata.as_ref().unwrap(); |
| 265 | let receipt = &meta["child_route"]; |
| 266 | assert_eq!(receipt["resolved_profile_id"], "bug-hunter"); |
| 267 | assert_eq!(receipt["profile_origin"], "project"); |
| 268 | assert_eq!(receipt["model_id"], row["route"]["model"]); |
| 269 | assert_eq!( |
| 270 | receipt["effective_reasoning"], |
| 271 | row["route"]["reasoning_effort"] |
| 272 | ); |
| 273 | assert_eq!(receipt["route_source"], "agent_profile.model"); |
| 274 | tokio::time::timeout(Duration::from_secs(5), async { |
| 275 | while calls.load(Ordering::SeqCst) == before { |
| 276 | tokio::time::sleep(Duration::from_millis(10)).await; |
| 277 | } |
| 278 | }) |
| 279 | .await |
| 280 | .expect("local provider receives the saved profile prompt"); |
| 281 | let body = bodies.lock().unwrap().last().unwrap().clone(); |
| 282 | assert!( |
| 283 | body.to_string().contains(instruction), |
| 284 | "saved instructions must reach the actual request" |
| 285 | ); |
| 286 | let id = meta["agent_id"].as_str().unwrap(); |
| 287 | let mut guard = manager.write().await; |
| 288 | let worker = guard.worker_records.get(id).unwrap(); |
| 289 | assert!( |
| 290 | worker |
| 291 | .spec |
| 292 | .launch_manifest |
| 293 | .as_ref() |
| 294 | .unwrap() |
| 295 | .prompt |
| 296 | .contains(instruction) |
| 297 | ); |
| 298 | assert!(!worker.spec.runtime_profile.permissions.write); |
| 299 | if guard.agents[id].status == SubAgentStatus::Running { |
| 300 | guard.cancel_agent(id).unwrap(); |
| 301 | } |
| 302 | } |
| 303 | crate::fleet::roster::set_project_agent_profiles_enabled(false); |
| 304 | let error = tool |
| 305 | .execute( |
| 306 | json!({"profile":"bug-hunter", "prompt":"Inspect."}), |
| 307 | &context, |
| 308 | ) |
| 309 | .await |
| 310 | .unwrap_err(); |
| 311 | assert!( |
| 312 | error.to_string().contains("Unknown Fleet role/profile"), |
| 313 | "{error}" |
| 314 | ); |
| 315 | } |
| 316 | |
| 317 | #[tokio::test] |
| 318 | async fn saved_provider_pin_reaches_actual_request_and_conflicts_fail_before_admission() { |
| 319 | let _env = crate::test_support::lock_test_env(); |
| 320 | let root = tempdir().unwrap(); |
| 321 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 322 | let _project = ProjectProfilesGuard::enabled(); |
| 323 | let profile_dir = root.path().join(".codewhale/agents"); |
| 324 | std::fs::create_dir_all(&profile_dir).unwrap(); |
| 325 | std::fs::write(profile_dir.join("router-review.toml"), "id = \"router-review\"\nbase_role = \"reviewer\"\nprovider = \"openrouter\"\nmodel = \"qwen/qwen3.7-plus\"\nreasoning_effort = \"low\"\n").unwrap(); |
| 326 | let (client, calls, bodies, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 327 | let mut config = crate::config::Config { |
| 328 | ..Default::default() |
| 329 | } |
| 330 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 331 | let router = config |
| 332 | .provider_config_for_mut(&config.test_identity_for_kind(ProviderKind::Openrouter)) |
| 333 | .unwrap(); |
| 334 | router.api_key = Some("test-router-key".into()); |
| 335 | router.base_url = Some(client.base_url().into()); |
| 336 | router.vendor = Some("cerebras".into()); |
| 337 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 2); |
| 338 | let context = ToolContext::new(root.path()).with_state_namespace("saved-provider-consumer"); |
| 339 | let runtime = SubAgentRuntime::new( |
| 340 | client, |
| 341 | "deepseek-v4-flash".into(), |
| 342 | context.clone(), |
| 343 | false, |
| 344 | None, |
| 345 | manager.clone(), |
| 346 | ) |
| 347 | .with_api_config(config); |
| 348 | let tool = AgentTool::new(manager.clone(), runtime); |
| 349 | for extra in [ |
| 350 | json!({"model":"deepseek-v4-pro"}), |
| 351 | json!({"model_strength":"faster"}), |
| 352 | json!({"type":"builder"}), |
| 353 | ] { |
| 354 | let mut input = json!({"profile":"router-review", "prompt":"Inspect."}); |
| 355 | input |
| 356 | .as_object_mut() |
| 357 | .unwrap() |
| 358 | .extend(extra.as_object().unwrap().clone()); |
| 359 | assert!(tool.execute(input, &context).await.is_err()); |
| 360 | } |
| 361 | assert!(manager.read().await.agents.is_empty()); |
| 362 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 363 | let discovered = tool |
| 364 | .execute(json!({"action":"roster"}), &context) |
| 365 | .await |
| 366 | .unwrap(); |
| 367 | let roster: Value = serde_json::from_str(&discovered.content).unwrap(); |
| 368 | let row = roster["profiles"] |
| 369 | .as_array() |
| 370 | .unwrap() |
| 371 | .iter() |
| 372 | .find(|row| row["member_id"] == "router-review") |
| 373 | .unwrap(); |
| 374 | assert_eq!(row["route"]["provider"], "openrouter"); |
| 375 | assert_eq!(row["route"]["openrouter_vendor"], "cerebras"); |
| 376 | assert_eq!(row["route"]["cost_class"], "unknown"); |
| 377 | let started = tool |
| 378 | .execute( |
| 379 | json!({"profile":"router-review", "prompt":"Say done.", "thinking":"high"}), |
| 380 | &context, |
| 381 | ) |
| 382 | .await |
| 383 | .unwrap(); |
| 384 | let meta = started.metadata.as_ref().unwrap(); |
| 385 | assert_eq!(meta["child_route"]["provider_id"], "openrouter"); |
| 386 | assert_eq!(meta["child_route"]["model_id"], "qwen/qwen3.7-plus"); |
| 387 | assert_eq!(meta["child_route"]["effective_reasoning"], "high"); |
| 388 | tokio::time::timeout(Duration::from_secs(5), async { |
| 389 | while calls.load(Ordering::SeqCst) == 0 { |
| 390 | tokio::time::sleep(Duration::from_millis(10)).await; |
| 391 | } |
| 392 | }) |
| 393 | .await |
| 394 | .expect("local provider fixture receives child request"); |
| 395 | let body = bodies.lock().unwrap()[0].clone(); |
| 396 | assert_eq!(body["model"], "qwen/qwen3.7-plus"); |
| 397 | assert_eq!(body["provider"]["order"], json!(["cerebras"])); |
| 398 | assert_eq!(body["provider"]["allow_fallbacks"], false); |
| 399 | assert!( |
| 400 | !serde_json::to_string(meta) |
| 401 | .unwrap() |
| 402 | .contains("test-router-key") |
| 403 | ); |
| 404 | let id = meta["agent_id"].as_str().unwrap(); |
| 405 | if manager.read().await.agents[id].status == SubAgentStatus::Running { |
| 406 | manager.write().await.cancel_agent(id).unwrap(); |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | #[tokio::test] |
| 411 | async fn saved_profile_cannot_widen_parent_posture_or_depth_and_missing_provider_fails_closed() { |
| 412 | let _env = crate::test_support::lock_test_env(); |
| 413 | let root = tempdir().unwrap(); |
| 414 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 415 | let (client, calls, _, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 416 | let mut profile = codewhale_config::FleetProfile::default(); |
| 417 | profile.role.name = "builder".into(); |
| 418 | profile.model = Some("deepseek-v4-flash".into()); |
| 419 | profile.delegation.max_spawn_depth = Some(0); |
| 420 | profile.permissions.allow_shell = true; |
| 421 | profile.permissions.trust = true; |
| 422 | let mut config = crate::config::Config { |
| 423 | ..Default::default() |
| 424 | } |
| 425 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 426 | let mut fleet = codewhale_config::FleetConfigToml::default(); |
| 427 | fleet |
| 428 | .profiles |
| 429 | .insert("bounded-builder".into(), profile.clone()); |
| 430 | profile.provider = Some("unconfigured-private-route".into()); |
| 431 | fleet.profiles.insert("missing-route".into(), profile); |
| 432 | config.fleet = Some(fleet); |
| 433 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 2); |
| 434 | let context = ToolContext::new(root.path()).with_state_namespace("saved-profile-ceiling"); |
| 435 | let mut runtime = SubAgentRuntime::new( |
| 436 | client, |
| 437 | "deepseek-v4-flash".into(), |
| 438 | context.clone(), |
| 439 | false, |
| 440 | None, |
| 441 | manager.clone(), |
| 442 | ) |
| 443 | .with_api_config(config); |
| 444 | runtime.worker_profile = WorkerRuntimeProfile::for_role(FleetRole::Scout); |
| 445 | runtime.worker_profile.shell = ShellPolicy::None; |
| 446 | let tool = AgentTool::new(manager.clone(), runtime); |
| 447 | assert!( |
| 448 | tool.execute( |
| 449 | json!({"profile":"missing-route", "prompt":"Inspect."}), |
| 450 | &context |
| 451 | ) |
| 452 | .await |
| 453 | .is_err() |
| 454 | ); |
| 455 | assert!(manager.read().await.agents.is_empty()); |
| 456 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 457 | let started = tool |
| 458 | .execute( |
| 459 | json!({"profile":"bounded-builder", "prompt":"Inspect only.", "max_depth":2}), |
| 460 | &context, |
| 461 | ) |
| 462 | .await |
| 463 | .unwrap(); |
| 464 | let id = started.metadata.as_ref().unwrap()["agent_id"] |
| 465 | .as_str() |
| 466 | .unwrap(); |
| 467 | let mut guard = manager.write().await; |
| 468 | let worker = guard.worker_records.get(id).unwrap(); |
| 469 | assert!(!worker.spec.runtime_profile.permissions.write); |
| 470 | assert_eq!(worker.spec.runtime_profile.shell, ShellPolicy::None); |
| 471 | assert_eq!(worker.spec.runtime_profile.max_spawn_depth, 1); |
| 472 | assert_eq!(worker.spec.runtime_profile.spawn_depth, 1); |
| 473 | assert!(!worker.spec.runtime_profile.can_spawn_child()); |
| 474 | guard.cancel_agent(id).unwrap(); |
| 475 | } |
| 476 | |
| 477 | #[tokio::test] |
| 478 | async fn selected_fleet_capability_and_broken_selection_refuse_actual_start() { |
| 479 | use crate::fleet::store::{FleetFile, FleetScope, save_fleet, set_selected}; |
| 480 | let _env = crate::test_support::lock_test_env(); |
| 481 | let root = tempdir().unwrap(); |
| 482 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 483 | let (client, calls, _, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 484 | let config = crate::config::Config { |
| 485 | ..Default::default() |
| 486 | } |
| 487 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 488 | let mut fleet = FleetFile::new("Capability fixture".into(), None).unwrap(); |
| 489 | fleet.members.push(serde_json::from_value(json!({ |
| 490 | "id":"visual-review", "role":"reviewer", "provider":"deepseek", "model":"deepseek-v4-flash", "requires":["vision"] |
| 491 | })).unwrap()); |
| 492 | let path = save_fleet(&fleet, FleetScope::Workspace, root.path()).unwrap(); |
| 493 | set_selected(&fleet.name, FleetScope::Workspace, root.path()).unwrap(); |
| 494 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 1); |
| 495 | let context = |
| 496 | ToolContext::new(root.path()).with_state_namespace("selected-capability-consumer"); |
| 497 | let runtime = SubAgentRuntime::new( |
| 498 | client, |
| 499 | "deepseek-v4-flash".into(), |
| 500 | context.clone(), |
| 501 | false, |
| 502 | None, |
| 503 | manager.clone(), |
| 504 | ) |
| 505 | .with_api_config(config); |
| 506 | let tool = AgentTool::new(manager.clone(), runtime); |
| 507 | let error = tool |
| 508 | .execute( |
| 509 | json!({"profile":"visual-review", "prompt":"Inspect image."}), |
| 510 | &context, |
| 511 | ) |
| 512 | .await |
| 513 | .unwrap_err(); |
| 514 | assert!(error.to_string().contains("requires vision"), "{error}"); |
| 515 | assert!(manager.read().await.agents.is_empty()); |
| 516 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 517 | std::fs::write(path, "this is not a Fleet document").unwrap(); |
| 518 | let error = tool |
| 519 | .execute(json!({"type":"reviewer", "prompt":"Inspect."}), &context) |
| 520 | .await |
| 521 | .unwrap_err(); |
| 522 | assert!(error.to_string().contains("Selected"), "{error}"); |
| 523 | assert!(manager.read().await.agents.is_empty()); |
| 524 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 525 | } |
| 526 | |
| 527 | #[tokio::test] |
| 528 | async fn selected_models_reach_exact_provider_and_off_list_refuses_before_admission() { |
| 529 | use crate::fleet::store::{FleetFile, FleetScope, save_fleet, set_selected}; |
| 530 | let _env = crate::test_support::lock_test_env(); |
| 531 | let root = tempdir().unwrap(); |
| 532 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 533 | let (client, calls, bodies, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 534 | let mut config = crate::config::Config { |
| 535 | ..Default::default() |
| 536 | } |
| 537 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 538 | let router = config |
| 539 | .provider_config_for_mut(&config.test_identity_for_kind(ProviderKind::Openrouter)) |
| 540 | .unwrap(); |
| 541 | router.api_key = Some("test-router-key".into()); |
| 542 | router.base_url = Some(client.base_url().into()); |
| 543 | router.vendor = Some("cerebras".into()); |
| 544 | let mut fleet = FleetFile::new("Selected routes".into(), None).unwrap(); |
| 545 | fleet.members.push(serde_json::from_value(json!({ |
| 546 | "id":"review-choice", "role":"reviewer", "provider":"openrouter", "model":"qwen/qwen3.7-plus" |
| 547 | })).unwrap()); |
| 548 | save_fleet(&fleet, FleetScope::Workspace, root.path()).unwrap(); |
| 549 | set_selected(&fleet.name, FleetScope::Workspace, root.path()).unwrap(); |
| 550 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 2); |
| 551 | let context = ToolContext::new(root.path()).with_state_namespace("shortlist-consumer"); |
| 552 | let runtime = SubAgentRuntime::new( |
| 553 | client, |
| 554 | "deepseek-v4-flash".into(), |
| 555 | context.clone(), |
| 556 | false, |
| 557 | None, |
| 558 | manager.clone(), |
| 559 | ) |
| 560 | .with_api_config(config); |
| 561 | let tool = AgentTool::new(manager.clone(), runtime); |
| 562 | let roster = tool |
| 563 | .execute(json!({"action":"roster"}), &context) |
| 564 | .await |
| 565 | .unwrap(); |
| 566 | let rows: Value = serde_json::from_str(&roster.content).unwrap(); |
| 567 | assert_eq!(rows["model_total_count"], 1); |
| 568 | assert_eq!( |
| 569 | rows["models"][0]["selector"]["model"], |
| 570 | "openrouter/qwen/qwen3.7-plus" |
| 571 | ); |
| 572 | assert_eq!(rows["models"][0]["route"]["provider"], "openrouter"); |
| 573 | assert_eq!(rows["models"][0]["route"]["openrouter_vendor"], "cerebras"); |
| 574 | assert_eq!(rows["models"][0]["route"]["reachability"], "unverified"); |
| 575 | assert_eq!(calls.load(Ordering::SeqCst), 0, "discovery must not infer"); |
| 576 | let error = tool |
| 577 | .execute( |
| 578 | json!({"type":"explore", "model":"deepseek-v4-pro", "prompt":"Inspect."}), |
| 579 | &context, |
| 580 | ) |
| 581 | .await |
| 582 | .unwrap_err(); |
| 583 | assert!( |
| 584 | error.to_string().contains("outside the selected Fleet"), |
| 585 | "{error}" |
| 586 | ); |
| 587 | assert!( |
| 588 | error.to_string().contains("openrouter/qwen/qwen3.7-plus"), |
| 589 | "{error}" |
| 590 | ); |
| 591 | assert!( |
| 592 | error.to_string().contains("deepseek/deepseek-v4-flash"), |
| 593 | "{error}" |
| 594 | ); |
| 595 | assert!(manager.read().await.agents.is_empty()); |
| 596 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 597 | let started = tool |
| 598 | .execute( |
| 599 | json!({"type":"explore", "model":"openrouter/qwen/qwen3.7-plus", "prompt":"Say done."}), |
| 600 | &context, |
| 601 | ) |
| 602 | .await |
| 603 | .unwrap(); |
| 604 | let meta = started.metadata.as_ref().unwrap(); |
| 605 | assert_eq!(meta["child_route"]["provider_id"], "openrouter"); |
| 606 | assert_eq!(meta["child_route"]["model_id"], "qwen/qwen3.7-plus"); |
| 607 | assert!( |
| 608 | meta["child_route"]["resolved_profile_id"].is_null(), |
| 609 | "model choice does not invent a saved profile" |
| 610 | ); |
| 611 | tokio::time::timeout(Duration::from_secs(5), async { |
| 612 | while calls.load(Ordering::SeqCst) == 0 { |
| 613 | tokio::time::sleep(Duration::from_millis(10)).await; |
| 614 | } |
| 615 | }) |
| 616 | .await |
| 617 | .expect("local fixture receives selected model request"); |
| 618 | let body = bodies.lock().unwrap()[0].clone(); |
| 619 | assert_eq!(body["model"], "qwen/qwen3.7-plus"); |
| 620 | assert_eq!(body["provider"]["order"], json!(["cerebras"])); |
| 621 | assert_eq!(body["provider"]["allow_fallbacks"], false); |
| 622 | let id = meta["agent_id"].as_str().unwrap(); |
| 623 | if manager.read().await.agents[id].status == SubAgentStatus::Running { |
| 624 | manager.write().await.cancel_agent(id).unwrap(); |
| 625 | } |
| 626 | let session = tool |
| 627 | .execute( |
| 628 | json!({"type":"explore", "model":"deepseek/deepseek-v4-flash", "prompt":"Say done."}), |
| 629 | &context, |
| 630 | ) |
| 631 | .await |
| 632 | .unwrap(); |
| 633 | assert_eq!( |
| 634 | session.metadata.as_ref().unwrap()["child_route"]["provider_id"], |
| 635 | "deepseek" |
| 636 | ); |
| 637 | let id = session.metadata.as_ref().unwrap()["agent_id"] |
| 638 | .as_str() |
| 639 | .unwrap(); |
| 640 | if manager.read().await.agents[id].status == SubAgentStatus::Running { |
| 641 | manager.write().await.cancel_agent(id).unwrap(); |
| 642 | } |
| 643 | fleet.members.clear(); |
| 644 | save_fleet(&fleet, FleetScope::Workspace, root.path()).unwrap(); |
| 645 | let empty = tool |
| 646 | .execute(json!({"action":"roster"}), &context) |
| 647 | .await |
| 648 | .unwrap(); |
| 649 | let rows: Value = serde_json::from_str(&empty.content).unwrap(); |
| 650 | assert_eq!( |
| 651 | rows["model_total_count"], 0, |
| 652 | "live removal must not preserve stale choices" |
| 653 | ); |
| 654 | let error = tool |
| 655 | .execute( |
| 656 | json!({"type":"explore", "model":"openrouter/qwen/qwen3.7-plus", "prompt":"Inspect."}), |
| 657 | &context, |
| 658 | ) |
| 659 | .await |
| 660 | .unwrap_err(); |
| 661 | assert!(error.to_string().contains("provider DeepSeek"), "{error}"); |
| 662 | } |
| 663 | |
| 664 | #[tokio::test] |
| 665 | async fn shortlisted_model_on_multiple_providers_requires_exact_selector() { |
| 666 | use crate::fleet::store::{FleetFile, FleetScope, save_fleet, set_selected}; |
| 667 | let _env = crate::test_support::lock_test_env(); |
| 668 | let root = tempdir().unwrap(); |
| 669 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 670 | let (client, calls, _, fixture_config) = delayed_chat_client(Duration::ZERO, "done").await; |
| 671 | let mut fleet = FleetFile::new("Ambiguous routes".into(), None).unwrap(); |
| 672 | for (id, provider) in [("review-a", "openrouter"), ("review-b", "openai")] { |
| 673 | fleet.members.push( |
| 674 | serde_json::from_value(json!({ |
| 675 | "id":id, "role":"reviewer", "provider":provider, "model":"shared-wire-model" |
| 676 | })) |
| 677 | .unwrap(), |
| 678 | ); |
| 679 | } |
| 680 | save_fleet(&fleet, FleetScope::Workspace, root.path()).unwrap(); |
| 681 | set_selected(&fleet.name, FleetScope::Workspace, root.path()).unwrap(); |
| 682 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 1); |
| 683 | let context = ToolContext::new(root.path()).with_state_namespace("ambiguous-model-consumer"); |
| 684 | let runtime = SubAgentRuntime::new( |
| 685 | client, |
| 686 | "deepseek-v4-flash".into(), |
| 687 | context.clone(), |
| 688 | false, |
| 689 | None, |
| 690 | manager.clone(), |
| 691 | ) |
| 692 | .with_api_config(fixture_config); |
| 693 | let tool = AgentTool::new(manager.clone(), runtime); |
| 694 | let error = tool |
| 695 | .execute( |
| 696 | json!({"type":"explore", "model":"shared-wire-model", "prompt":"Inspect."}), |
| 697 | &context, |
| 698 | ) |
| 699 | .await |
| 700 | .unwrap_err(); |
| 701 | assert!(error.to_string().contains("multiple providers"), "{error}"); |
| 702 | assert!(manager.read().await.agents.is_empty()); |
| 703 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 704 | } |
| 705 | |
| 706 | fn write_restart_route_config(path: &std::path::Path, base_url: &str, role_pins: &str) { |
| 707 | std::fs::write( |
| 708 | path, |
| 709 | format!( |
| 710 | r#" |
| 711 | provider = "deepseek" |
| 712 | model = "deepseek-v4-flash" |
| 713 | api_key = "fixture-key" |
| 714 | base_url = "{base_url}" |
| 715 | |
| 716 | [providers.ReviewerRoute] |
| 717 | kind = "openai-compatible" |
| 718 | api_key = "fixture-review-key" |
| 719 | base_url = "{base_url}" |
| 720 | model = "fixture-review-model" |
| 721 | |
| 722 | [providers.OtherRoute] |
| 723 | kind = "openai-compatible" |
| 724 | api_key = "fixture-other-key" |
| 725 | base_url = "{base_url}" |
| 726 | model = "fixture-review-model" |
| 727 | |
| 728 | {role_pins} |
| 729 | "# |
| 730 | ), |
| 731 | ) |
| 732 | .unwrap(); |
| 733 | } |
| 734 | |
| 735 | async fn assert_admitted_route( |
| 736 | manager: &SharedSubAgentManager, |
| 737 | started: &crate::tools::spec::ToolResult, |
| 738 | expected: Value, |
| 739 | ) -> String { |
| 740 | let content: Value = serde_json::from_str(&started.content).unwrap(); |
| 741 | let metadata = started.metadata.as_ref().unwrap(); |
| 742 | let receipt = &metadata["child_route"]; |
| 743 | assert_eq!(&content["child_route"], receipt); |
| 744 | for (field, value) in expected.as_object().unwrap() { |
| 745 | assert_eq!(&receipt[field], value, "admitted {field}: {receipt}"); |
| 746 | } |
| 747 | let id = metadata["agent_id"].as_str().unwrap().to_string(); |
| 748 | let manager = manager.read().await; |
| 749 | let spec = &manager.worker_records[&id].spec; |
| 750 | assert_eq!(serde_json::to_value(&spec.child_route).unwrap(), *receipt); |
| 751 | assert_eq!(spec.model, receipt["model_id"].as_str().unwrap()); |
| 752 | assert_eq!(spec.agent_type.as_str(), receipt["canonical_role"]); |
| 753 | assert_eq!(manager.agents[&id].model, spec.model); |
| 754 | let manifest = spec |
| 755 | .launch_manifest |
| 756 | .as_ref() |
| 757 | .expect("persisted launch authority"); |
| 758 | assert_eq!(manifest.child_id, id); |
| 759 | assert_eq!(manifest.profile, spec.runtime_profile); |
| 760 | assert_eq!(manifest.profile.role, spec.agent_type); |
| 761 | assert_eq!( |
| 762 | manifest.profile.model, |
| 763 | crate::worker_profile::ModelRoute::Fixed(spec.model.clone()) |
| 764 | ); |
| 765 | assert_eq!( |
| 766 | manifest.profile.provider.as_deref(), |
| 767 | receipt["provider_id"].as_str() |
| 768 | ); |
| 769 | assert_eq!( |
| 770 | manifest.profile.reasoning_effort.as_deref(), |
| 771 | receipt["effective_reasoning"].as_str() |
| 772 | ); |
| 773 | id |
| 774 | } |
| 775 | |
| 776 | async fn wait_for_queued_child(mailbox: &mut MailboxReceiver, id: &str) { |
| 777 | tokio::time::timeout(Duration::from_secs(2), async { |
| 778 | loop { |
| 779 | let envelope = mailbox.recv().await.expect("child progress channel"); |
| 780 | if matches!(envelope.message, MailboxMessage::Progress { agent_id, status } |
| 781 | if agent_id == id && status.contains("queued")) |
| 782 | { |
| 783 | break; |
| 784 | } |
| 785 | } |
| 786 | }) |
| 787 | .await |
| 788 | .expect("actual child reaches the held launch gate"); |
| 789 | } |
| 790 | |
| 791 | #[tokio::test] |
| 792 | async fn fleet_editor_save_reload_reaches_type_only_admission_without_a_model_request() { |
| 793 | use crate::fleet::store::{ |
| 794 | FleetFile, FleetScope, load_fleet_in_scope, save_fleet, set_selected, |
| 795 | }; |
| 796 | use crate::tui::views::fleet_detail::FleetDetailView; |
| 797 | use crate::tui::views::{ModalView, ViewAction, ViewEvent}; |
| 798 | use crossterm::event::{KeyCode, KeyEvent, KeyModifiers}; |
| 799 | |
| 800 | let _env = crate::test_support::lock_test_env(); |
| 801 | let root = tempdir().unwrap(); |
| 802 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 803 | let _project = ProjectProfilesGuard::enabled(); |
| 804 | let (fixture_client, calls, bodies, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 805 | let _provider = crate::test_support::EnvVarGuard::set("CODEWHALE_PROVIDER", "deepseek"); |
| 806 | let _endpoint = |
| 807 | crate::test_support::EnvVarGuard::set("CODEWHALE_BASE_URL", fixture_client.base_url()); |
| 808 | let _model = crate::test_support::EnvVarGuard::set("CODEWHALE_MODEL", "deepseek-v4-flash"); |
| 809 | let config_path = root.path().join("config.toml"); |
| 810 | write_restart_route_config(&config_path, fixture_client.base_url(), ""); |
| 811 | let config = crate::config::Config::load(Some(config_path.clone()), None).unwrap(); |
| 812 | let mut fleet = FleetFile::new("Editor restart acceptance".into(), None).unwrap(); |
| 813 | fleet.members.push( |
| 814 | serde_json::from_value(json!({ |
| 815 | "id":"review-pin", "role":"reviewer", "instructions":"SAVED_REVIEW_INSTRUCTION" |
| 816 | })) |
| 817 | .unwrap(), |
| 818 | ); |
| 819 | save_fleet(&fleet, FleetScope::Workspace, root.path()).unwrap(); |
| 820 | set_selected(&fleet.name, FleetScope::Workspace, root.path()).unwrap(); |
| 821 | |
| 822 | let mut app = |
| 823 | crate::tui::app::App::new(crate::test_support::test_tui_options(root.path()), &config); |
| 824 | app.workspace = root.path().to_path_buf(); |
| 825 | let mut view = FleetDetailView::open_for_member( |
| 826 | &app, |
| 827 | &config, |
| 828 | &fleet.name, |
| 829 | FleetScope::Workspace, |
| 830 | Some("review-pin"), |
| 831 | ) |
| 832 | .unwrap(); |
| 833 | let key = |code| KeyEvent::new(code, KeyModifiers::NONE); |
| 834 | view.handle_key(key(KeyCode::Char('e'))); |
| 835 | for ch in "ReviewerRoute".chars() { |
| 836 | view.handle_key(key(KeyCode::Char(ch))); |
| 837 | } |
| 838 | view.handle_key(key(KeyCode::Enter)); |
| 839 | // Off survives generic route normalization and distinguishes the saved |
| 840 | // choice from default reasoning without inventing fixture capabilities. |
| 841 | view.handle_key(key(KeyCode::Char('t'))); |
| 842 | assert!(matches!( |
| 843 | view.handle_key(key(KeyCode::Char('s'))), |
| 844 | ViewAction::EmitAndClose(ViewEvent::FleetStoreChanged { .. }) |
| 845 | )); |
| 846 | let (saved, _) = load_fleet_in_scope(&fleet.name, FleetScope::Workspace, root.path()).unwrap(); |
| 847 | let pin = &saved.members[0]; |
| 848 | assert!(!pin.shortlist); |
| 849 | assert_eq!(pin.role, "reviewer"); |
| 850 | assert_eq!(pin.provider.as_deref(), Some("ReviewerRoute")); |
| 851 | assert_eq!(pin.model.as_deref(), Some("fixture-review-model")); |
| 852 | assert_eq!(pin.reasoning.as_deref(), Some("off")); |
| 853 | drop((view, app, config, fixture_client)); |
| 854 | |
| 855 | // Restart from the ordinary file loader; no old UI state or runtime roster survives. |
| 856 | let reloaded = crate::config::Config::load(Some(config_path), None).unwrap(); |
| 857 | let client = CodewhaleClient::new(&reloaded).unwrap(); |
| 858 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 1); |
| 859 | let gate = manager.read().await.launch_gate.clone(); |
| 860 | let held_permit = gate.acquire().await; |
| 861 | let (mailbox, mut mailbox_rx) = Mailbox::new(CancellationToken::new()); |
| 862 | let context = ToolContext::new(root.path()).with_state_namespace("fleet-editor-restarted"); |
| 863 | let mut runtime = SubAgentRuntime::new( |
| 864 | client, |
| 865 | "deepseek-v4-flash".into(), |
| 866 | context.clone(), |
| 867 | false, |
| 868 | None, |
| 869 | manager.clone(), |
| 870 | ) |
| 871 | .with_api_config(reloaded); |
| 872 | runtime.mailbox = Some(mailbox); |
| 873 | let loaded_roster = spawn_roster(&runtime); |
| 874 | let loaded_pin = loaded_roster |
| 875 | .members() |
| 876 | .iter() |
| 877 | .find(|member| member.id == "review-pin") |
| 878 | .unwrap(); |
| 879 | assert_eq!(loaded_pin.profile.reasoning_effort.as_deref(), Some("off")); |
| 880 | let tool = AgentTool::new(manager.clone(), runtime); |
| 881 | for extra in [ |
| 882 | json!({"model":"OtherRoute/fixture-review-model"}), |
| 883 | json!({"model_strength":"faster"}), |
| 884 | ] { |
| 885 | let mut request = json!({"type":"reviewer", "prompt":"Review without execution."}); |
| 886 | request |
| 887 | .as_object_mut() |
| 888 | .unwrap() |
| 889 | .extend(extra.as_object().unwrap().clone()); |
| 890 | assert!(tool.execute(request, &context).await.is_err()); |
| 891 | assert!(manager.read().await.agents.is_empty()); |
| 892 | assert!(manager.read().await.worker_records.is_empty()); |
| 893 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 894 | } |
| 895 | let started = tool |
| 896 | .execute( |
| 897 | json!({"type":"reviewer", "prompt":"Review without execution."}), |
| 898 | &context, |
| 899 | ) |
| 900 | .await |
| 901 | .unwrap(); |
| 902 | let id = assert_admitted_route( |
| 903 | &manager, |
| 904 | &started, |
| 905 | json!({ |
| 906 | "requested_type":"reviewer", "requested_profile":null, |
| 907 | "resolved_profile_id":"review-pin", "profile_origin":"project", |
| 908 | "canonical_role":"reviewer", "provider_id":"ReviewerRoute", |
| 909 | "model_id":"fixture-review-model", "route_source":"agent_profile.model", |
| 910 | "requested_reasoning":"inherit", "effective_reasoning":"off" |
| 911 | }), |
| 912 | ) |
| 913 | .await; |
| 914 | assert!( |
| 915 | manager.read().await.worker_records[&id] |
| 916 | .spec |
| 917 | .launch_manifest |
| 918 | .as_ref() |
| 919 | .unwrap() |
| 920 | .prompt |
| 921 | .contains("SAVED_REVIEW_INSTRUCTION") |
| 922 | ); |
| 923 | wait_for_queued_child(&mut mailbox_rx, &id).await; |
| 924 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 925 | assert!(bodies.lock().unwrap().is_empty()); |
| 926 | let requested = manager.write().await.cancel_agent(&id).unwrap(); |
| 927 | let settled = settle_requested_child(&manager, requested).await; |
| 928 | assert_eq!(settled.status, SubAgentStatus::Cancelled); |
| 929 | assert_eq!( |
| 930 | manager.read().await.agents[&id].status, |
| 931 | SubAgentStatus::Cancelled |
| 932 | ); |
| 933 | drop(held_permit); |
| 934 | } |
| 935 | |
| 936 | #[tokio::test] |
| 937 | async fn reloaded_manual_role_pin_and_explicit_profile_keep_distinct_shortlist_receipts() { |
| 938 | use crate::fleet::store::{FleetFile, FleetScope, save_fleet, set_selected}; |
| 939 | let _env = crate::test_support::lock_test_env(); |
| 940 | let root = tempdir().unwrap(); |
| 941 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 942 | let _project = ProjectProfilesGuard::enabled(); |
| 943 | let (fixture_client, calls, bodies, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 944 | let _provider = crate::test_support::EnvVarGuard::set("CODEWHALE_PROVIDER", "deepseek"); |
| 945 | let _endpoint = |
| 946 | crate::test_support::EnvVarGuard::set("CODEWHALE_BASE_URL", fixture_client.base_url()); |
| 947 | let _model = crate::test_support::EnvVarGuard::set("CODEWHALE_MODEL", "deepseek-v4-flash"); |
| 948 | let config_path = root.path().join("config.toml"); |
| 949 | write_restart_route_config( |
| 950 | &config_path, |
| 951 | fixture_client.base_url(), |
| 952 | r#" |
| 953 | [subagents.models] |
| 954 | reviewer = "deepseek-v4-flash" |
| 955 | default = "deepseek-v4-pro" |
| 956 | [subagents.roles.reviewer] |
| 957 | model = "deepseek-v4-pro" |
| 958 | "#, |
| 959 | ); |
| 960 | let mut fleet = FleetFile::new("Manual pin precedence".into(), None).unwrap(); |
| 961 | for member in [ |
| 962 | json!({"id":"review-choice", "shortlist":true, "provider":"ReviewerRoute", "model":"fixture-review-model"}), |
| 963 | // Off stays distinct from the default on this generic custom route. |
| 964 | json!({"id":"review-pin", "role":"reviewer", "provider":"ReviewerRoute", "model":"fixture-review-model", "reasoning":"off"}), |
| 965 | ] { |
| 966 | fleet.members.push(serde_json::from_value(member).unwrap()); |
| 967 | } |
| 968 | save_fleet(&fleet, FleetScope::Workspace, root.path()).unwrap(); |
| 969 | set_selected(&fleet.name, FleetScope::Workspace, root.path()).unwrap(); |
| 970 | let config = crate::config::Config::load(Some(config_path), None).unwrap(); |
| 971 | let overrides = config.subagent_model_overrides(); |
| 972 | assert_eq!(overrides["reviewer"].model, "deepseek-v4-pro"); |
| 973 | assert_eq!(overrides["reviewer"].provider, None); |
| 974 | let client = CodewhaleClient::new(&config).unwrap(); |
| 975 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 1); |
| 976 | let gate = manager.read().await.launch_gate.clone(); |
| 977 | let held_permit = gate.acquire().await; |
| 978 | let (mailbox, mut mailbox_rx) = Mailbox::new(CancellationToken::new()); |
| 979 | let context = ToolContext::new(root.path()).with_state_namespace("manual-role-restarted"); |
| 980 | let mut runtime = SubAgentRuntime::new( |
| 981 | client, |
| 982 | "deepseek-v4-flash".into(), |
| 983 | context.clone(), |
| 984 | false, |
| 985 | None, |
| 986 | manager.clone(), |
| 987 | ) |
| 988 | .with_api_config(config); |
| 989 | runtime.mailbox = Some(mailbox); |
| 990 | let loaded_roster = spawn_roster(&runtime); |
| 991 | let loaded_pin = loaded_roster |
| 992 | .members() |
| 993 | .iter() |
| 994 | .find(|member| member.id == "review-pin") |
| 995 | .unwrap(); |
| 996 | assert_eq!(loaded_pin.profile.reasoning_effort.as_deref(), Some("off")); |
| 997 | let tool = AgentTool::new(manager.clone(), runtime); |
| 998 | let result = tool |
| 999 | .execute(json!({"action":"roster"}), &context) |
| 1000 | .await |
| 1001 | .unwrap(); |
| 1002 | let roster: Value = serde_json::from_str(&result.content).unwrap(); |
| 1003 | let model = roster["models"] |
| 1004 | .as_array() |
| 1005 | .unwrap() |
| 1006 | .iter() |
| 1007 | .find(|row| row["selector"]["model"] == "ReviewerRoute/fixture-review-model") |
| 1008 | .expect("shortlisted model row"); |
| 1009 | assert_eq!(model["route"]["provider"], "ReviewerRoute"); |
| 1010 | assert_eq!(model["route"]["model"], "fixture-review-model"); |
| 1011 | assert_ne!(model["route"]["source"], "role.pin"); |
| 1012 | for role in ["general", "reviewer"] { |
| 1013 | let row = roster["members"] |
| 1014 | .as_array() |
| 1015 | .unwrap() |
| 1016 | .iter() |
| 1017 | .find(|row| row["role"] == role) |
| 1018 | .unwrap(); |
| 1019 | assert_eq!(row["route"]["model"], "deepseek-v4-pro"); |
| 1020 | assert_eq!(row["route"]["source"], "role.pin"); |
| 1021 | } |
| 1022 | for extra in [ |
| 1023 | json!({"model":"ReviewerRoute/fixture-review-model"}), |
| 1024 | json!({"model_strength":"faster"}), |
| 1025 | ] { |
| 1026 | let mut request = json!({"type":"reviewer", "prompt":"Review."}); |
| 1027 | request |
| 1028 | .as_object_mut() |
| 1029 | .unwrap() |
| 1030 | .extend(extra.as_object().unwrap().clone()); |
| 1031 | assert!(tool.execute(request, &context).await.is_err()); |
| 1032 | assert!(manager.read().await.agents.is_empty()); |
| 1033 | assert!(manager.read().await.worker_records.is_empty()); |
| 1034 | } |
| 1035 | for (request, expected) in [ |
| 1036 | ( |
| 1037 | json!({"type":"reviewer", "prompt":"Review."}), |
| 1038 | json!({ |
| 1039 | "requested_profile":null, "resolved_profile_id":null, |
| 1040 | "provider_id":"deepseek", "model_id":"deepseek-v4-pro", "route_source":"role.pin" |
| 1041 | }), |
| 1042 | ), |
| 1043 | ( |
| 1044 | json!({"type":"reviewer", "model":"deepseek/deepseek-v4-pro", "prompt":"Review."}), |
| 1045 | json!({ |
| 1046 | "requested_profile":null, "resolved_profile_id":null, |
| 1047 | "provider_id":"deepseek", "model_id":"deepseek-v4-pro", "route_source":"role.pin" |
| 1048 | }), |
| 1049 | ), |
| 1050 | ( |
| 1051 | json!({"profile":"member:review-pin", "prompt":"Review."}), |
| 1052 | json!({ |
| 1053 | "requested_profile":"member:review-pin", "resolved_profile_id":"review-pin", |
| 1054 | "provider_id":"ReviewerRoute", "model_id":"fixture-review-model", |
| 1055 | "canonical_role":"reviewer", "requested_reasoning":"inherit", |
| 1056 | "effective_reasoning":"off", "route_source":"agent_profile.model" |
| 1057 | }), |
| 1058 | ), |
| 1059 | ( |
| 1060 | json!({"profile":"member:review-pin", "thinking":"low", "prompt":"Review."}), |
| 1061 | json!({ |
| 1062 | "requested_profile":"member:review-pin", "resolved_profile_id":"review-pin", |
| 1063 | "provider_id":"ReviewerRoute", "model_id":"fixture-review-model", |
| 1064 | "canonical_role":"reviewer", "requested_reasoning":"low", |
| 1065 | "effective_reasoning":"high", "route_source":"agent_profile.model" |
| 1066 | }), |
| 1067 | ), |
| 1068 | ] { |
| 1069 | let started = tool.execute(request, &context).await.unwrap(); |
| 1070 | let id = assert_admitted_route(&manager, &started, expected).await; |
| 1071 | wait_for_queued_child(&mut mailbox_rx, &id).await; |
| 1072 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 1073 | assert!(bodies.lock().unwrap().is_empty()); |
| 1074 | let requested = manager.write().await.cancel_agent(&id).unwrap(); |
| 1075 | let settled = settle_requested_child(&manager, requested).await; |
| 1076 | assert_eq!(settled.status, SubAgentStatus::Cancelled); |
| 1077 | assert_eq!( |
| 1078 | manager.read().await.agents[&id].status, |
| 1079 | SubAgentStatus::Cancelled |
| 1080 | ); |
| 1081 | } |
| 1082 | drop(held_permit); |
| 1083 | } |
| 1084 | |
| 1085 | #[tokio::test] |
| 1086 | async fn loaded_structured_role_routes_bind_exact_providers_and_legacy_namespaces_stay_opaque() { |
| 1087 | let _env = crate::test_support::lock_test_env(); |
| 1088 | for ( |
| 1089 | name, |
| 1090 | parent_provider, |
| 1091 | parent_model, |
| 1092 | declaration, |
| 1093 | pin_provider, |
| 1094 | pin_model, |
| 1095 | admitted_provider, |
| 1096 | ) in [ |
| 1097 | ( |
| 1098 | "structured-route", |
| 1099 | "deepseek", |
| 1100 | "deepseek-v4-flash", |
| 1101 | "[subagents.roles.reviewer]\nmodel = \"ReviewerRoute/fixture-review-model\"", |
| 1102 | Some("ReviewerRoute"), |
| 1103 | "fixture-review-model", |
| 1104 | Some("ReviewerRoute"), |
| 1105 | ), |
| 1106 | ( |
| 1107 | "unknown-route", |
| 1108 | "deepseek", |
| 1109 | "deepseek-v4-flash", |
| 1110 | "[subagents.roles.reviewer]\nmodel = \"MissingRoute/fixture-review-model\"", |
| 1111 | Some("MissingRoute"), |
| 1112 | "fixture-review-model", |
| 1113 | None, |
| 1114 | ), |
| 1115 | ( |
| 1116 | "legacy-namespace", |
| 1117 | "openrouter", |
| 1118 | "deepseek/deepseek-v4-flash", |
| 1119 | "[subagents.models]\nreviewer = \"deepseek/deepseek-v4-pro\"", |
| 1120 | None, |
| 1121 | "deepseek/deepseek-v4-pro", |
| 1122 | Some("openrouter"), |
| 1123 | ), |
| 1124 | ] { |
| 1125 | let root = tempdir().unwrap(); |
| 1126 | let _home = |
| 1127 | crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", root.path().join("state")); |
| 1128 | let (fixture_client, calls, bodies, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 1129 | let _provider = |
| 1130 | crate::test_support::EnvVarGuard::set("CODEWHALE_PROVIDER", parent_provider); |
| 1131 | let _endpoint = |
| 1132 | crate::test_support::EnvVarGuard::set("CODEWHALE_BASE_URL", fixture_client.base_url()); |
| 1133 | let _model = crate::test_support::EnvVarGuard::set("CODEWHALE_MODEL", parent_model); |
| 1134 | let config_path = root.path().join("config.toml"); |
| 1135 | // `deepseek` is a real configured provider as well as the legacy wire |
| 1136 | // namespace. The old map must not reinterpret that namespace as a pin. |
| 1137 | let declarations = format!( |
| 1138 | r#" |
| 1139 | [providers.deepseek] |
| 1140 | api_key = "fixture-deepseek-key" |
| 1141 | base_url = "{base_url}" |
| 1142 | model = "deepseek-v4-flash" |
| 1143 | [providers.openrouter] |
| 1144 | api_key = "fixture-router-key" |
| 1145 | base_url = "{base_url}" |
| 1146 | model = "deepseek/deepseek-v4-flash" |
| 1147 | {declaration} |
| 1148 | "#, |
| 1149 | base_url = fixture_client.base_url() |
| 1150 | ); |
| 1151 | write_restart_route_config(&config_path, fixture_client.base_url(), &declarations); |
| 1152 | let config = crate::config::Config::load(Some(config_path), None).unwrap(); |
| 1153 | let overrides = config.subagent_model_overrides(); |
| 1154 | assert_eq!( |
| 1155 | overrides["reviewer"].provider.as_deref(), |
| 1156 | pin_provider, |
| 1157 | "{name}" |
| 1158 | ); |
| 1159 | assert_eq!(overrides["reviewer"].model, pin_model, "{name}"); |
| 1160 | assert_eq!( |
| 1161 | config.active_provider_identity().unwrap().key.as_str(), |
| 1162 | parent_provider |
| 1163 | ); |
| 1164 | let client = CodewhaleClient::new(&config).unwrap(); |
| 1165 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 1); |
| 1166 | let gate = manager.read().await.launch_gate.clone(); |
| 1167 | let held_permit = gate.acquire().await; |
| 1168 | let (mailbox, mut mailbox_rx) = Mailbox::new(CancellationToken::new()); |
| 1169 | let context = ToolContext::new(root.path()).with_state_namespace(name); |
| 1170 | let mut runtime = SubAgentRuntime::new( |
| 1171 | client, |
| 1172 | parent_model.into(), |
| 1173 | context.clone(), |
| 1174 | false, |
| 1175 | None, |
| 1176 | manager.clone(), |
| 1177 | ) |
| 1178 | .with_api_config(config); |
| 1179 | runtime.mailbox = Some(mailbox); |
| 1180 | let tool = AgentTool::new(manager.clone(), runtime); |
| 1181 | let roster_result = tool |
| 1182 | .execute(json!({"action":"roster"}), &context) |
| 1183 | .await |
| 1184 | .unwrap(); |
| 1185 | let roster: Value = serde_json::from_str(&roster_result.content).unwrap(); |
| 1186 | let row = roster["members"] |
| 1187 | .as_array() |
| 1188 | .unwrap() |
| 1189 | .iter() |
| 1190 | .find(|row| row["role"] == "reviewer") |
| 1191 | .unwrap(); |
| 1192 | let request = json!({"type":"reviewer", "prompt":"Review before model execution."}); |
| 1193 | let Some(expected_provider) = admitted_provider else { |
| 1194 | assert!(row["route"].is_null()); |
| 1195 | assert!( |
| 1196 | row["route_error"] |
| 1197 | .as_str() |
| 1198 | .unwrap() |
| 1199 | .contains("MissingRoute") |
| 1200 | ); |
| 1201 | let error = tool.execute(request, &context).await.unwrap_err(); |
| 1202 | assert!(error.to_string().contains("MissingRoute"), "{error}"); |
| 1203 | assert!(manager.read().await.agents.is_empty()); |
| 1204 | assert!(manager.read().await.worker_records.is_empty()); |
| 1205 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 1206 | assert!(bodies.lock().unwrap().is_empty()); |
| 1207 | drop(held_permit); |
| 1208 | continue; |
| 1209 | }; |
| 1210 | assert_eq!(row["route"]["provider"], expected_provider, "{name}"); |
| 1211 | assert_eq!(row["route"]["model"], pin_model, "{name}"); |
| 1212 | assert_eq!(row["route"]["source"], "role.pin", "{name}"); |
| 1213 | if pin_provider.is_some() { |
| 1214 | for extra in [ |
| 1215 | json!({"model":"OtherRoute/fixture-review-model"}), |
| 1216 | json!({"model_strength":"faster"}), |
| 1217 | ] { |
| 1218 | let mut conflicting = request.clone(); |
| 1219 | conflicting |
| 1220 | .as_object_mut() |
| 1221 | .unwrap() |
| 1222 | .extend(extra.as_object().unwrap().clone()); |
| 1223 | let error = tool.execute(conflicting, &context).await.unwrap_err(); |
| 1224 | assert!(error.to_string().contains("pins"), "{error}"); |
| 1225 | assert!(manager.read().await.agents.is_empty()); |
| 1226 | assert!(manager.read().await.worker_records.is_empty()); |
| 1227 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 1228 | } |
| 1229 | } |
| 1230 | let mut matching = request.clone(); |
| 1231 | matching["model"] = json!(format!("{expected_provider}/{pin_model}")); |
| 1232 | for request in [request, matching] { |
| 1233 | let started = tool.execute(request, &context).await.unwrap(); |
| 1234 | let id = assert_admitted_route( |
| 1235 | &manager, |
| 1236 | &started, |
| 1237 | json!({ |
| 1238 | "requested_type":"reviewer", "canonical_role":"reviewer", |
| 1239 | "requested_profile":null, "resolved_profile_id":null, "profile_origin":null, |
| 1240 | "provider_id":expected_provider, "model_id":pin_model, "route_source":"role.pin" |
| 1241 | }), |
| 1242 | ) |
| 1243 | .await; |
| 1244 | wait_for_queued_child(&mut mailbox_rx, &id).await; |
| 1245 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 1246 | assert!(bodies.lock().unwrap().is_empty()); |
| 1247 | let requested = manager.write().await.cancel_agent(&id).unwrap(); |
| 1248 | let settled = settle_requested_child(&manager, requested).await; |
| 1249 | assert_eq!(settled.status, SubAgentStatus::Cancelled); |
| 1250 | assert_eq!( |
| 1251 | manager.read().await.agents[&id].status, |
| 1252 | SubAgentStatus::Cancelled |
| 1253 | ); |
| 1254 | } |
| 1255 | drop(held_permit); |
| 1256 | } |
| 1257 | } |
| 1258 | |
| 1259 | #[tokio::test] |
| 1260 | async fn issue_6117_invalid_personal_profile_is_visible_and_never_admitted_as_builtin() { |
| 1261 | let _env = crate::test_support::lock_test_env(); |
| 1262 | let root = tempdir().unwrap(); |
| 1263 | let home = root.path().join("state"); |
| 1264 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", &home); |
| 1265 | std::fs::create_dir_all(home.join("agents")).unwrap(); |
| 1266 | let profile = home.join("agents/scout.toml"); |
| 1267 | std::fs::write(&profile, "provider = \"openrouter\"\nmodel = \"qwen/qwen3.7-plus\"\nallow_shell = false\ntrust = false\n").unwrap(); |
| 1268 | let (client, calls, _, _) = delayed_chat_client(Duration::ZERO, "done").await; |
| 1269 | let config = crate::config::Config { |
| 1270 | ..Default::default() |
| 1271 | } |
| 1272 | .with_legacy_root(Some("test-key".into()), Some(client.base_url().into())); |
| 1273 | let manager = new_shared_subagent_manager(root.path().to_path_buf(), 2); |
| 1274 | let context = ToolContext::new(root.path()).with_state_namespace("issue-6117"); |
| 1275 | let runtime = SubAgentRuntime::new( |
| 1276 | client, |
| 1277 | "deepseek-v4-flash".into(), |
| 1278 | context.clone(), |
| 1279 | false, |
| 1280 | None, |
| 1281 | manager.clone(), |
| 1282 | ) |
| 1283 | .with_api_config(config); |
| 1284 | let tool = AgentTool::new(manager.clone(), runtime); |
| 1285 | let discovered = tool |
| 1286 | .execute(json!({"action":"roster"}), &context) |
| 1287 | .await |
| 1288 | .unwrap(); |
| 1289 | let roster: Value = serde_json::from_str(&discovered.content).unwrap(); |
| 1290 | assert_eq!(roster["profile_load_issue_count"], 1); |
| 1291 | assert_eq!(roster["profile_load_issues"][0]["id"], "scout"); |
| 1292 | for selector in ["scout", "explore", "member:SCOUT"] { |
| 1293 | let error = tool |
| 1294 | .execute( |
| 1295 | json!({"action":"start", "profile":selector, "prompt":"Inspect."}), |
| 1296 | &context, |
| 1297 | ) |
| 1298 | .await |
| 1299 | .unwrap_err(); |
| 1300 | assert!(error.to_string().contains("scout.toml"), "{error}"); |
| 1301 | } |
| 1302 | assert_eq!(calls.load(Ordering::SeqCst), 0); |
| 1303 | assert!(manager.read().await.agents.is_empty()); |
| 1304 | std::fs::write(&profile, "base_role = \"explore\"\nprovider = \"deepseek\"\nmodel = \"deepseek-v4-pro\"\nreasoning_effort = \"low\"\n[permissions]\nallow_shell = false\ntrust = false\n").unwrap(); |
| 1305 | let started = tool |
| 1306 | .execute( |
| 1307 | json!({"action":"start", "profile":"scout", "prompt":"Say done."}), |
| 1308 | &context, |
| 1309 | ) |
| 1310 | .await |
| 1311 | .unwrap(); |
| 1312 | let meta = started.metadata.as_ref().unwrap(); |
| 1313 | let receipt = &meta["child_route"]; |
| 1314 | assert_eq!(receipt["resolved_profile_id"], "scout"); |
| 1315 | assert_eq!(receipt["profile_origin"], "personal"); |
| 1316 | assert_eq!(receipt["provider_id"], "deepseek"); |
| 1317 | assert_eq!(receipt["model_id"], "deepseek-v4-pro"); |
| 1318 | assert_eq!(receipt["effective_reasoning"], "low"); |
| 1319 | assert_eq!(receipt["route_source"], "agent_profile.model"); |
| 1320 | let id = meta["agent_id"].as_str().unwrap(); |
| 1321 | let mut guard = manager.write().await; |
| 1322 | assert!( |
| 1323 | !guard.worker_records[id] |
| 1324 | .spec |
| 1325 | .runtime_profile |
| 1326 | .permissions |
| 1327 | .write |
| 1328 | ); |
| 1329 | if guard.agents[id].status == SubAgentStatus::Running { |
| 1330 | guard.cancel_agent(id).unwrap(); |
| 1331 | } |
| 1332 | } |
| 1333 |