| 1 | //! Cucumber acceptance test for the public LLM/tool lifecycle. |
| 2 | |
| 3 | use std::io::Read; |
| 4 | use std::process::{Command, Stdio}; |
| 5 | use std::time::Duration; |
| 6 | |
| 7 | use cucumber::{World as _, gherkin::Step, given, then, when, writer::Stats as _}; |
| 8 | use serde_json::{Value, json}; |
| 9 | use tempfile::TempDir; |
| 10 | use wait_timeout::ChildExt; |
| 11 | use wiremock::matchers::{method, path}; |
| 12 | use wiremock::{Mock, MockServer, Request, ResponseTemplate}; |
| 13 | |
| 14 | const FEATURE_NAME: &str = "Tool call lifecycle"; |
| 15 | const FEATURE_PATH: &str = concat!( |
| 16 | env!("CARGO_MANIFEST_DIR"), |
| 17 | "/tests/features/tool_lifecycle.feature" |
| 18 | ); |
| 19 | const HAPPY_PATH_SCENARIO: &str = "Happy path lists the current directory through a tool"; |
| 20 | const UNKNOWN_TOOL_SCENARIO: &str = "Unknown tool returns an error result"; |
| 21 | const MALFORMED_ARGUMENTS_SCENARIO: &str = "Malformed tool arguments return an error result"; |
| 22 | const REAL_TOOL_ERROR_SCENARIO: &str = "A real tool error is returned to the follow-up request"; |
| 23 | const EMPTY_TOOL_RESULT_SCENARIO: &str = |
| 24 | "An empty tool result is returned to the follow-up request"; |
| 25 | const MISSING_SUMMARY_SCENARIO: &str = |
| 26 | "A follow-up answer missing the expected summary is detected"; |
| 27 | const TOOL_CALL_ID: &str = "call_tool"; |
| 28 | const TEST_MODEL: &str = "acceptance-model"; |
| 29 | |
| 30 | #[derive(Debug, Default, cucumber::World)] |
| 31 | struct ToolLifecycleWorld { |
| 32 | workspace: Option<TempDir>, |
| 33 | home: Option<TempDir>, |
| 34 | llm_server: Option<MockServer>, |
| 35 | tool_name: Option<String>, |
| 36 | tool_arguments: Option<String>, |
| 37 | final_answer: Option<String>, |
| 38 | prompt: Option<String>, |
| 39 | stdout: String, |
| 40 | stderr: String, |
| 41 | events: Vec<Value>, |
| 42 | requests: Vec<Value>, |
| 43 | } |
| 44 | |
| 45 | #[given("an offline CodeWhale workspace containing:")] |
| 46 | fn offline_codewhale_workspace_containing(world: &mut ToolLifecycleWorld, step: &Step) { |
| 47 | let workspace = TempDir::new().expect("workspace tempdir"); |
| 48 | let home = TempDir::new().expect("home tempdir"); |
| 49 | |
| 50 | for row in data_table_rows(step) { |
| 51 | let relative_path = row_value(&row, "path"); |
| 52 | let kind = row_value(&row, "kind"); |
| 53 | let path = workspace.path().join(relative_path); |
| 54 | match kind.as_str() { |
| 55 | "file" => { |
| 56 | if let Some(parent) = path.parent() { |
| 57 | std::fs::create_dir_all(parent).expect("create workspace file parent"); |
| 58 | } |
| 59 | std::fs::write(&path, "").expect("write workspace file"); |
| 60 | } |
| 61 | "folder" => std::fs::create_dir_all(&path).expect("create workspace folder"), |
| 62 | other => panic!("unsupported workspace entry kind: {other}"), |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | world.workspace = Some(workspace); |
| 67 | world.home = Some(home); |
| 68 | } |
| 69 | |
| 70 | #[given(regex = r#"^the mocked LLM will request the "([^"]+)" tool with:$"#)] |
| 71 | fn mocked_llm_will_request_tool(world: &mut ToolLifecycleWorld, tool_name: String, step: &Step) { |
| 72 | let rows = data_table_rows(step); |
| 73 | assert_eq!(rows.len(), 1, "tool input table should contain one row"); |
| 74 | let input = Value::Object( |
| 75 | rows[0] |
| 76 | .iter() |
| 77 | .map(|(key, value)| (key.clone(), Value::String(value.clone()))) |
| 78 | .collect(), |
| 79 | ); |
| 80 | |
| 81 | world.tool_name = Some(tool_name); |
| 82 | world.tool_arguments = Some(serde_json::to_string(&input).expect("tool input arguments")); |
| 83 | } |
| 84 | |
| 85 | #[given( |
| 86 | regex = r#"^the mocked LLM will request the "([^"]+)" tool with malformed arguments "([^"]+)"$"# |
| 87 | )] |
| 88 | fn mocked_llm_will_request_tool_with_malformed_arguments( |
| 89 | world: &mut ToolLifecycleWorld, |
| 90 | tool_name: String, |
| 91 | arguments: String, |
| 92 | ) { |
| 93 | world.tool_name = Some(tool_name); |
| 94 | world.tool_arguments = Some(arguments); |
| 95 | } |
| 96 | |
| 97 | #[given("the mocked LLM will answer after the tool result:")] |
| 98 | fn mocked_llm_will_answer_after_tool_result(world: &mut ToolLifecycleWorld, step: &Step) { |
| 99 | let rows = data_table_rows(step); |
| 100 | assert_eq!(rows.len(), 1, "final answer table should contain one row"); |
| 101 | world.final_answer = Some(row_value(&rows[0], "content")); |
| 102 | } |
| 103 | |
| 104 | #[when(regex = r#"^the user asks "([^"]+)"$"#)] |
| 105 | async fn user_asks(world: &mut ToolLifecycleWorld, prompt: String) { |
| 106 | let server = start_mock_llm(world).await; |
| 107 | let output = run_codewhale_exec(world, &server, &prompt); |
| 108 | |
| 109 | world.prompt = Some(prompt); |
| 110 | world.stdout = String::from_utf8_lossy(&output.stdout).into_owned(); |
| 111 | world.stderr = String::from_utf8_lossy(&output.stderr).into_owned(); |
| 112 | assert!( |
| 113 | output.status.success(), |
| 114 | "codewhale exec failed\nstdout:\n{}\nstderr:\n{}", |
| 115 | world.stdout, |
| 116 | world.stderr |
| 117 | ); |
| 118 | |
| 119 | world.events = parse_stream_events(&world.stdout); |
| 120 | world.requests = server |
| 121 | .received_requests() |
| 122 | .await |
| 123 | .expect("mock server should record requests") |
| 124 | .into_iter() |
| 125 | .filter(|request| request.url.path().ends_with("/chat/completions")) |
| 126 | .map(|request| { |
| 127 | request |
| 128 | .body_json() |
| 129 | .expect("chat request body should be JSON") |
| 130 | }) |
| 131 | .collect(); |
| 132 | world.llm_server = Some(server); |
| 133 | } |
| 134 | |
| 135 | #[then("CodeWhale should send the user request to the mocked LLM")] |
| 136 | fn codewhale_should_send_user_request_to_mocked_llm(world: &mut ToolLifecycleWorld) { |
| 137 | let first_request = world |
| 138 | .requests |
| 139 | .first() |
| 140 | .expect("expected an initial chat request"); |
| 141 | |
| 142 | assert!( |
| 143 | request_contains_user_text( |
| 144 | first_request, |
| 145 | world |
| 146 | .prompt |
| 147 | .as_deref() |
| 148 | .expect("scenario prompt should be set") |
| 149 | ), |
| 150 | "initial request should include the user prompt:\n{first_request:#}" |
| 151 | ); |
| 152 | assert!( |
| 153 | !request_contains_tool_result(first_request), |
| 154 | "initial request should not include a tool result:\n{first_request:#}" |
| 155 | ); |
| 156 | } |
| 157 | |
| 158 | #[then("the public tool lifecycle should show a running tool:")] |
| 159 | fn public_tool_lifecycle_should_show_running_tool(world: &mut ToolLifecycleWorld, step: &Step) { |
| 160 | let expected = one_table_row(step); |
| 161 | assert_eq!(row_value(&expected, "status"), "running"); |
| 162 | assert_eq!(row_value(&expected, "marker"), "[~]"); |
| 163 | |
| 164 | let event = tool_use_event(world, &row_value(&expected, "tool")); |
| 165 | assert_eq!( |
| 166 | event.get("input").and_then(|input| input.get("path")), |
| 167 | Some(&json!(row_value(&expected, "input"))) |
| 168 | ); |
| 169 | assert_expected_action(event, &expected); |
| 170 | } |
| 171 | |
| 172 | #[then("the public tool result should return directory entries:")] |
| 173 | fn public_tool_result_should_return_directory_entries(world: &mut ToolLifecycleWorld, step: &Step) { |
| 174 | let output = tool_result_output(world); |
| 175 | let entries: Vec<Value> = |
| 176 | serde_json::from_str(output).expect("File.list result should be JSON entries"); |
| 177 | |
| 178 | for row in data_table_rows(step) { |
| 179 | let expected_name = row_value(&row, "entry"); |
| 180 | let expected_is_dir = match row_value(&row, "kind").as_str() { |
| 181 | "file" => false, |
| 182 | "folder" => true, |
| 183 | other => panic!("unsupported expected entry kind: {other}"), |
| 184 | }; |
| 185 | assert!( |
| 186 | entries.iter().any(|entry| { |
| 187 | entry.get("name").and_then(Value::as_str) == Some(expected_name.as_str()) |
| 188 | && entry.get("is_dir").and_then(Value::as_bool) == Some(expected_is_dir) |
| 189 | }), |
| 190 | "missing {expected_name} in File.list result:\n{output}" |
| 191 | ); |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | #[then("CodeWhale should send the tool result back to the mocked LLM")] |
| 196 | fn codewhale_should_send_tool_result_back_to_mocked_llm(world: &mut ToolLifecycleWorld) { |
| 197 | let request = world |
| 198 | .requests |
| 199 | .iter() |
| 200 | .find(|request| request_contains_tool_result(request)) |
| 201 | .expect("expected a follow-up chat request containing the tool result"); |
| 202 | let tool_result = tool_result_message(request).expect("tool result message"); |
| 203 | assert_eq!( |
| 204 | tool_result |
| 205 | .get("tool_call_id") |
| 206 | .and_then(serde_json::Value::as_str), |
| 207 | Some(TOOL_CALL_ID) |
| 208 | ); |
| 209 | |
| 210 | let content = tool_result |
| 211 | .get("content") |
| 212 | .and_then(serde_json::Value::as_str) |
| 213 | .expect("tool result content"); |
| 214 | assert_eq!( |
| 215 | content, |
| 216 | tool_result_output(world), |
| 217 | "follow-up request should preserve the exact public tool result" |
| 218 | ); |
| 219 | } |
| 220 | |
| 221 | #[then(regex = r#"^the public tool result should report an error for "([^"]+)"$"#)] |
| 222 | fn public_tool_result_should_report_error_for(world: &mut ToolLifecycleWorld, tool_name: String) { |
| 223 | let _ = tool_use_event(world, &tool_name); |
| 224 | let event = tool_result_event(world); |
| 225 | |
| 226 | assert_eq!(event.get("status").and_then(Value::as_str), Some("error")); |
| 227 | let output = event |
| 228 | .get("output") |
| 229 | .and_then(Value::as_str) |
| 230 | .expect("tool_result error output"); |
| 231 | assert!( |
| 232 | output.contains(&tool_name) && output.contains("not available"), |
| 233 | "tool_result error should name the unavailable tool:\n{output}" |
| 234 | ); |
| 235 | } |
| 236 | |
| 237 | #[then("CodeWhale should send the tool error back to the mocked LLM")] |
| 238 | fn codewhale_should_send_tool_error_back_to_mocked_llm(world: &mut ToolLifecycleWorld) { |
| 239 | let request = world |
| 240 | .requests |
| 241 | .iter() |
| 242 | .find(|request| request_contains_tool_result(request)) |
| 243 | .expect("expected a follow-up chat request containing the tool error"); |
| 244 | let tool_result = tool_result_message(request).expect("tool result message"); |
| 245 | assert_eq!( |
| 246 | tool_result |
| 247 | .get("tool_call_id") |
| 248 | .and_then(serde_json::Value::as_str), |
| 249 | Some(TOOL_CALL_ID) |
| 250 | ); |
| 251 | |
| 252 | let content = tool_result |
| 253 | .get("content") |
| 254 | .and_then(serde_json::Value::as_str) |
| 255 | .expect("tool result content"); |
| 256 | let tool_name = world.tool_name.as_deref().expect("tool name"); |
| 257 | assert!( |
| 258 | content.contains(tool_name) && content.contains("not available"), |
| 259 | "tool error sent to LLM should describe the unavailable tool:\n{content}" |
| 260 | ); |
| 261 | } |
| 262 | |
| 263 | #[then( |
| 264 | regex = r#"^the public tool lifecycle should show a running tool with raw input for "([^"]+)"$"# |
| 265 | )] |
| 266 | fn public_tool_lifecycle_should_show_running_tool_with_raw_input( |
| 267 | world: &mut ToolLifecycleWorld, |
| 268 | tool_name: String, |
| 269 | ) { |
| 270 | let event = tool_use_event(world, &tool_name); |
| 271 | assert!( |
| 272 | value_contains_text(event.get("input").expect("tool_use input"), "{not-json"), |
| 273 | "tool_use input should preserve malformed raw arguments:\n{event:#}" |
| 274 | ); |
| 275 | } |
| 276 | |
| 277 | #[then(regex = r#"^the public tool result should report malformed arguments for "([^"]+)"$"#)] |
| 278 | fn public_tool_result_should_report_malformed_arguments_for( |
| 279 | world: &mut ToolLifecycleWorld, |
| 280 | tool_name: String, |
| 281 | ) { |
| 282 | let _ = tool_use_event(world, &tool_name); |
| 283 | let event = tool_result_event(world); |
| 284 | |
| 285 | assert_eq!(event.get("status").and_then(Value::as_str), Some("error")); |
| 286 | let output = event |
| 287 | .get("output") |
| 288 | .and_then(Value::as_str) |
| 289 | .expect("tool_result error output"); |
| 290 | assert_malformed_arguments_text(output); |
| 291 | } |
| 292 | |
| 293 | #[then("CodeWhale should send the malformed argument error back to the mocked LLM")] |
| 294 | fn codewhale_should_send_malformed_argument_error_back_to_mocked_llm( |
| 295 | world: &mut ToolLifecycleWorld, |
| 296 | ) { |
| 297 | let request = world |
| 298 | .requests |
| 299 | .iter() |
| 300 | .find(|request| request_contains_tool_result(request)) |
| 301 | .expect("expected a follow-up chat request containing the malformed argument error"); |
| 302 | let tool_result = tool_result_message(request).expect("tool result message"); |
| 303 | assert_eq!( |
| 304 | tool_result |
| 305 | .get("tool_call_id") |
| 306 | .and_then(serde_json::Value::as_str), |
| 307 | Some(TOOL_CALL_ID) |
| 308 | ); |
| 309 | |
| 310 | let content = tool_result |
| 311 | .get("content") |
| 312 | .and_then(serde_json::Value::as_str) |
| 313 | .expect("tool result content"); |
| 314 | assert_malformed_arguments_text(content); |
| 315 | } |
| 316 | |
| 317 | #[then( |
| 318 | regex = r#"^the public tool result should report a real error for "([^"]+)" containing "([^"]+)"$"# |
| 319 | )] |
| 320 | fn public_tool_result_should_report_real_error( |
| 321 | world: &mut ToolLifecycleWorld, |
| 322 | tool_name: String, |
| 323 | expected: String, |
| 324 | ) { |
| 325 | let _ = tool_use_event(world, &tool_name); |
| 326 | let event = tool_result_event(world); |
| 327 | assert_eq!(event.get("status").and_then(Value::as_str), Some("error")); |
| 328 | |
| 329 | let output = event |
| 330 | .get("output") |
| 331 | .and_then(Value::as_str) |
| 332 | .expect("real tool error output"); |
| 333 | assert!( |
| 334 | output.contains(&expected) && output.contains("Failed to read"), |
| 335 | "real {tool_name} failure should preserve the path and execution error:\n{output}" |
| 336 | ); |
| 337 | } |
| 338 | |
| 339 | #[then("CodeWhale should send the real tool error back to the mocked LLM")] |
| 340 | fn codewhale_should_send_real_tool_error_back_to_mocked_llm(world: &mut ToolLifecycleWorld) { |
| 341 | let request = world |
| 342 | .requests |
| 343 | .iter() |
| 344 | .find(|request| request_contains_tool_result(request)) |
| 345 | .expect("expected a follow-up chat request containing the real tool error"); |
| 346 | let content = tool_result_message(request) |
| 347 | .and_then(|message| message.get("content")) |
| 348 | .and_then(Value::as_str) |
| 349 | .expect("real tool error content"); |
| 350 | assert!( |
| 351 | content.contains("missing.txt") && content.contains("Failed to read"), |
| 352 | "real tool error sent to the LLM should preserve the execution failure:\n{content}" |
| 353 | ); |
| 354 | } |
| 355 | |
| 356 | #[then("the public tool result should be an empty list")] |
| 357 | fn public_tool_result_should_be_an_empty_list(world: &mut ToolLifecycleWorld) { |
| 358 | let output = tool_result_output(world); |
| 359 | let value: Value = serde_json::from_str(output).expect("empty File.list result should be JSON"); |
| 360 | assert_eq!(value, json!([]), "empty workspace should return []"); |
| 361 | assert_eq!( |
| 362 | tool_result_event(world) |
| 363 | .get("status") |
| 364 | .and_then(Value::as_str), |
| 365 | Some("success") |
| 366 | ); |
| 367 | } |
| 368 | |
| 369 | #[then("CodeWhale should send the empty tool result back to the mocked LLM")] |
| 370 | fn codewhale_should_send_empty_tool_result_back_to_mocked_llm(world: &mut ToolLifecycleWorld) { |
| 371 | let request = world |
| 372 | .requests |
| 373 | .iter() |
| 374 | .find(|request| request_contains_tool_result(request)) |
| 375 | .expect("expected a follow-up chat request containing the empty tool result"); |
| 376 | let content = tool_result_message(request) |
| 377 | .and_then(|message| message.get("content")) |
| 378 | .and_then(Value::as_str) |
| 379 | .expect("empty tool result content"); |
| 380 | let value: Value = |
| 381 | serde_json::from_str(content).expect("forwarded empty result should be JSON"); |
| 382 | assert_eq!(value, json!([]), "follow-up request should preserve []"); |
| 383 | } |
| 384 | |
| 385 | #[then( |
| 386 | regex = r#"^the public tool lifecycle should show a failed tool with raw input for "([^"]+)"$"# |
| 387 | )] |
| 388 | fn public_tool_lifecycle_should_show_failed_tool_with_raw_input( |
| 389 | world: &mut ToolLifecycleWorld, |
| 390 | tool_name: String, |
| 391 | ) { |
| 392 | let event = tool_result_event(world); |
| 393 | assert_eq!(event.get("status").and_then(Value::as_str), Some("error")); |
| 394 | |
| 395 | let tool_use = tool_use_event(world, &tool_name); |
| 396 | assert!( |
| 397 | value_contains_text(tool_use.get("input").expect("tool_use input"), "{not-json"), |
| 398 | "failed tool_use input should preserve malformed raw arguments:\n{tool_use:#}" |
| 399 | ); |
| 400 | } |
| 401 | |
| 402 | #[then("the public tool lifecycle should show a completed tool:")] |
| 403 | fn public_tool_lifecycle_should_show_completed_tool(world: &mut ToolLifecycleWorld, step: &Step) { |
| 404 | let expected = one_table_row(step); |
| 405 | assert_eq!(row_value(&expected, "status"), "completed"); |
| 406 | assert_eq!(row_value(&expected, "marker"), "✓"); |
| 407 | |
| 408 | let event = tool_result_event(world); |
| 409 | assert_eq!(event.get("status").and_then(Value::as_str), Some("success")); |
| 410 | |
| 411 | let tool_use = tool_use_event(world, &row_value(&expected, "tool")); |
| 412 | assert_eq!( |
| 413 | tool_use.get("input").and_then(|input| input.get("path")), |
| 414 | Some(&json!(row_value(&expected, "input"))) |
| 415 | ); |
| 416 | assert_expected_action(tool_use, &expected); |
| 417 | } |
| 418 | |
| 419 | #[then("the public tool lifecycle should show a failed tool:")] |
| 420 | fn public_tool_lifecycle_should_show_failed_tool(world: &mut ToolLifecycleWorld, step: &Step) { |
| 421 | let expected = one_table_row(step); |
| 422 | assert_eq!(row_value(&expected, "status"), "error"); |
| 423 | assert_eq!(row_value(&expected, "marker"), "[!]"); |
| 424 | |
| 425 | let event = tool_result_event(world); |
| 426 | assert_eq!(event.get("status").and_then(Value::as_str), Some("error")); |
| 427 | |
| 428 | let tool_use = tool_use_event(world, &row_value(&expected, "tool")); |
| 429 | assert_eq!( |
| 430 | tool_use.get("input").and_then(|input| input.get("path")), |
| 431 | Some(&json!(row_value(&expected, "input"))) |
| 432 | ); |
| 433 | assert_expected_action(tool_use, &expected); |
| 434 | } |
| 435 | |
| 436 | #[then(regex = r#"^the public output should include "([^"]+)"$"#)] |
| 437 | fn public_output_should_include(world: &mut ToolLifecycleWorld, expected: String) { |
| 438 | let content = public_content_output(world); |
| 439 | assert!( |
| 440 | content.contains(&expected), |
| 441 | "public content output should include {expected:?}:\nstdout:\n{}\nstderr:\n{}", |
| 442 | world.stdout, |
| 443 | world.stderr |
| 444 | ); |
| 445 | } |
| 446 | |
| 447 | #[then(regex = r#"^acceptance should report the missing expected summary "([^"]+)"$"#)] |
| 448 | fn acceptance_should_report_missing_expected_summary( |
| 449 | world: &mut ToolLifecycleWorld, |
| 450 | expected: String, |
| 451 | ) { |
| 452 | let report = require_follow_up_summary(world, &expected) |
| 453 | .expect_err("fixture answer intentionally omits the expected summary"); |
| 454 | assert!( |
| 455 | report.contains(&expected) && report.contains("missing expected summary"), |
| 456 | "missing-summary oracle should name the absent contract:\n{report}" |
| 457 | ); |
| 458 | } |
| 459 | |
| 460 | #[tokio::test(flavor = "current_thread")] |
| 461 | async fn happy_path_lists_current_directory_through_tool() { |
| 462 | run_scenario(HAPPY_PATH_SCENARIO, 10).await; |
| 463 | } |
| 464 | |
| 465 | #[tokio::test(flavor = "current_thread")] |
| 466 | async fn unknown_tool_returns_error_result() { |
| 467 | run_scenario(UNKNOWN_TOOL_SCENARIO, 10).await; |
| 468 | } |
| 469 | |
| 470 | #[tokio::test(flavor = "current_thread")] |
| 471 | async fn malformed_tool_arguments_return_error_result() { |
| 472 | run_scenario(MALFORMED_ARGUMENTS_SCENARIO, 10).await; |
| 473 | } |
| 474 | |
| 475 | #[tokio::test(flavor = "current_thread")] |
| 476 | async fn real_tool_error_is_returned_to_follow_up_request() { |
| 477 | run_scenario(REAL_TOOL_ERROR_SCENARIO, 10).await; |
| 478 | } |
| 479 | |
| 480 | #[tokio::test(flavor = "current_thread")] |
| 481 | async fn empty_tool_result_is_returned_to_follow_up_request() { |
| 482 | run_scenario(EMPTY_TOOL_RESULT_SCENARIO, 10).await; |
| 483 | } |
| 484 | |
| 485 | #[tokio::test(flavor = "current_thread")] |
| 486 | async fn missing_follow_up_summary_is_detected() { |
| 487 | run_scenario(MISSING_SUMMARY_SCENARIO, 11).await; |
| 488 | } |
| 489 | |
| 490 | async fn run_scenario(name: &'static str, expected_steps: usize) { |
| 491 | let writer = ToolLifecycleWorld::cucumber() |
| 492 | .fail_on_skipped() |
| 493 | .with_default_cli() |
| 494 | .filter_run(FEATURE_PATH, move |feature, _, scenario| { |
| 495 | feature.name == FEATURE_NAME && scenario.name == name |
| 496 | }) |
| 497 | .await; |
| 498 | assert_eq!(writer.failed_steps(), 0, "scenario failed: {name}"); |
| 499 | assert_eq!(writer.skipped_steps(), 0, "scenario skipped steps: {name}"); |
| 500 | assert_eq!( |
| 501 | writer.passed_steps(), |
| 502 | expected_steps, |
| 503 | "scenario did not run: {name}" |
| 504 | ); |
| 505 | } |
| 506 | |
| 507 | async fn start_mock_llm(world: &ToolLifecycleWorld) -> MockServer { |
| 508 | let server = MockServer::start().await; |
| 509 | |
| 510 | Mock::given(method("GET")) |
| 511 | .and(path("/v1/models")) |
| 512 | .respond_with(json_response(json!({ |
| 513 | "object": "list", |
| 514 | "data": [{ "id": TEST_MODEL, "object": "model" }] |
| 515 | }))) |
| 516 | .mount(&server) |
| 517 | .await; |
| 518 | |
| 519 | Mock::given(method("POST")) |
| 520 | .and(path("/v1/chat/completions")) |
| 521 | .and(request_has_tool_result) |
| 522 | .respond_with(sse_response(&final_answer_sse( |
| 523 | world.final_answer.as_ref().expect("final LLM answer"), |
| 524 | ))) |
| 525 | .mount(&server) |
| 526 | .await; |
| 527 | |
| 528 | Mock::given(method("POST")) |
| 529 | .and(path("/v1/chat/completions")) |
| 530 | .and(request_has_no_tool_result) |
| 531 | .respond_with(sse_response(&tool_call_sse( |
| 532 | world.tool_name.as_ref().expect("tool name"), |
| 533 | world.tool_arguments.as_ref().expect("tool arguments"), |
| 534 | ))) |
| 535 | .mount(&server) |
| 536 | .await; |
| 537 | |
| 538 | server |
| 539 | } |
| 540 | |
| 541 | fn run_codewhale_exec( |
| 542 | world: &ToolLifecycleWorld, |
| 543 | server: &MockServer, |
| 544 | prompt: &str, |
| 545 | ) -> std::process::Output { |
| 546 | let workspace = world |
| 547 | .workspace |
| 548 | .as_ref() |
| 549 | .expect("workspace") |
| 550 | .path() |
| 551 | .to_path_buf(); |
| 552 | let home = world.home.as_ref().expect("home").path().to_path_buf(); |
| 553 | |
| 554 | let mut command = Command::new(crate::binary::codewhale()); |
| 555 | preserve_host_env(&mut command); |
| 556 | command |
| 557 | .current_dir(&workspace) |
| 558 | .arg("--workspace") |
| 559 | .arg(&workspace) |
| 560 | .arg("--no-project-config") |
| 561 | .arg("exec") |
| 562 | .arg("--auto") |
| 563 | .arg("--model") |
| 564 | .arg(TEST_MODEL) |
| 565 | .arg("--output-format") |
| 566 | .arg("stream-json") |
| 567 | .arg(prompt) |
| 568 | .env("HOME", &home) |
| 569 | .env("USERPROFILE", &home) |
| 570 | .env("XDG_CONFIG_HOME", home.join(".config")) |
| 571 | .env("XDG_DATA_HOME", home.join(".local").join("share")) |
| 572 | .env("XDG_CACHE_HOME", home.join(".cache")) |
| 573 | .env( |
| 574 | "CODEWHALE_CONFIG_PATH", |
| 575 | home.join(".codewhale").join("config.toml"), |
| 576 | ) |
| 577 | .env( |
| 578 | "DEEPSEEK_CONFIG_PATH", |
| 579 | home.join(".deepseek").join("config.toml"), |
| 580 | ) |
| 581 | .env("DEEPSEEK_API_KEY", "ci-test-key-not-real") |
| 582 | .env("DEEPSEEK_BASE_URL", server.uri()) |
| 583 | .env("CODEWHALE_BASE_URL", server.uri()) |
| 584 | .env("DEEPSEEK_MODEL", TEST_MODEL) |
| 585 | .env("CODEWHALE_MODEL", TEST_MODEL) |
| 586 | .env("RUST_LOG", "warn") |
| 587 | .stdout(Stdio::piped()) |
| 588 | .stderr(Stdio::piped()); |
| 589 | |
| 590 | std::fs::create_dir_all(home.join(".codewhale")).expect("create codewhale home config dir"); |
| 591 | std::fs::create_dir_all(home.join(".deepseek")).expect("create deepseek home config dir"); |
| 592 | |
| 593 | run_with_timeout(command, Duration::from_secs(45)) |
| 594 | } |
| 595 | |
| 596 | fn run_with_timeout(mut command: Command, timeout: Duration) -> std::process::Output { |
| 597 | let mut child = command.spawn().expect("spawn codewhale exec"); |
| 598 | let stdout_reader = read_pipe_in_background(child.stdout.take().expect("stdout pipe")); |
| 599 | let stderr_reader = read_pipe_in_background(child.stderr.take().expect("stderr pipe")); |
| 600 | |
| 601 | let status = match child.wait_timeout(timeout).expect("wait for codewhale") { |
| 602 | Some(status) => status, |
| 603 | None => { |
| 604 | let _ = child.kill(); |
| 605 | let _ = child.wait(); |
| 606 | let stdout = join_pipe_reader(stdout_reader, "stdout"); |
| 607 | let stderr = join_pipe_reader(stderr_reader, "stderr"); |
| 608 | panic!( |
| 609 | "codewhale exec timed out after {timeout:?}\nstdout:\n{}\nstderr:\n{}", |
| 610 | String::from_utf8_lossy(&stdout), |
| 611 | String::from_utf8_lossy(&stderr) |
| 612 | ); |
| 613 | } |
| 614 | }; |
| 615 | |
| 616 | let stdout = join_pipe_reader(stdout_reader, "stdout"); |
| 617 | let stderr = join_pipe_reader(stderr_reader, "stderr"); |
| 618 | |
| 619 | std::process::Output { |
| 620 | status, |
| 621 | stdout, |
| 622 | stderr, |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | fn read_pipe_in_background<R>(mut reader: R) -> std::thread::JoinHandle<std::io::Result<Vec<u8>>> |
| 627 | where |
| 628 | R: Read + Send + 'static, |
| 629 | { |
| 630 | std::thread::spawn(move || { |
| 631 | let mut output = Vec::new(); |
| 632 | reader.read_to_end(&mut output).map(|_| output) |
| 633 | }) |
| 634 | } |
| 635 | |
| 636 | fn join_pipe_reader( |
| 637 | handle: std::thread::JoinHandle<std::io::Result<Vec<u8>>>, |
| 638 | stream_name: &str, |
| 639 | ) -> Vec<u8> { |
| 640 | handle |
| 641 | .join() |
| 642 | .unwrap_or_else(|_| panic!("{stream_name} reader thread panicked")) |
| 643 | .unwrap_or_else(|err| panic!("read {stream_name}: {err}")) |
| 644 | } |
| 645 | |
| 646 | fn preserve_host_env(command: &mut Command) { |
| 647 | command.env_clear(); |
| 648 | for key in [ |
| 649 | "PATH", |
| 650 | "PATHEXT", |
| 651 | "SystemRoot", |
| 652 | "SystemDrive", |
| 653 | "WINDIR", |
| 654 | "COMSPEC", |
| 655 | "TEMP", |
| 656 | "TMP", |
| 657 | "TERM", |
| 658 | "COLORTERM", |
| 659 | "LANG", |
| 660 | "LC_ALL", |
| 661 | ] { |
| 662 | if let Some(value) = std::env::var_os(key) { |
| 663 | command.env(key, value); |
| 664 | } |
| 665 | } |
| 666 | } |
| 667 | |
| 668 | fn tool_call_sse(tool_name: &str, arguments: &str) -> String { |
| 669 | [ |
| 670 | sse_chunk(json!({ |
| 671 | "id": "chatcmpl-tool", |
| 672 | "object": "chat.completion.chunk", |
| 673 | "model": TEST_MODEL, |
| 674 | "choices": [{ |
| 675 | "index": 0, |
| 676 | "delta": { |
| 677 | "tool_calls": [{ |
| 678 | "index": 0, |
| 679 | "id": TOOL_CALL_ID, |
| 680 | "type": "function", |
| 681 | "function": { |
| 682 | "name": tool_name, |
| 683 | "arguments": arguments |
| 684 | } |
| 685 | }] |
| 686 | }, |
| 687 | "finish_reason": null |
| 688 | }] |
| 689 | })), |
| 690 | sse_chunk(json!({ |
| 691 | "id": "chatcmpl-tool", |
| 692 | "object": "chat.completion.chunk", |
| 693 | "model": TEST_MODEL, |
| 694 | "choices": [{ |
| 695 | "index": 0, |
| 696 | "delta": {}, |
| 697 | "finish_reason": "tool_calls" |
| 698 | }], |
| 699 | "usage": { |
| 700 | "prompt_tokens": 10, |
| 701 | "completion_tokens": 2, |
| 702 | "total_tokens": 12 |
| 703 | } |
| 704 | })), |
| 705 | "data: [DONE]\n\n".to_string(), |
| 706 | ] |
| 707 | .join("") |
| 708 | } |
| 709 | |
| 710 | fn final_answer_sse(answer: &str) -> String { |
| 711 | [ |
| 712 | sse_chunk(json!({ |
| 713 | "id": "chatcmpl-final", |
| 714 | "object": "chat.completion.chunk", |
| 715 | "model": TEST_MODEL, |
| 716 | "choices": [{ |
| 717 | "index": 0, |
| 718 | "delta": { "content": answer }, |
| 719 | "finish_reason": null |
| 720 | }] |
| 721 | })), |
| 722 | sse_chunk(json!({ |
| 723 | "id": "chatcmpl-final", |
| 724 | "object": "chat.completion.chunk", |
| 725 | "model": TEST_MODEL, |
| 726 | "choices": [{ |
| 727 | "index": 0, |
| 728 | "delta": {}, |
| 729 | "finish_reason": "stop" |
| 730 | }], |
| 731 | "usage": { |
| 732 | "prompt_tokens": 20, |
| 733 | "completion_tokens": 8, |
| 734 | "total_tokens": 28 |
| 735 | } |
| 736 | })), |
| 737 | "data: [DONE]\n\n".to_string(), |
| 738 | ] |
| 739 | .join("") |
| 740 | } |
| 741 | |
| 742 | fn assert_malformed_arguments_text(text: &str) { |
| 743 | let lower = text.to_ascii_lowercase(); |
| 744 | assert!( |
| 745 | lower.contains("argument") |
| 746 | && (lower.contains("malformed") |
| 747 | || lower.contains("parse") |
| 748 | || lower.contains("json") |
| 749 | || lower.contains("invalid")), |
| 750 | "expected malformed argument error text:\n{text}" |
| 751 | ); |
| 752 | } |
| 753 | |
| 754 | fn sse_chunk(value: Value) -> String { |
| 755 | format!( |
| 756 | "data: {}\n\n", |
| 757 | serde_json::to_string(&value).expect("SSE JSON") |
| 758 | ) |
| 759 | } |
| 760 | |
| 761 | fn sse_response(body: &str) -> ResponseTemplate { |
| 762 | ResponseTemplate::new(200) |
| 763 | .insert_header("content-type", "text/event-stream") |
| 764 | .insert_header("cache-control", "no-cache") |
| 765 | .set_body_string(body.to_string()) |
| 766 | } |
| 767 | |
| 768 | fn json_response(value: Value) -> ResponseTemplate { |
| 769 | ResponseTemplate::new(200) |
| 770 | .insert_header("content-type", "application/json") |
| 771 | .set_body_json(value) |
| 772 | } |
| 773 | |
| 774 | fn request_has_tool_result(request: &Request) -> bool { |
| 775 | request |
| 776 | .body_json::<Value>() |
| 777 | .is_ok_and(|body| request_contains_tool_result(&body)) |
| 778 | } |
| 779 | |
| 780 | fn request_has_no_tool_result(request: &Request) -> bool { |
| 781 | !request_has_tool_result(request) |
| 782 | } |
| 783 | |
| 784 | fn request_contains_tool_result(request: &Value) -> bool { |
| 785 | tool_result_message(request).is_some() |
| 786 | } |
| 787 | |
| 788 | fn tool_result_message(request: &Value) -> Option<&Value> { |
| 789 | request |
| 790 | .get("messages") |
| 791 | .and_then(Value::as_array)? |
| 792 | .iter() |
| 793 | .find(|message| message.get("role").and_then(Value::as_str) == Some("tool")) |
| 794 | } |
| 795 | |
| 796 | fn request_contains_user_text(request: &Value, expected: &str) -> bool { |
| 797 | request |
| 798 | .get("messages") |
| 799 | .and_then(Value::as_array) |
| 800 | .into_iter() |
| 801 | .flatten() |
| 802 | .any(|message| { |
| 803 | message.get("role").and_then(Value::as_str) == Some("user") |
| 804 | && message |
| 805 | .get("content") |
| 806 | .is_some_and(|content| value_contains_text(content, expected)) |
| 807 | }) |
| 808 | } |
| 809 | |
| 810 | fn value_contains_text(value: &Value, expected: &str) -> bool { |
| 811 | match value { |
| 812 | Value::String(text) => text.contains(expected), |
| 813 | Value::Array(values) => values |
| 814 | .iter() |
| 815 | .any(|value| value_contains_text(value, expected)), |
| 816 | Value::Object(values) => values |
| 817 | .values() |
| 818 | .any(|value| value_contains_text(value, expected)), |
| 819 | _ => false, |
| 820 | } |
| 821 | } |
| 822 | |
| 823 | fn public_content_output(world: &ToolLifecycleWorld) -> String { |
| 824 | world |
| 825 | .events |
| 826 | .iter() |
| 827 | .filter(|event| event.get("type").and_then(Value::as_str) == Some("content")) |
| 828 | .filter_map(|event| event.get("content").and_then(Value::as_str)) |
| 829 | .collect() |
| 830 | } |
| 831 | |
| 832 | fn require_follow_up_summary(world: &ToolLifecycleWorld, expected: &str) -> Result<(), String> { |
| 833 | let content = public_content_output(world); |
| 834 | if content.contains(expected) { |
| 835 | Ok(()) |
| 836 | } else { |
| 837 | Err(format!( |
| 838 | "missing expected summary {expected:?} in follow-up answer {content:?}" |
| 839 | )) |
| 840 | } |
| 841 | } |
| 842 | |
| 843 | fn parse_stream_events(stdout: &str) -> Vec<Value> { |
| 844 | stdout |
| 845 | .lines() |
| 846 | .filter(|line| !line.trim().is_empty()) |
| 847 | .filter_map(|line| { |
| 848 | let json_start = line.find('{')?; |
| 849 | let json_line = &line[json_start..]; |
| 850 | Some(serde_json::from_str(json_line).unwrap_or_else(|err| { |
| 851 | panic!( |
| 852 | "stream-json line should parse: {err}\nline: {line}\njson: {json_line}\nstdout:\n{stdout}" |
| 853 | ) |
| 854 | })) |
| 855 | }) |
| 856 | .collect() |
| 857 | } |
| 858 | |
| 859 | fn tool_use_event<'a>(world: &'a ToolLifecycleWorld, expected_tool: &str) -> &'a Value { |
| 860 | world |
| 861 | .events |
| 862 | .iter() |
| 863 | .find(|event| { |
| 864 | event.get("type").and_then(Value::as_str) == Some("tool_use") |
| 865 | && event.get("name").and_then(Value::as_str) == Some(expected_tool) |
| 866 | }) |
| 867 | .unwrap_or_else(|| { |
| 868 | panic!( |
| 869 | "expected tool_use event for {expected_tool}\nstdout:\n{}\nstderr:\n{}", |
| 870 | world.stdout, world.stderr |
| 871 | ) |
| 872 | }) |
| 873 | } |
| 874 | |
| 875 | fn tool_result_event(world: &ToolLifecycleWorld) -> &Value { |
| 876 | world |
| 877 | .events |
| 878 | .iter() |
| 879 | .find(|event| event.get("type").and_then(Value::as_str) == Some("tool_result")) |
| 880 | .unwrap_or_else(|| { |
| 881 | panic!( |
| 882 | "expected tool_result event\nstdout:\n{}\nstderr:\n{}", |
| 883 | world.stdout, world.stderr |
| 884 | ) |
| 885 | }) |
| 886 | } |
| 887 | |
| 888 | fn tool_result_output(world: &ToolLifecycleWorld) -> &str { |
| 889 | tool_result_event(world) |
| 890 | .get("output") |
| 891 | .and_then(Value::as_str) |
| 892 | .expect("tool_result output") |
| 893 | } |
| 894 | |
| 895 | fn one_table_row(step: &Step) -> Vec<(String, String)> { |
| 896 | let rows = data_table_rows(step); |
| 897 | assert_eq!(rows.len(), 1, "expected exactly one data table row"); |
| 898 | rows.into_iter().next().expect("one row") |
| 899 | } |
| 900 | |
| 901 | fn data_table_rows(step: &Step) -> Vec<Vec<(String, String)>> { |
| 902 | let table = step |
| 903 | .table |
| 904 | .as_ref() |
| 905 | .expect("step should include a data table"); |
| 906 | let mut rows = table.rows.iter(); |
| 907 | let headers = rows |
| 908 | .next() |
| 909 | .expect("data table should include a header") |
| 910 | .clone(); |
| 911 | |
| 912 | let values: Vec<Vec<(String, String)>> = rows |
| 913 | .map(|row| { |
| 914 | headers |
| 915 | .iter() |
| 916 | .zip(row.iter()) |
| 917 | .map(|(header, value)| (header.clone(), value.clone())) |
| 918 | .collect() |
| 919 | }) |
| 920 | .collect(); |
| 921 | assert!( |
| 922 | !values.is_empty(), |
| 923 | "data table should include at least one row" |
| 924 | ); |
| 925 | values |
| 926 | } |
| 927 | |
| 928 | fn row_value(row: &[(String, String)], header: &str) -> String { |
| 929 | row.iter() |
| 930 | .find_map(|(key, value)| (key == header).then(|| value.clone())) |
| 931 | .unwrap_or_else(|| panic!("data table row missing {header} value")) |
| 932 | } |
| 933 | |
| 934 | fn assert_expected_action(event: &Value, expected: &[(String, String)]) { |
| 935 | let Some(action) = expected |
| 936 | .iter() |
| 937 | .find_map(|(key, value)| (key == "action").then_some(value)) |
| 938 | else { |
| 939 | return; |
| 940 | }; |
| 941 | assert_eq!( |
| 942 | event.get("input").and_then(|input| input.get("action")), |
| 943 | Some(&json!(action)), |
| 944 | "canonical tool action should be visible in the lifecycle event" |
| 945 | ); |
| 946 | } |
| 947 |