| 1 | /// A fence is admitted through the same planning as a `code_execution` |
| 2 | /// call, so an Auto-Review block rule on `code_execution` also stops the |
| 3 | /// fence, even under Full Access where no card would open. |
| 4 | #[tokio::test] |
| 5 | async fn repl_fence_obeys_auto_review_block_under_full_access() { |
| 6 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 7 | use codewhale_models::{ContentBlock, Message}; |
| 8 | |
| 9 | let workspace = tempdir().expect("tempdir"); |
| 10 | let marker = workspace.path().join("fence-ran"); |
| 11 | let fence = format!( |
| 12 | "```repl\nopen({:?}, 'w').write('x')\nfinalize('done')\n```", |
| 13 | marker.display().to_string() |
| 14 | ); |
| 15 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 16 | canned::simple_text_turn(&fence), |
| 17 | canned::simple_text_turn("Done."), |
| 18 | ])); |
| 19 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 20 | let mut config = deterministic_engine_config(workspace.path()); |
| 21 | // Built through the config entry point, as production does. |
| 22 | config.auto_review_policy = Config { |
| 23 | auto_review: Some(crate::config::AutoReviewConfig { |
| 24 | block: vec![crate::config::AutoReviewRuleConfig { |
| 25 | id: Some("no-code".to_string()), |
| 26 | tool: Some(CODE_EXECUTION_TOOL_NAME.to_string()), |
| 27 | reason: Some("code is blocked here".to_string()), |
| 28 | ..Default::default() |
| 29 | }], |
| 30 | ..Default::default() |
| 31 | }), |
| 32 | ..Config::default() |
| 33 | } |
| 34 | .auto_review_policy(); |
| 35 | let (mut engine, handle) = Engine::new_with_model_client(config, &Config::default(), client); |
| 36 | engine.session.auto_approve = true; |
| 37 | engine.session.add_message(Message { |
| 38 | role: Role::User, |
| 39 | content: vec![ContentBlock::Text { |
| 40 | text: "Compute.".to_string(), |
| 41 | cache_control: None, |
| 42 | }], |
| 43 | }); |
| 44 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 45 | workspace.path().to_path_buf(), |
| 46 | )); |
| 47 | let policy = test_tool_surface( |
| 48 | &engine, |
| 49 | registry, |
| 50 | Some(vec![catalog_tool(CODE_EXECUTION_TOOL_NAME)]), |
| 51 | AppMode::Agent, |
| 52 | ); |
| 53 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 54 | let (status, error) = engine.run_turn(&mut turn, policy, None, None).await; |
| 55 | |
| 56 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 57 | assert!( |
| 58 | engine.repl_kernel.is_none(), |
| 59 | "blocked fence starts no kernel" |
| 60 | ); |
| 61 | assert!(!marker.exists(), "blocked fence must not run"); |
| 62 | let note = { |
| 63 | let mut rx = handle.rx_event.write().await; |
| 64 | std::iter::from_fn(|| rx.try_recv().ok()).any(|event| { |
| 65 | matches!(event, Event::Status { message } |
| 66 | if message.starts_with("REPL block not run:") && message.contains("code is blocked here")) |
| 67 | }) |
| 68 | }; |
| 69 | assert!(note, "the block reason is shown"); |
| 70 | } |
| 71 | |
| 72 | /// Under Auto-Review a fence goes through the same review as a |
| 73 | /// `code_execution` call instead of an approval card that the posture can |
| 74 | /// only auto-deny: with the reviewer allowing it, the fence runs. |
| 75 | #[tokio::test] |
| 76 | async fn repl_fence_runs_through_auto_review_without_a_card() { |
| 77 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 78 | use codewhale_models::{ContentBlock, Message}; |
| 79 | |
| 80 | let workspace = tempdir().expect("tempdir"); |
| 81 | let marker = workspace.path().join("fence-ran"); |
| 82 | let fence = format!( |
| 83 | "```repl\nopen({:?}, 'w').write('x')\nfinalize('done')\n```", |
| 84 | marker.display().to_string() |
| 85 | ); |
| 86 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 87 | canned::simple_text_turn(&fence), |
| 88 | canned::simple_text_turn("Done."), |
| 89 | ])); |
| 90 | mock.push_message_response(guardian_fixture_response( |
| 91 | r#"{"risk_level":"low","decision":"allow","reason":"isolated fixture write"}"#, |
| 92 | )); |
| 93 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 94 | let (mut engine, handle) = Engine::new_with_model_client( |
| 95 | deterministic_engine_config(workspace.path()), |
| 96 | &Config::default(), |
| 97 | client, |
| 98 | ); |
| 99 | engine.session.auto_approve = false; |
| 100 | engine.session.approval_mode = ApprovalMode::Auto; |
| 101 | engine.session.add_message(Message { |
| 102 | role: Role::User, |
| 103 | content: vec![ContentBlock::Text { |
| 104 | text: "Compute.".to_string(), |
| 105 | cache_control: None, |
| 106 | }], |
| 107 | }); |
| 108 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 109 | workspace.path().to_path_buf(), |
| 110 | )); |
| 111 | let policy = test_tool_surface( |
| 112 | &engine, |
| 113 | registry, |
| 114 | Some(vec![catalog_tool(CODE_EXECUTION_TOOL_NAME)]), |
| 115 | AppMode::Agent, |
| 116 | ); |
| 117 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 118 | let (status, error) = tokio::time::timeout( |
| 119 | Duration::from_secs(30), |
| 120 | engine.run_turn(&mut turn, policy, None, None), |
| 121 | ) |
| 122 | .await |
| 123 | .expect("an Auto-Review fence must not wait on a card"); |
| 124 | |
| 125 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 126 | let asked = { |
| 127 | let mut rx = handle.rx_event.write().await; |
| 128 | std::iter::from_fn(|| rx.try_recv().ok()) |
| 129 | .any(|event| matches!(event, Event::ApprovalRequired { .. })) |
| 130 | }; |
| 131 | assert!(!asked, "Auto-Review opens no card for a reviewed fence"); |
| 132 | assert!(marker.exists(), "the reviewed fence runs"); |
| 133 | } |
| 134 | |
| 135 | async fn snapshot_for_catalog( |
| 136 | workspace: &Path, |
| 137 | catalog: Option<Vec<codewhale_models::Tool>>, |
| 138 | ) -> crate::tool_inspection::ToolInspectionSnapshot { |
| 139 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 140 | |
| 141 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![canned::simple_text_turn("Done.")])); |
| 142 | let client: crate::core::model_client::SharedModelClient = mock; |
| 143 | let (mut engine, handle) = Engine::new_with_model_client( |
| 144 | deterministic_engine_config(workspace), |
| 145 | &Config::default(), |
| 146 | client, |
| 147 | ); |
| 148 | let registry = |
| 149 | crate::tools::ToolRegistry::new(crate::tools::ToolContext::new(workspace.to_path_buf())); |
| 150 | let surface = test_tool_surface(&engine, registry, catalog, AppMode::Agent); |
| 151 | let mut turn = crate::core::turn::TurnContext::new(2); |
| 152 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 153 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 154 | let mut events = handle.rx_event.write().await; |
| 155 | std::iter::from_fn(|| events.try_recv().ok()) |
| 156 | .find_map(|event| match event { |
| 157 | Event::ToolRequestSnapshot { snapshot } => Some(snapshot), |
| 158 | _ => None, |
| 159 | }) |
| 160 | .expect("request snapshot") |
| 161 | } |
| 162 | |
| 163 | #[tokio::test] |
| 164 | async fn request_selector_distinguishes_absent_tools_from_present_empty_tools() { |
| 165 | let workspace = tempdir().expect("tempdir"); |
| 166 | let absent = snapshot_for_catalog(workspace.path(), None).await; |
| 167 | let deferred_only = codewhale_models::Tool { |
| 168 | tool_type: Some("function".to_string()), |
| 169 | name: "deferred_fixture".to_string(), |
| 170 | description: "Deferred fixture".to_string(), |
| 171 | input_schema: json!({"type": "object"}), |
| 172 | allowed_callers: None, |
| 173 | defer_loading: Some(true), |
| 174 | input_examples: None, |
| 175 | strict: None, |
| 176 | cache_control: None, |
| 177 | }; |
| 178 | let selected = active_tools_for_request(&[deferred_only], &HashSet::new(), false); |
| 179 | let present_empty = crate::tool_inspection::ToolInspectionSnapshot::from_prepared_request( |
| 180 | "turn", |
| 181 | 0, |
| 182 | selected.as_deref(), |
| 183 | ); |
| 184 | |
| 185 | assert!(!absent.tools_field_present); |
| 186 | assert_eq!(absent.tool_count, 0); |
| 187 | assert!(present_empty.tools_field_present); |
| 188 | assert_eq!(present_empty.tool_count, 0); |
| 189 | assert_eq!(present_empty.payload_json_bytes, Some(2)); |
| 190 | } |
| 191 | |
| 192 | #[tokio::test] |
| 193 | async fn terminal_diagnostics_distinguish_narration_from_missing_protocol_tool_calls() { |
| 194 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 195 | use crate::tool_inspection::TurnStopReason; |
| 196 | |
| 197 | for finish in ["end_turn", "tool_calls", "tool_use"] { |
| 198 | let workspace = tempdir().expect("tempdir"); |
| 199 | let response = vec![ |
| 200 | canned::message_start("stop_diagnostic_fixture"), |
| 201 | canned::text_block_start(0), |
| 202 | canned::text_delta(0, "I will edit the file. 我现在修改文件。"), |
| 203 | canned::block_stop(0), |
| 204 | canned::message_delta( |
| 205 | finish, |
| 206 | Some(Usage { |
| 207 | input_tokens: 100, |
| 208 | output_tokens: 12, |
| 209 | reasoning_tokens: Some(4), |
| 210 | prompt_cache_hit_tokens: Some(80), |
| 211 | prompt_cache_miss_tokens: Some(20), |
| 212 | ..Usage::default() |
| 213 | }), |
| 214 | ), |
| 215 | canned::message_stop(), |
| 216 | ]; |
| 217 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![response])); |
| 218 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 219 | let (mut engine, handle) = Engine::new_with_model_client( |
| 220 | deterministic_engine_config(workspace.path()), |
| 221 | &Config::default(), |
| 222 | client, |
| 223 | ); |
| 224 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 225 | workspace.path().to_path_buf(), |
| 226 | )); |
| 227 | let surface = test_tool_surface(&engine, registry, None, AppMode::Agent); |
| 228 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 229 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 230 | let expected = if finish == "end_turn" { |
| 231 | assert!( |
| 232 | error.is_none(), |
| 233 | "ordinary narration is not a protocol error" |
| 234 | ); |
| 235 | TurnStopReason::ProviderNoToolCall |
| 236 | } else { |
| 237 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 238 | assert!( |
| 239 | error |
| 240 | .as_deref() |
| 241 | .is_some_and(|error| error.contains("supplied no tool call")) |
| 242 | ); |
| 243 | TurnStopReason::ProviderToolCallMissing |
| 244 | }; |
| 245 | let snapshot = turn |
| 246 | .terminal_request_snapshot(status) |
| 247 | .expect("terminal request snapshot"); |
| 248 | let terminal = snapshot.terminal.as_ref().expect("terminal facts"); |
| 249 | assert_eq!(terminal.reason, Some(expected)); |
| 250 | assert_eq!(terminal.model_requests_started, 1); |
| 251 | assert_eq!(terminal.last_reported_input_tokens, Some(100)); |
| 252 | assert_eq!(terminal.last_response_tool_calls, Some(0)); |
| 253 | assert_eq!(terminal.last_response_tool_calls_suppressed, Some(0)); |
| 254 | assert_eq!( |
| 255 | terminal.last_provider_finish_reason.as_ref().unwrap().value, |
| 256 | finish |
| 257 | ); |
| 258 | assert_eq!( |
| 259 | mock.captured_requests().len(), |
| 260 | 1, |
| 261 | "narration must not synthesize continuation" |
| 262 | ); |
| 263 | assert_eq!(turn.usage.input_tokens, 100); |
| 264 | assert_eq!( |
| 265 | turn.usage.output_tokens, 12, |
| 266 | "reasoning is an output subset, not additional output" |
| 267 | ); |
| 268 | assert!(snapshot.render_text().contains("Terminal diagnostics")); |
| 269 | let json = serde_json::to_value(&snapshot).expect("serialize snapshot"); |
| 270 | assert_eq!(json["terminal"]["model_requests_started"], 1); |
| 271 | let mut events = handle.rx_event.write().await; |
| 272 | assert!( |
| 273 | !std::iter::from_fn(|| events.try_recv().ok()) |
| 274 | .any(|event| matches!(event, Event::ToolCallStarted { .. })) |
| 275 | ); |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | #[tokio::test] |
| 280 | async fn terminal_diagnostics_merge_cumulative_usage_within_one_request() { |
| 281 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 282 | |
| 283 | let workspace = tempdir().expect("tempdir"); |
| 284 | let usage = Usage { |
| 285 | input_tokens: 100, |
| 286 | output_tokens: 12, |
| 287 | reasoning_tokens: Some(4), |
| 288 | prompt_cache_hit_tokens: Some(80), |
| 289 | prompt_cache_miss_tokens: Some(20), |
| 290 | ..Usage::default() |
| 291 | }; |
| 292 | let mut start = canned::message_start("cumulative_usage_fixture"); |
| 293 | if let StreamEvent::MessageStart { message } = &mut start { |
| 294 | message.usage = usage.clone(); |
| 295 | } |
| 296 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![vec![ |
| 297 | start, |
| 298 | canned::text_block_start(0), |
| 299 | canned::text_delta(0, "Done."), |
| 300 | canned::block_stop(0), |
| 301 | canned::message_delta("end_turn", Some(usage.clone())), |
| 302 | canned::message_delta("end_turn", Some(usage.clone())), |
| 303 | canned::message_stop(), |
| 304 | ]])); |
| 305 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 306 | let (mut engine, handle) = Engine::new_with_model_client( |
| 307 | deterministic_engine_config(workspace.path()), |
| 308 | &Config::default(), |
| 309 | client, |
| 310 | ); |
| 311 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 312 | workspace.path().to_path_buf(), |
| 313 | )); |
| 314 | let surface = test_tool_surface(&engine, registry, None, AppMode::Agent); |
| 315 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 316 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 317 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 318 | assert_eq!(mock.captured_requests().len(), 1); |
| 319 | assert_eq!( |
| 320 | turn.usage, usage, |
| 321 | "repeated cumulative receipts are counted once" |
| 322 | ); |
| 323 | let mut events = handle.rx_event.write().await; |
| 324 | let usages = std::iter::from_fn(|| events.try_recv().ok()) |
| 325 | .filter_map(|event| match event { |
| 326 | Event::TurnUsage { usage, .. } => Some(usage), |
| 327 | _ => None, |
| 328 | }) |
| 329 | .collect::<Vec<_>>(); |
| 330 | assert_eq!(usages, vec![usage]); |
| 331 | } |
| 332 | |
| 333 | #[tokio::test] |
| 334 | async fn request_snapshots_advance_to_the_latest_tool_step() { |
| 335 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 336 | |
| 337 | let workspace = tempdir().expect("tempdir"); |
| 338 | fs::write(workspace.path().join("README.md"), "fixture\n").expect("write fixture"); |
| 339 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 340 | canned::tool_call_turn("call-read", "read_file", r#"{"path":"README.md"}"#), |
| 341 | canned::simple_text_turn("Done."), |
| 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(4); |
| 355 | |
| 356 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 357 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 358 | let mut events = handle.rx_event.write().await; |
| 359 | let snapshots = std::iter::from_fn(|| events.try_recv().ok()) |
| 360 | .filter_map(|event| match event { |
| 361 | Event::ToolRequestSnapshot { snapshot } => Some(snapshot), |
| 362 | _ => None, |
| 363 | }) |
| 364 | .collect::<Vec<_>>(); |
| 365 | |
| 366 | assert_eq!(snapshots.len(), 2); |
| 367 | assert_eq!(snapshots[0].step, 0); |
| 368 | assert_eq!(snapshots[1].step, 1); |
| 369 | assert_eq!(snapshots[1].turn_id.value, turn.id); |
| 370 | } |
| 371 | |
| 372 | #[tokio::test] |
| 373 | async fn tool_result_followed_by_terminal_empty_assistant_fails_turn() { |
| 374 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 375 | |
| 376 | let workspace = tempdir().expect("tempdir"); |
| 377 | fs::write(workspace.path().join("README.md"), "fixture\n").expect("write fixture"); |
| 378 | let empty_terminal_turn = vec![ |
| 379 | canned::message_start("mock_empty_after_tool"), |
| 380 | canned::message_delta("stop", None), |
| 381 | canned::message_stop(), |
| 382 | ]; |
| 383 | // #6310: an answerless clean stop is retried (exact prefix, then nudged) |
| 384 | // before the turn fails, so the fixture stays empty for every attempt. |
| 385 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 386 | canned::tool_call_turn("call-read", "read_file", r#"{"path":"README.md"}"#), |
| 387 | empty_terminal_turn.clone(), |
| 388 | empty_terminal_turn.clone(), |
| 389 | empty_terminal_turn, |
| 390 | ])); |
| 391 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 392 | let (mut engine, handle) = Engine::new_with_model_client( |
| 393 | deterministic_engine_config(workspace.path()), |
| 394 | &Config::default(), |
| 395 | client, |
| 396 | ); |
| 397 | let context = crate::tools::ToolContext::new(workspace.path().to_path_buf()); |
| 398 | let mut registry = crate::tools::ToolRegistry::new(context); |
| 399 | registry.register(std::sync::Arc::new(crate::tools::file::ReadFileTool)); |
| 400 | let tools = Some(registry.to_api_tools_with_cache(true)); |
| 401 | let surface = test_tool_surface(&engine, registry, tools, AppMode::Agent); |
| 402 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 403 | |
| 404 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 405 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 406 | assert_eq!( |
| 407 | mock.call_count(), |
| 408 | 4, |
| 409 | "tool step, empty provider step, then exactly two bounded retries" |
| 410 | ); |
| 411 | assert_eq!(turn.stop_diagnostics.empty_stop_retries, 2); |
| 412 | assert!( |
| 413 | error |
| 414 | .as_deref() |
| 415 | .is_some_and(|message| message.contains("terminal stop reason `stop`") |
| 416 | && message.contains("after 2 retries")), |
| 417 | "terminal empty response must produce a precise failure: {error:?}" |
| 418 | ); |
| 419 | |
| 420 | let mut events = handle.rx_event.write().await; |
| 421 | let events = std::iter::from_fn(|| events.try_recv().ok()).collect::<Vec<_>>(); |
| 422 | assert!( |
| 423 | events.iter().any(|event| matches!( |
| 424 | event, |
| 425 | Event::ToolCallComplete { model_call: Some(model_call), result, .. } |
| 426 | if model_call.provider_id == "call-read" && result.is_ok() |
| 427 | )), |
| 428 | "the successful tool result must remain durable: {events:?}" |
| 429 | ); |
| 430 | assert!( |
| 431 | events |
| 432 | .iter() |
| 433 | .any(|event| matches!(event, Event::Error { .. })), |
| 434 | "the empty provider response must be visible as an error: {events:?}" |
| 435 | ); |
| 436 | assert!( |
| 437 | engine |
| 438 | .session |
| 439 | .messages |
| 440 | .iter() |
| 441 | .all(|message| { message.role != Role::Assistant || !message.content.is_empty() }), |
| 442 | "the engine must not fabricate an empty assistant message" |
| 443 | ); |
| 444 | } |
| 445 | |
| 446 | fn empty_clean_stop_turn() -> Vec<StreamEvent> { |
| 447 | use crate::llm_client::mock::canned; |
| 448 | vec![ |
| 449 | canned::message_start("mock_empty_clean_stop"), |
| 450 | canned::message_delta("stop", None), |
| 451 | canned::message_stop(), |
| 452 | ] |
| 453 | } |
| 454 | |
| 455 | async fn run_empty_stop_fixture( |
| 456 | turns: Vec<Vec<StreamEvent>>, |
| 457 | ) -> ( |
| 458 | std::sync::Arc<crate::llm_client::mock::MockLlmClient>, |
| 459 | Engine, |
| 460 | crate::core::turn::TurnContext, |
| 461 | TurnOutcomeStatus, |
| 462 | Option<String>, |
| 463 | Vec<Event>, |
| 464 | ) { |
| 465 | let workspace = tempdir().expect("tempdir"); |
| 466 | let mock = std::sync::Arc::new(crate::llm_client::mock::MockLlmClient::new(turns)); |
| 467 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 468 | let (mut engine, handle) = Engine::new_with_model_client( |
| 469 | deterministic_engine_config(workspace.path()), |
| 470 | &Config::default(), |
| 471 | client, |
| 472 | ); |
| 473 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 474 | workspace.path().to_path_buf(), |
| 475 | )); |
| 476 | let surface = test_tool_surface(&engine, registry, None, AppMode::Agent); |
| 477 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 478 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 479 | let mut rx = handle.rx_event.write().await; |
| 480 | let events = std::iter::from_fn(|| rx.try_recv().ok()).collect::<Vec<_>>(); |
| 481 | (mock, engine, turn, status, error, events) |
| 482 | } |
| 483 | |
| 484 | /// #6310: one clean `stop` with no text, reasoning or tool call is retried |
| 485 | /// with the identical request and the turn completes on the real answer. |
| 486 | #[tokio::test] |
| 487 | async fn empty_clean_stop_is_retried_once_and_the_turn_completes() { |
| 488 | use crate::llm_client::mock::canned; |
| 489 | |
| 490 | let (mock, engine, turn, status, error, events) = run_empty_stop_fixture(vec![ |
| 491 | empty_clean_stop_turn(), |
| 492 | canned::simple_text_turn("the recovered answer"), |
| 493 | ]) |
| 494 | .await; |
| 495 | |
| 496 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 497 | assert_eq!(mock.call_count(), 2, "exactly one retry"); |
| 498 | assert_eq!(turn.stop_diagnostics.empty_stop_retries, 1); |
| 499 | let requests = mock.captured_requests(); |
| 500 | assert_eq!( |
| 501 | requests[0].messages.len(), |
| 502 | requests[1].messages.len(), |
| 503 | "the first retry is an exact-prefix re-request" |
| 504 | ); |
| 505 | let transcript = |
| 506 | serde_json::to_string(&engine.session.messages.iter().collect::<Vec<_>>()).unwrap(); |
| 507 | assert_eq!(transcript.matches("the recovered answer").count(), 1); |
| 508 | assert!( |
| 509 | engine |
| 510 | .session |
| 511 | .messages |
| 512 | .iter() |
| 513 | .all(|message| message.role != Role::Assistant || !message.content.is_empty()), |
| 514 | "the empty response must not be persisted" |
| 515 | ); |
| 516 | let retry_receipts = events |
| 517 | .iter() |
| 518 | .filter_map(|event| match event { |
| 519 | Event::Status { message } if message.starts_with("Retry attempt: empty-stop ") => { |
| 520 | Some(message) |
| 521 | } |
| 522 | _ => None, |
| 523 | }) |
| 524 | .collect::<Vec<_>>(); |
| 525 | assert_eq!(retry_receipts.len(), 1); |
| 526 | for (index, receipt) in retry_receipts.iter().enumerate() { |
| 527 | assert!(receipt.starts_with(&format!("Retry attempt: empty-stop {}/2;", index + 1))); |
| 528 | } |
| 529 | assert_eq!(events.iter().filter(|event| matches!(event, Event::Status { message } if message == "Retry recovery: empty-stop used 1/2 retries; turn completed")).count(), 1); |
| 530 | } |
| 531 | |
| 532 | /// #6310: the second retry carries the request-scoped nudge, which never |
| 533 | /// joins the session; the retry after that budget is not attempted. |
| 534 | #[tokio::test] |
| 535 | async fn empty_clean_stop_second_retry_is_nudged_and_never_persisted() { |
| 536 | use crate::llm_client::mock::canned; |
| 537 | |
| 538 | let (mock, engine, turn, status, error, events) = run_empty_stop_fixture(vec![ |
| 539 | empty_clean_stop_turn(), |
| 540 | empty_clean_stop_turn(), |
| 541 | canned::simple_text_turn("answer after nudge"), |
| 542 | ]) |
| 543 | .await; |
| 544 | |
| 545 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 546 | assert_eq!(mock.call_count(), 3); |
| 547 | assert_eq!(turn.stop_diagnostics.empty_stop_retries, 2); |
| 548 | let requests = mock.captured_requests(); |
| 549 | let nudge = crate::config::DEFAULT_REASONING_ONLY_REPROMPT_MESSAGE; |
| 550 | let carries_nudge = |request: &codewhale_models::MessageRequest| { |
| 551 | serde_json::to_string(&request.messages) |
| 552 | .unwrap() |
| 553 | .contains(nudge) |
| 554 | }; |
| 555 | assert!(!carries_nudge(&requests[0])); |
| 556 | assert!(!carries_nudge(&requests[1]), "first retry is exact-prefix"); |
| 557 | assert!(carries_nudge(&requests[2]), "second retry is nudged"); |
| 558 | assert_eq!(requests[2].messages.len(), requests[0].messages.len() + 1); |
| 559 | assert!( |
| 560 | !serde_json::to_string(&engine.session.messages.iter().collect::<Vec<_>>()) |
| 561 | .unwrap() |
| 562 | .contains(nudge), |
| 563 | "the nudge is request-scoped and never written to the session" |
| 564 | ); |
| 565 | let retry_receipts = events |
| 566 | .iter() |
| 567 | .filter_map(|event| match event { |
| 568 | Event::Status { message } if message.starts_with("Retry attempt: empty-stop ") => { |
| 569 | Some(message) |
| 570 | } |
| 571 | _ => None, |
| 572 | }) |
| 573 | .collect::<Vec<_>>(); |
| 574 | assert_eq!(retry_receipts.len(), 2); |
| 575 | for (index, receipt) in retry_receipts.iter().enumerate() { |
| 576 | assert!(receipt.starts_with(&format!("Retry attempt: empty-stop {}/2;", index + 1))); |
| 577 | } |
| 578 | assert_eq!(events.iter().filter(|event| matches!(event, Event::Status { message } if message == "Retry recovery: empty-stop used 2/2 retries; turn completed")).count(), 1); |
| 579 | } |
| 580 | |
| 581 | /// #6310: an empty response on every attempt fails visibly once the budget |
| 582 | /// is spent, with the retries recorded in stop diagnostics. |
| 583 | #[tokio::test] |
| 584 | async fn empty_clean_stop_every_time_fails_after_the_retry_budget() { |
| 585 | let (mock, _engine, turn, status, error, events) = run_empty_stop_fixture(vec![ |
| 586 | empty_clean_stop_turn(), |
| 587 | empty_clean_stop_turn(), |
| 588 | empty_clean_stop_turn(), |
| 589 | ]) |
| 590 | .await; |
| 591 | |
| 592 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 593 | assert_eq!( |
| 594 | mock.call_count(), |
| 595 | 1 + crate::core::engine::turn_loop::EMPTY_STOP_MAX_RETRIES as usize |
| 596 | ); |
| 597 | assert_eq!( |
| 598 | turn.stop_diagnostics.empty_stop_retries, |
| 599 | crate::core::engine::turn_loop::EMPTY_STOP_MAX_RETRIES |
| 600 | ); |
| 601 | assert!( |
| 602 | error |
| 603 | .as_deref() |
| 604 | .is_some_and(|message| message.contains("terminal stop reason `stop`") |
| 605 | && message.contains("after 2 retries")), |
| 606 | "{error:?}" |
| 607 | ); |
| 608 | let retry_receipts = events |
| 609 | .iter() |
| 610 | .filter_map(|event| match event { |
| 611 | Event::Status { message } if message.starts_with("Retry attempt: empty-stop ") => { |
| 612 | Some(message) |
| 613 | } |
| 614 | _ => None, |
| 615 | }) |
| 616 | .collect::<Vec<_>>(); |
| 617 | assert_eq!(retry_receipts.len(), 2); |
| 618 | for (index, receipt) in retry_receipts.iter().enumerate() { |
| 619 | assert!(receipt.starts_with(&format!("Retry attempt: empty-stop {}/2;", index + 1))); |
| 620 | } |
| 621 | assert_eq!(events.iter().filter(|event| matches!(event, Event::Status { message } if message == "Retry stopped: empty-stop used 2/2 retries; turn failed")).count(), 1); |
| 622 | } |
| 623 | |
| 624 | #[tokio::test] |
| 625 | async fn request_snapshot_reports_registry_provenance_for_the_transmitted_catalog() { |
| 626 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 627 | |
| 628 | let workspace = tempdir().expect("tempdir"); |
| 629 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![canned::simple_text_turn("Done.")])); |
| 630 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 631 | // Pin `read_file` loaded so this step's request actually carries it; the |
| 632 | // point of the test is what a *transmitted* tool is reported as. |
| 633 | let mut engine_config = deterministic_engine_config(workspace.path()); |
| 634 | engine_config.tools_always_load = HashSet::from(["read_file".to_string()]); |
| 635 | let (mut engine, handle) = |
| 636 | Engine::new_with_model_client(engine_config, &Config::default(), client); |
| 637 | let context = crate::tools::ToolContext::new(workspace.path().to_path_buf()); |
| 638 | let mut registry = crate::tools::ToolRegistry::new(context); |
| 639 | registry.register(std::sync::Arc::new(crate::tools::file::ReadFileTool)); |
| 640 | // `read_file` is a hidden compatibility alias, so `to_api_tools` would hand |
| 641 | // the turn an empty catalog and nothing would be transmitted. Hand the |
| 642 | // engine an explicit catalog instead: the registry is still the source of |
| 643 | // the *facts*, including the fact that this tool is not model-visible. |
| 644 | let tools = Some(vec![codewhale_models::Tool { |
| 645 | tool_type: None, |
| 646 | name: "read_file".to_string(), |
| 647 | description: "Read a file".to_string(), |
| 648 | input_schema: json!({"type": "object"}), |
| 649 | allowed_callers: Some(vec!["direct".to_string()]), |
| 650 | defer_loading: Some(false), |
| 651 | input_examples: None, |
| 652 | strict: None, |
| 653 | cache_control: None, |
| 654 | }]); |
| 655 | |
| 656 | // The same surface context `handle_send_message` resolves for a real turn: |
| 657 | // real registry facts, real (empty) MCP attribution, the engine's own |
| 658 | // synthetic-name list, and the resolved model client's receipt. |
| 659 | let synthetic_names = super::tool_catalog::default_synthetic_catalog_tool_names(); |
| 660 | let surface = crate::tool_inspection::ToolSurfaceContext { |
| 661 | registry: registry.registry_facts(&HashSet::new()), |
| 662 | mcp_servers: std::collections::BTreeMap::new(), |
| 663 | synthetic_names: synthetic_names.clone(), |
| 664 | provider: engine.tool_surface_provider_receipt(), |
| 665 | }; |
| 666 | let policy = test_tool_surface(&engine, registry, tools, AppMode::Agent); |
| 667 | |
| 668 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 669 | let (status, error) = engine |
| 670 | .run_turn(&mut turn, policy, None, Some(surface)) |
| 671 | .await; |
| 672 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 673 | |
| 674 | let mut events = handle.rx_event.write().await; |
| 675 | let snapshot = std::iter::from_fn(|| events.try_recv().ok()) |
| 676 | .find_map(|event| match event { |
| 677 | Event::ToolRequestSnapshot { snapshot } => Some(snapshot), |
| 678 | _ => None, |
| 679 | }) |
| 680 | .expect("request snapshot"); |
| 681 | |
| 682 | let transmitted = mock |
| 683 | .last_request() |
| 684 | .expect("captured request") |
| 685 | .tools |
| 686 | .unwrap_or_default(); |
| 687 | assert!(!transmitted.is_empty(), "the turn must carry tools"); |
| 688 | |
| 689 | // The digest is the request path's, over what was prepared. |
| 690 | assert_eq!( |
| 691 | snapshot.active_tool_catalog_sha256.as_deref(), |
| 692 | Some(crate::core::engine::preview::active_tool_catalog_sha256(&transmitted).as_str()) |
| 693 | ); |
| 694 | |
| 695 | // Provenance is registry-derived truth, not "unavailable". |
| 696 | assert!(snapshot.registry_facts_present); |
| 697 | assert!(snapshot.provider.is_available()); |
| 698 | assert_eq!( |
| 699 | snapshot.unavailable_for_this_request, |
| 700 | vec!["provider_wire_payload"] |
| 701 | ); |
| 702 | |
| 703 | let read_file = snapshot |
| 704 | .tools |
| 705 | .iter() |
| 706 | .find(|entry| entry.name.value == "read_file") |
| 707 | .expect("read_file projected"); |
| 708 | assert_eq!( |
| 709 | read_file.provenance, |
| 710 | crate::tool_inspection::Evidence::Known { |
| 711 | value: crate::tool_inspection::ToolProvenance::Builtin |
| 712 | } |
| 713 | ); |
| 714 | assert!(matches!( |
| 715 | read_file.approval, |
| 716 | crate::tool_inspection::Evidence::Known { .. } |
| 717 | )); |
| 718 | // Registry truth, not an inference from the request: this alias is hidden |
| 719 | // from the model catalog even though this catalog carried it explicitly. |
| 720 | assert_eq!( |
| 721 | read_file.model_visible, |
| 722 | crate::tool_inspection::Evidence::Known { value: false } |
| 723 | ); |
| 724 | assert!(read_file.visibility.in_request()); |
| 725 | |
| 726 | // Anything the engine injected rather than registered is reported as |
| 727 | // synthetic, from the engine's own list — never guessed from the name. |
| 728 | for entry in &snapshot.tools { |
| 729 | if synthetic_names.contains(&entry.name.value) { |
| 730 | assert_eq!( |
| 731 | entry.provenance, |
| 732 | crate::tool_inspection::Evidence::Known { |
| 733 | value: crate::tool_inspection::ToolProvenance::Synthetic |
| 734 | }, |
| 735 | "'{}' is engine-injected, not registry-backed", |
| 736 | entry.name.value |
| 737 | ); |
| 738 | // Not in the registry, so capabilities are unknown, not "none". |
| 739 | assert!(matches!( |
| 740 | entry.capabilities, |
| 741 | crate::tool_inspection::Evidence::Unknown { .. } |
| 742 | )); |
| 743 | } |
| 744 | } |
| 745 | } |
| 746 | |
| 747 | fn deterministic_engine_config(workspace: &Path) -> EngineConfig { |
| 748 | EngineConfig { |
| 749 | workspace: workspace.to_path_buf(), |
| 750 | snapshots_enabled: false, |
| 751 | subagents_enabled: false, |
| 752 | ..EngineConfig::default() |
| 753 | } |
| 754 | } |
| 755 | |
| 756 | /// The compaction budget is measured against the real system prompt, and the |
| 757 | /// skills block in that prompt is discovered from the developer's home as well |
| 758 | /// as from the workspace. Left ambient, this test counts whatever skills the |
| 759 | /// machine happens to have installed into its token budget: 39 of them trip a |
| 760 | /// seventh compaction pass on a developer box while CI, with an empty home, |
| 761 | /// sees six and passes. That is the #5359 leak class, and the isolated home is |
| 762 | /// what makes `deterministic_engine_config` actually deterministic here. |
| 763 | #[test] |
| 764 | fn automatic_compaction_continues_one_task_and_suppresses_failed_passes() { |
| 765 | let _env = lock_test_env(); |
| 766 | let home = tempdir().unwrap(); |
| 767 | let _codewhale_home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 768 | let _user_home = EnvVarGuard::set("HOME", home.path()); |
| 769 | let _user_profile = EnvVarGuard::set("USERPROFILE", home.path()); |
| 770 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 771 | .enable_all() |
| 772 | .build() |
| 773 | .unwrap(); |
| 774 | runtime.block_on(async { |
| 775 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 776 | for fail_summary in [false, true] { |
| 777 | let workspace = tempdir().unwrap(); |
| 778 | fs::write( |
| 779 | workspace.path().join("README.md"), |
| 780 | "verified fixture evidence", |
| 781 | ) |
| 782 | .unwrap(); |
| 783 | let mock = std::sync::Arc::new(MockLlmClient::new(Vec::new())); |
| 784 | for step in 0..16 { |
| 785 | mock.push_turn(vec![ |
| 786 | canned::message_start(&format!("response-{step}")), |
| 787 | canned::text_block_start(0), |
| 788 | canned::text_delta(0, &format!("Step {step}: {}", "x".repeat(32_000))), |
| 789 | canned::block_stop(0), |
| 790 | canned::tool_use_block_start(1, &format!("read-{step}"), "File"), |
| 791 | canned::tool_input_delta(1, r#"{"action":"read","path":"README.md"}"#), |
| 792 | canned::block_stop(1), |
| 793 | canned::message_delta("tool_use", None), |
| 794 | canned::message_stop(), |
| 795 | ]); |
| 796 | } |
| 797 | mock.push_turn(canned::simple_text_turn( |
| 798 | "All sixteen reads verified; task complete.", |
| 799 | )); |
| 800 | for checkpoint in 0..8 { |
| 801 | let content = if fail_summary { |
| 802 | json!([{"type":"tool_use","id":"unexpected","name":"File","input":{}}]) |
| 803 | } else { |
| 804 | json!([{"type":"text","text":format!("Current objective: complete all sixteen reads. Checkpoint {checkpoint}: earlier reads verified. Preserve the user's no-publication constraint. Continue the remaining File reads, then report the observed evidence.")}]) |
| 805 | }; |
| 806 | mock.push_message_response(serde_json::from_value(json!({ |
| 807 | "id":format!("summary-{checkpoint}"), "type":"message", "role":"assistant", |
| 808 | "content":content, "model":"mock-model", "usage":{"input_tokens":0,"output_tokens":0} |
| 809 | })).unwrap()); |
| 810 | } |
| 811 | let config = Config::default(); |
| 812 | let (engine, handle) = Engine::new_with_model_client( |
| 813 | deterministic_engine_config(workspace.path()), |
| 814 | &config, |
| 815 | mock.clone(), |
| 816 | ); |
| 817 | let task = tokio::spawn(engine.run()); |
| 818 | let mut op = external_user_message_op( |
| 819 | "Complete all sixteen reads; do not publish.", |
| 820 | AppMode::Agent, |
| 821 | &config, |
| 822 | ); |
| 823 | if let Op::SendMessage(TurnSpec { |
| 824 | compaction, |
| 825 | auto_approve, |
| 826 | .. |
| 827 | }) = &mut op |
| 828 | { |
| 829 | compaction.token_threshold = 40_000; |
| 830 | *auto_approve = true; |
| 831 | } |
| 832 | handle.send(op).await.unwrap(); |
| 833 | let mut completed = 0; |
| 834 | let mut failed = 0; |
| 835 | { |
| 836 | let mut rx = handle.rx_event.write().await; |
| 837 | loop { |
| 838 | match tokio::time::timeout(Duration::from_secs(30), rx.recv()) |
| 839 | .await |
| 840 | .unwrap() |
| 841 | .unwrap() |
| 842 | { |
| 843 | Event::CompactionCompleted { auto: true, .. } => completed += 1, |
| 844 | Event::CompactionFailed { auto: true, .. } => failed += 1, |
| 845 | Event::TurnComplete { status, error, .. } => { |
| 846 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 847 | break; |
| 848 | } |
| 849 | _ => {} |
| 850 | } |
| 851 | } |
| 852 | } |
| 853 | let requests = mock.captured_requests(); |
| 854 | let streaming = requests |
| 855 | .iter() |
| 856 | .filter(|r| r.stream == Some(true)) |
| 857 | .collect::<Vec<_>>(); |
| 858 | assert_eq!( |
| 859 | streaming.len(), |
| 860 | 17, |
| 861 | "one user request must continue through all tool steps" |
| 862 | ); |
| 863 | if fail_summary { |
| 864 | assert_eq!( |
| 865 | (completed, failed), |
| 866 | (0, 1), |
| 867 | "failed compaction must not loop at every tool boundary" |
| 868 | ); |
| 869 | } else { |
| 870 | assert!( |
| 871 | (2..=6).contains(&completed), |
| 872 | "expected repeated useful compaction: {completed}" |
| 873 | ); |
| 874 | assert_eq!(failed, 0); |
| 875 | } |
| 876 | for request in &requests { |
| 877 | assert_eq!( |
| 878 | request.system, streaming[0].system, |
| 879 | "the stable system prefix must survive every pass" |
| 880 | ); |
| 881 | assert_eq!( |
| 882 | request.tools, streaming[0].tools, |
| 883 | "summarizing must reuse the tool prefix" |
| 884 | ); |
| 885 | if request.stream == Some(false) { |
| 886 | assert_eq!(request.tool_choice, Some(json!("none"))); |
| 887 | } |
| 888 | let mut calls = HashSet::new(); |
| 889 | for message in &request.messages { |
| 890 | for block in &message.content { |
| 891 | match block { |
| 892 | ContentBlock::ToolUse { id, .. } => { |
| 893 | calls.insert(id); |
| 894 | } |
| 895 | ContentBlock::ToolResult { tool_use_id, .. } => assert!( |
| 896 | calls.contains(tool_use_id), |
| 897 | "orphan tool result after compaction" |
| 898 | ), |
| 899 | _ => {} |
| 900 | } |
| 901 | } |
| 902 | } |
| 903 | } |
| 904 | let snapshot = handle.get_session_snapshot().await.unwrap(); |
| 905 | assert!(snapshot.messages.iter().any(|m| m.content.iter().any(|b| matches!(b, ContentBlock::Text {text,..} if text.contains("All sixteen reads verified"))))); |
| 906 | handle.send(Op::Shutdown).await.unwrap(); |
| 907 | task.await.unwrap(); |
| 908 | } |
| 909 | }); |
| 910 | } |
| 911 | |
| 912 | #[tokio::test] |
| 913 | async fn initial_routed_usage_is_total_only_emitted_once_and_keeps_parent_route_separate() { |
| 914 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 915 | |
| 916 | let _cost_scope = crate::cost_status::test_scope(); |
| 917 | let workspace = tempdir().expect("tempdir"); |
| 918 | let parent_usage = Usage { |
| 919 | input_tokens: 11, |
| 920 | output_tokens: 3, |
| 921 | ..Usage::default() |
| 922 | }; |
| 923 | let classifier_usage = Usage { |
| 924 | input_tokens: 7, |
| 925 | output_tokens: 5, |
| 926 | ..Usage::default() |
| 927 | }; |
| 928 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![vec![ |
| 929 | canned::message_start("parent-response"), |
| 930 | canned::text_block_start(0), |
| 931 | canned::text_delta(0, "done"), |
| 932 | canned::block_stop(0), |
| 933 | canned::message_delta("end_turn", Some(parent_usage.clone())), |
| 934 | canned::message_stop(), |
| 935 | ]])); |
| 936 | let client: crate::core::model_client::SharedModelClient = mock; |
| 937 | let api_config = Config::default(); |
| 938 | let (engine, handle) = Engine::new_with_model_client( |
| 939 | deterministic_engine_config(workspace.path()), |
| 940 | &api_config, |
| 941 | client, |
| 942 | ); |
| 943 | let task = tokio::spawn(engine.run()); |
| 944 | |
| 945 | let mut op = external_user_message_op("account for classifier", AppMode::Agent, &api_config); |
| 946 | let Op::SendMessage(TurnSpec { |
| 947 | initial_routed_usage, |
| 948 | .. |
| 949 | }) = &mut op |
| 950 | else { |
| 951 | unreachable!("external_user_message_op always builds SendMessage"); |
| 952 | }; |
| 953 | let mut missing_usage_route = crate::cost_status::EffectiveRouteEnvelope::capture( |
| 954 | None, |
| 955 | ProviderKind::Openai, |
| 956 | "openai", |
| 957 | "classifier-model", |
| 958 | Some(ProviderKind::Openai.provider().default_base_url()), |
| 959 | chrono::Utc::now(), |
| 960 | ); |
| 961 | missing_usage_route.billing_mode = crate::cost_status::RouteBillingMode::Metered; |
| 962 | **initial_routed_usage = crate::cost_status::RuntimeUsageBatch { |
| 963 | decisions: Vec::new(), |
| 964 | records: vec![crate::cost_status::RuntimeUsageRecord { |
| 965 | source_id: "auto-router:engine-fixture".to_string(), |
| 966 | usage: crate::cost_status::EffectiveRouteUsage { |
| 967 | route: crate::cost_status::EffectiveRouteEnvelope::capture( |
| 968 | None, |
| 969 | ProviderKind::Openai, |
| 970 | "openai", |
| 971 | "classifier-model", |
| 972 | Some(ProviderKind::Openai.provider().default_base_url()), |
| 973 | chrono::Utc::now(), |
| 974 | ), |
| 975 | usage: classifier_usage.clone(), |
| 976 | }, |
| 977 | }], |
| 978 | drop_records: vec![crate::cost_status::RuntimeUsageDropRecord { |
| 979 | reason: crate::cost_status::RuntimeUsageMissingReason::default(), |
| 980 | source_id: "auto-router:engine-fixture:missing-usage".to_string(), |
| 981 | route: missing_usage_route, |
| 982 | }], |
| 983 | // One exact route-aware missing receipt plus two residual gaps whose |
| 984 | // route identity was truncated upstream. |
| 985 | dropped_records: 3, |
| 986 | }; |
| 987 | handle.send(op).await.expect("send routed-usage turn"); |
| 988 | |
| 989 | let mut routed_events = Vec::new(); |
| 990 | let mut rx = handle.rx_event.write().await; |
| 991 | let (total_usage, terminal_parent_usage, dropped_records) = loop { |
| 992 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 993 | .await |
| 994 | .expect("timed out waiting for routed-usage turn") |
| 995 | .expect("engine event stream closed"); |
| 996 | match event { |
| 997 | Event::RoutedTurnUsage { usage, .. } => routed_events.push(usage), |
| 998 | Event::TurnComplete { |
| 999 | usage, |
| 1000 | parent_route_usage, |
| 1001 | routed_usage_dropped_records, |
| 1002 | status, |
| 1003 | error, |
| 1004 | .. |
| 1005 | } => { |
| 1006 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1007 | break (usage, parent_route_usage, routed_usage_dropped_records); |
| 1008 | } |
| 1009 | _ => {} |
| 1010 | } |
| 1011 | }; |
| 1012 | drop(rx); |
| 1013 | |
| 1014 | assert_eq!(routed_events, vec![classifier_usage.clone()]); |
| 1015 | assert_eq!(terminal_parent_usage, parent_usage); |
| 1016 | assert_eq!(total_usage.input_tokens, 18); |
| 1017 | assert_eq!(total_usage.output_tokens, 8); |
| 1018 | assert_eq!(dropped_records, 2); |
| 1019 | let initial_cost = crate::cost_status::drain(); |
| 1020 | assert!( |
| 1021 | initial_cost.usage_source_fingerprints.contains( |
| 1022 | &crate::cost_status::usage_source_fingerprint( |
| 1023 | "auto-router:engine-fixture:missing-usage" |
| 1024 | ) |
| 1025 | ), |
| 1026 | "exact missing-usage response identity was not settled" |
| 1027 | ); |
| 1028 | assert!( |
| 1029 | initial_cost |
| 1030 | .unpriced_reasons |
| 1031 | .contains("provider_success_missing_usage"), |
| 1032 | "metered missing-usage route was not marked incomplete" |
| 1033 | ); |
| 1034 | |
| 1035 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1036 | task.await.expect("engine task"); |
| 1037 | } |
| 1038 | |
| 1039 | #[tokio::test] |
| 1040 | async fn isolated_runtime_chat_provider_request_contains_no_host_context_or_tools() { |
| 1041 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1042 | |
| 1043 | const PROJECT_CANARY: &str = "PRIVATE_PROJECT_AGENTS_CANARY"; |
| 1044 | const MEMORY_CANARY: &str = "PRIVATE_USER_MEMORY_CANARY"; |
| 1045 | const SKILL_CANARY: &str = "PRIVATE_SKILL_CANARY"; |
| 1046 | const MCP_CANARY: &str = "PRIVATE_MCP_CANARY"; |
| 1047 | const INSTRUCTION_CANARY: &str = "PRIVATE_INSTRUCTION_CANARY"; |
| 1048 | const ATTACHMENT_CANARY: &str = "PRIVATE_ATTACHMENT_BYTES_CANARY"; |
| 1049 | |
| 1050 | let root = tempdir().expect("private Runtime Chat root"); |
| 1051 | let workspace = root.path().join("private-host-workspace-canary"); |
| 1052 | let skills = root.path().join("private-skills-canary"); |
| 1053 | fs::create_dir_all(&workspace).expect("create private workspace"); |
| 1054 | fs::create_dir_all(&skills).expect("create private skills"); |
| 1055 | fs::write(workspace.join("AGENTS.md"), PROJECT_CANARY).expect("write AGENTS canary"); |
| 1056 | let memory = root.path().join("private-memory.md"); |
| 1057 | fs::write(&memory, MEMORY_CANARY).expect("write memory canary"); |
| 1058 | fs::write(skills.join("SKILL.md"), SKILL_CANARY).expect("write skill canary"); |
| 1059 | let mcp = root.path().join("private-mcp.json"); |
| 1060 | fs::write(&mcp, format!(r#"{{"server":"{MCP_CANARY}"}}"#)).expect("write MCP canary"); |
| 1061 | let instructions = root.path().join("private-instructions.md"); |
| 1062 | fs::write(&instructions, INSTRUCTION_CANARY).expect("write instruction canary"); |
| 1063 | let attachment = root.path().join("private-attachment.png"); |
| 1064 | fs::write(&attachment, ATTACHMENT_CANARY).expect("write attachment canary"); |
| 1065 | |
| 1066 | let api_config = Config { |
| 1067 | runtime_chat_isolated: true, |
| 1068 | ..Config::default() |
| 1069 | }; |
| 1070 | let mut engine_config = deterministic_engine_config(&workspace); |
| 1071 | engine_config.instructions = vec![instructions.into()]; |
| 1072 | engine_config.project_context_pack_enabled = true; |
| 1073 | engine_config.memory_enabled = true; |
| 1074 | engine_config.memory_path = memory; |
| 1075 | engine_config.skills_dir = skills; |
| 1076 | engine_config.mcp_config_path = mcp; |
| 1077 | engine_config.subagents_enabled = true; |
| 1078 | engine_config.allowed_tools = None; |
| 1079 | |
| 1080 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![canned::simple_text_turn( |
| 1081 | "Hello from isolated Chat.", |
| 1082 | )])); |
| 1083 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1084 | let (engine, handle) = Engine::new_with_model_client(engine_config, &api_config, client); |
| 1085 | let task = tokio::spawn(engine.run()); |
| 1086 | handle |
| 1087 | .send(external_user_message_op( |
| 1088 | &format!("Say hello.\n[Attached image: {}]", attachment.display()), |
| 1089 | AppMode::Agent, |
| 1090 | &api_config, |
| 1091 | )) |
| 1092 | .await |
| 1093 | .expect("send isolated Chat turn"); |
| 1094 | |
| 1095 | let mut rx = handle.rx_event.write().await; |
| 1096 | loop { |
| 1097 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 1098 | .await |
| 1099 | .expect("isolated Chat turn timed out") |
| 1100 | .expect("isolated Chat event stream closed"); |
| 1101 | if let Event::TurnComplete { status, error, .. } = event { |
| 1102 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1103 | break; |
| 1104 | } |
| 1105 | } |
| 1106 | |
| 1107 | let request = mock.last_request().expect("captured provider request"); |
| 1108 | assert!( |
| 1109 | request.tools.as_ref().is_none_or(Vec::is_empty), |
| 1110 | "isolated Chat must expose no provider tools" |
| 1111 | ); |
| 1112 | // #6517: the engine and the Runtime Chat relay once carried two different |
| 1113 | // isolated-chat prompts. The engine sends the one shared constant; the |
| 1114 | // relay's own test pins `dedicated_chat_system_prompt(None)` to the same |
| 1115 | // constant, so core tests need no edge into the relay (runtime ratchet). |
| 1116 | let system = match request.system.as_ref() { |
| 1117 | Some(SystemPrompt::Text(text)) => text.clone(), |
| 1118 | other => panic!("isolated Chat should send one text system prompt: {other:?}"), |
| 1119 | }; |
| 1120 | assert_eq!(system, ISOLATED_CHAT_SYSTEM_PROMPT); |
| 1121 | let serialized = serde_json::to_string(&request).expect("serialize captured request"); |
| 1122 | assert!(serialized.contains("Say hello."), "{serialized}"); |
| 1123 | assert!(serialized.contains("Attachment omitted"), "{serialized}"); |
| 1124 | for forbidden in [ |
| 1125 | workspace.to_string_lossy().as_ref(), |
| 1126 | root.path().to_string_lossy().as_ref(), |
| 1127 | PROJECT_CANARY, |
| 1128 | MEMORY_CANARY, |
| 1129 | SKILL_CANARY, |
| 1130 | MCP_CANARY, |
| 1131 | INSTRUCTION_CANARY, |
| 1132 | ATTACHMENT_CANARY, |
| 1133 | "<turn_meta>", |
| 1134 | "<user_memory>", |
| 1135 | "<available_skills>", |
| 1136 | "Git workspace:", |
| 1137 | ] { |
| 1138 | assert!( |
| 1139 | !serialized.contains(forbidden), |
| 1140 | "leaked {forbidden}: {serialized}" |
| 1141 | ); |
| 1142 | } |
| 1143 | |
| 1144 | task.abort(); |
| 1145 | } |
| 1146 | |
| 1147 | #[tokio::test] |
| 1148 | async fn injected_model_drives_real_engine_navigation_trajectory() { |
| 1149 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1150 | |
| 1151 | let workspace = tempdir().expect("tempdir"); |
| 1152 | fs::write( |
| 1153 | workspace.path().join("README.md"), |
| 1154 | "navigation-seam-proof\n", |
| 1155 | ) |
| 1156 | .expect("write fixture"); |
| 1157 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 1158 | canned::tool_call_turn( |
| 1159 | "call-read", |
| 1160 | "File", |
| 1161 | r#"{"action":"read","path":"README.md"}"#, |
| 1162 | ), |
| 1163 | canned::simple_text_turn("Navigation complete."), |
| 1164 | ])); |
| 1165 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1166 | let (engine, handle) = Engine::new_with_model_client( |
| 1167 | deterministic_engine_config(workspace.path()), |
| 1168 | &Config::default(), |
| 1169 | client, |
| 1170 | ); |
| 1171 | let task = tokio::spawn(engine.run()); |
| 1172 | handle |
| 1173 | .send(external_user_message_op( |
| 1174 | "Read README.md and report what it contains.", |
| 1175 | AppMode::Agent, |
| 1176 | &Config::default(), |
| 1177 | )) |
| 1178 | .await |
| 1179 | .expect("send deterministic navigation turn"); |
| 1180 | |
| 1181 | let mut saw_read = false; |
| 1182 | let mut saw_answer = false; |
| 1183 | let mut saw_unreceipted_injected_route = false; |
| 1184 | let mut saw_unattributed_injected_completion = false; |
| 1185 | let mut rx = handle.rx_event.write().await; |
| 1186 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 1187 | .await |
| 1188 | .expect("timed out waiting for deterministic navigation") |
| 1189 | { |
| 1190 | match event { |
| 1191 | Event::TurnStarted { route, .. } => { |
| 1192 | let route = route.expect("injected model turn route"); |
| 1193 | assert!( |
| 1194 | route.receipt.is_none(), |
| 1195 | "an auxiliary route client must not receipt injected model I/O" |
| 1196 | ); |
| 1197 | saw_unreceipted_injected_route = true; |
| 1198 | } |
| 1199 | Event::ToolCallComplete { name, result, .. } if name == "File" => { |
| 1200 | let result = result.expect("File.read result"); |
| 1201 | assert!(result.success, "{result:?}"); |
| 1202 | assert!(result.content.contains("navigation-seam-proof")); |
| 1203 | saw_read = true; |
| 1204 | } |
| 1205 | Event::MessageDelta { content, .. } => { |
| 1206 | saw_answer |= content.contains("Navigation complete"); |
| 1207 | } |
| 1208 | Event::TurnComplete { |
| 1209 | status, |
| 1210 | error, |
| 1211 | base_url, |
| 1212 | .. |
| 1213 | } => { |
| 1214 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1215 | assert!( |
| 1216 | base_url.is_none(), |
| 1217 | "an auxiliary route client must not attribute an injected completion" |
| 1218 | ); |
| 1219 | saw_unattributed_injected_completion = true; |
| 1220 | break; |
| 1221 | } |
| 1222 | _ => {} |
| 1223 | } |
| 1224 | } |
| 1225 | drop(rx); |
| 1226 | assert!( |
| 1227 | saw_read, |
| 1228 | "real registry must execute the mock-requested read" |
| 1229 | ); |
| 1230 | assert!( |
| 1231 | saw_answer, |
| 1232 | "real stream projection must emit the final answer" |
| 1233 | ); |
| 1234 | assert!(saw_unreceipted_injected_route); |
| 1235 | assert!(saw_unattributed_injected_completion); |
| 1236 | assert_eq!(mock.call_count(), 2); |
| 1237 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1238 | task.await.expect("engine task"); |
| 1239 | } |
| 1240 | |
| 1241 | #[tokio::test] |
| 1242 | async fn injected_model_sandbox_escalation_applies_only_after_exact_call_approval() { |
| 1243 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1244 | |
| 1245 | const COMMAND: &str = "echo elevated > escalation.txt"; |
| 1246 | for (tool, source) in [ |
| 1247 | ("bash", json!({"command": COMMAND})), |
| 1248 | ( |
| 1249 | CODE_EXECUTION_TOOL_NAME, |
| 1250 | json!({"code": "open('escalation.txt', 'w').write('approved')"}), |
| 1251 | ), |
| 1252 | ( |
| 1253 | JS_EXECUTION_TOOL_NAME, |
| 1254 | json!({"code": "require('fs').writeFileSync('escalation.txt', 'approved')"}), |
| 1255 | ), |
| 1256 | ] { |
| 1257 | let mut input = source; |
| 1258 | input["sandbox_permissions"] = json!("workspace-write"); |
| 1259 | input["justification"] = json!("the exact command writes the requested workspace proof"); |
| 1260 | let workspace = tempdir().expect("tempdir"); |
| 1261 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 1262 | canned::tool_call_turn("call-escalated-bash", tool, &input.to_string()), |
| 1263 | canned::simple_text_turn("Escalated command complete."), |
| 1264 | ])); |
| 1265 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1266 | let config = Config { |
| 1267 | sandbox_mode: Some("read-only".to_string()), |
| 1268 | ..Config::default() |
| 1269 | }; |
| 1270 | let mut engine_config = deterministic_engine_config(workspace.path()); |
| 1271 | engine_config.exec_policy_engine = ask_rule_engine(COMMAND); |
| 1272 | let (engine, handle) = Engine::new_with_model_client(engine_config, &config, client); |
| 1273 | let task = tokio::spawn(engine.run()); |
| 1274 | handle |
| 1275 | .send(external_user_message_op( |
| 1276 | "Create the escalation proof after approval.", |
| 1277 | AppMode::Agent, |
| 1278 | &config, |
| 1279 | )) |
| 1280 | .await |
| 1281 | .expect("send escalation journey"); |
| 1282 | |
| 1283 | let mut approved_result = None; |
| 1284 | let mut rx = handle.rx_event.write().await; |
| 1285 | loop { |
| 1286 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 1287 | .await |
| 1288 | .expect("timed out waiting for escalation journey") |
| 1289 | .expect("engine event stream closed"); |
| 1290 | match event { |
| 1291 | Event::ApprovalRequired { |
| 1292 | id, |
| 1293 | input, |
| 1294 | description, |
| 1295 | .. |
| 1296 | } => { |
| 1297 | assert_eq!(input["sandbox_permissions"], "workspace-write"); |
| 1298 | assert!( |
| 1299 | description |
| 1300 | .contains("the exact command writes the requested workspace proof"), |
| 1301 | "{description}" |
| 1302 | ); |
| 1303 | if tool == "bash" { |
| 1304 | assert!( |
| 1305 | description.contains("Additional approval gate"), |
| 1306 | "the sandbox grant must not hide the typed ask rule: {description}" |
| 1307 | ); |
| 1308 | assert!( |
| 1309 | description.contains("Typed ask rule"), |
| 1310 | "the typed ask rule must not hide the sandbox grant: {description}" |
| 1311 | ); |
| 1312 | } |
| 1313 | assert!( |
| 1314 | !workspace.path().join("escalation.txt").exists(), |
| 1315 | "approval must happen before execution" |
| 1316 | ); |
| 1317 | handle |
| 1318 | .approve_tool_call(&id) |
| 1319 | .await |
| 1320 | .expect("approve exact escalated call"); |
| 1321 | } |
| 1322 | Event::ToolCallComplete { |
| 1323 | model_call: Some(model_call), |
| 1324 | result, |
| 1325 | .. |
| 1326 | } if model_call.provider_id == "call-escalated-bash" => { |
| 1327 | approved_result = Some(result.expect("approved escalation result")); |
| 1328 | } |
| 1329 | Event::TurnComplete { status, error, .. } => { |
| 1330 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1331 | break; |
| 1332 | } |
| 1333 | _ => {} |
| 1334 | } |
| 1335 | } |
| 1336 | drop(rx); |
| 1337 | |
| 1338 | let result = approved_result.expect("paired escalated tool result"); |
| 1339 | assert!(result.success, "{result:?}"); |
| 1340 | assert!( |
| 1341 | result |
| 1342 | .content |
| 1343 | .contains("approved by the user with an adjusted execution policy"), |
| 1344 | "{}", |
| 1345 | result.content |
| 1346 | ); |
| 1347 | assert!(workspace.path().join("escalation.txt").exists()); |
| 1348 | assert_eq!(mock.call_count(), 2); |
| 1349 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1350 | task.await.expect("engine task"); |
| 1351 | } |
| 1352 | } |
| 1353 | |
| 1354 | #[tokio::test] |
| 1355 | async fn sandbox_escalation_fails_closed_when_the_posture_cannot_prompt() { |
| 1356 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1357 | |
| 1358 | for (approval_mode, auto_approve, posture, expected_denial) in [ |
| 1359 | ( |
| 1360 | ApprovalMode::Auto, |
| 1361 | false, |
| 1362 | "Auto-Review", |
| 1363 | "Sandbox escalation requires a one-shot user approval", |
| 1364 | ), |
| 1365 | ( |
| 1366 | ApprovalMode::Suggest, |
| 1367 | true, |
| 1368 | "Full Access", |
| 1369 | "requires a one-shot user approval", |
| 1370 | ), |
| 1371 | ] { |
| 1372 | for tool in ["bash", CODE_EXECUTION_TOOL_NAME, JS_EXECUTION_TOOL_NAME] { |
| 1373 | let mut input = if tool == CODE_EXECUTION_TOOL_NAME { |
| 1374 | json!({"code": "open('escalation.txt', 'w').write('denied')"}) |
| 1375 | } else if tool == JS_EXECUTION_TOOL_NAME { |
| 1376 | json!({"code": "require('fs').writeFileSync('escalation.txt', 'denied')"}) |
| 1377 | } else { |
| 1378 | json!({"command": "echo denied > escalation.txt"}) |
| 1379 | }; |
| 1380 | input["sandbox_permissions"] = json!("workspace-write"); |
| 1381 | input["justification"] = json!("this execution needs workspace write access"); |
| 1382 | let workspace = tempdir().expect("tempdir"); |
| 1383 | let mock = std::sync::Arc::new(MockLlmClient::new(vec![ |
| 1384 | canned::tool_call_turn("call-unattended-escalation", tool, &input.to_string()), |
| 1385 | canned::simple_text_turn("Escalation was unavailable."), |
| 1386 | ])); |
| 1387 | if matches!(approval_mode, ApprovalMode::Auto) { |
| 1388 | // Let Auto-Review's independent guardian approve the bounded |
| 1389 | // fixture call so this test reaches the separate rule under test: |
| 1390 | // unattended postures still cannot mint a sandbox escalation. |
| 1391 | mock.push_message_response(guardian_fixture_response( |
| 1392 | r#"{"risk_level":"low","decision":"allow","reason":"isolated fixture write"}"#, |
| 1393 | )); |
| 1394 | } |
| 1395 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1396 | let config = Config { |
| 1397 | sandbox_mode: Some("read-only".to_string()), |
| 1398 | ..Config::default() |
| 1399 | }; |
| 1400 | let (engine, handle) = Engine::new_with_model_client( |
| 1401 | deterministic_engine_config(workspace.path()), |
| 1402 | &config, |
| 1403 | client, |
| 1404 | ); |
| 1405 | let task = tokio::spawn(engine.run()); |
| 1406 | let mut op = external_user_message_op( |
| 1407 | "Do not pause for an unattended escalation.", |
| 1408 | AppMode::Agent, |
| 1409 | &config, |
| 1410 | ); |
| 1411 | let Op::SendMessage(TurnSpec { |
| 1412 | approval_mode: op_approval_mode, |
| 1413 | auto_approve: op_auto_approve, |
| 1414 | .. |
| 1415 | }) = &mut op |
| 1416 | else { |
| 1417 | panic!("user message op") |
| 1418 | }; |
| 1419 | *op_approval_mode = approval_mode; |
| 1420 | *op_auto_approve = auto_approve; |
| 1421 | handle.send(op).await.expect("send unattended escalation"); |
| 1422 | |
| 1423 | let mut saw_denial = false; |
| 1424 | let mut rx = handle.rx_event.write().await; |
| 1425 | loop { |
| 1426 | let event = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 1427 | .await |
| 1428 | .expect("timed out waiting for unattended escalation") |
| 1429 | .expect("engine event stream closed"); |
| 1430 | match event { |
| 1431 | Event::ApprovalRequired { .. } => { |
| 1432 | panic!("{posture} must not open an escalation prompt") |
| 1433 | } |
| 1434 | Event::ToolCallComplete { |
| 1435 | model_call: Some(model_call), |
| 1436 | result, |
| 1437 | .. |
| 1438 | } if model_call.provider_id == "call-unattended-escalation" => { |
| 1439 | let error = result.expect_err("unattended escalation must be denied"); |
| 1440 | assert!(error.to_string().contains(expected_denial), "{error}"); |
| 1441 | assert!(error.to_string().contains(posture), "{error}"); |
| 1442 | saw_denial = true; |
| 1443 | } |
| 1444 | Event::TurnComplete { status, error, .. } => { |
| 1445 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1446 | break; |
| 1447 | } |
| 1448 | _ => {} |
| 1449 | } |
| 1450 | } |
| 1451 | drop(rx); |
| 1452 | |
| 1453 | assert!(saw_denial); |
| 1454 | assert!(!workspace.path().join("escalation.txt").exists()); |
| 1455 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1456 | task.await.expect("engine task"); |
| 1457 | } |
| 1458 | } |
| 1459 | } |
| 1460 | |
| 1461 | #[tokio::test] |
| 1462 | async fn productive_tool_results_do_not_hit_no_user_input_backstop() { |
| 1463 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1464 | |
| 1465 | const TOOL_ROUNDS: usize = 201; |
| 1466 | const FINAL_ANSWER: &str = "All productive tool rounds completed."; |
| 1467 | |
| 1468 | let workspace = tempdir().expect("tempdir"); |
| 1469 | let mut turns = Vec::with_capacity(TOOL_ROUNDS + 1); |
| 1470 | for index in 1..=TOOL_ROUNDS { |
| 1471 | let fixture = format!("fixture-{index}.txt"); |
| 1472 | fs::write( |
| 1473 | workspace.path().join(&fixture), |
| 1474 | format!("productive-round-{index}\n"), |
| 1475 | ) |
| 1476 | .expect("write distinct read fixture"); |
| 1477 | turns.push(canned::tool_call_turn( |
| 1478 | &format!("call-read-{index}"), |
| 1479 | "File", |
| 1480 | &format!(r#"{{"action":"read","path":"{fixture}"}}"#), |
| 1481 | )); |
| 1482 | } |
| 1483 | turns.push(canned::simple_text_turn(FINAL_ANSWER)); |
| 1484 | |
| 1485 | let mock = std::sync::Arc::new(MockLlmClient::new(turns)); |
| 1486 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1487 | let (engine, handle) = Engine::new_with_model_client( |
| 1488 | deterministic_engine_config(workspace.path()), |
| 1489 | &Config::default(), |
| 1490 | client, |
| 1491 | ); |
| 1492 | let task = tokio::spawn(engine.run()); |
| 1493 | handle |
| 1494 | .send(external_user_message_op( |
| 1495 | "Read every distinct fixture, then report completion.", |
| 1496 | AppMode::Agent, |
| 1497 | &Config::default(), |
| 1498 | )) |
| 1499 | .await |
| 1500 | .expect("send productive tool trajectory"); |
| 1501 | |
| 1502 | let mut successful_tool_ids = HashSet::new(); |
| 1503 | let mut saw_final_answer = false; |
| 1504 | let mut rx = handle.rx_event.write().await; |
| 1505 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 1506 | .await |
| 1507 | .expect("timed out waiting for productive tool trajectory") |
| 1508 | { |
| 1509 | match event { |
| 1510 | Event::ToolCallComplete { |
| 1511 | id, name, result, .. |
| 1512 | } if name == "File" => { |
| 1513 | let result = result.expect("File.read result"); |
| 1514 | assert!(result.success, "{id}: {result:?}"); |
| 1515 | assert!( |
| 1516 | successful_tool_ids.insert(id.clone()), |
| 1517 | "tool id completed twice: {id}" |
| 1518 | ); |
| 1519 | } |
| 1520 | Event::MessageDelta { content, .. } => { |
| 1521 | saw_final_answer |= content.contains(FINAL_ANSWER); |
| 1522 | } |
| 1523 | Event::TurnComplete { status, error, .. } => { |
| 1524 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1525 | assert_eq!( |
| 1526 | successful_tool_ids.len(), |
| 1527 | TOOL_ROUNDS, |
| 1528 | "turn completed before all productive tool rounds" |
| 1529 | ); |
| 1530 | assert_eq!( |
| 1531 | mock.call_count(), |
| 1532 | TOOL_ROUNDS + 1, |
| 1533 | "productive work must finish beyond the former 200-step default" |
| 1534 | ); |
| 1535 | assert!( |
| 1536 | saw_final_answer, |
| 1537 | "turn completed before the final assistant text arrived" |
| 1538 | ); |
| 1539 | break; |
| 1540 | } |
| 1541 | _ => {} |
| 1542 | } |
| 1543 | } |
| 1544 | drop(rx); |
| 1545 | |
| 1546 | assert_eq!(mock.remaining_turns(), 0, "the final turn must be consumed"); |
| 1547 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1548 | task.await.expect("engine task"); |
| 1549 | } |
| 1550 | |
| 1551 | #[tokio::test] |
| 1552 | async fn eight_identical_sequential_reads_all_execute_before_final_answer() { |
| 1553 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 1554 | |
| 1555 | const READ_ROUNDS: usize = 8; |
| 1556 | const FINAL_ANSWER: &str = "Repeated observation complete."; |
| 1557 | |
| 1558 | let workspace = tempdir().expect("tempdir"); |
| 1559 | fs::write(workspace.path().join("state.txt"), "stable-observation\n") |
| 1560 | .expect("write repeated-read fixture"); |
| 1561 | let mut turns = Vec::with_capacity(READ_ROUNDS + 1); |
| 1562 | for index in 1..=READ_ROUNDS { |
| 1563 | turns.push(canned::tool_call_turn( |
| 1564 | &format!("call-identical-read-{index}"), |
| 1565 | "File", |
| 1566 | r#"{"action":"read","path":"state.txt"}"#, |
| 1567 | )); |
| 1568 | } |
| 1569 | turns.push(canned::simple_text_turn(FINAL_ANSWER)); |
| 1570 | |
| 1571 | let mock = std::sync::Arc::new(MockLlmClient::new(turns)); |
| 1572 | let client: crate::core::model_client::SharedModelClient = mock.clone(); |
| 1573 | let (engine, handle) = Engine::new_with_model_client( |
| 1574 | deterministic_engine_config(workspace.path()), |
| 1575 | &Config::default(), |
| 1576 | client, |
| 1577 | ); |
| 1578 | let task = tokio::spawn(engine.run()); |
| 1579 | handle |
| 1580 | .send(external_user_message_op( |
| 1581 | "Read state.txt until you have enough evidence, then report completion.", |
| 1582 | AppMode::Agent, |
| 1583 | &Config::default(), |
| 1584 | )) |
| 1585 | .await |
| 1586 | .expect("send repeated-read trajectory"); |
| 1587 | |
| 1588 | let mut completed_ids = HashSet::new(); |
| 1589 | let mut saw_final_answer = false; |
| 1590 | let mut rx = handle.rx_event.write().await; |
| 1591 | while let Some(event) = tokio::time::timeout(model_turn_event_timeout(), rx.recv()) |
| 1592 | .await |
| 1593 | .expect("timed out waiting for repeated-read trajectory") |
| 1594 | { |
| 1595 | match event { |
| 1596 | Event::ToolCallComplete { |
| 1597 | id, name, result, .. |
| 1598 | } if name == "File" => { |
| 1599 | let result = result.expect("File.read result"); |
| 1600 | assert!(result.success, "{id}: {result:?}"); |
| 1601 | assert!(result.content.contains("stable-observation")); |
| 1602 | assert!(completed_ids.insert(id), "tool id completed twice"); |
| 1603 | } |
| 1604 | Event::MessageDelta { content, .. } => { |
| 1605 | saw_final_answer |= content.contains(FINAL_ANSWER); |
| 1606 | } |
| 1607 | Event::TurnComplete { status, error, .. } => { |
| 1608 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1609 | assert_eq!(completed_ids.len(), READ_ROUNDS); |
| 1610 | assert_eq!(mock.call_count(), READ_ROUNDS + 1); |
| 1611 | assert!(saw_final_answer); |
| 1612 | break; |
| 1613 | } |
| 1614 | _ => {} |
| 1615 | } |
| 1616 | } |
| 1617 | drop(rx); |
| 1618 | |
| 1619 | assert_eq!(mock.remaining_turns(), 0); |
| 1620 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1621 | task.await.expect("engine task"); |
| 1622 | } |
| 1623 |