| 1 | //! Parallel verifier ensemble tool: `run_verifiers`. |
| 2 | //! |
| 3 | //! This is the agent-facing path for "parallelize the verifier, not the |
| 4 | //! generator": one tool call fans out to independent project checks across |
| 5 | //! common ecosystems and returns a single structured verdict. |
| 6 | |
| 7 | use std::collections::{BTreeSet, HashMap}; |
| 8 | use std::fs; |
| 9 | use std::path::{Path, PathBuf}; |
| 10 | use std::process::{Command, Stdio}; |
| 11 | use std::time::Instant; |
| 12 | |
| 13 | use async_trait::async_trait; |
| 14 | use serde::{Deserialize, Serialize}; |
| 15 | use serde_json::{Value, json}; |
| 16 | use shlex::try_join; |
| 17 | |
| 18 | use crate::dependencies::ExternalTool; |
| 19 | |
| 20 | use super::spec::{ |
| 21 | ApprovalRequirement, ToolCapability, ToolContext, ToolError, ToolResult, ToolSpec, |
| 22 | }; |
| 23 | |
| 24 | const MAX_GATE_OUTPUT_CHARS: usize = 16_000; |
| 25 | const DEFAULT_MAX_PYTHON_FILES: usize = 200; |
| 26 | const MAX_CUSTOM_GATES: usize = 12; |
| 27 | const BACKGROUND_GATE_TIMEOUT_MS: u64 = 600_000; |
| 28 | |
| 29 | /// Tool for running independent verifier gates concurrently. |
| 30 | pub struct RunVerifiersTool; |
| 31 | |
| 32 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)] |
| 33 | #[serde(rename_all = "snake_case")] |
| 34 | enum VerifierProfile { |
| 35 | Auto, |
| 36 | Rust, |
| 37 | Node, |
| 38 | Python, |
| 39 | Go, |
| 40 | } |
| 41 | |
| 42 | impl VerifierProfile { |
| 43 | fn parse(raw: &str) -> Result<Self, ToolError> { |
| 44 | match raw { |
| 45 | "auto" => Ok(Self::Auto), |
| 46 | "rust" => Ok(Self::Rust), |
| 47 | "node" => Ok(Self::Node), |
| 48 | "python" => Ok(Self::Python), |
| 49 | "go" => Ok(Self::Go), |
| 50 | other => Err(ToolError::invalid_input(format!( |
| 51 | "Unsupported profile '{other}'. Expected one of: auto, rust, node, python, go" |
| 52 | ))), |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | fn as_str(self) -> &'static str { |
| 57 | match self { |
| 58 | Self::Auto => "auto", |
| 59 | Self::Rust => "rust", |
| 60 | Self::Node => "node", |
| 61 | Self::Python => "python", |
| 62 | Self::Go => "go", |
| 63 | } |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)] |
| 68 | #[serde(rename_all = "snake_case")] |
| 69 | enum VerifierLevel { |
| 70 | Quick, |
| 71 | Full, |
| 72 | } |
| 73 | |
| 74 | impl VerifierLevel { |
| 75 | fn parse(raw: &str) -> Result<Self, ToolError> { |
| 76 | match raw { |
| 77 | "quick" => Ok(Self::Quick), |
| 78 | "full" => Ok(Self::Full), |
| 79 | other => Err(ToolError::invalid_input(format!( |
| 80 | "Unsupported level '{other}'. Expected one of: quick, full" |
| 81 | ))), |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | fn as_str(self) -> &'static str { |
| 86 | match self { |
| 87 | Self::Quick => "quick", |
| 88 | Self::Full => "full", |
| 89 | } |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | #[derive(Debug, Clone, Deserialize)] |
| 94 | #[serde(default, deny_unknown_fields)] |
| 95 | struct RunVerifiersInput { |
| 96 | profile: String, |
| 97 | level: String, |
| 98 | max_python_files: usize, |
| 99 | commands: Vec<CustomVerifierInput>, |
| 100 | background: bool, |
| 101 | } |
| 102 | |
| 103 | impl Default for RunVerifiersInput { |
| 104 | fn default() -> Self { |
| 105 | Self { |
| 106 | profile: "auto".to_string(), |
| 107 | level: "quick".to_string(), |
| 108 | max_python_files: DEFAULT_MAX_PYTHON_FILES, |
| 109 | commands: Vec::new(), |
| 110 | background: false, |
| 111 | } |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | #[derive(Debug, Clone, Default, Deserialize)] |
| 116 | #[serde(default, deny_unknown_fields)] |
| 117 | struct CustomVerifierInput { |
| 118 | name: String, |
| 119 | program: String, |
| 120 | args: Vec<String>, |
| 121 | cwd: Option<String>, |
| 122 | } |
| 123 | |
| 124 | #[derive(Debug, Clone)] |
| 125 | struct VerifierGate { |
| 126 | name: String, |
| 127 | ecosystem: String, |
| 128 | cwd: PathBuf, |
| 129 | program: Option<String>, |
| 130 | args: Vec<String>, |
| 131 | env: Vec<(String, String)>, |
| 132 | skipped_reason: Option<String>, |
| 133 | } |
| 134 | |
| 135 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 136 | struct GateResult { |
| 137 | name: String, |
| 138 | ecosystem: String, |
| 139 | status: GateStatus, |
| 140 | command: String, |
| 141 | cwd: String, |
| 142 | exit_code: Option<i32>, |
| 143 | duration_ms: u64, |
| 144 | stdout: String, |
| 145 | stderr: String, |
| 146 | stdout_truncated: bool, |
| 147 | stderr_truncated: bool, |
| 148 | skipped_reason: Option<String>, |
| 149 | } |
| 150 | |
| 151 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 152 | #[serde(rename_all = "snake_case")] |
| 153 | enum GateStatus { |
| 154 | Passed, |
| 155 | Failed, |
| 156 | Skipped, |
| 157 | } |
| 158 | |
| 159 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 160 | #[serde(rename_all = "snake_case")] |
| 161 | enum VerifierVerdict { |
| 162 | Pass, |
| 163 | Partial, |
| 164 | Fail, |
| 165 | } |
| 166 | |
| 167 | impl VerifierVerdict { |
| 168 | fn from_counts(gate_count: usize, failed: usize, skipped: usize) -> Self { |
| 169 | if failed > 0 { |
| 170 | Self::Fail |
| 171 | } else if skipped > 0 || gate_count == 0 { |
| 172 | Self::Partial |
| 173 | } else { |
| 174 | Self::Pass |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | fn hunt_verdict(self) -> &'static str { |
| 179 | match self { |
| 180 | Self::Pass => "hunted", |
| 181 | Self::Partial => "wounded", |
| 182 | Self::Fail => "escaped", |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | fn goal_status(self) -> &'static str { |
| 187 | match self { |
| 188 | Self::Pass => "complete", |
| 189 | Self::Partial => "paused", |
| 190 | Self::Fail => "blocked", |
| 191 | } |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 196 | struct RunVerifiersOutput { |
| 197 | success: bool, |
| 198 | profile: String, |
| 199 | level: String, |
| 200 | workspace: String, |
| 201 | gate_count: usize, |
| 202 | passed: usize, |
| 203 | failed: usize, |
| 204 | skipped: usize, |
| 205 | verifier_verdict: VerifierVerdict, |
| 206 | hunt_verdict: String, |
| 207 | goal_status: String, |
| 208 | summary: String, |
| 209 | gates: Vec<GateResult>, |
| 210 | } |
| 211 | |
| 212 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 213 | struct BackgroundGateJob { |
| 214 | name: String, |
| 215 | ecosystem: String, |
| 216 | status: String, |
| 217 | command: String, |
| 218 | cwd: String, |
| 219 | task_id: Option<String>, |
| 220 | skipped_reason: Option<String>, |
| 221 | error: Option<String>, |
| 222 | } |
| 223 | |
| 224 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 225 | struct RunVerifiersBackgroundOutput { |
| 226 | success: bool, |
| 227 | profile: String, |
| 228 | level: String, |
| 229 | workspace: String, |
| 230 | background: bool, |
| 231 | gate_count: usize, |
| 232 | started: usize, |
| 233 | skipped: usize, |
| 234 | failed_to_start: usize, |
| 235 | summary: String, |
| 236 | jobs: Vec<BackgroundGateJob>, |
| 237 | } |
| 238 | |
| 239 | #[async_trait] |
| 240 | impl ToolSpec for RunVerifiersTool { |
| 241 | fn name(&self) -> &'static str { |
| 242 | "run_verifiers" |
| 243 | } |
| 244 | |
| 245 | fn model_visible(&self) -> bool { |
| 246 | false |
| 247 | } |
| 248 | |
| 249 | fn description(&self) -> &'static str { |
| 250 | "Run independent verifier gates in parallel across detected Rust, Node, Python, and Go projects. Supports explicit custom verifier commands as program+args without requiring Bash." |
| 251 | } |
| 252 | |
| 253 | fn input_schema(&self) -> Value { |
| 254 | json!({ |
| 255 | "type": "object", |
| 256 | "properties": { |
| 257 | "profile": { |
| 258 | "type": "string", |
| 259 | "enum": ["auto", "rust", "node", "python", "go"], |
| 260 | "default": "auto", |
| 261 | "description": "Which ecosystem verifier set to run. 'auto' detects all supported project types in the workspace." |
| 262 | }, |
| 263 | "level": { |
| 264 | "type": "string", |
| 265 | "enum": ["quick", "full"], |
| 266 | "default": "quick", |
| 267 | "description": "Quick runs fast syntax/drift/build checks. Full adds heavier test/lint gates where available." |
| 268 | }, |
| 269 | "max_python_files": { |
| 270 | "type": "integer", |
| 271 | "minimum": 1, |
| 272 | "maximum": 1000, |
| 273 | "default": DEFAULT_MAX_PYTHON_FILES, |
| 274 | "description": "Maximum Python files to syntax-parse in the built-in python-syntax gate." |
| 275 | }, |
| 276 | "commands": { |
| 277 | "type": "array", |
| 278 | "description": "Optional explicit verifier gates. Commands run directly as program+args, not through a shell. Use program='bash', args=['-lc', '...'] only when Bash is intentionally part of the verifier.", |
| 279 | "items": { |
| 280 | "type": "object", |
| 281 | "properties": { |
| 282 | "name": { |
| 283 | "type": "string", |
| 284 | "description": "Short unique gate name." |
| 285 | }, |
| 286 | "program": { |
| 287 | "type": "string", |
| 288 | "description": "Executable to spawn, for example 'uv', 'pytest', 'npm', 'make', 'cmd', 'powershell', or 'bash'." |
| 289 | }, |
| 290 | "args": { |
| 291 | "type": "array", |
| 292 | "items": { "type": "string" }, |
| 293 | "description": "Arguments passed directly to the executable." |
| 294 | }, |
| 295 | "cwd": { |
| 296 | "type": "string", |
| 297 | "description": "Optional working directory relative to the workspace." |
| 298 | } |
| 299 | }, |
| 300 | "required": ["name", "program"], |
| 301 | "additionalProperties": false |
| 302 | }, |
| 303 | }, |
| 304 | "background": { |
| 305 | "type": "boolean", |
| 306 | "default": false, |
| 307 | "description": "Start verifier gates as background shell jobs and return task_ids immediately. Use for long build/test/lint gates; completion is tracked in task/status state, and `Bash` with action 'wait' / task_shell_wait are only for early output, final output, or true dependency barriers." |
| 308 | } |
| 309 | }, |
| 310 | "additionalProperties": false |
| 311 | }) |
| 312 | } |
| 313 | |
| 314 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 315 | vec![ToolCapability::ExecutesCode, ToolCapability::Sandboxable] |
| 316 | } |
| 317 | |
| 318 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 319 | ApprovalRequirement::Required |
| 320 | } |
| 321 | |
| 322 | fn starts_detached_for(&self, input: &Value) -> bool { |
| 323 | input.get("background").and_then(Value::as_bool) == Some(true) |
| 324 | } |
| 325 | |
| 326 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 327 | let input: RunVerifiersInput = serde_json::from_value(input) |
| 328 | .map_err(|err| ToolError::invalid_input(err.to_string()))?; |
| 329 | let profile = VerifierProfile::parse(input.profile.as_str())?; |
| 330 | let level = VerifierLevel::parse(input.level.as_str())?; |
| 331 | if input.max_python_files == 0 || input.max_python_files > 1000 { |
| 332 | return Err(ToolError::invalid_input( |
| 333 | "max_python_files must be between 1 and 1000", |
| 334 | )); |
| 335 | } |
| 336 | if input.commands.len() > MAX_CUSTOM_GATES { |
| 337 | return Err(ToolError::invalid_input(format!( |
| 338 | "commands may contain at most {MAX_CUSTOM_GATES} custom gates" |
| 339 | ))); |
| 340 | } |
| 341 | |
| 342 | let gates = build_gate_plan( |
| 343 | context, |
| 344 | profile, |
| 345 | level, |
| 346 | input.max_python_files, |
| 347 | &input.commands, |
| 348 | )?; |
| 349 | if gates.is_empty() { |
| 350 | let verifier_verdict = VerifierVerdict::from_counts(0, 0, 0); |
| 351 | let output = RunVerifiersOutput { |
| 352 | success: false, |
| 353 | profile: profile.as_str().to_string(), |
| 354 | level: level.as_str().to_string(), |
| 355 | workspace: context.workspace.display().to_string(), |
| 356 | gate_count: 0, |
| 357 | passed: 0, |
| 358 | failed: 0, |
| 359 | skipped: 0, |
| 360 | verifier_verdict, |
| 361 | hunt_verdict: verifier_verdict.hunt_verdict().to_string(), |
| 362 | goal_status: verifier_verdict.goal_status().to_string(), |
| 363 | summary: "No verifier gates were detected. Provide custom commands or choose a profile that matches this workspace.".to_string(), |
| 364 | gates: Vec::new(), |
| 365 | }; |
| 366 | return verifier_tool_result(&output); |
| 367 | } |
| 368 | |
| 369 | if input.background { |
| 370 | return start_background_gates(context, profile, level, gates); |
| 371 | } |
| 372 | |
| 373 | let mut handles = Vec::with_capacity(gates.len()); |
| 374 | for gate in gates { |
| 375 | handles.push(tokio::task::spawn_blocking(move || run_gate(gate))); |
| 376 | } |
| 377 | |
| 378 | let mut results = Vec::with_capacity(handles.len()); |
| 379 | for handle in handles { |
| 380 | match handle.await { |
| 381 | Ok(result) => results.push(result), |
| 382 | Err(err) => results.push(GateResult { |
| 383 | name: "internal-join".to_string(), |
| 384 | ecosystem: "internal".to_string(), |
| 385 | status: GateStatus::Failed, |
| 386 | command: "tokio::task::spawn_blocking".to_string(), |
| 387 | cwd: context.workspace.display().to_string(), |
| 388 | exit_code: None, |
| 389 | duration_ms: 0, |
| 390 | stdout: String::new(), |
| 391 | stderr: format!("Verifier task join failed: {err}"), |
| 392 | stdout_truncated: false, |
| 393 | stderr_truncated: false, |
| 394 | skipped_reason: None, |
| 395 | }), |
| 396 | } |
| 397 | } |
| 398 | results.sort_by(|a, b| a.name.cmp(&b.name)); |
| 399 | |
| 400 | let passed = results |
| 401 | .iter() |
| 402 | .filter(|result| result.status == GateStatus::Passed) |
| 403 | .count(); |
| 404 | let failed = results |
| 405 | .iter() |
| 406 | .filter(|result| result.status == GateStatus::Failed) |
| 407 | .count(); |
| 408 | let skipped = results |
| 409 | .iter() |
| 410 | .filter(|result| result.status == GateStatus::Skipped) |
| 411 | .count(); |
| 412 | let success = failed == 0 && skipped == 0; |
| 413 | let verifier_verdict = VerifierVerdict::from_counts(results.len(), failed, skipped); |
| 414 | let summary = if success { |
| 415 | format!("All {passed} verifier gates passed.") |
| 416 | } else { |
| 417 | format!("{passed} passed, {failed} failed, {skipped} skipped.") |
| 418 | }; |
| 419 | |
| 420 | let output = RunVerifiersOutput { |
| 421 | success, |
| 422 | profile: profile.as_str().to_string(), |
| 423 | level: level.as_str().to_string(), |
| 424 | workspace: context.workspace.display().to_string(), |
| 425 | gate_count: results.len(), |
| 426 | passed, |
| 427 | failed, |
| 428 | skipped, |
| 429 | verifier_verdict, |
| 430 | hunt_verdict: verifier_verdict.hunt_verdict().to_string(), |
| 431 | goal_status: verifier_verdict.goal_status().to_string(), |
| 432 | summary, |
| 433 | gates: results, |
| 434 | }; |
| 435 | |
| 436 | verifier_tool_result(&output) |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | /// Run quick auto verifier gates after a successful workflow completion (#4013). |
| 441 | pub(crate) async fn run_workflow_completion_gates( |
| 442 | context: &ToolContext, |
| 443 | ) -> Result<Value, ToolError> { |
| 444 | let gates = build_gate_plan( |
| 445 | context, |
| 446 | VerifierProfile::Auto, |
| 447 | VerifierLevel::Quick, |
| 448 | DEFAULT_MAX_PYTHON_FILES, |
| 449 | &[], |
| 450 | )?; |
| 451 | if gates.is_empty() { |
| 452 | return Ok(json!({ |
| 453 | "success": false, |
| 454 | "profile": "auto", |
| 455 | "level": "quick", |
| 456 | "gate_count": 0, |
| 457 | "summary": "No verifier gates detected for this workspace.", |
| 458 | "gates": [], |
| 459 | })); |
| 460 | } |
| 461 | |
| 462 | let workspace = context.workspace.display().to_string(); |
| 463 | let mut handles = Vec::with_capacity(gates.len()); |
| 464 | for gate in gates { |
| 465 | handles.push(tokio::task::spawn_blocking(move || run_gate(gate))); |
| 466 | } |
| 467 | |
| 468 | let mut results = Vec::with_capacity(handles.len()); |
| 469 | for handle in handles { |
| 470 | match handle.await { |
| 471 | Ok(result) => results.push(result), |
| 472 | Err(err) => results.push(GateResult { |
| 473 | name: "internal-join".to_string(), |
| 474 | ecosystem: "internal".to_string(), |
| 475 | status: GateStatus::Failed, |
| 476 | command: "tokio::task::spawn_blocking".to_string(), |
| 477 | cwd: workspace.clone(), |
| 478 | exit_code: None, |
| 479 | duration_ms: 0, |
| 480 | stdout: String::new(), |
| 481 | stderr: format!("Verifier task join failed: {err}"), |
| 482 | stdout_truncated: false, |
| 483 | stderr_truncated: false, |
| 484 | skipped_reason: None, |
| 485 | }), |
| 486 | } |
| 487 | } |
| 488 | results.sort_by(|a, b| a.name.cmp(&b.name)); |
| 489 | |
| 490 | let passed = results |
| 491 | .iter() |
| 492 | .filter(|result| result.status == GateStatus::Passed) |
| 493 | .count(); |
| 494 | let failed = results |
| 495 | .iter() |
| 496 | .filter(|result| result.status == GateStatus::Failed) |
| 497 | .count(); |
| 498 | let skipped = results |
| 499 | .iter() |
| 500 | .filter(|result| result.status == GateStatus::Skipped) |
| 501 | .count(); |
| 502 | let success = failed == 0 && skipped == 0; |
| 503 | if !success { |
| 504 | return Err(ToolError::execution_failed(format!( |
| 505 | "{passed} passed, {failed} failed, {skipped} skipped" |
| 506 | ))); |
| 507 | } |
| 508 | Ok(json!({ |
| 509 | "success": true, |
| 510 | "profile": "auto", |
| 511 | "level": "quick", |
| 512 | "gate_count": results.len(), |
| 513 | "passed": passed, |
| 514 | "failed": failed, |
| 515 | "skipped": skipped, |
| 516 | "summary": format!("All {passed} verifier gates passed."), |
| 517 | "gates": results, |
| 518 | })) |
| 519 | } |
| 520 | |
| 521 | fn verifier_tool_result(output: &RunVerifiersOutput) -> Result<ToolResult, ToolError> { |
| 522 | ToolResult::json(output) |
| 523 | .map_err(|err| ToolError::execution_failed(err.to_string())) |
| 524 | .map(|result| { |
| 525 | result.with_metadata(json!({ |
| 526 | "verifier_verdict": output.verifier_verdict, |
| 527 | "hunt_verdict": output.hunt_verdict, |
| 528 | "goal_status": output.goal_status, |
| 529 | "task_updates": { |
| 530 | "hunt_verdict": output.hunt_verdict |
| 531 | } |
| 532 | })) |
| 533 | }) |
| 534 | } |
| 535 | |
| 536 | fn start_background_gates( |
| 537 | context: &ToolContext, |
| 538 | profile: VerifierProfile, |
| 539 | level: VerifierLevel, |
| 540 | gates: Vec<VerifierGate>, |
| 541 | ) -> Result<ToolResult, ToolError> { |
| 542 | let mut jobs = Vec::with_capacity(gates.len()); |
| 543 | let mut started = 0usize; |
| 544 | let mut skipped = 0usize; |
| 545 | let mut failed_to_start = 0usize; |
| 546 | |
| 547 | for gate in gates { |
| 548 | let cwd = gate.cwd.display().to_string(); |
| 549 | let Some(program) = gate.program.as_deref() else { |
| 550 | skipped += 1; |
| 551 | jobs.push(BackgroundGateJob { |
| 552 | name: gate.name, |
| 553 | ecosystem: gate.ecosystem, |
| 554 | status: "skipped".to_string(), |
| 555 | command: String::new(), |
| 556 | cwd, |
| 557 | task_id: None, |
| 558 | skipped_reason: gate.skipped_reason, |
| 559 | error: None, |
| 560 | }); |
| 561 | continue; |
| 562 | }; |
| 563 | |
| 564 | let command = render_gate_command(program, &gate.args)?; |
| 565 | let env: HashMap<String, String> = gate.env.into_iter().collect(); |
| 566 | let spawn_result = { |
| 567 | let mut manager = context |
| 568 | .shell_manager |
| 569 | .lock() |
| 570 | .map_err(|_| ToolError::execution_failed("shell manager lock poisoned"))?; |
| 571 | manager.execute_with_options_env( |
| 572 | &command, |
| 573 | Some(&cwd), |
| 574 | BACKGROUND_GATE_TIMEOUT_MS, |
| 575 | true, |
| 576 | None, |
| 577 | false, |
| 578 | context.elevated_sandbox_policy.clone(), |
| 579 | env, |
| 580 | ) |
| 581 | }; |
| 582 | |
| 583 | match spawn_result { |
| 584 | Ok(result) => { |
| 585 | started += 1; |
| 586 | jobs.push(BackgroundGateJob { |
| 587 | name: gate.name, |
| 588 | ecosystem: gate.ecosystem, |
| 589 | status: "running".to_string(), |
| 590 | command, |
| 591 | cwd, |
| 592 | task_id: result.task_id, |
| 593 | skipped_reason: None, |
| 594 | error: None, |
| 595 | }); |
| 596 | } |
| 597 | Err(err) => { |
| 598 | failed_to_start += 1; |
| 599 | jobs.push(BackgroundGateJob { |
| 600 | name: gate.name, |
| 601 | ecosystem: gate.ecosystem, |
| 602 | status: "failed_to_start".to_string(), |
| 603 | command, |
| 604 | cwd, |
| 605 | task_id: None, |
| 606 | skipped_reason: None, |
| 607 | error: Some(err.to_string()), |
| 608 | }); |
| 609 | } |
| 610 | } |
| 611 | } |
| 612 | |
| 613 | jobs.sort_by(|a, b| a.name.cmp(&b.name)); |
| 614 | let success = failed_to_start == 0 && started > 0; |
| 615 | let summary = if failed_to_start == 0 { |
| 616 | format!( |
| 617 | "Started {started} verifier gate(s) in the background; {skipped} skipped. Completion is tracked in task/status state. Continue inspecting or implementing while they run." |
| 618 | ) |
| 619 | } else { |
| 620 | format!( |
| 621 | "Started {started} verifier gate(s), failed to start {failed_to_start}, and skipped {skipped}. Completion is tracked in task/status state. Continue inspecting or implementing while they run." |
| 622 | ) |
| 623 | }; |
| 624 | let task_ids = jobs |
| 625 | .iter() |
| 626 | .filter_map(|job| job.task_id.clone()) |
| 627 | .collect::<Vec<_>>(); |
| 628 | let output = RunVerifiersBackgroundOutput { |
| 629 | success, |
| 630 | profile: profile.as_str().to_string(), |
| 631 | level: level.as_str().to_string(), |
| 632 | workspace: context.workspace.display().to_string(), |
| 633 | background: true, |
| 634 | gate_count: jobs.len(), |
| 635 | started, |
| 636 | skipped, |
| 637 | failed_to_start, |
| 638 | summary, |
| 639 | jobs, |
| 640 | }; |
| 641 | |
| 642 | let mut result = |
| 643 | ToolResult::json(&output).map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 644 | result.success = success; |
| 645 | Ok(result.with_metadata(json!({ |
| 646 | "backgrounded": true, |
| 647 | "detached_start": true, |
| 648 | "verifier_background": true, |
| 649 | "auto_resume_on_completion": false, |
| 650 | "completion_surface": "task_status", |
| 651 | "background_policy": "nonblocking", |
| 652 | "task_ids": task_ids, |
| 653 | "poll_with": ["exec_shell_wait", "task_shell_wait"] |
| 654 | }))) |
| 655 | } |
| 656 | |
| 657 | fn render_gate_command(program: &str, args: &[String]) -> Result<String, ToolError> { |
| 658 | try_join(std::iter::once(program).chain(args.iter().map(String::as_str))) |
| 659 | .map_err(|err| ToolError::execution_failed(format!("failed to render gate command: {err}"))) |
| 660 | } |
| 661 | |
| 662 | fn build_gate_plan( |
| 663 | context: &ToolContext, |
| 664 | profile: VerifierProfile, |
| 665 | level: VerifierLevel, |
| 666 | max_python_files: usize, |
| 667 | custom_commands: &[CustomVerifierInput], |
| 668 | ) -> Result<Vec<VerifierGate>, ToolError> { |
| 669 | let workspace = &context.workspace; |
| 670 | let mut gates = Vec::new(); |
| 671 | |
| 672 | if profile == VerifierProfile::Auto && workspace.join(".git").exists() { |
| 673 | gates.push(gate( |
| 674 | "git-whitespace", |
| 675 | "git", |
| 676 | workspace, |
| 677 | "git", |
| 678 | ["diff", "--check"], |
| 679 | )); |
| 680 | } |
| 681 | |
| 682 | if profile_matches(profile, VerifierProfile::Rust) && workspace.join("Cargo.toml").exists() { |
| 683 | add_rust_gates(&mut gates, workspace, level); |
| 684 | } |
| 685 | if profile_matches(profile, VerifierProfile::Node) && workspace.join("package.json").exists() { |
| 686 | add_node_gates(&mut gates, workspace, level); |
| 687 | } |
| 688 | if profile_matches(profile, VerifierProfile::Python) && has_python_project(workspace) { |
| 689 | add_python_gates(&mut gates, workspace, level, max_python_files); |
| 690 | } |
| 691 | if profile_matches(profile, VerifierProfile::Go) && workspace.join("go.mod").exists() { |
| 692 | add_go_gates(&mut gates, workspace, level); |
| 693 | } |
| 694 | |
| 695 | for custom in custom_commands { |
| 696 | gates.push(custom_gate(context, custom)?); |
| 697 | } |
| 698 | |
| 699 | Ok(gates) |
| 700 | } |
| 701 | |
| 702 | fn profile_matches(selected: VerifierProfile, candidate: VerifierProfile) -> bool { |
| 703 | selected == VerifierProfile::Auto || selected == candidate |
| 704 | } |
| 705 | |
| 706 | fn add_rust_gates(gates: &mut Vec<VerifierGate>, workspace: &Path, level: VerifierLevel) { |
| 707 | let locked = workspace.join("Cargo.lock").exists(); |
| 708 | gates.push(gate( |
| 709 | "rust-fmt", |
| 710 | "rust", |
| 711 | workspace, |
| 712 | "cargo", |
| 713 | ["fmt", "--all", "--", "--check"], |
| 714 | )); |
| 715 | |
| 716 | let metadata_args = if locked { |
| 717 | vec!["metadata", "--locked", "--format-version", "1", "--no-deps"] |
| 718 | } else { |
| 719 | vec!["metadata", "--format-version", "1", "--no-deps"] |
| 720 | }; |
| 721 | gates.push(gate_vec( |
| 722 | "rust-metadata", |
| 723 | "rust", |
| 724 | workspace, |
| 725 | "cargo", |
| 726 | metadata_args, |
| 727 | )); |
| 728 | |
| 729 | let mut check_args = vec!["check", "--workspace", "--all-targets"]; |
| 730 | if locked { |
| 731 | check_args.push("--locked"); |
| 732 | } |
| 733 | gates.push(gate_vec( |
| 734 | "rust-check", |
| 735 | "rust", |
| 736 | workspace, |
| 737 | "cargo", |
| 738 | check_args, |
| 739 | )); |
| 740 | |
| 741 | if level == VerifierLevel::Full { |
| 742 | let mut clippy_args = vec!["clippy", "--workspace", "--all-targets", "--all-features"]; |
| 743 | if locked { |
| 744 | clippy_args.push("--locked"); |
| 745 | } |
| 746 | clippy_args.extend(["--", "-D", "warnings"]); |
| 747 | gates.push(gate_vec( |
| 748 | "rust-clippy", |
| 749 | "rust", |
| 750 | workspace, |
| 751 | "cargo", |
| 752 | clippy_args, |
| 753 | )); |
| 754 | |
| 755 | let mut test_args = vec!["test", "--workspace", "--all-features"]; |
| 756 | if locked { |
| 757 | test_args.push("--locked"); |
| 758 | } |
| 759 | gates.push(gate_vec("rust-test", "rust", workspace, "cargo", test_args)); |
| 760 | } |
| 761 | } |
| 762 | |
| 763 | fn add_node_gates(gates: &mut Vec<VerifierGate>, workspace: &Path, level: VerifierLevel) { |
| 764 | let scripts = package_json_scripts(workspace); |
| 765 | let Some(scripts) = scripts else { |
| 766 | gates.push(skipped_gate( |
| 767 | "node-package-json", |
| 768 | "node", |
| 769 | workspace, |
| 770 | "package.json is missing or could not be parsed", |
| 771 | )); |
| 772 | return; |
| 773 | }; |
| 774 | let package_manager = detect_node_package_manager(workspace); |
| 775 | for script in ["format:check", "check", "typecheck", "lint"] { |
| 776 | if has_meaningful_script(&scripts, script) { |
| 777 | gates.push(node_script_gate(workspace, &package_manager, script)); |
| 778 | } |
| 779 | } |
| 780 | if level == VerifierLevel::Full && has_meaningful_script(&scripts, "test") { |
| 781 | gates.push(node_script_gate(workspace, &package_manager, "test")); |
| 782 | } |
| 783 | } |
| 784 | |
| 785 | fn add_python_gates( |
| 786 | gates: &mut Vec<VerifierGate>, |
| 787 | workspace: &Path, |
| 788 | level: VerifierLevel, |
| 789 | max_python_files: usize, |
| 790 | ) { |
| 791 | let python_files = collect_python_files(workspace, max_python_files); |
| 792 | match python_files { |
| 793 | PythonFiles::Files(files) if !files.is_empty() => { |
| 794 | gates.push(python_syntax_gate(workspace, &files)); |
| 795 | } |
| 796 | PythonFiles::TooMany { limit, found } => gates.push(skipped_gate( |
| 797 | "python-syntax", |
| 798 | "python", |
| 799 | workspace, |
| 800 | format!( |
| 801 | "found more than {limit} Python files ({found}); raise max_python_files to verify them" |
| 802 | ), |
| 803 | )), |
| 804 | PythonFiles::Files(_) => {} |
| 805 | } |
| 806 | |
| 807 | if level == VerifierLevel::Full && has_pytest_signal(workspace) { |
| 808 | gates.push(python_module_gate( |
| 809 | "python-pytest", |
| 810 | workspace, |
| 811 | ["-m", "pytest"], |
| 812 | )); |
| 813 | } |
| 814 | } |
| 815 | |
| 816 | fn add_go_gates(gates: &mut Vec<VerifierGate>, workspace: &Path, level: VerifierLevel) { |
| 817 | gates.push(gate("go-test", "go", workspace, "go", ["test", "./..."])); |
| 818 | if level == VerifierLevel::Full { |
| 819 | gates.push(gate("go-vet", "go", workspace, "go", ["vet", "./..."])); |
| 820 | } |
| 821 | } |
| 822 | |
| 823 | fn gate<const N: usize>( |
| 824 | name: &str, |
| 825 | ecosystem: &str, |
| 826 | cwd: &Path, |
| 827 | program: &str, |
| 828 | args: [&str; N], |
| 829 | ) -> VerifierGate { |
| 830 | gate_vec(name, ecosystem, cwd, program, args) |
| 831 | } |
| 832 | |
| 833 | fn gate_vec<I, S>(name: &str, ecosystem: &str, cwd: &Path, program: &str, args: I) -> VerifierGate |
| 834 | where |
| 835 | I: IntoIterator<Item = S>, |
| 836 | S: AsRef<str>, |
| 837 | { |
| 838 | VerifierGate { |
| 839 | name: name.to_string(), |
| 840 | ecosystem: ecosystem.to_string(), |
| 841 | cwd: cwd.to_path_buf(), |
| 842 | program: Some(program.to_string()), |
| 843 | args: args |
| 844 | .into_iter() |
| 845 | .map(|arg| arg.as_ref().to_string()) |
| 846 | .collect(), |
| 847 | env: Vec::new(), |
| 848 | skipped_reason: None, |
| 849 | } |
| 850 | } |
| 851 | |
| 852 | fn skipped_gate( |
| 853 | name: &str, |
| 854 | ecosystem: &str, |
| 855 | cwd: &Path, |
| 856 | reason: impl Into<String>, |
| 857 | ) -> VerifierGate { |
| 858 | VerifierGate { |
| 859 | name: name.to_string(), |
| 860 | ecosystem: ecosystem.to_string(), |
| 861 | cwd: cwd.to_path_buf(), |
| 862 | program: None, |
| 863 | args: Vec::new(), |
| 864 | env: Vec::new(), |
| 865 | skipped_reason: Some(reason.into()), |
| 866 | } |
| 867 | } |
| 868 | |
| 869 | fn custom_gate( |
| 870 | context: &ToolContext, |
| 871 | custom: &CustomVerifierInput, |
| 872 | ) -> Result<VerifierGate, ToolError> { |
| 873 | if custom.name.trim().is_empty() { |
| 874 | return Err(ToolError::invalid_input( |
| 875 | "Custom verifier command is missing 'name'", |
| 876 | )); |
| 877 | } |
| 878 | if custom.program.trim().is_empty() { |
| 879 | return Err(ToolError::invalid_input(format!( |
| 880 | "Custom verifier '{}' is missing 'program'", |
| 881 | custom.name |
| 882 | ))); |
| 883 | } |
| 884 | let cwd = match custom.cwd.as_deref() { |
| 885 | Some(raw) if !raw.trim().is_empty() => context.resolve_path(raw)?, |
| 886 | _ => context.workspace.clone(), |
| 887 | }; |
| 888 | Ok(VerifierGate { |
| 889 | name: custom.name.clone(), |
| 890 | ecosystem: "custom".to_string(), |
| 891 | cwd, |
| 892 | program: Some(custom.program.clone()), |
| 893 | args: custom.args.clone(), |
| 894 | env: Vec::new(), |
| 895 | skipped_reason: None, |
| 896 | }) |
| 897 | } |
| 898 | |
| 899 | fn node_script_gate( |
| 900 | workspace: &Path, |
| 901 | package_manager: &NodePackageManager, |
| 902 | script: &str, |
| 903 | ) -> VerifierGate { |
| 904 | let (program, args) = package_manager.command_for_script(script); |
| 905 | gate_vec(&format!("node-{script}"), "node", workspace, program, args) |
| 906 | } |
| 907 | |
| 908 | fn python_syntax_gate(workspace: &Path, files: &[PathBuf]) -> VerifierGate { |
| 909 | let Some((program, mut args)) = python_command_parts() else { |
| 910 | return skipped_gate( |
| 911 | "python-syntax", |
| 912 | "python", |
| 913 | workspace, |
| 914 | "Python interpreter is not installed or not in PATH", |
| 915 | ); |
| 916 | }; |
| 917 | args.push("-c".to_string()); |
| 918 | args.push(PYTHON_SYNTAX_SCRIPT.to_string()); |
| 919 | args.extend(files.iter().map(|path| path.display().to_string())); |
| 920 | let mut gate = gate_vec("python-syntax", "python", workspace, &program, args); |
| 921 | gate.env |
| 922 | .push(("PYTHONDONTWRITEBYTECODE".to_string(), "1".to_string())); |
| 923 | gate |
| 924 | } |
| 925 | |
| 926 | fn python_module_gate<const N: usize>( |
| 927 | name: &str, |
| 928 | workspace: &Path, |
| 929 | module_args: [&str; N], |
| 930 | ) -> VerifierGate { |
| 931 | let Some((program, mut args)) = python_command_parts() else { |
| 932 | return skipped_gate( |
| 933 | name, |
| 934 | "python", |
| 935 | workspace, |
| 936 | "Python interpreter is not installed or not in PATH", |
| 937 | ); |
| 938 | }; |
| 939 | args.extend(module_args.into_iter().map(str::to_string)); |
| 940 | gate_vec(name, "python", workspace, &program, args) |
| 941 | } |
| 942 | |
| 943 | fn python_command_parts() -> Option<(String, Vec<String>)> { |
| 944 | let spec = crate::dependencies::Python::resolve()?; |
| 945 | Some(crate::dependencies::split_interpreter_spec(&spec)) |
| 946 | } |
| 947 | |
| 948 | const PYTHON_SYNTAX_SCRIPT: &str = r#" |
| 949 | import ast |
| 950 | import pathlib |
| 951 | import sys |
| 952 | |
| 953 | failures = [] |
| 954 | for raw in sys.argv[1:]: |
| 955 | path = pathlib.Path(raw) |
| 956 | try: |
| 957 | source = path.read_text(encoding="utf-8") |
| 958 | ast.parse(source, filename=raw) |
| 959 | except Exception as exc: |
| 960 | failures.append(f"{raw}: {exc.__class__.__name__}: {exc}") |
| 961 | |
| 962 | if failures: |
| 963 | print("\n".join(failures), file=sys.stderr) |
| 964 | sys.exit(1) |
| 965 | |
| 966 | print(f"parsed {len(sys.argv) - 1} Python file(s)") |
| 967 | "#; |
| 968 | |
| 969 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 970 | enum NodePackageManager { |
| 971 | Npm, |
| 972 | Pnpm, |
| 973 | Yarn, |
| 974 | Bun, |
| 975 | } |
| 976 | |
| 977 | impl NodePackageManager { |
| 978 | fn command_for_script(self, script: &str) -> (&'static str, Vec<String>) { |
| 979 | match self { |
| 980 | Self::Npm => ("npm", vec!["run".to_string(), script.to_string()]), |
| 981 | Self::Pnpm => ("pnpm", vec!["run".to_string(), script.to_string()]), |
| 982 | Self::Yarn => ("yarn", vec!["run".to_string(), script.to_string()]), |
| 983 | Self::Bun => ("bun", vec!["run".to_string(), script.to_string()]), |
| 984 | } |
| 985 | } |
| 986 | } |
| 987 | |
| 988 | fn detect_node_package_manager(workspace: &Path) -> NodePackageManager { |
| 989 | if workspace.join("pnpm-lock.yaml").exists() { |
| 990 | NodePackageManager::Pnpm |
| 991 | } else if workspace.join("yarn.lock").exists() { |
| 992 | NodePackageManager::Yarn |
| 993 | } else if workspace.join("bun.lock").exists() || workspace.join("bun.lockb").exists() { |
| 994 | NodePackageManager::Bun |
| 995 | } else { |
| 996 | NodePackageManager::Npm |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | fn package_json_scripts(workspace: &Path) -> Option<HashMap<String, String>> { |
| 1001 | let raw = fs::read_to_string(workspace.join("package.json")).ok()?; |
| 1002 | let parsed = serde_json::from_str::<serde_json::Value>(&raw).ok()?; |
| 1003 | let scripts = parsed.get("scripts")?.as_object()?; |
| 1004 | Some( |
| 1005 | scripts |
| 1006 | .iter() |
| 1007 | .filter_map(|(key, value)| { |
| 1008 | value |
| 1009 | .as_str() |
| 1010 | .map(|script| (key.clone(), script.to_string())) |
| 1011 | }) |
| 1012 | .collect(), |
| 1013 | ) |
| 1014 | } |
| 1015 | |
| 1016 | fn has_meaningful_script(scripts: &HashMap<String, String>, name: &str) -> bool { |
| 1017 | let Some(script) = scripts.get(name).map(|value| value.trim()) else { |
| 1018 | return false; |
| 1019 | }; |
| 1020 | !(script.is_empty() |
| 1021 | || name == "test" |
| 1022 | && script.contains("Error: no test specified") |
| 1023 | && script.contains("exit 1")) |
| 1024 | } |
| 1025 | |
| 1026 | fn has_python_project(workspace: &Path) -> bool { |
| 1027 | workspace.join("pyproject.toml").exists() |
| 1028 | || workspace.join("setup.py").exists() |
| 1029 | || workspace.join("setup.cfg").exists() |
| 1030 | || workspace.join("requirements.txt").exists() |
| 1031 | || match collect_python_files(workspace, 1) { |
| 1032 | PythonFiles::Files(files) => !files.is_empty(), |
| 1033 | PythonFiles::TooMany { .. } => true, |
| 1034 | } |
| 1035 | } |
| 1036 | |
| 1037 | fn has_pytest_signal(workspace: &Path) -> bool { |
| 1038 | if workspace.join("pytest.ini").exists() |
| 1039 | || workspace.join("tox.ini").exists() |
| 1040 | || workspace.join("tests").is_dir() |
| 1041 | { |
| 1042 | return true; |
| 1043 | } |
| 1044 | let pyproject = workspace.join("pyproject.toml"); |
| 1045 | fs::read_to_string(pyproject) |
| 1046 | .map(|raw| raw.contains("pytest") || raw.contains("[tool.pytest")) |
| 1047 | .unwrap_or(false) |
| 1048 | } |
| 1049 | |
| 1050 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 1051 | enum PythonFiles { |
| 1052 | Files(Vec<PathBuf>), |
| 1053 | TooMany { limit: usize, found: usize }, |
| 1054 | } |
| 1055 | |
| 1056 | fn collect_python_files(workspace: &Path, limit: usize) -> PythonFiles { |
| 1057 | let mut files = BTreeSet::new(); |
| 1058 | collect_python_files_inner(workspace, workspace, limit, &mut files); |
| 1059 | let found = files.len(); |
| 1060 | if found > limit { |
| 1061 | PythonFiles::TooMany { limit, found } |
| 1062 | } else { |
| 1063 | PythonFiles::Files(files.into_iter().collect()) |
| 1064 | } |
| 1065 | } |
| 1066 | |
| 1067 | fn collect_python_files_inner( |
| 1068 | root: &Path, |
| 1069 | dir: &Path, |
| 1070 | limit: usize, |
| 1071 | files: &mut BTreeSet<PathBuf>, |
| 1072 | ) { |
| 1073 | if files.len() > limit { |
| 1074 | return; |
| 1075 | } |
| 1076 | let Ok(entries) = fs::read_dir(dir) else { |
| 1077 | return; |
| 1078 | }; |
| 1079 | for entry in entries.flatten() { |
| 1080 | if files.len() > limit { |
| 1081 | return; |
| 1082 | } |
| 1083 | let path = entry.path(); |
| 1084 | let name = entry.file_name(); |
| 1085 | if path.is_dir() { |
| 1086 | if should_skip_dir_name(&name.to_string_lossy()) { |
| 1087 | continue; |
| 1088 | } |
| 1089 | collect_python_files_inner(root, &path, limit, files); |
| 1090 | } else if path.extension().and_then(|ext| ext.to_str()) == Some("py") |
| 1091 | && let Ok(relative) = path.strip_prefix(root) |
| 1092 | { |
| 1093 | files.insert(relative.to_path_buf()); |
| 1094 | } |
| 1095 | } |
| 1096 | } |
| 1097 | |
| 1098 | fn should_skip_dir_name(name: &str) -> bool { |
| 1099 | matches!( |
| 1100 | name, |
| 1101 | ".git" |
| 1102 | | ".hg" |
| 1103 | | ".svn" |
| 1104 | | ".venv" |
| 1105 | | "venv" |
| 1106 | | "env" |
| 1107 | | "__pycache__" |
| 1108 | | ".mypy_cache" |
| 1109 | | ".pytest_cache" |
| 1110 | | ".tox" |
| 1111 | | "node_modules" |
| 1112 | | "target" |
| 1113 | | "dist" |
| 1114 | | "build" |
| 1115 | ) |
| 1116 | } |
| 1117 | |
| 1118 | fn run_gate(gate: VerifierGate) -> GateResult { |
| 1119 | let command = render_command(gate.program.as_deref(), &gate.args); |
| 1120 | if let Some(reason) = gate.skipped_reason { |
| 1121 | return GateResult { |
| 1122 | name: gate.name, |
| 1123 | ecosystem: gate.ecosystem, |
| 1124 | status: GateStatus::Skipped, |
| 1125 | command, |
| 1126 | cwd: gate.cwd.display().to_string(), |
| 1127 | exit_code: None, |
| 1128 | duration_ms: 0, |
| 1129 | stdout: String::new(), |
| 1130 | stderr: String::new(), |
| 1131 | stdout_truncated: false, |
| 1132 | stderr_truncated: false, |
| 1133 | skipped_reason: Some(reason), |
| 1134 | }; |
| 1135 | } |
| 1136 | |
| 1137 | let Some(program) = gate.program else { |
| 1138 | return GateResult { |
| 1139 | name: gate.name, |
| 1140 | ecosystem: gate.ecosystem, |
| 1141 | status: GateStatus::Skipped, |
| 1142 | command, |
| 1143 | cwd: gate.cwd.display().to_string(), |
| 1144 | exit_code: None, |
| 1145 | duration_ms: 0, |
| 1146 | stdout: String::new(), |
| 1147 | stderr: String::new(), |
| 1148 | stdout_truncated: false, |
| 1149 | stderr_truncated: false, |
| 1150 | skipped_reason: Some("verifier has no executable program".to_string()), |
| 1151 | }; |
| 1152 | }; |
| 1153 | |
| 1154 | let started = Instant::now(); |
| 1155 | let mut cmd = Command::new(&program); |
| 1156 | cmd.args(&gate.args) |
| 1157 | .current_dir(&gate.cwd) |
| 1158 | .stdout(Stdio::piped()) |
| 1159 | .stderr(Stdio::piped()); |
| 1160 | for (key, value) in &gate.env { |
| 1161 | cmd.env(key, value); |
| 1162 | } |
| 1163 | |
| 1164 | let output = match cmd.output() { |
| 1165 | Ok(output) => output, |
| 1166 | Err(err) if err.kind() == std::io::ErrorKind::NotFound => { |
| 1167 | return GateResult { |
| 1168 | name: gate.name, |
| 1169 | ecosystem: gate.ecosystem, |
| 1170 | status: GateStatus::Skipped, |
| 1171 | command, |
| 1172 | cwd: gate.cwd.display().to_string(), |
| 1173 | exit_code: None, |
| 1174 | duration_ms: started.elapsed().as_millis() as u64, |
| 1175 | stdout: String::new(), |
| 1176 | stderr: String::new(), |
| 1177 | stdout_truncated: false, |
| 1178 | stderr_truncated: false, |
| 1179 | skipped_reason: Some(format!("{program} is not installed or not in PATH")), |
| 1180 | }; |
| 1181 | } |
| 1182 | Err(err) => { |
| 1183 | return GateResult { |
| 1184 | name: gate.name, |
| 1185 | ecosystem: gate.ecosystem, |
| 1186 | status: GateStatus::Failed, |
| 1187 | command, |
| 1188 | cwd: gate.cwd.display().to_string(), |
| 1189 | exit_code: None, |
| 1190 | duration_ms: started.elapsed().as_millis() as u64, |
| 1191 | stdout: String::new(), |
| 1192 | stderr: format!("Failed to spawn verifier: {err}"), |
| 1193 | stdout_truncated: false, |
| 1194 | stderr_truncated: false, |
| 1195 | skipped_reason: None, |
| 1196 | }; |
| 1197 | } |
| 1198 | }; |
| 1199 | |
| 1200 | let (stdout, stdout_truncated) = truncate_with_note( |
| 1201 | &String::from_utf8_lossy(&output.stdout), |
| 1202 | MAX_GATE_OUTPUT_CHARS, |
| 1203 | ); |
| 1204 | let (stderr, stderr_truncated) = truncate_with_note( |
| 1205 | &String::from_utf8_lossy(&output.stderr), |
| 1206 | MAX_GATE_OUTPUT_CHARS, |
| 1207 | ); |
| 1208 | GateResult { |
| 1209 | name: gate.name, |
| 1210 | ecosystem: gate.ecosystem, |
| 1211 | status: if output.status.success() { |
| 1212 | GateStatus::Passed |
| 1213 | } else { |
| 1214 | GateStatus::Failed |
| 1215 | }, |
| 1216 | command, |
| 1217 | cwd: gate.cwd.display().to_string(), |
| 1218 | exit_code: output.status.code(), |
| 1219 | duration_ms: started.elapsed().as_millis() as u64, |
| 1220 | stdout, |
| 1221 | stderr, |
| 1222 | stdout_truncated, |
| 1223 | stderr_truncated, |
| 1224 | skipped_reason: None, |
| 1225 | } |
| 1226 | } |
| 1227 | |
| 1228 | fn render_command(program: Option<&str>, args: &[String]) -> String { |
| 1229 | let mut parts = Vec::new(); |
| 1230 | parts.push(program.unwrap_or("<unavailable>").to_string()); |
| 1231 | parts.extend(args.iter().cloned()); |
| 1232 | parts.join(" ") |
| 1233 | } |
| 1234 | |
| 1235 | fn truncate_with_note(text: &str, max_chars: usize) -> (String, bool) { |
| 1236 | if text.chars().count() <= max_chars { |
| 1237 | return (text.to_string(), false); |
| 1238 | } |
| 1239 | let end = char_boundary_index(text, max_chars); |
| 1240 | let truncated = &text[..end]; |
| 1241 | let omitted_chars = text |
| 1242 | .chars() |
| 1243 | .count() |
| 1244 | .saturating_sub(truncated.chars().count()); |
| 1245 | ( |
| 1246 | format!( |
| 1247 | "{truncated}\n\n[output truncated to {max_chars} characters; {omitted_chars} characters omitted]" |
| 1248 | ), |
| 1249 | true, |
| 1250 | ) |
| 1251 | } |
| 1252 | |
| 1253 | fn char_boundary_index(text: &str, max_chars: usize) -> usize { |
| 1254 | if max_chars == 0 { |
| 1255 | return 0; |
| 1256 | } |
| 1257 | for (count, (idx, _)) in text.char_indices().enumerate() { |
| 1258 | if count == max_chars { |
| 1259 | return idx; |
| 1260 | } |
| 1261 | } |
| 1262 | text.len() |
| 1263 | } |
| 1264 | |
| 1265 | #[cfg(test)] |
| 1266 | mod tests { |
| 1267 | use super::*; |
| 1268 | use crate::tools::shell::ShellStatus; |
| 1269 | use std::time::Duration; |
| 1270 | use tempfile::tempdir; |
| 1271 | |
| 1272 | const BACKGROUND_COMPLETION_WAIT_MS: u64 = 30_000; |
| 1273 | |
| 1274 | fn wait_for_completed_shell( |
| 1275 | manager: &mut crate::tools::shell::ShellManager, |
| 1276 | task_id: &str, |
| 1277 | ) -> crate::tools::shell::ShellResult { |
| 1278 | let deadline = Instant::now() + Duration::from_millis(BACKGROUND_COMPLETION_WAIT_MS); |
| 1279 | |
| 1280 | loop { |
| 1281 | let result = manager |
| 1282 | .get_output(task_id, true, 1_000) |
| 1283 | .expect("background output"); |
| 1284 | if result.status != ShellStatus::Running || Instant::now() >= deadline { |
| 1285 | return result; |
| 1286 | } |
| 1287 | std::thread::sleep(Duration::from_millis(50)); |
| 1288 | } |
| 1289 | } |
| 1290 | |
| 1291 | #[test] |
| 1292 | fn run_verifiers_requires_user_approval() { |
| 1293 | let tool = RunVerifiersTool; |
| 1294 | assert_eq!( |
| 1295 | tool.approval_requirement(), |
| 1296 | ApprovalRequirement::Required, |
| 1297 | "run_verifiers executes project code and must require approval" |
| 1298 | ); |
| 1299 | } |
| 1300 | |
| 1301 | #[test] |
| 1302 | fn run_verifiers_background_advertises_detached_start() { |
| 1303 | let tool = RunVerifiersTool; |
| 1304 | let schema = tool.input_schema(); |
| 1305 | let background_description = schema["properties"]["background"]["description"] |
| 1306 | .as_str() |
| 1307 | .expect("background description"); |
| 1308 | |
| 1309 | assert!(background_description.contains("Bash")); |
| 1310 | assert!(background_description.contains("task_shell_wait")); |
| 1311 | assert!( |
| 1312 | !background_description.contains("exec_shell"), |
| 1313 | "live descriptions must not teach the retired exec_shell name" |
| 1314 | ); |
| 1315 | assert!(tool.starts_detached_for(&json!({"background": true}))); |
| 1316 | assert!(!tool.starts_detached_for(&json!({"profile": "auto"}))); |
| 1317 | } |
| 1318 | |
| 1319 | #[test] |
| 1320 | fn auto_profile_detects_multiple_ecosystems_without_bash() { |
| 1321 | let tmp = tempdir().expect("tempdir"); |
| 1322 | fs::write(tmp.path().join("Cargo.toml"), "[workspace]\n").expect("cargo manifest"); |
| 1323 | fs::write( |
| 1324 | tmp.path().join("package.json"), |
| 1325 | r#"{"scripts":{"lint":"eslint .","test":"echo ok"}}"#, |
| 1326 | ) |
| 1327 | .expect("package json"); |
| 1328 | fs::write(tmp.path().join("main.py"), "print('ok')\n").expect("python file"); |
| 1329 | fs::write(tmp.path().join("go.mod"), "module example.com/app\n").expect("go mod"); |
| 1330 | |
| 1331 | let ctx = ToolContext::new(tmp.path()); |
| 1332 | let gates = build_gate_plan( |
| 1333 | &ctx, |
| 1334 | VerifierProfile::Auto, |
| 1335 | VerifierLevel::Quick, |
| 1336 | DEFAULT_MAX_PYTHON_FILES, |
| 1337 | &[], |
| 1338 | ) |
| 1339 | .expect("plan"); |
| 1340 | let names: BTreeSet<&str> = gates.iter().map(|gate| gate.name.as_str()).collect(); |
| 1341 | |
| 1342 | assert!(names.contains("rust-fmt")); |
| 1343 | assert!(names.contains("node-lint")); |
| 1344 | assert!(names.contains("python-syntax")); |
| 1345 | assert!(names.contains("go-test")); |
| 1346 | assert!( |
| 1347 | gates |
| 1348 | .iter() |
| 1349 | .filter_map(|gate| gate.program.as_deref()) |
| 1350 | .all(|program| program != "bash"), |
| 1351 | "built-in verifier gates must not require bash" |
| 1352 | ); |
| 1353 | } |
| 1354 | |
| 1355 | #[test] |
| 1356 | fn custom_commands_can_choose_bash_explicitly() { |
| 1357 | let tmp = tempdir().expect("tempdir"); |
| 1358 | let ctx = ToolContext::new(tmp.path()); |
| 1359 | let custom = CustomVerifierInput { |
| 1360 | name: "shell-check".to_string(), |
| 1361 | program: "bash".to_string(), |
| 1362 | args: vec!["-lc".to_string(), "echo ok".to_string()], |
| 1363 | cwd: None, |
| 1364 | }; |
| 1365 | |
| 1366 | let gate = custom_gate(&ctx, &custom).expect("custom gate"); |
| 1367 | |
| 1368 | assert_eq!(gate.program.as_deref(), Some("bash")); |
| 1369 | assert_eq!(gate.args, vec!["-lc", "echo ok"]); |
| 1370 | } |
| 1371 | |
| 1372 | #[test] |
| 1373 | fn node_default_npm_init_test_script_is_not_a_verifier() { |
| 1374 | let mut scripts = HashMap::new(); |
| 1375 | scripts.insert( |
| 1376 | "test".to_string(), |
| 1377 | "echo \"Error: no test specified\" && exit 1".to_string(), |
| 1378 | ); |
| 1379 | |
| 1380 | assert!(!has_meaningful_script(&scripts, "test")); |
| 1381 | } |
| 1382 | |
| 1383 | #[tokio::test] |
| 1384 | async fn run_verifiers_executes_custom_direct_command() { |
| 1385 | if !crate::dependencies::RustC::available() { |
| 1386 | return; |
| 1387 | } |
| 1388 | let tmp = tempdir().expect("tempdir"); |
| 1389 | let ctx = ToolContext::new(tmp.path()); |
| 1390 | let tool = RunVerifiersTool; |
| 1391 | let result = tool |
| 1392 | .execute( |
| 1393 | json!({ |
| 1394 | "profile": "auto", |
| 1395 | "commands": [ |
| 1396 | { |
| 1397 | "name": "rustc-version", |
| 1398 | "program": crate::dependencies::RustC::resolve().expect("rustc"), |
| 1399 | "args": ["--version"] |
| 1400 | } |
| 1401 | ] |
| 1402 | }), |
| 1403 | &ctx, |
| 1404 | ) |
| 1405 | .await |
| 1406 | .expect("execute"); |
| 1407 | |
| 1408 | let parsed: RunVerifiersOutput = |
| 1409 | serde_json::from_str(&result.content).expect("verifier output json"); |
| 1410 | assert!(parsed.success, "result: {}", result.content); |
| 1411 | assert_eq!(parsed.passed, 1); |
| 1412 | assert_eq!(parsed.failed, 0); |
| 1413 | assert_eq!(parsed.skipped, 0); |
| 1414 | assert!( |
| 1415 | parsed.gates[0].stdout.contains("rustc"), |
| 1416 | "stdout should include rustc version: {:?}", |
| 1417 | parsed.gates[0].stdout |
| 1418 | ); |
| 1419 | } |
| 1420 | |
| 1421 | #[tokio::test] |
| 1422 | async fn run_verifiers_emits_hunt_verdict_mapping() { |
| 1423 | let tmp = tempdir().expect("tempdir"); |
| 1424 | let ctx = ToolContext::new(tmp.path()); |
| 1425 | let tool = RunVerifiersTool; |
| 1426 | |
| 1427 | let partial = tool |
| 1428 | .execute(json!({"profile": "auto"}), &ctx) |
| 1429 | .await |
| 1430 | .expect("execute partial verifier"); |
| 1431 | assert_hunt_mapping(&partial.content, "partial", "wounded", "paused"); |
| 1432 | assert_hunt_metadata(&partial, "partial", "wounded", "paused"); |
| 1433 | |
| 1434 | if !crate::dependencies::RustC::available() { |
| 1435 | return; |
| 1436 | } |
| 1437 | |
| 1438 | let pass = tool |
| 1439 | .execute( |
| 1440 | json!({ |
| 1441 | "profile": "auto", |
| 1442 | "commands": [ |
| 1443 | { |
| 1444 | "name": "rustc-version", |
| 1445 | "program": crate::dependencies::RustC::resolve().expect("rustc"), |
| 1446 | "args": ["--version"] |
| 1447 | } |
| 1448 | ] |
| 1449 | }), |
| 1450 | &ctx, |
| 1451 | ) |
| 1452 | .await |
| 1453 | .expect("execute passing verifier"); |
| 1454 | assert_hunt_mapping(&pass.content, "pass", "hunted", "complete"); |
| 1455 | assert_hunt_metadata(&pass, "pass", "hunted", "complete"); |
| 1456 | |
| 1457 | let fail = tool |
| 1458 | .execute( |
| 1459 | json!({ |
| 1460 | "profile": "auto", |
| 1461 | "commands": [ |
| 1462 | { |
| 1463 | "name": "rustc-bad-flag", |
| 1464 | "program": crate::dependencies::RustC::resolve().expect("rustc"), |
| 1465 | "args": ["--definitely-not-a-rustc-flag"] |
| 1466 | } |
| 1467 | ] |
| 1468 | }), |
| 1469 | &ctx, |
| 1470 | ) |
| 1471 | .await |
| 1472 | .expect("execute failing verifier"); |
| 1473 | assert_hunt_mapping(&fail.content, "fail", "escaped", "blocked"); |
| 1474 | assert_hunt_metadata(&fail, "fail", "escaped", "blocked"); |
| 1475 | } |
| 1476 | |
| 1477 | fn assert_hunt_mapping(content: &str, verifier: &str, hunt: &str, goal: &str) { |
| 1478 | let parsed: Value = serde_json::from_str(content).expect("verifier output json"); |
| 1479 | assert_eq!(parsed["verifier_verdict"], verifier, "{content}"); |
| 1480 | assert_eq!(parsed["hunt_verdict"], hunt, "{content}"); |
| 1481 | assert_eq!(parsed["goal_status"], goal, "{content}"); |
| 1482 | } |
| 1483 | |
| 1484 | fn assert_hunt_metadata(result: &ToolResult, verifier: &str, hunt: &str, goal: &str) { |
| 1485 | let metadata = result.metadata.as_ref().expect("hunt metadata"); |
| 1486 | assert_eq!(metadata["verifier_verdict"], verifier, "{metadata}"); |
| 1487 | assert_eq!(metadata["hunt_verdict"], hunt, "{metadata}"); |
| 1488 | assert_eq!(metadata["goal_status"], goal, "{metadata}"); |
| 1489 | assert_eq!(metadata["task_updates"]["hunt_verdict"], hunt, "{metadata}"); |
| 1490 | } |
| 1491 | |
| 1492 | #[tokio::test] |
| 1493 | #[allow(clippy::await_holding_lock)] |
| 1494 | async fn run_verifiers_background_starts_shell_jobs_and_returns_task_ids() { |
| 1495 | if !crate::dependencies::RustC::available() { |
| 1496 | return; |
| 1497 | } |
| 1498 | // The spawned `rustc` is usually the rustup shim, which resolves its |
| 1499 | // toolchain through $HOME. Hold the process-wide env mutex so tests |
| 1500 | // that temporarily swap HOME cannot break the child process. |
| 1501 | let _env_lock = crate::test_support::lock_test_env(); |
| 1502 | let tmp = tempdir().expect("tempdir"); |
| 1503 | let ctx = ToolContext::new(tmp.path()); |
| 1504 | let tool = RunVerifiersTool; |
| 1505 | let result = tool |
| 1506 | .execute( |
| 1507 | json!({ |
| 1508 | "profile": "auto", |
| 1509 | "background": true, |
| 1510 | "commands": [ |
| 1511 | { |
| 1512 | "name": "rustc-version", |
| 1513 | "program": crate::dependencies::RustC::resolve().expect("rustc"), |
| 1514 | "args": ["--version"] |
| 1515 | } |
| 1516 | ] |
| 1517 | }), |
| 1518 | &ctx, |
| 1519 | ) |
| 1520 | .await |
| 1521 | .expect("execute"); |
| 1522 | |
| 1523 | let parsed: RunVerifiersBackgroundOutput = |
| 1524 | serde_json::from_str(&result.content).expect("background verifier output json"); |
| 1525 | assert!(parsed.success, "result: {}", result.content); |
| 1526 | assert!(parsed.background); |
| 1527 | assert_eq!(parsed.started, 1); |
| 1528 | assert_eq!(parsed.failed_to_start, 0); |
| 1529 | assert!(parsed.summary.contains("Completion is tracked")); |
| 1530 | let task_id = parsed.jobs[0] |
| 1531 | .task_id |
| 1532 | .as_deref() |
| 1533 | .expect("background task id"); |
| 1534 | let metadata = result.metadata.as_ref().expect("metadata"); |
| 1535 | assert!( |
| 1536 | metadata |
| 1537 | .get("verifier_background") |
| 1538 | .and_then(Value::as_bool) |
| 1539 | .unwrap_or(false), |
| 1540 | "metadata should mark verifier background start" |
| 1541 | ); |
| 1542 | assert_eq!( |
| 1543 | metadata |
| 1544 | .get("auto_notify_on_completion") |
| 1545 | .and_then(Value::as_bool), |
| 1546 | None |
| 1547 | ); |
| 1548 | assert_eq!( |
| 1549 | metadata |
| 1550 | .get("auto_resume_on_completion") |
| 1551 | .and_then(Value::as_bool), |
| 1552 | Some(false) |
| 1553 | ); |
| 1554 | assert_eq!( |
| 1555 | metadata.get("completion_surface").and_then(Value::as_str), |
| 1556 | Some("task_status") |
| 1557 | ); |
| 1558 | assert_eq!( |
| 1559 | metadata.get("background_policy").and_then(Value::as_str), |
| 1560 | Some("nonblocking") |
| 1561 | ); |
| 1562 | |
| 1563 | let output = wait_for_completed_shell( |
| 1564 | &mut ctx.shell_manager.lock().expect("shell manager"), |
| 1565 | task_id, |
| 1566 | ); |
| 1567 | assert_eq!( |
| 1568 | output.status, |
| 1569 | ShellStatus::Completed, |
| 1570 | "stdout: {:?} stderr: {:?}", |
| 1571 | output.stdout, |
| 1572 | output.stderr |
| 1573 | ); |
| 1574 | assert!( |
| 1575 | output.stdout.contains("rustc"), |
| 1576 | "stdout should include rustc version: {:?}", |
| 1577 | output.stdout |
| 1578 | ); |
| 1579 | } |
| 1580 | } |
| 1581 |