| 1 | #[tokio::test] |
| 2 | async fn max_steps_exhaustion_fails_as_budget_never_completed() { |
| 3 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 4 | |
| 5 | let workspace = tempdir().expect("tempdir"); |
| 6 | fs::write(workspace.path().join("state.txt"), "still-working\n").expect("write fixture"); |
| 7 | // The model keeps tool-calling past the 1-step budget; it never gets to |
| 8 | // produce a final answer. |
| 9 | let turns = vec![ |
| 10 | canned::tool_call_turn( |
| 11 | "call-step-1", |
| 12 | "File", |
| 13 | r#"{"action":"read","path":"state.txt"}"#, |
| 14 | ), |
| 15 | canned::tool_call_turn( |
| 16 | "call-step-2", |
| 17 | "File", |
| 18 | r#"{"action":"read","path":"state.txt"}"#, |
| 19 | ), |
| 20 | ]; |
| 21 | let mock = std::sync::Arc::new(MockLlmClient::new(turns)); |
| 22 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 23 | let engine_config = EngineConfig { |
| 24 | max_steps: 1, |
| 25 | ..deterministic_engine_config(workspace.path()) |
| 26 | }; |
| 27 | let (engine, handle) = Engine::new_with_model_client(engine_config, &Config::default(), client); |
| 28 | let task = tokio::spawn(engine.run()); |
| 29 | handle |
| 30 | .send(external_user_message_op( |
| 31 | "Keep reading until done.", |
| 32 | AppMode::Agent, |
| 33 | &Config::default(), |
| 34 | )) |
| 35 | .await |
| 36 | .expect("send step-budget trajectory"); |
| 37 | |
| 38 | let mut rx = handle.rx_event.write().await; |
| 39 | let (status, error) = loop { |
| 40 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 41 | .await |
| 42 | .expect("timed out waiting for step-budget trajectory") |
| 43 | .expect("engine event"); |
| 44 | if let Event::TurnComplete { status, error, .. } = event { |
| 45 | break (status, error); |
| 46 | } |
| 47 | }; |
| 48 | drop(rx); |
| 49 | |
| 50 | assert_eq!( |
| 51 | status, |
| 52 | TurnOutcomeStatus::Failed, |
| 53 | "step-budget exhaustion must never report Completed" |
| 54 | ); |
| 55 | let error = error.expect("step-budget exhaustion must carry a terminal error"); |
| 56 | assert!(error.contains("Maximum model steps"), "{error}"); |
| 57 | |
| 58 | // The terminal error reduces to BudgetExhausted for machine consumers — |
| 59 | // this is the same reduction the headless exec receipt applies, and it is |
| 60 | // what gates persistent-service release on Completed-only turns. |
| 61 | let category = crate::error_taxonomy::classify_error_message(&error); |
| 62 | assert_eq!(category, ErrorCategory::Budget); |
| 63 | assert_eq!( |
| 64 | crate::core::termination::classify_turn_termination(status, Some(category), false, false), |
| 65 | crate::core::termination::RunTerminationReason::BudgetExhausted |
| 66 | ); |
| 67 | |
| 68 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 69 | task.await.expect("engine task"); |
| 70 | } |
| 71 | |
| 72 | #[tokio::test] |
| 73 | async fn goal_turn_uses_goal_step_allowance_and_pauses_budget_limit_after_final_report() { |
| 74 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 75 | |
| 76 | let config = goal_custom_route_config(); |
| 77 | // The model makes one tool step (consuming the 1-step goal allowance), |
| 78 | // then writes its bounded final report when granted it. |
| 79 | let turns = vec![ |
| 80 | canned::tool_call_turn( |
| 81 | "call-step-1", |
| 82 | "File", |
| 83 | r#"{"action":"read","path":"state.txt"}"#, |
| 84 | ), |
| 85 | canned::simple_text_turn("final report: one step of progress made"), |
| 86 | ]; |
| 87 | let mock = std::sync::Arc::new(MockLlmClient::new(turns)); |
| 88 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 89 | let workspace = tempdir().expect("tempdir"); |
| 90 | fs::write(workspace.path().join("state.txt"), "still-working\n").expect("write fixture"); |
| 91 | let (engine, handle) = Engine::new_with_model_client( |
| 92 | EngineConfig { |
| 93 | model: "local-model".to_string(), |
| 94 | workspace: workspace.path().to_path_buf(), |
| 95 | max_steps: 200, |
| 96 | goal_max_steps: Some(1), |
| 97 | goal_objective: Some("finish the migration".to_string()), |
| 98 | snapshots_enabled: false, |
| 99 | terminal_chrome_enabled: false, |
| 100 | ..EngineConfig::default() |
| 101 | }, |
| 102 | &config, |
| 103 | client, |
| 104 | ); |
| 105 | let goal_state = engine.config.goal_state.clone(); |
| 106 | let task = tokio::spawn(engine.run()); |
| 107 | handle |
| 108 | .send(active_goal_message_op( |
| 109 | &config, |
| 110 | "Work on the goal.", |
| 111 | "finish the migration", |
| 112 | None, |
| 113 | )) |
| 114 | .await |
| 115 | .expect("send goal-budget trajectory"); |
| 116 | |
| 117 | let mut rx = handle.rx_event.write().await; |
| 118 | let (status, _) = loop { |
| 119 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 120 | .await |
| 121 | .expect("timed out waiting for goal-budget trajectory") |
| 122 | .expect("engine event"); |
| 123 | if let Event::TurnComplete { status, error, .. } = event { |
| 124 | break (status, error); |
| 125 | } |
| 126 | }; |
| 127 | drop(rx); |
| 128 | |
| 129 | // The final report closes the turn cleanly; the unfinished goal then |
| 130 | // pauses BudgetLimit instead of re-arming another full goal turn (#5994). |
| 131 | assert_eq!(status, TurnOutcomeStatus::Completed); |
| 132 | let snapshot = goal_state.lock().expect("goal lock").snapshot(); |
| 133 | assert_eq!(snapshot.status, "paused"); |
| 134 | assert_eq!( |
| 135 | snapshot.pause_reason, |
| 136 | Some(crate::tools::goal::GoalPauseReason::BudgetLimit) |
| 137 | ); |
| 138 | // The mock served exactly the tool step plus the final report; any |
| 139 | // re-armed continuation would have needed a third provider turn. |
| 140 | assert_eq!(mock.call_count(), 2); |
| 141 | |
| 142 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 143 | task.await.expect("engine task"); |
| 144 | } |
| 145 | |
| 146 | #[tokio::test] |
| 147 | async fn interactive_turn_keeps_ordinary_ceiling_when_goal_allowance_is_configured() { |
| 148 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 149 | |
| 150 | let workspace = tempdir().expect("tempdir"); |
| 151 | fs::write(workspace.path().join("state.txt"), "still-working\n").expect("write fixture"); |
| 152 | // No active goal: the [goal] allowance must never raise the ordinary |
| 153 | // interactive ceiling. |
| 154 | let turns = vec![ |
| 155 | canned::tool_call_turn( |
| 156 | "call-step-1", |
| 157 | "File", |
| 158 | r#"{"action":"read","path":"state.txt"}"#, |
| 159 | ), |
| 160 | canned::tool_call_turn( |
| 161 | "call-step-2", |
| 162 | "File", |
| 163 | r#"{"action":"read","path":"state.txt"}"#, |
| 164 | ), |
| 165 | ]; |
| 166 | let mock = std::sync::Arc::new(MockLlmClient::new(turns)); |
| 167 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 168 | let engine_config = EngineConfig { |
| 169 | max_steps: 1, |
| 170 | goal_max_steps: Some(1_000), |
| 171 | ..deterministic_engine_config(workspace.path()) |
| 172 | }; |
| 173 | let (engine, handle) = Engine::new_with_model_client(engine_config, &Config::default(), client); |
| 174 | let task = tokio::spawn(engine.run()); |
| 175 | handle |
| 176 | .send(external_user_message_op( |
| 177 | "Keep reading until done.", |
| 178 | AppMode::Agent, |
| 179 | &Config::default(), |
| 180 | )) |
| 181 | .await |
| 182 | .expect("send interactive trajectory"); |
| 183 | |
| 184 | let mut rx = handle.rx_event.write().await; |
| 185 | let (status, error) = loop { |
| 186 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 187 | .await |
| 188 | .expect("timed out waiting for interactive trajectory") |
| 189 | .expect("engine event"); |
| 190 | if let Event::TurnComplete { status, error, .. } = event { |
| 191 | break (status, error); |
| 192 | } |
| 193 | }; |
| 194 | drop(rx); |
| 195 | |
| 196 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 197 | let error = error.expect("budget exhaustion carries a terminal error"); |
| 198 | assert!(error.contains("limit: 1"), "{error}"); |
| 199 | assert!(error.contains("max_steps"), "{error}"); |
| 200 | |
| 201 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 202 | task.await.expect("engine task"); |
| 203 | } |
| 204 | |
| 205 | #[test] |
| 206 | fn synthetic_resume_paths_have_no_hidden_default_ceiling() { |
| 207 | let turn_loop = include_str!("../turn_loop.rs"); |
| 208 | |
| 209 | for legacy_marker in ["no_user_input_continues", "no-user-input resume backstop"] { |
| 210 | assert!( |
| 211 | !turn_loop.contains(legacy_marker), |
| 212 | "turn_loop.rs reintroduced the hidden synthetic-resume ceiling marker {legacy_marker:?}" |
| 213 | ); |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | #[tokio::test] |
| 218 | async fn injected_model_duplicate_reads_both_execute_and_close_both_tool_ids() { |
| 219 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 220 | |
| 221 | let workspace = tempdir().expect("tempdir"); |
| 222 | fs::write(workspace.path().join("README.md"), "duplicate-read-proof\n").expect("write fixture"); |
| 223 | let duplicate_read_turn = vec![ |
| 224 | canned::message_start("mock_msg_duplicate_read"), |
| 225 | canned::tool_use_block_start(0, "call-read-1", "File"), |
| 226 | canned::tool_input_delta(0, r#"{"action":"read","path":"README.md"}"#), |
| 227 | canned::block_stop(0), |
| 228 | canned::tool_use_block_start(1, "call-read-2", "File"), |
| 229 | canned::tool_input_delta(1, r#"{"action":"read","path":"README.md"}"#), |
| 230 | canned::block_stop(1), |
| 231 | canned::message_delta("tool_use", None), |
| 232 | canned::message_stop(), |
| 233 | ]; |
| 234 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 235 | duplicate_read_turn, |
| 236 | canned::simple_text_turn("Duplicate read complete."), |
| 237 | ])); |
| 238 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 239 | let (engine, handle) = Engine::new_with_model_client( |
| 240 | deterministic_engine_config(workspace.path()), |
| 241 | &Config::default(), |
| 242 | client, |
| 243 | ); |
| 244 | let task = tokio::spawn(engine.run()); |
| 245 | handle |
| 246 | .send(external_user_message_op( |
| 247 | "Issue the duplicate read batch.", |
| 248 | AppMode::Agent, |
| 249 | &Config::default(), |
| 250 | )) |
| 251 | .await |
| 252 | .expect("send duplicate-read trajectory"); |
| 253 | |
| 254 | let mut results = HashMap::new(); |
| 255 | let mut rx = handle.rx_event.write().await; |
| 256 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 257 | .await |
| 258 | .expect("timed out waiting for duplicate-read trajectory") |
| 259 | { |
| 260 | match event { |
| 261 | Event::ToolCallComplete { |
| 262 | model_call: Some(model_call), |
| 263 | name, |
| 264 | result, |
| 265 | .. |
| 266 | } if name == "File" => { |
| 267 | results.insert(model_call.provider_id, result.expect("read result")); |
| 268 | } |
| 269 | Event::TurnComplete { status, error, .. } => { |
| 270 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 271 | break; |
| 272 | } |
| 273 | _ => {} |
| 274 | } |
| 275 | } |
| 276 | drop(rx); |
| 277 | |
| 278 | assert_eq!(results.len(), 2, "every tool ID needs one terminal result"); |
| 279 | assert!( |
| 280 | results["call-read-1"] |
| 281 | .content |
| 282 | .contains("duplicate-read-proof") |
| 283 | ); |
| 284 | assert!( |
| 285 | results["call-read-2"] |
| 286 | .content |
| 287 | .contains("duplicate-read-proof") |
| 288 | ); |
| 289 | assert!( |
| 290 | results.values().all(|result| result |
| 291 | .metadata |
| 292 | .as_ref() |
| 293 | .is_none_or(|metadata| { metadata.get("executed").is_none() })), |
| 294 | "neither model-requested read may be replaced with a synthetic receipt" |
| 295 | ); |
| 296 | |
| 297 | let requests = mock.captured_requests(); |
| 298 | assert_eq!(requests.len(), 2); |
| 299 | let result_ids = requests[1] |
| 300 | .messages |
| 301 | .iter() |
| 302 | .flat_map(|message| message.content.iter()) |
| 303 | .filter_map(|block| match block { |
| 304 | ContentBlock::ToolResult { tool_use_id, .. } => Some(tool_use_id.as_str()), |
| 305 | _ => None, |
| 306 | }) |
| 307 | .collect::<Vec<_>>(); |
| 308 | assert!(result_ids.contains(&"call-read-1")); |
| 309 | assert!(result_ids.contains(&"call-read-2")); |
| 310 | |
| 311 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 312 | task.await.expect("engine task"); |
| 313 | } |
| 314 | |
| 315 | #[tokio::test] |
| 316 | async fn duplicate_raw_read_errors_each_touch_the_working_set() { |
| 317 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 318 | |
| 319 | async fn missing_read_touches(read_count: usize) -> u32 { |
| 320 | let workspace = tempdir().expect("tempdir"); |
| 321 | let mut read_turn = vec![canned::message_start("mock_msg_missing_read")]; |
| 322 | for index in 0..read_count { |
| 323 | let block_index = u32::try_from(index).expect("test read count fits u32"); |
| 324 | let tool_id = format!("call-missing-{}", index + 1); |
| 325 | read_turn.push(canned::tool_use_block_start( |
| 326 | block_index, |
| 327 | &tool_id, |
| 328 | "read_file", |
| 329 | )); |
| 330 | read_turn.push(canned::tool_input_delta( |
| 331 | block_index, |
| 332 | r#"{"path":"missing.rs"}"#, |
| 333 | )); |
| 334 | read_turn.push(canned::block_stop(block_index)); |
| 335 | } |
| 336 | read_turn.push(canned::message_delta("tool_use", None)); |
| 337 | read_turn.push(canned::message_stop()); |
| 338 | |
| 339 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 340 | read_turn, |
| 341 | canned::simple_text_turn("Missing read handled."), |
| 342 | ])); |
| 343 | let client: crate::core::model_client::SharedModelClient = mock; |
| 344 | let (mut engine, _handle) = Engine::new_with_model_client( |
| 345 | deterministic_engine_config(workspace.path()), |
| 346 | &Config::default(), |
| 347 | client, |
| 348 | ); |
| 349 | let context = crate::tools::ToolContext::new(workspace.path().to_path_buf()); |
| 350 | let mut registry = crate::tools::ToolRegistry::new(context); |
| 351 | registry.register(std::sync::Arc::new(crate::tools::file::ReadFileTool)); |
| 352 | let tools = Some(registry.to_api_tools_with_cache(true)); |
| 353 | let surface = test_tool_surface(&engine, registry, tools, AppMode::Agent); |
| 354 | let mut turn = crate::core::turn::TurnContext::new(8); |
| 355 | |
| 356 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 357 | |
| 358 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 359 | engine |
| 360 | .session |
| 361 | .working_set |
| 362 | .entries |
| 363 | .get("missing.rs") |
| 364 | .expect("leader error should record the attempted path") |
| 365 | .touches |
| 366 | } |
| 367 | |
| 368 | let baseline_touches = missing_read_touches(1).await; |
| 369 | let duplicate_touches = missing_read_touches(2).await; |
| 370 | assert_eq!( |
| 371 | duplicate_touches, |
| 372 | baseline_touches.saturating_mul(2), |
| 373 | "each model-requested read must execute and record its own observation" |
| 374 | ); |
| 375 | } |
| 376 | |
| 377 | #[tokio::test] |
| 378 | async fn truncated_response_continues_turn() { |
| 379 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 380 | |
| 381 | let workspace = tempdir().expect("tempdir"); |
| 382 | let truncated_turn = vec![ |
| 383 | canned::message_start("mock_msg_truncated_continue"), |
| 384 | canned::text_block_start(0), |
| 385 | canned::text_delta(0, "Partial answer before the output budget ran out"), |
| 386 | canned::block_stop(0), |
| 387 | canned::message_delta( |
| 388 | "max_output_tokens", |
| 389 | Some(Usage { |
| 390 | input_tokens: 41, |
| 391 | output_tokens: 7, |
| 392 | reasoning_tokens: Some(3), |
| 393 | ..Default::default() |
| 394 | }), |
| 395 | ), |
| 396 | canned::message_stop(), |
| 397 | ]; |
| 398 | let followup_turn = canned::simple_text_turn("Continued after the truncation."); |
| 399 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![truncated_turn, followup_turn])); |
| 400 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 401 | let (engine, handle) = Engine::new_with_model_client( |
| 402 | deterministic_engine_config(workspace.path()), |
| 403 | &Config::default(), |
| 404 | client, |
| 405 | ); |
| 406 | let task = tokio::spawn(engine.run()); |
| 407 | handle |
| 408 | .send(external_user_message_op( |
| 409 | "Answer the question.", |
| 410 | AppMode::Agent, |
| 411 | &Config::default(), |
| 412 | )) |
| 413 | .await |
| 414 | .expect("send truncated-then-continue trajectory"); |
| 415 | |
| 416 | let mut saw_turn_usage = false; |
| 417 | let mut saw_truncation_observation = false; |
| 418 | let mut last_session_messages = None; |
| 419 | let mut rx = handle.rx_event.write().await; |
| 420 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 421 | .await |
| 422 | .expect("timed out waiting for truncated-then-continue trajectory") |
| 423 | { |
| 424 | match event { |
| 425 | Event::TurnUsage { |
| 426 | usage: reported, .. |
| 427 | } => { |
| 428 | assert_eq!(reported.input_tokens, 41); |
| 429 | assert_eq!(reported.output_tokens, 7); |
| 430 | assert_eq!(reported.reasoning_tokens, Some(3)); |
| 431 | saw_turn_usage = true; |
| 432 | } |
| 433 | Event::SessionUpdated { messages, .. } => { |
| 434 | saw_truncation_observation |= messages.iter().any(|message| { |
| 435 | message.content.iter().any(|block| { |
| 436 | matches!( |
| 437 | block, |
| 438 | ContentBlock::Text { text, .. } |
| 439 | if text.contains("output limit") |
| 440 | && text.contains("Continue from where you left off") |
| 441 | ) |
| 442 | }) |
| 443 | }); |
| 444 | last_session_messages = Some(messages); |
| 445 | } |
| 446 | Event::TurnComplete { status, error, .. } => { |
| 447 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 448 | assert!(error.is_none(), "no terminal error expected: {error:?}"); |
| 449 | break; |
| 450 | } |
| 451 | _ => {} |
| 452 | } |
| 453 | } |
| 454 | drop(rx); |
| 455 | |
| 456 | assert!( |
| 457 | saw_turn_usage, |
| 458 | "reported usage must be accounted before the turn continues" |
| 459 | ); |
| 460 | assert!( |
| 461 | saw_truncation_observation, |
| 462 | "the truncation must be surfaced to the model as a bounded observation" |
| 463 | ); |
| 464 | let messages = last_session_messages.expect("session updated after truncation"); |
| 465 | assert!( |
| 466 | messages.iter().any(|message| { |
| 467 | message.role == "assistant" |
| 468 | && message.content.iter().any(|block| { |
| 469 | matches!( |
| 470 | block, |
| 471 | ContentBlock::Text { text, .. } |
| 472 | if text.contains("Partial answer before the output budget ran out") |
| 473 | ) |
| 474 | }) |
| 475 | }), |
| 476 | "the partial text must be accepted as a completed assistant message" |
| 477 | ); |
| 478 | |
| 479 | let requests = mock.captured_requests(); |
| 480 | assert_eq!( |
| 481 | requests.len(), |
| 482 | 2, |
| 483 | "the loop must continue with a follow-up request" |
| 484 | ); |
| 485 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 486 | task.await.expect("engine task"); |
| 487 | } |
| 488 | |
| 489 | #[tokio::test] |
| 490 | async fn injected_chat_content_filter_never_becomes_a_completed_answer() { |
| 491 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 492 | |
| 493 | let workspace = tempdir().expect("tempdir"); |
| 494 | let filtered_turn = vec![ |
| 495 | canned::message_start("mock_msg_content_filter"), |
| 496 | canned::text_block_start(0), |
| 497 | canned::text_delta(0, "Provider returned only a partial fragment"), |
| 498 | canned::block_stop(0), |
| 499 | canned::message_delta( |
| 500 | "content_filter", |
| 501 | Some(Usage { |
| 502 | input_tokens: 17, |
| 503 | output_tokens: 4, |
| 504 | ..Default::default() |
| 505 | }), |
| 506 | ), |
| 507 | canned::message_stop(), |
| 508 | ]; |
| 509 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![filtered_turn])); |
| 510 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 511 | let (engine, handle) = Engine::new_with_model_client( |
| 512 | deterministic_engine_config(workspace.path()), |
| 513 | &Config::default(), |
| 514 | client, |
| 515 | ); |
| 516 | let task = tokio::spawn(engine.run()); |
| 517 | handle |
| 518 | .send(external_user_message_op( |
| 519 | "Answer the question.", |
| 520 | AppMode::Agent, |
| 521 | &Config::default(), |
| 522 | )) |
| 523 | .await |
| 524 | .expect("send content-filter trajectory"); |
| 525 | |
| 526 | let mut rx = handle.rx_event.write().await; |
| 527 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 528 | .await |
| 529 | .expect("timed out waiting for content-filter trajectory") |
| 530 | { |
| 531 | match event { |
| 532 | Event::MessageComplete { .. } => { |
| 533 | panic!("content-filtered text must not be marked completed") |
| 534 | } |
| 535 | Event::TurnComplete { status, error, .. } => { |
| 536 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 537 | let error = error.expect("content-filtered turn needs a terminal error"); |
| 538 | assert!(error.contains("Model response incomplete"), "{error}"); |
| 539 | assert!(error.contains("content_filter"), "{error}"); |
| 540 | break; |
| 541 | } |
| 542 | _ => {} |
| 543 | } |
| 544 | } |
| 545 | drop(rx); |
| 546 | |
| 547 | assert_eq!(mock.captured_requests().len(), 1); |
| 548 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 549 | task.await.expect("engine task"); |
| 550 | } |
| 551 | |
| 552 | #[tokio::test] |
| 553 | async fn injected_model_complete_tool_block_at_max_output_tokens_executes() { |
| 554 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 555 | |
| 556 | let workspace = tempdir().expect("tempdir"); |
| 557 | fs::write( |
| 558 | workspace.path().join("truncated-tool-ran.txt"), |
| 559 | "truncated-tool-executed", |
| 560 | ) |
| 561 | .expect("write fixture"); |
| 562 | let usage = Usage { |
| 563 | input_tokens: 52, |
| 564 | output_tokens: 23, |
| 565 | ..Default::default() |
| 566 | }; |
| 567 | let truncated_tool_turn = vec![ |
| 568 | canned::message_start("mock_msg_truncated_tool"), |
| 569 | canned::tool_use_block_start(0, "call-truncated", "File"), |
| 570 | canned::tool_input_delta(0, r#"{"action":"read","path":"truncated-tool-ran.txt"}"#), |
| 571 | canned::block_stop(0), |
| 572 | canned::message_delta("max_output_tokens", Some(usage.clone())), |
| 573 | canned::message_stop(), |
| 574 | ]; |
| 575 | let followup_turn = canned::simple_text_turn("Done."); |
| 576 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![truncated_tool_turn, followup_turn])); |
| 577 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 578 | let (engine, handle) = Engine::new_with_model_client( |
| 579 | deterministic_engine_config(workspace.path()), |
| 580 | &Config::default(), |
| 581 | client, |
| 582 | ); |
| 583 | let task = tokio::spawn(engine.run()); |
| 584 | handle |
| 585 | .send(external_user_message_op( |
| 586 | "Read the fixture file.", |
| 587 | AppMode::Agent, |
| 588 | &Config::default(), |
| 589 | )) |
| 590 | .await |
| 591 | .expect("send truncated-tool trajectory"); |
| 592 | |
| 593 | let mut saw_turn_usage = false; |
| 594 | let mut saw_tool_start = false; |
| 595 | let mut saw_tool_success = false; |
| 596 | let mut rx = handle.rx_event.write().await; |
| 597 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 598 | .await |
| 599 | .expect("timed out waiting for truncated-tool trajectory") |
| 600 | { |
| 601 | match event { |
| 602 | Event::TurnUsage { .. } => saw_turn_usage = true, |
| 603 | Event::ToolCallStarted { |
| 604 | id, |
| 605 | name, |
| 606 | model_call, |
| 607 | .. |
| 608 | } => { |
| 609 | assert_ne!(id, "call-truncated"); |
| 610 | assert_eq!(model_call.unwrap().provider_id, "call-truncated"); |
| 611 | assert_eq!(name, "File"); |
| 612 | saw_tool_start = true; |
| 613 | } |
| 614 | Event::ToolCallComplete { |
| 615 | id, |
| 616 | name, |
| 617 | result, |
| 618 | model_call, |
| 619 | } => { |
| 620 | assert_ne!(id, "call-truncated"); |
| 621 | assert_eq!(model_call.unwrap().provider_id, "call-truncated"); |
| 622 | assert_eq!(name, "File"); |
| 623 | let result = result.expect("complete tool call closes with a tool result"); |
| 624 | assert!( |
| 625 | result.success, |
| 626 | "complete tool call must be accepted and executed: {result:?}" |
| 627 | ); |
| 628 | saw_tool_success = true; |
| 629 | } |
| 630 | Event::TurnComplete { status, error, .. } => { |
| 631 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 632 | assert!(error.is_none(), "no terminal error expected: {error:?}"); |
| 633 | break; |
| 634 | } |
| 635 | _ => {} |
| 636 | } |
| 637 | } |
| 638 | drop(rx); |
| 639 | |
| 640 | assert!(saw_turn_usage, "reported usage must be emitted"); |
| 641 | assert!(saw_tool_start, "the streamed tool lifecycle must open"); |
| 642 | assert!( |
| 643 | saw_tool_success, |
| 644 | "the complete tool call must be accepted and executed" |
| 645 | ); |
| 646 | let requests = mock.captured_requests(); |
| 647 | assert_eq!( |
| 648 | requests.len(), |
| 649 | 2, |
| 650 | "the loop must continue with a follow-up request after the tool result" |
| 651 | ); |
| 652 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 653 | task.await.expect("engine task"); |
| 654 | } |
| 655 | |
| 656 | #[tokio::test] |
| 657 | async fn injected_model_receives_malformed_tool_feedback_and_recovers() { |
| 658 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 659 | |
| 660 | let workspace = tempdir().expect("tempdir"); |
| 661 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 662 | canned::tool_call_turn("call-bad-read", "File", r#"{"action":"read"}"#), |
| 663 | canned::simple_text_turn("Recovered after validation feedback."), |
| 664 | ])); |
| 665 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 666 | let (engine, handle) = Engine::new_with_model_client( |
| 667 | deterministic_engine_config(workspace.path()), |
| 668 | &Config::default(), |
| 669 | client, |
| 670 | ); |
| 671 | let task = tokio::spawn(engine.run()); |
| 672 | handle |
| 673 | .send(external_user_message_op( |
| 674 | "Exercise malformed tool feedback.", |
| 675 | AppMode::Agent, |
| 676 | &Config::default(), |
| 677 | )) |
| 678 | .await |
| 679 | .expect("send malformed trajectory"); |
| 680 | |
| 681 | let mut validation_feedback = None; |
| 682 | let mut recovered = false; |
| 683 | let mut rx = handle.rx_event.write().await; |
| 684 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 685 | .await |
| 686 | .expect("timed out waiting for malformed trajectory") |
| 687 | { |
| 688 | match event { |
| 689 | Event::ToolCallComplete { name, result, .. } if name == "File" => { |
| 690 | validation_feedback = Some(match result { |
| 691 | Ok(result) => result.content, |
| 692 | Err(error) => error.to_string(), |
| 693 | }); |
| 694 | } |
| 695 | Event::MessageDelta { content, .. } => { |
| 696 | recovered |= content.contains("Recovered after validation feedback"); |
| 697 | } |
| 698 | Event::TurnComplete { status, error, .. } => { |
| 699 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 700 | break; |
| 701 | } |
| 702 | _ => {} |
| 703 | } |
| 704 | } |
| 705 | drop(rx); |
| 706 | let feedback = validation_feedback.expect("validation feedback event"); |
| 707 | assert!(feedback.to_ascii_lowercase().contains("path"), "{feedback}"); |
| 708 | assert!( |
| 709 | recovered, |
| 710 | "model must get a follow-up turn after tool failure" |
| 711 | ); |
| 712 | assert_eq!(mock.call_count(), 2); |
| 713 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 714 | task.await.expect("engine task"); |
| 715 | } |
| 716 | |
| 717 | #[tokio::test] |
| 718 | async fn engine_cancellation_drops_active_injected_model_request() { |
| 719 | let workspace = tempdir().expect("tempdir"); |
| 720 | let entered = std::sync::Arc::new(tokio::sync::Notify::new()); |
| 721 | let request_dropped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)); |
| 722 | let client: crate::core::model_client::SharedModelClient = |
| 723 | std::sync::Arc::new(BlockingModelClient { |
| 724 | entered: std::sync::Arc::clone(&entered), |
| 725 | request_dropped: std::sync::Arc::clone(&request_dropped), |
| 726 | }); |
| 727 | let (engine, handle) = Engine::new_with_model_client( |
| 728 | deterministic_engine_config(workspace.path()), |
| 729 | &Config::default(), |
| 730 | client, |
| 731 | ); |
| 732 | let task = tokio::spawn(engine.run()); |
| 733 | handle |
| 734 | .send(external_user_message_op( |
| 735 | "Block until explicitly cancelled.", |
| 736 | AppMode::Agent, |
| 737 | &Config::default(), |
| 738 | )) |
| 739 | .await |
| 740 | .expect("send cancellation trajectory"); |
| 741 | tokio::time::timeout(model_turn_event_timeout(), entered.notified()) |
| 742 | .await |
| 743 | .expect("model request was never entered"); |
| 744 | |
| 745 | let mut rx = handle.rx_event.write().await; |
| 746 | let pending_snapshot = loop { |
| 747 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 748 | .await |
| 749 | .expect("timed out waiting for pending request snapshot") |
| 750 | .expect("engine event"); |
| 751 | if let Event::ToolRequestSnapshot { snapshot } = event { |
| 752 | break snapshot; |
| 753 | } |
| 754 | }; |
| 755 | assert!(pending_snapshot.delivery_status.starts_with("unknown")); |
| 756 | assert!(pending_snapshot.tools_field_present); |
| 757 | drop(rx); |
| 758 | handle.cancel(); |
| 759 | |
| 760 | let mut rx = handle.rx_event.write().await; |
| 761 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 762 | .await |
| 763 | .expect("timed out waiting for cancellation") |
| 764 | { |
| 765 | if let Event::TurnComplete { status, error, .. } = event { |
| 766 | assert_eq!(status, TurnOutcomeStatus::Interrupted, "{error:?}"); |
| 767 | break; |
| 768 | } |
| 769 | } |
| 770 | drop(rx); |
| 771 | assert!( |
| 772 | request_dropped.load(std::sync::atomic::Ordering::SeqCst), |
| 773 | "cancellation must drop the active provider future" |
| 774 | ); |
| 775 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 776 | task.await.expect("engine task"); |
| 777 | } |
| 778 | |
| 779 | fn guardian_fixture_response(text: &str) -> codewhale_models::MessageResponse { |
| 780 | codewhale_models::MessageResponse { |
| 781 | id: "guardian-fixture".to_string(), |
| 782 | r#type: "message".to_string(), |
| 783 | role: "assistant".to_string(), |
| 784 | content: vec![ContentBlock::Text { |
| 785 | text: text.to_string(), |
| 786 | cache_control: None, |
| 787 | }], |
| 788 | model: "mock-model".to_string(), |
| 789 | stop_reason: Some("end_turn".to_string()), |
| 790 | stop_sequence: None, |
| 791 | container: None, |
| 792 | usage: Usage { |
| 793 | input_tokens: 17, |
| 794 | output_tokens: 3, |
| 795 | ..Usage::default() |
| 796 | }, |
| 797 | } |
| 798 | } |
| 799 | |
| 800 | fn guardian_tool_results<'a>( |
| 801 | request: &'a codewhale_models::MessageRequest, |
| 802 | call_id: &str, |
| 803 | ) -> Vec<(&'a str, Option<bool>)> { |
| 804 | request |
| 805 | .messages |
| 806 | .iter() |
| 807 | .flat_map(|message| &message.content) |
| 808 | .filter_map(|block| match block { |
| 809 | ContentBlock::ToolResult { |
| 810 | tool_use_id, |
| 811 | content, |
| 812 | is_error, |
| 813 | .. |
| 814 | } if tool_use_id == call_id => Some((content.as_str(), *is_error)), |
| 815 | _ => None, |
| 816 | }) |
| 817 | .collect() |
| 818 | } |
| 819 | |
| 820 | /// One transcript-visible gate receipt observed on the event stream. |
| 821 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 822 | struct GateReceipt { |
| 823 | gate: crate::core::events::ToolGate, |
| 824 | decision: crate::core::events::ToolGateVerdict, |
| 825 | risk: Option<String>, |
| 826 | reason: String, |
| 827 | } |
| 828 | |
| 829 | async fn collect_guardian_journey_with_receipts( |
| 830 | handle: &EngineHandle, |
| 831 | call_id: &str, |
| 832 | ) -> ( |
| 833 | Result<crate::tools::spec::ToolResult, crate::tools::spec::ToolError>, |
| 834 | Vec<Usage>, |
| 835 | Usage, |
| 836 | Vec<GateReceipt>, |
| 837 | ) { |
| 838 | let mut completion = None; |
| 839 | let mut execution_id = None; |
| 840 | let mut usage_events = Vec::new(); |
| 841 | let mut receipts = Vec::new(); |
| 842 | let mut rx = handle.rx_event.write().await; |
| 843 | let terminal_usage = loop { |
| 844 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 845 | .await |
| 846 | .expect("timed out waiting for Auto-Review journey") |
| 847 | .expect("engine event stream closed"); |
| 848 | match event { |
| 849 | Event::ToolCallStarted { |
| 850 | id, |
| 851 | model_call: Some(model_call), |
| 852 | .. |
| 853 | } if model_call.provider_id == call_id => { |
| 854 | execution_id = Some(id); |
| 855 | } |
| 856 | Event::ToolCallComplete { |
| 857 | model_call: Some(model_call), |
| 858 | result, |
| 859 | .. |
| 860 | } if model_call.provider_id == call_id => { |
| 861 | assert!( |
| 862 | completion.replace(result).is_none(), |
| 863 | "duplicate tool result" |
| 864 | ); |
| 865 | } |
| 866 | // Guardian consults carry their own routed receipt; both the |
| 867 | // parent-route and routed per-call telemetry count as reaching |
| 868 | // the cost UI. |
| 869 | Event::TurnUsage { usage, .. } | Event::RoutedTurnUsage { usage, .. } => { |
| 870 | usage_events.push(usage); |
| 871 | } |
| 872 | Event::ToolGateDecision { |
| 873 | tool_id, |
| 874 | gate, |
| 875 | decision, |
| 876 | risk, |
| 877 | reason, |
| 878 | .. |
| 879 | } if execution_id.as_deref() == Some(tool_id.as_str()) => receipts.push(GateReceipt { |
| 880 | gate, |
| 881 | decision, |
| 882 | risk, |
| 883 | reason, |
| 884 | }), |
| 885 | Event::TurnComplete { |
| 886 | status, |
| 887 | error, |
| 888 | usage, |
| 889 | .. |
| 890 | } => { |
| 891 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 892 | break usage; |
| 893 | } |
| 894 | _ => {} |
| 895 | } |
| 896 | }; |
| 897 | drop(rx); |
| 898 | ( |
| 899 | completion.expect("held call must have one paired result"), |
| 900 | usage_events, |
| 901 | terminal_usage, |
| 902 | receipts, |
| 903 | ) |
| 904 | } |
| 905 | |
| 906 | #[cfg(windows)] |
| 907 | #[test] |
| 908 | fn windows_full_access_node_image_kill_is_denied_before_any_shell_effect() { |
| 909 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 910 | |
| 911 | with_artifact_home(|home| { |
| 912 | tokio::runtime::Builder::new_current_thread() |
| 913 | .enable_all() |
| 914 | .build() |
| 915 | .unwrap() |
| 916 | .block_on(async { |
| 917 | let workspace = home.join("windows-node-floor"); |
| 918 | fs::create_dir(&workspace).expect("fixture workspace"); |
| 919 | let sentinel = workspace.join("must-not-run.txt"); |
| 920 | // The scanner deliberately holds commands inside conditionals. |
| 921 | // Keep the kill unreachable even if this regression breaks: |
| 922 | // failing the guard may write the sentinel, but must never kill |
| 923 | // the developer's other Node or Codewhale processes. |
| 924 | let command = "powershell -NoProfile -Command \"Set-Content -LiteralPath 'must-not-run.txt' -Value 'ran'; if ($false) { taskkill /F /IM node.exe }\""; |
| 925 | let arguments = json!({"command": command}).to_string(); |
| 926 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![canned::tool_call_turn( |
| 927 | "windows-node-kill", |
| 928 | "bash", |
| 929 | &arguments, |
| 930 | )])); |
| 931 | mock.push_factory(|request| { |
| 932 | let results = guardian_tool_results(request, "windows-node-kill"); |
| 933 | assert_eq!(results.len(), 1, "one provider ID gets one result"); |
| 934 | assert_eq!(results[0].1, Some(true)); |
| 935 | assert!(results[0].0.contains("Node launcher"), "{results:?}"); |
| 936 | assert!(results[0].0.contains("do not work around"), "{results:?}"); |
| 937 | canned::simple_text_turn("Use an owned server PID instead.") |
| 938 | }); |
| 939 | let config = Config { allow_shell: Some(true), ..Config::default() }; |
| 940 | let (engine, handle) = Engine::new_with_model_client( |
| 941 | deterministic_engine_config(&workspace), &config, mock.clone(), |
| 942 | ); |
| 943 | let mut op = external_user_message_op("Attempt the supplied cleanup.", AppMode::Agent, &config); |
| 944 | let Op::SendMessage(turn) = &mut op else { unreachable!("model turn fixture") }; |
| 945 | turn.auto_approve = true; |
| 946 | turn.approval_mode = ApprovalMode::Bypass; |
| 947 | let task = tokio::spawn(engine.run()); |
| 948 | handle.send(op).await.expect("send Full Access trajectory"); |
| 949 | let (completion, _, _, receipts) = collect_guardian_journey_with_receipts(&handle, "windows-node-kill").await; |
| 950 | let error = completion.expect_err("runtime safety floor must deny before shell execution"); |
| 951 | assert!(matches!(error, crate::tools::spec::ToolError::PermissionDenied { .. })); |
| 952 | assert!(error.to_string().contains("Node launcher"), "{error}"); |
| 953 | assert!(!sentinel.exists(), "no part of the composed shell call may run"); |
| 954 | assert_eq!(receipts.len(), 1, "{receipts:?}"); |
| 955 | assert_eq!(receipts[0].gate, crate::core::events::ToolGate::AutoReviewDeterministic); |
| 956 | assert_eq!(receipts[0].decision, crate::core::events::ToolGateVerdict::Denied); |
| 957 | assert!(receipts[0].reason.contains("Node launcher")); |
| 958 | assert_eq!(mock.captured_requests().len(), 2, "main and paired-result follow-up only; no guardian"); |
| 959 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 960 | tokio::time::timeout(model_turn_event_timeout(), task).await.expect("bounded shutdown").expect("engine task"); |
| 961 | }); |
| 962 | }); |
| 963 | } |
| 964 | |
| 965 | #[tokio::test] |
| 966 | async fn auto_review_guardian_allow_executes_once_and_accounts_usage_without_prompt_leak() { |
| 967 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 968 | |
| 969 | const REVIEW_REASON: &str = "bounded fixture write is reversible"; |
| 970 | let workspace = tempdir().expect("tempdir"); |
| 971 | fs::create_dir(workspace.path().join(".git")).expect("git marker"); |
| 972 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![canned::tool_call_turn( |
| 973 | "call-guardian-allow", |
| 974 | "File", |
| 975 | r#"{"action":"write","path":".env","content":"assembled=true\n"}"#, |
| 976 | )])); |
| 977 | mock.push_factory(|request| { |
| 978 | let tool_results = guardian_tool_results(request, "call-guardian-allow"); |
| 979 | assert_eq!(tool_results.len(), 1, "one call must produce one result"); |
| 980 | assert_ne!(tool_results[0].1, Some(true)); |
| 981 | let request_json = serde_json::to_string(request).expect("serialize follow-up request"); |
| 982 | assert!(!request_json.contains(REVIEW_REASON), "{request_json}"); |
| 983 | assert!( |
| 984 | !request_json.contains("deterministic_observations"), |
| 985 | "{request_json}" |
| 986 | ); |
| 987 | assert!(!request_json.contains("hold_reason"), "{request_json}"); |
| 988 | canned::simple_text_turn("Guardian-approved write complete.") |
| 989 | }); |
| 990 | mock.push_message_response(guardian_fixture_response(&format!( |
| 991 | r#"{{"risk_level":"low","decision":"allow","reason":"{REVIEW_REASON}"}}"# |
| 992 | ))); |
| 993 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 994 | let config = Config::default(); |
| 995 | let (engine, handle) = Engine::new_with_model_client( |
| 996 | deterministic_engine_config(workspace.path()), |
| 997 | &config, |
| 998 | client, |
| 999 | ); |
| 1000 | let task = tokio::spawn(engine.run()); |
| 1001 | handle |
| 1002 | .send(auto_review_message_op( |
| 1003 | "Write the isolated fixture.", |
| 1004 | &config, |
| 1005 | )) |
| 1006 | .await |
| 1007 | .expect("send Auto-Review allow journey"); |
| 1008 | |
| 1009 | let (completion, usage_events, terminal_usage, receipts) = |
| 1010 | collect_guardian_journey_with_receipts(&handle, "call-guardian-allow").await; |
| 1011 | assert!(completion.expect("guardian-approved tool result").success); |
| 1012 | // The person never saw a prompt, so the transcript gets exactly one |
| 1013 | // receipt naming the guardian's verdict and risk tier. |
| 1014 | assert_eq!(receipts.len(), 1, "{receipts:?}"); |
| 1015 | assert_eq!( |
| 1016 | receipts[0].gate, |
| 1017 | crate::core::events::ToolGate::AutoReviewGuardian |
| 1018 | ); |
| 1019 | assert_eq!( |
| 1020 | receipts[0].decision, |
| 1021 | crate::core::events::ToolGateVerdict::Allowed |
| 1022 | ); |
| 1023 | assert!(receipts[0].risk.is_some(), "{receipts:?}"); |
| 1024 | assert!(!receipts[0].reason.contains('\n')); |
| 1025 | assert!( |
| 1026 | usage_events |
| 1027 | .iter() |
| 1028 | .any(|usage| usage.input_tokens == 17 && usage.output_tokens == 3), |
| 1029 | "guardian usage must reach the cost UI" |
| 1030 | ); |
| 1031 | assert_eq!(terminal_usage.input_tokens, 17); |
| 1032 | assert_eq!(terminal_usage.output_tokens, 3); |
| 1033 | assert_eq!( |
| 1034 | fs::read_to_string(workspace.path().join(".env")).expect("written fixture"), |
| 1035 | "assembled=true\n" |
| 1036 | ); |
| 1037 | let requests = mock.captured_requests(); |
| 1038 | assert_eq!(requests.len(), 3, "main, guardian, follow-up"); |
| 1039 | assert_eq!(requests[1].stream, Some(false)); |
| 1040 | assert!(requests[1].tools.is_none()); |
| 1041 | let guardian_json = serde_json::to_string(&requests[1]).expect("guardian request JSON"); |
| 1042 | assert!(guardian_json.contains("call") || guardian_json.contains("proposed_tool_call")); |
| 1043 | assert!(guardian_json.contains(".env")); |
| 1044 | |
| 1045 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1046 | task.await.expect("engine task"); |
| 1047 | } |
| 1048 | |
| 1049 | #[tokio::test] |
| 1050 | async fn auto_review_guardian_deny_returns_one_paired_failed_result() { |
| 1051 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1052 | |
| 1053 | const DENIAL: &str = "sensitive configuration must remain untouched"; |
| 1054 | let workspace = tempdir().expect("tempdir"); |
| 1055 | fs::create_dir(workspace.path().join(".git")).expect("git marker"); |
| 1056 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![canned::tool_call_turn( |
| 1057 | "call-guardian-deny", |
| 1058 | "File", |
| 1059 | r#"{"action":"write","path":".env","content":"must-not-run\n"}"#, |
| 1060 | )])); |
| 1061 | mock.push_factory(|request| { |
| 1062 | let tool_results = guardian_tool_results(request, "call-guardian-deny"); |
| 1063 | assert_eq!(tool_results.len(), 1, "denied call must not be orphaned"); |
| 1064 | assert_eq!(tool_results[0].1, Some(true)); |
| 1065 | assert!(tool_results[0].0.contains(DENIAL), "{tool_results:?}"); |
| 1066 | assert!( |
| 1067 | tool_results[0].0.contains("Do not work around this denial"), |
| 1068 | "{tool_results:?}" |
| 1069 | ); |
| 1070 | let request_json = serde_json::to_string(request).expect("serialize follow-up request"); |
| 1071 | assert!( |
| 1072 | !request_json.contains("deterministic_observations"), |
| 1073 | "{request_json}" |
| 1074 | ); |
| 1075 | assert!(!request_json.contains("hold_reason"), "{request_json}"); |
| 1076 | canned::simple_text_turn("Stopped after the guardian denial.") |
| 1077 | }); |
| 1078 | mock.push_message_response(guardian_fixture_response(&format!( |
| 1079 | r#"{{"risk_level":"medium","decision":"deny","reason":"{DENIAL}"}}"# |
| 1080 | ))); |
| 1081 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1082 | let config = Config::default(); |
| 1083 | let (engine, handle) = Engine::new_with_model_client( |
| 1084 | deterministic_engine_config(workspace.path()), |
| 1085 | &config, |
| 1086 | client, |
| 1087 | ); |
| 1088 | let task = tokio::spawn(engine.run()); |
| 1089 | handle |
| 1090 | .send(auto_review_message_op("Attempt the held write.", &config)) |
| 1091 | .await |
| 1092 | .expect("send Auto-Review deny journey"); |
| 1093 | |
| 1094 | let (completion, _, _, receipts) = |
| 1095 | collect_guardian_journey_with_receipts(&handle, "call-guardian-deny").await; |
| 1096 | let error = completion.expect_err("guardian denial must fail the tool call"); |
| 1097 | assert!(error.to_string().contains(DENIAL), "{error}"); |
| 1098 | assert_eq!(receipts.len(), 1, "{receipts:?}"); |
| 1099 | assert_eq!( |
| 1100 | receipts[0].decision, |
| 1101 | crate::core::events::ToolGateVerdict::Denied |
| 1102 | ); |
| 1103 | assert!(receipts[0].reason.contains(DENIAL), "{receipts:?}"); |
| 1104 | assert!(!workspace.path().join(".env").exists()); |
| 1105 | assert_eq!(mock.captured_requests().len(), 3); |
| 1106 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1107 | task.await.expect("engine task"); |
| 1108 | } |
| 1109 | |
| 1110 | #[tokio::test] |
| 1111 | async fn auto_review_guardian_parse_and_transport_failures_deny_closed() { |
| 1112 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1113 | |
| 1114 | for failure in ["parse", "transport"] { |
| 1115 | let workspace = tempdir().expect("tempdir"); |
| 1116 | fs::create_dir(workspace.path().join(".git")).expect("git marker"); |
| 1117 | let call_id = format!("call-guardian-{failure}"); |
| 1118 | let initial = canned::tool_call_turn( |
| 1119 | &call_id, |
| 1120 | "File", |
| 1121 | r#"{"action":"write","path":".env","content":"must-not-run\n"}"#, |
| 1122 | ); |
| 1123 | let follow_up_id = call_id.clone(); |
| 1124 | let follow_up = move |request: &codewhale_models::MessageRequest| { |
| 1125 | let results = guardian_tool_results(request, &follow_up_id); |
| 1126 | assert_eq!(results.len(), 1, "reviewer failure must pair one result"); |
| 1127 | let result = results[0]; |
| 1128 | assert_eq!(result.1, Some(true)); |
| 1129 | assert!(result.0.contains("denied (fail closed)"), "{result:?}"); |
| 1130 | assert!(!result.0.contains("fixture guardian transport failure")); |
| 1131 | canned::simple_text_turn("Stopped after reviewer failure.") |
| 1132 | }; |
| 1133 | |
| 1134 | let config = Config::default(); |
| 1135 | let client: crate::core::model_client::SharedModelClient = if failure == "parse" { |
| 1136 | let mock = MockLlmClient::new(vec![initial]); |
| 1137 | mock.push_factory(follow_up); |
| 1138 | mock.push_message_response(guardian_fixture_response("not valid guardian JSON")); |
| 1139 | std::sync::Arc::new(mock) |
| 1140 | } else { |
| 1141 | let mock = MockLlmClient::new(vec![initial]); |
| 1142 | mock.push_factory(follow_up); |
| 1143 | std::sync::Arc::new(FailingGuardianModelClient { inner: mock }) |
| 1144 | }; |
| 1145 | let (engine, handle) = Engine::new_with_model_client( |
| 1146 | deterministic_engine_config(workspace.path()), |
| 1147 | &config, |
| 1148 | client, |
| 1149 | ); |
| 1150 | let task = tokio::spawn(engine.run()); |
| 1151 | handle |
| 1152 | .send(auto_review_message_op("Attempt the held write.", &config)) |
| 1153 | .await |
| 1154 | .expect("send reviewer failure journey"); |
| 1155 | |
| 1156 | let (completion, _, _, receipts) = |
| 1157 | collect_guardian_journey_with_receipts(&handle, &call_id).await; |
| 1158 | let error = completion.expect_err("reviewer failure must deny"); |
| 1159 | assert!(error.to_string().contains("fail closed"), "{error}"); |
| 1160 | assert_eq!(receipts.len(), 1, "{receipts:?}"); |
| 1161 | assert_eq!( |
| 1162 | receipts[0].decision, |
| 1163 | crate::core::events::ToolGateVerdict::Unavailable, |
| 1164 | "{receipts:?}" |
| 1165 | ); |
| 1166 | assert!(receipts[0].risk.is_none()); |
| 1167 | assert!(!workspace.path().join(".env").exists()); |
| 1168 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1169 | task.await.expect("engine task"); |
| 1170 | } |
| 1171 | } |
| 1172 | |
| 1173 | #[tokio::test] |
| 1174 | async fn auto_review_cancellation_promptly_drops_the_guardian_request() { |
| 1175 | let workspace = tempdir().expect("tempdir"); |
| 1176 | fs::create_dir(workspace.path().join(".git")).expect("git marker"); |
| 1177 | let guardian_entered = std::sync::Arc::new(tokio::sync::Notify::new()); |
| 1178 | let guardian_dropped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)); |
| 1179 | let model = std::sync::Arc::new(BlockingGuardianModelClient { |
| 1180 | guardian_entered: std::sync::Arc::clone(&guardian_entered), |
| 1181 | guardian_dropped: std::sync::Arc::clone(&guardian_dropped), |
| 1182 | streaming_calls: std::sync::atomic::AtomicUsize::new(0), |
| 1183 | }); |
| 1184 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 1185 | let config = Config::default(); |
| 1186 | let (engine, handle) = Engine::new_with_model_client( |
| 1187 | deterministic_engine_config(workspace.path()), |
| 1188 | &config, |
| 1189 | client, |
| 1190 | ); |
| 1191 | let task = tokio::spawn(engine.run()); |
| 1192 | handle |
| 1193 | .send(auto_review_message_op("Attempt the held write.", &config)) |
| 1194 | .await |
| 1195 | .expect("send blocking guardian journey"); |
| 1196 | tokio::time::timeout(model_turn_event_timeout(), guardian_entered.notified()) |
| 1197 | .await |
| 1198 | .expect("guardian request was never entered"); |
| 1199 | |
| 1200 | handle.cancel(); |
| 1201 | let mut rx = handle.rx_event.write().await; |
| 1202 | loop { |
| 1203 | let event = tokio::time::timeout(Duration::from_secs(1), rx.recv()) |
| 1204 | .await |
| 1205 | .expect("guardian cancellation did not settle promptly") |
| 1206 | .expect("engine event stream closed"); |
| 1207 | if let Event::TurnComplete { status, error, .. } = event { |
| 1208 | assert_eq!(status, TurnOutcomeStatus::Interrupted, "{error:?}"); |
| 1209 | break; |
| 1210 | } |
| 1211 | } |
| 1212 | drop(rx); |
| 1213 | |
| 1214 | assert!( |
| 1215 | guardian_dropped.load(std::sync::atomic::Ordering::SeqCst), |
| 1216 | "cancellation must drop the guardian provider future" |
| 1217 | ); |
| 1218 | assert_eq!( |
| 1219 | model |
| 1220 | .streaming_calls |
| 1221 | .load(std::sync::atomic::Ordering::SeqCst), |
| 1222 | 1 |
| 1223 | ); |
| 1224 | assert!(!workspace.path().join(".env").exists()); |
| 1225 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1226 | task.await.expect("engine task"); |
| 1227 | } |
| 1228 | |
| 1229 | #[tokio::test] |
| 1230 | #[allow(clippy::await_holding_lock)] |
| 1231 | async fn operate_conversation_reaches_provider_when_workers_are_disabled() { |
| 1232 | use wiremock::matchers::{method, path}; |
| 1233 | use wiremock::{Mock, MockServer, ResponseTemplate}; |
| 1234 | |
| 1235 | let _lock = lock_test_env(); |
| 1236 | let workspace = tempdir().expect("tempdir"); |
| 1237 | let server = MockServer::start().await; |
| 1238 | let done_sse = concat!( |
| 1239 | "data: {\"id\":\"chatcmpl-operate\",\"choices\":[{\"index\":0,", |
| 1240 | "\"delta\":{\"content\":\"I can still answer normally.\"},\"finish_reason\":null}]}\n\n", |
| 1241 | "data: {\"id\":\"chatcmpl-operate\",\"choices\":[{\"index\":0,", |
| 1242 | "\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n", |
| 1243 | "data: [DONE]\n\n", |
| 1244 | ); |
| 1245 | Mock::given(method("POST")) |
| 1246 | .and(path("/v1/chat/completions")) |
| 1247 | .respond_with( |
| 1248 | ResponseTemplate::new(200) |
| 1249 | .insert_header("content-type", "text/event-stream") |
| 1250 | .set_body_string(done_sse), |
| 1251 | ) |
| 1252 | .expect(1) |
| 1253 | .mount(&server) |
| 1254 | .await; |
| 1255 | |
| 1256 | let api_config = Config { |
| 1257 | ..Config::default() |
| 1258 | } |
| 1259 | .with_legacy_root(Some("test-key".to_string()), Some(server.uri())); |
| 1260 | let engine_config = EngineConfig { |
| 1261 | workspace: workspace.path().to_path_buf(), |
| 1262 | snapshots_enabled: false, |
| 1263 | subagents_enabled: false, |
| 1264 | ..EngineConfig::default() |
| 1265 | }; |
| 1266 | let (operate_engine, operate_handle) = Engine::new(engine_config, &api_config); |
| 1267 | let operate_task = tokio::spawn(operate_engine.run()); |
| 1268 | operate_handle |
| 1269 | .send(external_user_message_op( |
| 1270 | "what is a Rust worktree?", |
| 1271 | AppMode::Operate, |
| 1272 | &api_config, |
| 1273 | )) |
| 1274 | .await |
| 1275 | .expect("send Operate turn"); |
| 1276 | |
| 1277 | let mut saw_operate_complete = false; |
| 1278 | let mut saw_operate_route = false; |
| 1279 | let mut operate_rx = operate_handle.rx_event.write().await; |
| 1280 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), operate_rx.recv()) |
| 1281 | .await |
| 1282 | .expect("timed out waiting for Operate completion") |
| 1283 | { |
| 1284 | match event { |
| 1285 | Event::RouteDispatched { route, .. } => { |
| 1286 | assert_eq!(route.provider, ProviderKind::Deepseek); |
| 1287 | assert_eq!(route.model, crate::config::DEFAULT_TEXT_MODEL); |
| 1288 | assert!(!route.auto_model); |
| 1289 | saw_operate_route = true; |
| 1290 | } |
| 1291 | Event::Error { envelope, .. } => { |
| 1292 | panic!("ordinary Operate conversation emitted an error: {envelope:?}"); |
| 1293 | } |
| 1294 | Event::TurnComplete { status, error, .. } => { |
| 1295 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1296 | saw_operate_complete = true; |
| 1297 | break; |
| 1298 | } |
| 1299 | _ => {} |
| 1300 | } |
| 1301 | } |
| 1302 | drop(operate_rx); |
| 1303 | |
| 1304 | assert!( |
| 1305 | saw_operate_route, |
| 1306 | "model turns must publish route provenance" |
| 1307 | ); |
| 1308 | assert!( |
| 1309 | saw_operate_complete, |
| 1310 | "Operate conversation must complete without worker readiness" |
| 1311 | ); |
| 1312 | let requests = server |
| 1313 | .received_requests() |
| 1314 | .await |
| 1315 | .expect("recorded requests after Operate"); |
| 1316 | assert_eq!(requests.len(), 1, "Operate must reach the provider"); |
| 1317 | operate_handle |
| 1318 | .send(Op::Shutdown) |
| 1319 | .await |
| 1320 | .expect("shutdown Operate engine"); |
| 1321 | operate_task.await.expect("Operate engine task"); |
| 1322 | } |
| 1323 | |
| 1324 | fn auto_review_plan_decision( |
| 1325 | policy: &crate::tui::auto_review::AutoReviewPolicy, |
| 1326 | tool_name: &str, |
| 1327 | tool_input: &Value, |
| 1328 | run_origin: crate::tui::auto_review::RunOrigin, |
| 1329 | approval_mode: ApprovalMode, |
| 1330 | workspace_trusted: bool, |
| 1331 | workspace: Option<&Path>, |
| 1332 | ) -> (AutoReviewPlanDecision, Value) { |
| 1333 | let context = crate::tui::auto_review::AutoReviewContext::from_tool_call( |
| 1334 | tool_name, |
| 1335 | tool_input, |
| 1336 | run_origin, |
| 1337 | approval_mode, |
| 1338 | workspace_trusted, |
| 1339 | workspace, |
| 1340 | ); |
| 1341 | auto_review_plan_decision_for_context(policy, &context) |
| 1342 | } |
| 1343 | |
| 1344 | #[test] |
| 1345 | fn auto_review_scenario() { |
| 1346 | // Scenario consolidation of: auto_review_classifies_publish_and_holds_without_prompting, auto_review_classifier_allow_executes_without_prompting, auto_review_allows_ordinary_shell_probe_without_prompting, auto_review_routes_unknown_tool_to_reviewer_in_auto, auto_review_policy_blocks_publish_when_approval_is_never, auto_review_allows_ordinary_test_command_without_prompting, auto_review_allows_ordinary_workspace_write_without_prompting, auto_review_routes_unbounded_or_sensitive_workspace_writes_to_reviewer |
| 1347 | // from auto_review_classifies_publish_and_holds_without_prompting |
| 1348 | { |
| 1349 | let (decision, audit) = auto_review_plan_decision( |
| 1350 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1351 | "exec_shell", |
| 1352 | &json!({"command": "git push origin main"}), |
| 1353 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1354 | ApprovalMode::Auto, |
| 1355 | true, |
| 1356 | None, |
| 1357 | ); |
| 1358 | |
| 1359 | assert_eq!( |
| 1360 | decision, |
| 1361 | AutoReviewPlanDecision::Block( |
| 1362 | "Built-in safety gate requires approval: publish-like action requires durable review" |
| 1363 | .to_string() |
| 1364 | ) |
| 1365 | ); |
| 1366 | assert_eq!(audit["action_kind"], "publish"); |
| 1367 | assert_eq!(audit["decision"], "hold_for_review"); |
| 1368 | } |
| 1369 | // from auto_review_classifier_allow_executes_without_prompting |
| 1370 | { |
| 1371 | let (decision, audit) = auto_review_plan_decision( |
| 1372 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1373 | "read_file", |
| 1374 | &json!({"path": "Cargo.toml"}), |
| 1375 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1376 | ApprovalMode::Auto, |
| 1377 | true, |
| 1378 | None, |
| 1379 | ); |
| 1380 | |
| 1381 | assert_eq!(decision, AutoReviewPlanDecision::Allow); |
| 1382 | assert_eq!(audit["decision"], "allow"); |
| 1383 | } |
| 1384 | // from auto_review_allows_ordinary_shell_probe_without_prompting |
| 1385 | { |
| 1386 | let (decision, audit) = auto_review_plan_decision( |
| 1387 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1388 | "exec_shell", |
| 1389 | &json!({"command": "git remote -v && git rev-parse --show-toplevel && git branch --show-current && git rev-parse HEAD && git tag --list 'v0.8.65'"}), |
| 1390 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1391 | ApprovalMode::Auto, |
| 1392 | true, |
| 1393 | None, |
| 1394 | ); |
| 1395 | |
| 1396 | assert_eq!(decision, AutoReviewPlanDecision::Allow); |
| 1397 | assert_eq!(audit["decision"], "allow"); |
| 1398 | assert_eq!(audit["action_kind"], "shell"); |
| 1399 | } |
| 1400 | // from auto_review_routes_unknown_tool_to_reviewer_in_auto |
| 1401 | { |
| 1402 | let (decision, audit) = auto_review_plan_decision( |
| 1403 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1404 | "mystery_tool", |
| 1405 | &json!({"value": true}), |
| 1406 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1407 | ApprovalMode::Auto, |
| 1408 | true, |
| 1409 | None, |
| 1410 | ); |
| 1411 | |
| 1412 | assert_eq!( |
| 1413 | decision, |
| 1414 | AutoReviewPlanDecision::ConsultReviewer( |
| 1415 | "unknown tool category requires explicit review".to_string() |
| 1416 | ) |
| 1417 | ); |
| 1418 | assert_eq!(audit["decision"], "ask_user"); |
| 1419 | } |
| 1420 | // from auto_review_policy_blocks_publish_when_approval_is_never |
| 1421 | { |
| 1422 | let (decision, audit) = auto_review_plan_decision( |
| 1423 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1424 | "github_publish_release", |
| 1425 | &json!({"tag": "v0.8.64"}), |
| 1426 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1427 | ApprovalMode::Never, |
| 1428 | true, |
| 1429 | None, |
| 1430 | ); |
| 1431 | |
| 1432 | assert_eq!( |
| 1433 | decision, |
| 1434 | AutoReviewPlanDecision::Block( |
| 1435 | "Built-in safety gate requires approval: publish-like action requires durable review" |
| 1436 | .to_string() |
| 1437 | ) |
| 1438 | ); |
| 1439 | assert_eq!(audit["approval_mode"], "NEVER"); |
| 1440 | assert_eq!(audit["decision"], "hold_for_review"); |
| 1441 | } |
| 1442 | // from auto_review_allows_ordinary_test_command_without_prompting |
| 1443 | { |
| 1444 | let (decision, audit) = auto_review_plan_decision( |
| 1445 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1446 | "exec_shell", |
| 1447 | &json!({"command": "cargo test"}), |
| 1448 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1449 | ApprovalMode::Auto, |
| 1450 | true, |
| 1451 | None, |
| 1452 | ); |
| 1453 | |
| 1454 | assert_eq!(decision, AutoReviewPlanDecision::Allow); |
| 1455 | assert_eq!(audit["decision"], "allow"); |
| 1456 | assert_eq!(audit["risk"], "destructive"); |
| 1457 | } |
| 1458 | // from auto_review_allows_ordinary_workspace_write_without_prompting |
| 1459 | { |
| 1460 | let tmp = tempdir().expect("tempdir"); |
| 1461 | std::fs::create_dir(tmp.path().join(".git")).expect("git marker"); |
| 1462 | std::fs::create_dir(tmp.path().join("src")).expect("source directory"); |
| 1463 | let (decision, audit) = auto_review_plan_decision( |
| 1464 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1465 | "write_file", |
| 1466 | &json!({"path": "src/lib.rs", "content": "pub fn ready() {}\n"}), |
| 1467 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1468 | ApprovalMode::Auto, |
| 1469 | true, |
| 1470 | Some(tmp.path()), |
| 1471 | ); |
| 1472 | |
| 1473 | assert_eq!(decision, AutoReviewPlanDecision::Allow); |
| 1474 | assert_eq!(audit["decision"], "allow"); |
| 1475 | assert_eq!(audit["action_kind"], "write"); |
| 1476 | } |
| 1477 | // from auto_review_routes_unbounded_or_sensitive_workspace_writes_to_reviewer |
| 1478 | { |
| 1479 | let tmp = tempdir().expect("tempdir"); |
| 1480 | std::fs::create_dir(tmp.path().join(".git")).expect("git marker"); |
| 1481 | for path in ["../outside.rs", "/etc/hostname", ".env", ".git/config"] { |
| 1482 | let (decision, audit) = auto_review_plan_decision( |
| 1483 | &crate::tui::auto_review::AutoReviewPolicy::default(), |
| 1484 | "write_file", |
| 1485 | &json!({"path": path, "content": "blocked"}), |
| 1486 | crate::tui::auto_review::RunOrigin::Interactive, |
| 1487 | ApprovalMode::Auto, |
| 1488 | true, |
| 1489 | Some(tmp.path()), |
| 1490 | ); |
| 1491 | assert!( |
| 1492 | matches!(decision, AutoReviewPlanDecision::ConsultReviewer(_)), |
| 1493 | "Auto-Review must not auto-approve {path} without reviewer judgment" |
| 1494 | ); |
| 1495 | assert_eq!(audit["decision"], "ask_user", "unexpected audit for {path}"); |
| 1496 | } |
| 1497 | } |
| 1498 | } |
| 1499 | |
| 1500 | #[test] |
| 1501 | fn repo_law_asks_only_in_ask_posture() { |
| 1502 | use ApprovalMode; |
| 1503 | |
| 1504 | assert!(!repo_law_must_block_without_prompt( |
| 1505 | ApprovalMode::Suggest, |
| 1506 | false |
| 1507 | )); |
| 1508 | for mode in [ |
| 1509 | ApprovalMode::Auto, |
| 1510 | ApprovalMode::Never, |
| 1511 | ApprovalMode::Bypass, |
| 1512 | ] { |
| 1513 | assert!( |
| 1514 | repo_law_must_block_without_prompt(mode, false), |
| 1515 | "{} must not open a human repo-law approval", |
| 1516 | mode.permission_chip_label() |
| 1517 | ); |
| 1518 | } |
| 1519 | assert!(repo_law_must_block_without_prompt( |
| 1520 | ApprovalMode::Suggest, |
| 1521 | true |
| 1522 | )); |
| 1523 | } |
| 1524 | |
| 1525 | #[test] |
| 1526 | fn rlm_eval_required_approval_is_auto_approved_in_full_access() { |
| 1527 | assert!(!registered_tool_approval_required( |
| 1528 | "rlm_eval", |
| 1529 | ApprovalRequirement::Required, |
| 1530 | true |
| 1531 | )); |
| 1532 | } |
| 1533 | |
| 1534 | /// A remembered grant for a Computer Use consent must not answer a later, |
| 1535 | /// identical call: the prompt is forced, so only a card decides it. |
| 1536 | #[test] |
| 1537 | fn computer_use_decisions_always_force_the_card() { |
| 1538 | let consent = serde_json::json!({"app": "Safari", "bundle_id": "com.apple.Safari"}); |
| 1539 | for name in [ |
| 1540 | "mcp_codewhale-cu_consent_allow", |
| 1541 | "mcp_codewhale-cu_consent_revoke", |
| 1542 | ] { |
| 1543 | assert!( |
| 1544 | call_forces_prompt(name, &consent, ApprovalRequirement::Required), |
| 1545 | "{name}" |
| 1546 | ); |
| 1547 | } |
| 1548 | assert!(call_forces_prompt( |
| 1549 | "mcp_codewhale-cu_app_script", |
| 1550 | &serde_json::json!({"script": "tell application \"Finder\" to activate"}), |
| 1551 | ApprovalRequirement::Required, |
| 1552 | )); |
| 1553 | assert!(!call_forces_prompt( |
| 1554 | "mcp_codewhale-cu_consent_status", |
| 1555 | &serde_json::json!({}), |
| 1556 | ApprovalRequirement::Required, |
| 1557 | )); |
| 1558 | let ask = crate::core::authority::TurnAuthority::from_effective_fields( |
| 1559 | AppMode::Agent, |
| 1560 | true, |
| 1561 | false, |
| 1562 | false, |
| 1563 | ApprovalMode::Suggest, |
| 1564 | ); |
| 1565 | assert_eq!( |
| 1566 | crate::core::authority::resolve_approval_request_disposition( |
| 1567 | &ask, true, false, true, false |
| 1568 | ), |
| 1569 | crate::core::authority::ApprovalRequestDisposition::Prompt, |
| 1570 | "a session grant must not pre-answer a forced prompt" |
| 1571 | ); |
| 1572 | } |
| 1573 | |
| 1574 | #[test] |
| 1575 | fn non_bypassable_registered_tools_auto_approve_in_full_access() { |
| 1576 | // #3866 reversed (owner decision, 2026-08-10): Full Access already grants |
| 1577 | // everything these calls can do — shell included — so a hold that cannot |
| 1578 | // open its own approval modal auto-approves instead of stranding the |
| 1579 | // call. Ask, which can open the modal, still gates every one of these. |
| 1580 | // Registry launcher is host-constructed and cache-bound (no free-form |
| 1581 | // command), so Full Access auto-approves it: `--auto` automation must |
| 1582 | // be able to complete the discovery flow end to end. Ask still gates it. |
| 1583 | assert!(!registered_tool_approval_required( |
| 1584 | "start_registry_mcp_server", |
| 1585 | ApprovalRequirement::Required, |
| 1586 | true |
| 1587 | )); |
| 1588 | assert!(registered_tool_approval_required( |
| 1589 | "start_registry_mcp_server", |
| 1590 | ApprovalRequirement::Required, |
| 1591 | false |
| 1592 | )); |
| 1593 | assert!(!registered_tool_approval_required( |
| 1594 | "start_mcp_server", |
| 1595 | ApprovalRequirement::Required, |
| 1596 | true |
| 1597 | )); |
| 1598 | assert!(!registered_tool_approval_required( |
| 1599 | "rlm_eval", |
| 1600 | ApprovalRequirement::Required, |
| 1601 | true |
| 1602 | )); |
| 1603 | assert!(registered_tool_forces_prompt( |
| 1604 | "start_mcp_server", |
| 1605 | ApprovalRequirement::Required, |
| 1606 | )); |
| 1607 | assert!(!registered_tool_forces_prompt( |
| 1608 | "start_registry_mcp_server", |
| 1609 | ApprovalRequirement::Required, |
| 1610 | )); |
| 1611 | assert!(registered_tool_forces_prompt( |
| 1612 | "rlm_eval", |
| 1613 | ApprovalRequirement::Required, |
| 1614 | )); |
| 1615 | assert!(!registered_tool_approval_required( |
| 1616 | "exec_shell", |
| 1617 | ApprovalRequirement::Required, |
| 1618 | true |
| 1619 | )); |
| 1620 | assert!( |
| 1621 | registered_tool_approval_required("start_mcp_server", ApprovalRequirement::Required, false), |
| 1622 | "start_mcp_server must require approval when auto_approve is disabled" |
| 1623 | ); |
| 1624 | // Sanity contrast: an ordinary Required tool is bypassable under auto-approve. |
| 1625 | assert!(!registered_tool_approval_required( |
| 1626 | "exec_shell", |
| 1627 | ApprovalRequirement::Required, |
| 1628 | true |
| 1629 | )); |
| 1630 | } |
| 1631 |