| 1 | //! Typed task-spec loading, artifact refs, deterministic scorers, and receipts. |
| 2 | |
| 3 | #![allow(dead_code)] |
| 4 | |
| 5 | use std::collections::{BTreeMap, BTreeSet}; |
| 6 | use std::path::{Path, PathBuf}; |
| 7 | |
| 8 | use anyhow::{Context, Result, bail}; |
| 9 | use chrono::{SecondsFormat, Utc}; |
| 10 | use codewhale_protocol::fleet::*; |
| 11 | use regex::Regex; |
| 12 | use serde::{Deserialize, Serialize}; |
| 13 | use serde_json::{Value, json}; |
| 14 | |
| 15 | use super::ledger::FleetLedger; |
| 16 | |
| 17 | const MAX_SCORER_READ_BYTES: u64 = 1_000_000; |
| 18 | const MAX_FLEET_ID_BYTES: usize = 128; |
| 19 | const MAX_FLEET_NAME_BYTES: usize = 256; |
| 20 | |
| 21 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 22 | pub struct FleetTaskSpecDocument { |
| 23 | #[serde(default)] |
| 24 | pub name: Option<String>, |
| 25 | #[serde(default)] |
| 26 | pub labels: BTreeMap<String, String>, |
| 27 | #[serde(default)] |
| 28 | #[serde(skip_serializing_if = "Option::is_none")] |
| 29 | /// Legacy replay/input compatibility only. New run validation rejects it; |
| 30 | /// Runtime policy owns execution authority. |
| 31 | pub security_policy: Option<FleetSecurityPolicy>, |
| 32 | #[serde(default, alias = "worker_specs")] |
| 33 | pub workers: Vec<FleetWorkerSpec>, |
| 34 | #[serde(default)] |
| 35 | pub tasks: Vec<FleetTaskSpec>, |
| 36 | /// Optional run-wide usage ceiling (R6, #5567), e.g. |
| 37 | /// `usage_ceiling = { max_total_tokens = 2_000_000 }`. |
| 38 | #[serde(default)] |
| 39 | #[serde(skip_serializing_if = "Option::is_none")] |
| 40 | pub usage_ceiling: Option<codewhale_protocol::fleet::FleetUsageCeiling>, |
| 41 | } |
| 42 | |
| 43 | #[derive(Debug, Clone, Deserialize)] |
| 44 | #[serde(untagged)] |
| 45 | enum FleetTaskSpecFile { |
| 46 | Document(FleetTaskSpecDocument), |
| 47 | Tasks(Vec<FleetTaskSpec>), |
| 48 | Single(Box<FleetTaskSpec>), |
| 49 | } |
| 50 | |
| 51 | impl FleetTaskSpecFile { |
| 52 | fn into_document(self, fallback_name: String) -> FleetTaskSpecDocument { |
| 53 | match self { |
| 54 | Self::Document(mut doc) => { |
| 55 | if doc.name.as_deref().is_none_or(str::is_empty) { |
| 56 | doc.name = Some(fallback_name); |
| 57 | } |
| 58 | doc |
| 59 | } |
| 60 | Self::Tasks(tasks) => FleetTaskSpecDocument { |
| 61 | name: Some(fallback_name), |
| 62 | labels: BTreeMap::new(), |
| 63 | security_policy: None, |
| 64 | workers: Vec::new(), |
| 65 | tasks, |
| 66 | usage_ceiling: None, |
| 67 | }, |
| 68 | Self::Single(task) => FleetTaskSpecDocument { |
| 69 | name: Some(fallback_name), |
| 70 | labels: BTreeMap::new(), |
| 71 | security_policy: None, |
| 72 | workers: Vec::new(), |
| 73 | tasks: vec![*task], |
| 74 | usage_ceiling: None, |
| 75 | }, |
| 76 | } |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | /// The worker's visible final answer as carried by the terminal exec |
| 81 | /// `metadata` receipt: `excerpt` is already bounded and secret-redacted by the |
| 82 | /// emitter (`visible_final_answer_excerpt`), `chars` is the real |
| 83 | /// pre-truncation length (`visible_final_answer_chars`). |
| 84 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 85 | pub struct FleetWorkerFinalAnswer { |
| 86 | pub excerpt: String, |
| 87 | pub chars: usize, |
| 88 | } |
| 89 | |
| 90 | impl FleetWorkerFinalAnswer { |
| 91 | /// The receipt note for a task whose only deliverable is its answer text. |
| 92 | /// The excerpt is used verbatim: it was bounded and redacted once at the |
| 93 | /// emitter, and the ledger redacts receipt notes again on write. |
| 94 | pub fn receipt_note(&self) -> String { |
| 95 | format!( |
| 96 | "worker produced {} characters of deliverable: {}", |
| 97 | self.chars, self.excerpt |
| 98 | ) |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | #[derive(Debug, Clone)] |
| 103 | pub struct FleetTaskVerificationInput { |
| 104 | pub run_id: FleetRunId, |
| 105 | pub task_id: String, |
| 106 | pub worker_id: String, |
| 107 | /// Durable lease generation whose result is being verified. |
| 108 | pub attempt: u32, |
| 109 | pub exit_code: Option<i32>, |
| 110 | pub artifacts: Vec<FleetArtifactRef>, |
| 111 | /// The worker's visible final answer, as reported by its terminal exec |
| 112 | /// receipt. Report/summary tasks with no scorer and no file artifact |
| 113 | /// surface this as their deliverable instead of "no verifiable output". |
| 114 | pub final_answer: Option<FleetWorkerFinalAnswer>, |
| 115 | /// Saved exec session id holding the worker's full transcript, when the |
| 116 | /// worker persisted one on completion. |
| 117 | pub saved_session_id: Option<String>, |
| 118 | /// Resolved-route snapshot to persist on the receipt (#3154). |
| 119 | pub resolved_route: Option<FleetResolvedRoute>, |
| 120 | /// Effective worker authority snapshot to persist on the receipt (#3211). |
| 121 | pub effective_permissions: Option<FleetEffectivePermissions>, |
| 122 | } |
| 123 | |
| 124 | #[derive(Debug, Clone)] |
| 125 | pub struct FleetTaskVerification { |
| 126 | pub result: FleetTaskResult, |
| 127 | pub failure_kind: Option<FleetTaskFailureKind>, |
| 128 | pub score: FleetScore, |
| 129 | pub evidence: Vec<String>, |
| 130 | } |
| 131 | |
| 132 | pub fn load_task_spec_document(path: &Path) -> Result<FleetTaskSpecDocument> { |
| 133 | let raw = std::fs::read_to_string(path) |
| 134 | .with_context(|| format!("reading fleet task spec {}", path.display()))?; |
| 135 | let fallback_name = path |
| 136 | .file_stem() |
| 137 | .and_then(|s| s.to_str()) |
| 138 | .filter(|s| !s.is_empty()) |
| 139 | .unwrap_or("fleet-run") |
| 140 | .to_string(); |
| 141 | let parsed = match path.extension().and_then(|s| s.to_str()) { |
| 142 | Some("toml") => toml::from_str::<FleetTaskSpecFile>(&raw) |
| 143 | .with_context(|| format!("parsing TOML fleet task spec {}", path.display()))?, |
| 144 | _ => serde_json::from_str::<FleetTaskSpecFile>(&raw) |
| 145 | .with_context(|| format!("parsing JSON fleet task spec {}", path.display()))?, |
| 146 | }; |
| 147 | let doc = parsed.into_document(fallback_name); |
| 148 | validate_task_spec_document(&doc)?; |
| 149 | Ok(doc) |
| 150 | } |
| 151 | |
| 152 | pub fn validate_task_spec_document(doc: &FleetTaskSpecDocument) -> Result<()> { |
| 153 | if doc.security_policy.is_some() { |
| 154 | bail!( |
| 155 | "fleet task spec security_policy is a legacy compatibility field, not executable Fleet identity; configure trust, secrets, approvals, sandboxing, and tool authority through Runtime policy" |
| 156 | ); |
| 157 | } |
| 158 | if doc.tasks.is_empty() { |
| 159 | bail!("fleet task spec must include at least one task"); |
| 160 | } |
| 161 | let mut ids = BTreeSet::new(); |
| 162 | for task in &doc.tasks { |
| 163 | validate_fleet_identity("task id", &task.id)?; |
| 164 | if !ids.insert(task.id.clone()) { |
| 165 | bail!("duplicate fleet task id {}", task.id); |
| 166 | } |
| 167 | validate_fleet_name(&format!("task {} name", task.id), &task.name)?; |
| 168 | if task.instructions.trim().is_empty() { |
| 169 | bail!("fleet task {} instructions cannot be empty", task.id); |
| 170 | } |
| 171 | if let Some(objective) = &task.objective |
| 172 | && objective.trim().is_empty() |
| 173 | { |
| 174 | bail!("fleet task {} objective cannot be empty", task.id); |
| 175 | } |
| 176 | validate_worker_profile(&task.id, task.worker.as_ref())?; |
| 177 | if task |
| 178 | .metadata |
| 179 | .contains_key(super::worker_runtime::FROZEN_FLEET_MEMBER_METADATA_KEY) |
| 180 | { |
| 181 | bail!( |
| 182 | "fleet task {} metadata key {} is reserved for the durable Runtime selection receipt", |
| 183 | task.id, |
| 184 | super::worker_runtime::FROZEN_FLEET_MEMBER_METADATA_KEY |
| 185 | ); |
| 186 | } |
| 187 | validate_tags(&task.id, &task.tags)?; |
| 188 | validate_workspace_requirements(task)?; |
| 189 | } |
| 190 | let mut worker_ids = BTreeSet::new(); |
| 191 | for worker in &doc.workers { |
| 192 | validate_fleet_identity("worker id", &worker.id)?; |
| 193 | if !worker_ids.insert(worker.id.clone()) { |
| 194 | bail!("duplicate fleet worker id {}", worker.id); |
| 195 | } |
| 196 | validate_fleet_name(&format!("worker {} name", worker.id), &worker.name)?; |
| 197 | if worker.trust_level.is_some() { |
| 198 | bail!( |
| 199 | "fleet worker {} trust_level is a legacy compatibility field, not Fleet identity; configure execution authority through Runtime policy", |
| 200 | worker.id |
| 201 | ); |
| 202 | } |
| 203 | } |
| 204 | Ok(()) |
| 205 | } |
| 206 | |
| 207 | fn validate_fleet_identity(field: &str, value: &str) -> Result<()> { |
| 208 | if value.is_empty() { |
| 209 | bail!("fleet {field} cannot be empty"); |
| 210 | } |
| 211 | if value.len() > MAX_FLEET_ID_BYTES || !value.chars().all(is_worker_token_char) { |
| 212 | bail!( |
| 213 | "fleet {field} must be a simple ASCII token no longer than {MAX_FLEET_ID_BYTES} bytes" |
| 214 | ); |
| 215 | } |
| 216 | Ok(()) |
| 217 | } |
| 218 | |
| 219 | fn validate_fleet_name(field: &str, value: &str) -> Result<()> { |
| 220 | if value.trim().is_empty() { |
| 221 | bail!("fleet {field} cannot be empty"); |
| 222 | } |
| 223 | if value.len() > MAX_FLEET_NAME_BYTES || value.chars().any(char::is_control) { |
| 224 | bail!( |
| 225 | "fleet {field} must be one printable line no longer than {MAX_FLEET_NAME_BYTES} bytes" |
| 226 | ); |
| 227 | } |
| 228 | Ok(()) |
| 229 | } |
| 230 | |
| 231 | fn validate_worker_profile(task_id: &str, worker: Option<&FleetTaskWorkerProfile>) -> Result<()> { |
| 232 | let Some(worker) = worker else { |
| 233 | return Ok(()); |
| 234 | }; |
| 235 | validate_worker_selector( |
| 236 | task_id, |
| 237 | "worker.agent_profile", |
| 238 | worker.agent_profile.as_deref(), |
| 239 | )?; |
| 240 | validate_worker_token(task_id, "worker.loadout", worker.loadout.as_deref())?; |
| 241 | validate_worker_token(task_id, "worker.model_class", worker.model_class.as_deref())?; |
| 242 | validate_worker_model(task_id, worker.model.as_deref())?; |
| 243 | Ok(()) |
| 244 | } |
| 245 | |
| 246 | fn validate_worker_selector(task_id: &str, field: &str, value: Option<&str>) -> Result<()> { |
| 247 | let Some(value) = value else { |
| 248 | return Ok(()); |
| 249 | }; |
| 250 | let trimmed = value.trim(); |
| 251 | if trimmed.is_empty() { |
| 252 | bail!("fleet task {task_id} {field} cannot be empty"); |
| 253 | } |
| 254 | if trimmed != value || value.len() > MAX_FLEET_NAME_BYTES || value.chars().any(char::is_control) |
| 255 | { |
| 256 | bail!( |
| 257 | "fleet task {task_id} {field} must be one printable selector no longer than {MAX_FLEET_NAME_BYTES} bytes" |
| 258 | ); |
| 259 | } |
| 260 | Ok(()) |
| 261 | } |
| 262 | |
| 263 | fn validate_worker_token(task_id: &str, field: &str, value: Option<&str>) -> Result<()> { |
| 264 | let Some(value) = value else { |
| 265 | return Ok(()); |
| 266 | }; |
| 267 | let trimmed = value.trim(); |
| 268 | if trimmed.is_empty() { |
| 269 | bail!("fleet task {task_id} {field} cannot be empty"); |
| 270 | } |
| 271 | if trimmed != value || !trimmed.chars().all(is_worker_token_char) { |
| 272 | bail!( |
| 273 | "fleet task {task_id} {field} must be a simple token, not a path or provider/model id" |
| 274 | ); |
| 275 | } |
| 276 | Ok(()) |
| 277 | } |
| 278 | |
| 279 | fn is_worker_token_char(ch: char) -> bool { |
| 280 | ch.is_ascii_alphanumeric() || matches!(ch, '-' | '_' | '.') |
| 281 | } |
| 282 | |
| 283 | fn validate_worker_model(task_id: &str, value: Option<&str>) -> Result<()> { |
| 284 | let Some(value) = value else { |
| 285 | return Ok(()); |
| 286 | }; |
| 287 | let trimmed = value.trim(); |
| 288 | if trimmed.is_empty() { |
| 289 | bail!("fleet task {task_id} worker.model cannot be empty"); |
| 290 | } |
| 291 | if trimmed != value |
| 292 | || !trimmed |
| 293 | .chars() |
| 294 | .all(|ch| ch.is_ascii_graphic() && !matches!(ch, '=' | '\'' | '"')) |
| 295 | { |
| 296 | bail!( |
| 297 | "fleet task {task_id} worker.model must be a visible model id without whitespace or secrets" |
| 298 | ); |
| 299 | } |
| 300 | Ok(()) |
| 301 | } |
| 302 | |
| 303 | #[allow(clippy::too_many_arguments)] |
| 304 | pub fn write_fleet_artifact_ref( |
| 305 | workspace: &Path, |
| 306 | run_id: &FleetRunId, |
| 307 | task_id: &str, |
| 308 | worker_id: &str, |
| 309 | kind: FleetArtifactKind, |
| 310 | filename: &str, |
| 311 | contents: &[u8], |
| 312 | mime_type: Option<&str>, |
| 313 | ) -> Result<FleetArtifactRef> { |
| 314 | let rel_path = PathBuf::from(".codewhale") |
| 315 | .join("fleet") |
| 316 | .join(safe_path_segment(&run_id.0)) |
| 317 | .join(safe_path_segment(task_id)) |
| 318 | .join(safe_path_segment(worker_id)) |
| 319 | .join(safe_path_segment(filename)); |
| 320 | super::artifacts::write(workspace, &rel_path, contents)?; |
| 321 | Ok(FleetArtifactRef { |
| 322 | kind, |
| 323 | path: rel_path, |
| 324 | checksum: Some(format!("sha256:{}", crate::hashing::sha256_hex(contents))), |
| 325 | mime_type: mime_type.map(str::to_string), |
| 326 | size_bytes: Some(contents.len() as u64), |
| 327 | }) |
| 328 | } |
| 329 | |
| 330 | pub fn verify_task_result( |
| 331 | workspace: &Path, |
| 332 | task: &FleetTaskSpec, |
| 333 | input: &FleetTaskVerificationInput, |
| 334 | ) -> FleetTaskVerification { |
| 335 | match &task.scorer { |
| 336 | Some(FleetScorerSpec::ExitCode) => verify_exit_code(input.exit_code), |
| 337 | Some(FleetScorerSpec::FileExists { path }) => verify_file_exists(workspace, path), |
| 338 | Some(FleetScorerSpec::RegexMatch { path, pattern }) => { |
| 339 | verify_regex_match(workspace, path, pattern) |
| 340 | } |
| 341 | Some(FleetScorerSpec::JsonPath { path, expression }) => { |
| 342 | verify_json_path(workspace, path, expression) |
| 343 | } |
| 344 | Some(FleetScorerSpec::Command { command, .. }) => partial( |
| 345 | format!("external scorer command configured: {command}"), |
| 346 | "run the configured scorer command to finalize this receipt", |
| 347 | ), |
| 348 | Some(FleetScorerSpec::CodeWhaleVerifierPrompt { .. }) => partial( |
| 349 | "Codewhale verifier prompt configured", |
| 350 | "run a verifier prompt pass to finalize this receipt", |
| 351 | ), |
| 352 | Some(FleetScorerSpec::Manual) => partial( |
| 353 | "manual scorer configured", |
| 354 | "manual verification is required to finalize this receipt", |
| 355 | ), |
| 356 | None if !has_verifiable_artifact(input) => match input |
| 357 | .final_answer |
| 358 | .as_ref() |
| 359 | .filter(|answer| !answer.excerpt.trim().is_empty()) |
| 360 | { |
| 361 | Some(answer) => partial( |
| 362 | "no scorer configured; worker produced a summary deliverable", |
| 363 | answer.receipt_note(), |
| 364 | ), |
| 365 | None => partial( |
| 366 | "no scorer configured and no verifiable artifacts recorded", |
| 367 | "worker exited successfully but produced no verifiable output", |
| 368 | ), |
| 369 | }, |
| 370 | None => partial( |
| 371 | "no scorer configured", |
| 372 | "task has artifacts but no deterministic scorer", |
| 373 | ), |
| 374 | } |
| 375 | } |
| 376 | |
| 377 | pub fn prepare_verification_receipt( |
| 378 | workspace: &Path, |
| 379 | input: &FleetTaskVerificationInput, |
| 380 | verification: FleetTaskVerification, |
| 381 | ) -> Result<FleetReceipt> { |
| 382 | let evidence = json!({ |
| 383 | "run_id": input.run_id.0.clone(), |
| 384 | "task_id": input.task_id.clone(), |
| 385 | "worker_id": input.worker_id.clone(), |
| 386 | "attempt": input.attempt, |
| 387 | "result": verification.result.clone(), |
| 388 | "failure_kind": verification.failure_kind.clone(), |
| 389 | "score": verification.score.clone(), |
| 390 | "evidence": verification.evidence.clone(), |
| 391 | "artifacts": input.artifacts.clone(), |
| 392 | }); |
| 393 | let bytes = |
| 394 | serde_json::to_vec_pretty(&evidence).context("serializing fleet receipt evidence")?; |
| 395 | // Content-address the evidence as well as namespacing it by attempt. A |
| 396 | // stale verifier may finish after a retry has started; it is allowed to |
| 397 | // leave an orphaned evidence file, but it must never overwrite the file a |
| 398 | // winning attempt's durable receipt references. |
| 399 | let evidence_hash = crate::hashing::sha256_hex(&bytes); |
| 400 | let filename = format!( |
| 401 | "verification-receipt-attempt-{:010}-{}.json", |
| 402 | input.attempt, evidence_hash |
| 403 | ); |
| 404 | let receipt_artifact = write_fleet_artifact_ref( |
| 405 | workspace, |
| 406 | &input.run_id, |
| 407 | &input.task_id, |
| 408 | &input.worker_id, |
| 409 | FleetArtifactKind::Receipt, |
| 410 | &filename, |
| 411 | &bytes, |
| 412 | Some("application/json"), |
| 413 | )?; |
| 414 | let mut artifacts = input.artifacts.clone(); |
| 415 | artifacts.push(receipt_artifact); |
| 416 | let receipt = FleetReceipt { |
| 417 | run_id: input.run_id.clone(), |
| 418 | task_id: input.task_id.clone(), |
| 419 | worker_id: input.worker_id.clone(), |
| 420 | attempt: Some(input.attempt), |
| 421 | terminal_seq: None, |
| 422 | completed_at: timestamp(), |
| 423 | result: verification.result, |
| 424 | failure_kind: verification.failure_kind, |
| 425 | artifacts, |
| 426 | score: Some(verification.score), |
| 427 | resolved_route: input.resolved_route.clone(), |
| 428 | saved_session_id: input.saved_session_id.clone(), |
| 429 | effective_permissions: input.effective_permissions.clone(), |
| 430 | }; |
| 431 | Ok(receipt) |
| 432 | } |
| 433 | |
| 434 | pub fn record_verification_receipt( |
| 435 | ledger: &FleetLedger, |
| 436 | workspace: &Path, |
| 437 | input: &FleetTaskVerificationInput, |
| 438 | verification: FleetTaskVerification, |
| 439 | ) -> Result<FleetReceipt> { |
| 440 | let receipt = prepare_verification_receipt(workspace, input, verification)?; |
| 441 | ledger.record_receipt(receipt.clone())?; |
| 442 | Ok(receipt) |
| 443 | } |
| 444 | |
| 445 | fn validate_tags(task_id: &str, tags: &[String]) -> Result<()> { |
| 446 | let mut seen = BTreeSet::new(); |
| 447 | for tag in tags { |
| 448 | if tag.trim().is_empty() { |
| 449 | bail!("fleet task {task_id} tag cannot be empty"); |
| 450 | } |
| 451 | if !seen.insert(tag) { |
| 452 | bail!("fleet task {task_id} has duplicate tag {tag}"); |
| 453 | } |
| 454 | } |
| 455 | Ok(()) |
| 456 | } |
| 457 | |
| 458 | fn validate_workspace_requirements(task: &FleetTaskSpec) -> Result<()> { |
| 459 | let Some(workspace) = &task.workspace else { |
| 460 | return Ok(()); |
| 461 | }; |
| 462 | let env = workspace.environment.as_ref(); |
| 463 | for name in env |
| 464 | .into_iter() |
| 465 | .flat_map(|env| env.required.iter().chain(env.allowlist.iter())) |
| 466 | { |
| 467 | if name.trim().is_empty() { |
| 468 | bail!( |
| 469 | "fleet task {} environment variable name cannot be empty", |
| 470 | task.id |
| 471 | ); |
| 472 | } |
| 473 | } |
| 474 | Ok(()) |
| 475 | } |
| 476 | |
| 477 | fn verify_exit_code(exit_code: Option<i32>) -> FleetTaskVerification { |
| 478 | match exit_code { |
| 479 | Some(0) => pass("exit_code=0"), |
| 480 | Some(code) => fail( |
| 481 | FleetTaskFailureKind::Task, |
| 482 | 0.0, |
| 483 | format!("exit_code={code}"), |
| 484 | "worker task exited unsuccessfully", |
| 485 | ), |
| 486 | None => fail( |
| 487 | FleetTaskFailureKind::Transport, |
| 488 | 0.0, |
| 489 | "missing exit code", |
| 490 | "worker transport did not report a process result", |
| 491 | ), |
| 492 | } |
| 493 | } |
| 494 | |
| 495 | fn verify_file_exists(workspace: &Path, path: &Path) -> FleetTaskVerification { |
| 496 | let abs_path = resolve_workspace_path(workspace, path); |
| 497 | if abs_path.is_file() { |
| 498 | pass(format!("file exists: {}", path.display())) |
| 499 | } else { |
| 500 | fail( |
| 501 | FleetTaskFailureKind::Task, |
| 502 | 0.0, |
| 503 | format!("missing file: {}", path.display()), |
| 504 | "expected artifact file was not produced", |
| 505 | ) |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | fn verify_regex_match(workspace: &Path, path: &Path, pattern: &str) -> FleetTaskVerification { |
| 510 | let regex = match Regex::new(pattern) { |
| 511 | Ok(regex) => regex, |
| 512 | Err(err) => { |
| 513 | return fail( |
| 514 | FleetTaskFailureKind::Verifier, |
| 515 | 0.0, |
| 516 | format!("invalid regex: {err}"), |
| 517 | "regex scorer could not be compiled", |
| 518 | ); |
| 519 | } |
| 520 | }; |
| 521 | let contents = match read_bounded_to_string(workspace, path) { |
| 522 | Ok(contents) => contents, |
| 523 | Err(err) => { |
| 524 | return fail( |
| 525 | err.failure_kind, |
| 526 | 0.0, |
| 527 | err.evidence, |
| 528 | "regex scorer could not read bounded evidence", |
| 529 | ); |
| 530 | } |
| 531 | }; |
| 532 | if regex.is_match(&contents) { |
| 533 | pass(format!("regex matched {}: {pattern}", path.display())) |
| 534 | } else { |
| 535 | fail( |
| 536 | FleetTaskFailureKind::Task, |
| 537 | 0.0, |
| 538 | format!("regex did not match {}: {pattern}", path.display()), |
| 539 | "worker output did not satisfy the regex scorer", |
| 540 | ) |
| 541 | } |
| 542 | } |
| 543 | |
| 544 | fn verify_json_path(workspace: &Path, path: &Path, expression: &str) -> FleetTaskVerification { |
| 545 | let Some(segments) = json_path_segments(expression) else { |
| 546 | return fail( |
| 547 | FleetTaskFailureKind::Verifier, |
| 548 | 0.0, |
| 549 | format!("unsupported JSON path expression: {expression}"), |
| 550 | "json_path scorer supports $.field or .field paths", |
| 551 | ); |
| 552 | }; |
| 553 | let contents = match read_bounded_to_string(workspace, path) { |
| 554 | Ok(contents) => contents, |
| 555 | Err(err) => { |
| 556 | return fail( |
| 557 | err.failure_kind, |
| 558 | 0.0, |
| 559 | err.evidence, |
| 560 | "json_path scorer could not read bounded evidence", |
| 561 | ); |
| 562 | } |
| 563 | }; |
| 564 | let value: Value = match serde_json::from_str(&contents) { |
| 565 | Ok(value) => value, |
| 566 | Err(err) => { |
| 567 | return fail( |
| 568 | FleetTaskFailureKind::Task, |
| 569 | 0.0, |
| 570 | format!("invalid JSON in {}: {err}", path.display()), |
| 571 | "worker artifact was not valid JSON", |
| 572 | ); |
| 573 | } |
| 574 | }; |
| 575 | match json_path_lookup(&value, &segments) { |
| 576 | Some(found) if json_truthy(found) => pass(format!( |
| 577 | "json_path matched {}: {expression}", |
| 578 | path.display() |
| 579 | )), |
| 580 | _ => fail( |
| 581 | FleetTaskFailureKind::Task, |
| 582 | 0.0, |
| 583 | format!( |
| 584 | "json_path missing or false in {}: {expression}", |
| 585 | path.display() |
| 586 | ), |
| 587 | "worker JSON artifact did not satisfy the scorer", |
| 588 | ), |
| 589 | } |
| 590 | } |
| 591 | |
| 592 | fn pass(evidence: impl Into<String>) -> FleetTaskVerification { |
| 593 | let evidence = evidence.into(); |
| 594 | FleetTaskVerification { |
| 595 | result: FleetTaskResult::Pass, |
| 596 | failure_kind: None, |
| 597 | score: FleetScore { |
| 598 | value: 1.0, |
| 599 | max: Some(1.0), |
| 600 | notes: Some(evidence.clone()), |
| 601 | }, |
| 602 | evidence: vec![evidence], |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | fn partial(evidence: impl Into<String>, notes: impl Into<String>) -> FleetTaskVerification { |
| 607 | let evidence = evidence.into(); |
| 608 | let notes = notes.into(); |
| 609 | FleetTaskVerification { |
| 610 | result: FleetTaskResult::Partial, |
| 611 | failure_kind: None, |
| 612 | score: FleetScore { |
| 613 | value: 0.5, |
| 614 | max: Some(1.0), |
| 615 | notes: Some(notes), |
| 616 | }, |
| 617 | evidence: vec![evidence], |
| 618 | } |
| 619 | } |
| 620 | |
| 621 | fn fail( |
| 622 | failure_kind: FleetTaskFailureKind, |
| 623 | value: f64, |
| 624 | evidence: impl Into<String>, |
| 625 | notes: impl Into<String>, |
| 626 | ) -> FleetTaskVerification { |
| 627 | let evidence = evidence.into(); |
| 628 | FleetTaskVerification { |
| 629 | result: FleetTaskResult::Fail, |
| 630 | failure_kind: Some(failure_kind), |
| 631 | score: FleetScore { |
| 632 | value, |
| 633 | max: Some(1.0), |
| 634 | notes: Some(notes.into()), |
| 635 | }, |
| 636 | evidence: vec![evidence], |
| 637 | } |
| 638 | } |
| 639 | |
| 640 | fn has_verifiable_artifact(input: &FleetTaskVerificationInput) -> bool { |
| 641 | input.artifacts.iter().any(|artifact| { |
| 642 | !matches!( |
| 643 | artifact.kind, |
| 644 | FleetArtifactKind::Log | FleetArtifactKind::Receipt |
| 645 | ) |
| 646 | }) |
| 647 | } |
| 648 | |
| 649 | #[derive(Debug)] |
| 650 | struct EvidenceReadError { |
| 651 | failure_kind: FleetTaskFailureKind, |
| 652 | evidence: String, |
| 653 | } |
| 654 | |
| 655 | fn read_bounded_to_string( |
| 656 | workspace: &Path, |
| 657 | path: &Path, |
| 658 | ) -> std::result::Result<String, EvidenceReadError> { |
| 659 | let abs_path = resolve_workspace_path(workspace, path); |
| 660 | let metadata = std::fs::metadata(&abs_path).map_err(|err| EvidenceReadError { |
| 661 | failure_kind: if err.kind() == std::io::ErrorKind::NotFound { |
| 662 | FleetTaskFailureKind::Task |
| 663 | } else { |
| 664 | FleetTaskFailureKind::Verifier |
| 665 | }, |
| 666 | evidence: format!("cannot read {}: {err}", path.display()), |
| 667 | })?; |
| 668 | if metadata.len() > MAX_SCORER_READ_BYTES { |
| 669 | return Err(EvidenceReadError { |
| 670 | failure_kind: FleetTaskFailureKind::Verifier, |
| 671 | evidence: format!( |
| 672 | "refusing to read oversized evidence {}: {} bytes", |
| 673 | path.display(), |
| 674 | metadata.len() |
| 675 | ), |
| 676 | }); |
| 677 | } |
| 678 | std::fs::read_to_string(&abs_path).map_err(|err| EvidenceReadError { |
| 679 | failure_kind: FleetTaskFailureKind::Verifier, |
| 680 | evidence: format!("cannot decode {} as UTF-8: {err}", path.display()), |
| 681 | }) |
| 682 | } |
| 683 | |
| 684 | fn resolve_workspace_path(workspace: &Path, path: &Path) -> PathBuf { |
| 685 | if path.is_absolute() { |
| 686 | path.to_path_buf() |
| 687 | } else { |
| 688 | workspace.join(path) |
| 689 | } |
| 690 | } |
| 691 | |
| 692 | fn json_path_segments(expression: &str) -> Option<Vec<&str>> { |
| 693 | let trimmed = expression.trim(); |
| 694 | let path = trimmed |
| 695 | .strip_prefix("$.") |
| 696 | .or_else(|| trimmed.strip_prefix('.'))?; |
| 697 | if path.is_empty() { |
| 698 | return None; |
| 699 | } |
| 700 | let segments: Vec<_> = path.split('.').collect(); |
| 701 | if segments.iter().any(|segment| segment.is_empty()) { |
| 702 | return None; |
| 703 | } |
| 704 | Some(segments) |
| 705 | } |
| 706 | |
| 707 | fn json_path_lookup<'a>(value: &'a Value, segments: &[&str]) -> Option<&'a Value> { |
| 708 | let mut current = value; |
| 709 | for segment in segments { |
| 710 | current = current.as_object()?.get(*segment)?; |
| 711 | } |
| 712 | Some(current) |
| 713 | } |
| 714 | |
| 715 | fn json_truthy(value: &Value) -> bool { |
| 716 | !matches!(value, Value::Null | Value::Bool(false)) |
| 717 | } |
| 718 | |
| 719 | fn timestamp() -> String { |
| 720 | Utc::now().to_rfc3339_opts(SecondsFormat::Secs, true) |
| 721 | } |
| 722 | |
| 723 | fn safe_path_segment(value: &str) -> String { |
| 724 | value |
| 725 | .chars() |
| 726 | .map(|ch| { |
| 727 | if ch.is_ascii_alphanumeric() || matches!(ch, '-' | '_' | '.') { |
| 728 | ch |
| 729 | } else { |
| 730 | '_' |
| 731 | } |
| 732 | }) |
| 733 | .collect() |
| 734 | } |
| 735 | |
| 736 | #[cfg(test)] |
| 737 | mod tests { |
| 738 | use super::*; |
| 739 | use serde_json::json; |
| 740 | use tempfile::TempDir; |
| 741 | |
| 742 | fn task(id: &str, scorer: Option<FleetScorerSpec>) -> FleetTaskSpec { |
| 743 | FleetTaskSpec { |
| 744 | id: id.to_string(), |
| 745 | name: id.to_string(), |
| 746 | description: None, |
| 747 | objective: Some(format!("Verify {id}")), |
| 748 | instructions: format!("do {id}"), |
| 749 | worker: Some(FleetTaskWorkerProfile { |
| 750 | agent_profile: None, |
| 751 | role: Some("reviewer".to_string()), |
| 752 | loadout: None, |
| 753 | model_class: None, |
| 754 | model: None, |
| 755 | tool_profile: Some("read-only".to_string()), |
| 756 | tools: vec!["git".to_string()], |
| 757 | capabilities: vec!["rust".to_string()], |
| 758 | }), |
| 759 | workspace: Some(FleetWorkspaceRequirements { |
| 760 | root: Some(PathBuf::from(".")), |
| 761 | required_files: vec![PathBuf::from("Cargo.toml")], |
| 762 | writable_paths: vec![PathBuf::from(".codewhale/fleet")], |
| 763 | environment: Some(FleetEnvironmentRequirements { |
| 764 | required: vec!["PATH".to_string()], |
| 765 | allowlist: vec!["RUST_LOG".to_string()], |
| 766 | }), |
| 767 | }), |
| 768 | input_files: vec![PathBuf::from("Cargo.toml")], |
| 769 | context: vec!["fleet verifier test".to_string()], |
| 770 | budget: Some(FleetTaskBudget { |
| 771 | max_tokens: Some(4000), |
| 772 | max_steps: None, |
| 773 | max_tool_calls: Some(12), |
| 774 | max_seconds: Some(120), |
| 775 | }), |
| 776 | expected_artifacts: vec![FleetArtifactKind::Log, FleetArtifactKind::Receipt], |
| 777 | scorer, |
| 778 | retry_policy: Some(FleetRetryPolicy::default()), |
| 779 | alert_policy: None, |
| 780 | timeout_seconds: Some(120), |
| 781 | tags: vec!["review".to_string()], |
| 782 | metadata: BTreeMap::new(), |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | #[test] |
| 787 | fn fleet_task_spec_document_parses_multi_task_verified_shape() { |
| 788 | let tmp = TempDir::new().unwrap(); |
| 789 | let path = tmp.path().join("fleet-tasks.json"); |
| 790 | let doc = json!({ |
| 791 | "name": "release triage", |
| 792 | "labels": {"milestone": "v0.8.60"}, |
| 793 | "tasks": [ |
| 794 | task("release-notes", Some(FleetScorerSpec::ExitCode)), |
| 795 | task("risk-review", Some(FleetScorerSpec::Manual)) |
| 796 | ] |
| 797 | }); |
| 798 | std::fs::write(&path, serde_json::to_string_pretty(&doc).unwrap()).unwrap(); |
| 799 | |
| 800 | let parsed = load_task_spec_document(&path).unwrap(); |
| 801 | |
| 802 | assert_eq!(parsed.name.as_deref(), Some("release triage")); |
| 803 | assert_eq!(parsed.tasks.len(), 2); |
| 804 | assert_eq!( |
| 805 | parsed.tasks[0].objective.as_deref(), |
| 806 | Some("Verify release-notes") |
| 807 | ); |
| 808 | assert_eq!( |
| 809 | parsed.tasks[0].worker.as_ref().unwrap().role.as_deref(), |
| 810 | Some("reviewer") |
| 811 | ); |
| 812 | assert_eq!(parsed.tasks[1].tags, vec!["review"]); |
| 813 | } |
| 814 | |
| 815 | #[test] |
| 816 | fn fleet_task_spec_document_parses_worker_profile_loadout_intent() { |
| 817 | let tmp = TempDir::new().unwrap(); |
| 818 | let path = tmp.path().join("fleet-profile-task.json"); |
| 819 | let doc = json!({ |
| 820 | "name": "profile loadout smoke", |
| 821 | "tasks": [{ |
| 822 | "id": "review", |
| 823 | "name": "review", |
| 824 | "instructions": "review the patch", |
| 825 | "worker": { |
| 826 | "profile": "DeepSeek V4 Flash", |
| 827 | "role": "reviewer", |
| 828 | "loadout": "auto", |
| 829 | "model_class": "balanced", |
| 830 | "model": "deepseek-v4-pro", |
| 831 | "tool_profile": "read-only", |
| 832 | "tools": ["read_file", "grep_files"], |
| 833 | "capabilities": ["rust"] |
| 834 | } |
| 835 | }] |
| 836 | }); |
| 837 | std::fs::write(&path, serde_json::to_string_pretty(&doc).unwrap()).unwrap(); |
| 838 | |
| 839 | let parsed = load_task_spec_document(&path).unwrap(); |
| 840 | let worker = parsed.tasks[0].worker.as_ref().unwrap(); |
| 841 | |
| 842 | assert_eq!(worker.agent_profile.as_deref(), Some("DeepSeek V4 Flash")); |
| 843 | assert_eq!(worker.role.as_deref(), Some("reviewer")); |
| 844 | assert_eq!(worker.loadout.as_deref(), Some("auto")); |
| 845 | assert_eq!(worker.model_class.as_deref(), Some("balanced")); |
| 846 | assert_eq!(worker.model.as_deref(), Some("deepseek-v4-pro")); |
| 847 | assert_eq!(worker.tool_profile.as_deref(), Some("read-only")); |
| 848 | } |
| 849 | |
| 850 | #[test] |
| 851 | fn fleet_task_spec_rejects_unsafe_worker_profile_intent_tokens() { |
| 852 | let tmp = TempDir::new().unwrap(); |
| 853 | let path = tmp.path().join("unsafe-profile-task.json"); |
| 854 | let doc = json!({ |
| 855 | "tasks": [{ |
| 856 | "id": "review", |
| 857 | "name": "review", |
| 858 | "instructions": "review the patch", |
| 859 | "worker": { |
| 860 | "profile": "reviewer\n../../secrets", |
| 861 | "loadout": "openrouter/deepseek", |
| 862 | "model_class": "", |
| 863 | "model": "deepseek/deepseek-v4-pro" |
| 864 | } |
| 865 | }] |
| 866 | }); |
| 867 | std::fs::write(&path, serde_json::to_string_pretty(&doc).unwrap()).unwrap(); |
| 868 | |
| 869 | let err = load_task_spec_document(&path).unwrap_err().to_string(); |
| 870 | |
| 871 | assert!( |
| 872 | err.contains("worker.agent_profile must be one printable selector"), |
| 873 | "unexpected error: {err}" |
| 874 | ); |
| 875 | } |
| 876 | |
| 877 | #[test] |
| 878 | fn fleet_task_spec_rejects_secret_like_worker_model() { |
| 879 | let tmp = TempDir::new().unwrap(); |
| 880 | let path = tmp.path().join("unsafe-worker-model.json"); |
| 881 | let doc = json!({ |
| 882 | "tasks": [{ |
| 883 | "id": "review", |
| 884 | "name": "review", |
| 885 | "instructions": "review the patch", |
| 886 | "worker": { |
| 887 | "model": "deepseek-v4-pro api_key=secret" |
| 888 | } |
| 889 | }] |
| 890 | }); |
| 891 | std::fs::write(&path, serde_json::to_string_pretty(&doc).unwrap()).unwrap(); |
| 892 | |
| 893 | let err = load_task_spec_document(&path).unwrap_err().to_string(); |
| 894 | |
| 895 | assert!( |
| 896 | err.contains("worker.model must be a visible model id"), |
| 897 | "unexpected error: {err}" |
| 898 | ); |
| 899 | } |
| 900 | |
| 901 | #[test] |
| 902 | fn fleet_task_spec_rejects_unbounded_or_multiline_task_and_worker_identities() { |
| 903 | let tmp = TempDir::new().unwrap(); |
| 904 | let path = tmp.path().join("unsafe-identities.json"); |
| 905 | let doc = json!({ |
| 906 | "workers": [{ |
| 907 | "id": "worker\r\nforged", |
| 908 | "name": "forged worker", |
| 909 | "host": {"kind": "local"} |
| 910 | }], |
| 911 | "tasks": [{ |
| 912 | "id": "review", |
| 913 | "name": "review", |
| 914 | "instructions": "review the patch" |
| 915 | }] |
| 916 | }); |
| 917 | std::fs::write(&path, serde_json::to_string_pretty(&doc).unwrap()).unwrap(); |
| 918 | |
| 919 | let err = load_task_spec_document(&path).unwrap_err().to_string(); |
| 920 | assert!( |
| 921 | err.contains("worker id must be a simple ASCII token"), |
| 922 | "unexpected error: {err}" |
| 923 | ); |
| 924 | |
| 925 | let mut doc = task("review", None); |
| 926 | doc.id = "a".repeat(MAX_FLEET_ID_BYTES + 1); |
| 927 | let err = validate_task_spec_document(&FleetTaskSpecDocument { |
| 928 | name: None, |
| 929 | labels: BTreeMap::new(), |
| 930 | security_policy: None, |
| 931 | workers: Vec::new(), |
| 932 | tasks: vec![doc], |
| 933 | usage_ceiling: None, |
| 934 | }) |
| 935 | .unwrap_err() |
| 936 | .to_string(); |
| 937 | assert!( |
| 938 | err.contains("task id must be a simple ASCII token"), |
| 939 | "unexpected error: {err}" |
| 940 | ); |
| 941 | } |
| 942 | |
| 943 | #[test] |
| 944 | fn fleet_task_spec_rejects_runtime_authority_fields_for_new_runs() { |
| 945 | let mut security_doc = FleetTaskSpecDocument { |
| 946 | name: None, |
| 947 | labels: BTreeMap::new(), |
| 948 | security_policy: Some(FleetSecurityPolicy::default()), |
| 949 | workers: Vec::new(), |
| 950 | tasks: vec![task("review", None)], |
| 951 | usage_ceiling: None, |
| 952 | }; |
| 953 | let error = validate_task_spec_document(&security_doc) |
| 954 | .unwrap_err() |
| 955 | .to_string(); |
| 956 | assert!( |
| 957 | error.contains("security_policy is a legacy compatibility field"), |
| 958 | "unexpected error: {error}" |
| 959 | ); |
| 960 | |
| 961 | security_doc.security_policy = None; |
| 962 | security_doc.workers.push(FleetWorkerSpec { |
| 963 | id: "worker-1".to_string(), |
| 964 | name: "Worker 1".to_string(), |
| 965 | host: FleetHostSpec::Local, |
| 966 | trust_level: Some(FleetTrustLevel::Local), |
| 967 | labels: BTreeMap::new(), |
| 968 | capabilities: vec!["local".to_string()], |
| 969 | max_concurrent_tasks: Some(1), |
| 970 | }); |
| 971 | let error = validate_task_spec_document(&security_doc) |
| 972 | .unwrap_err() |
| 973 | .to_string(); |
| 974 | assert!( |
| 975 | error.contains("trust_level is a legacy compatibility field"), |
| 976 | "unexpected error: {error}" |
| 977 | ); |
| 978 | } |
| 979 | |
| 980 | #[cfg(unix)] |
| 981 | #[test] |
| 982 | fn fleet_artifact_publication_rejects_symlinked_workspace_paths() { |
| 983 | use std::os::unix::fs::symlink; |
| 984 | let workspace = TempDir::new().unwrap(); |
| 985 | let outside = TempDir::new().unwrap(); |
| 986 | std::fs::create_dir_all(workspace.path().join(".codewhale")).unwrap(); |
| 987 | symlink(outside.path(), workspace.path().join(".codewhale/fleet")).unwrap(); |
| 988 | let result = write_fleet_artifact_ref( |
| 989 | workspace.path(), |
| 990 | &FleetRunId::from("run-1"), |
| 991 | "task-a", |
| 992 | "worker-1", |
| 993 | FleetArtifactKind::Receipt, |
| 994 | "receipt.json", |
| 995 | b"synthetic receipt", |
| 996 | Some("application/json"), |
| 997 | ); |
| 998 | assert!( |
| 999 | result.is_err(), |
| 1000 | "publication must reject a symlinked parent" |
| 1001 | ); |
| 1002 | assert!(!outside.path().join("run-1").exists()); |
| 1003 | } |
| 1004 | |
| 1005 | #[test] |
| 1006 | fn fleet_task_spec_artifact_refs_are_bounded_paths() { |
| 1007 | let tmp = TempDir::new().unwrap(); |
| 1008 | let artifact = write_fleet_artifact_ref( |
| 1009 | tmp.path(), |
| 1010 | &FleetRunId::from("run-1"), |
| 1011 | "task-a", |
| 1012 | "worker-1", |
| 1013 | FleetArtifactKind::Log, |
| 1014 | "worker.log", |
| 1015 | b"this is artifact content", |
| 1016 | Some("text/plain"), |
| 1017 | ) |
| 1018 | .unwrap(); |
| 1019 | |
| 1020 | let json = serde_json::to_string(&artifact).unwrap(); |
| 1021 | assert!(!json.contains("this is artifact content")); |
| 1022 | assert!(json.contains("worker.log")); |
| 1023 | assert_eq!(artifact.size_bytes, Some(24)); |
| 1024 | assert!(artifact.checksum.as_deref().unwrap().starts_with("sha256:")); |
| 1025 | assert!(tmp.path().join(&artifact.path).exists()); |
| 1026 | } |
| 1027 | |
| 1028 | #[test] |
| 1029 | fn fleet_task_spec_scorers_record_pass_fail_partial_evidence() { |
| 1030 | let tmp = TempDir::new().unwrap(); |
| 1031 | std::fs::write(tmp.path().join("result.txt"), "status=ok\n").unwrap(); |
| 1032 | std::fs::write(tmp.path().join("result.json"), r#"{"status":"ok"}"#).unwrap(); |
| 1033 | let input = FleetTaskVerificationInput { |
| 1034 | run_id: FleetRunId::from("run-1"), |
| 1035 | task_id: "task-a".to_string(), |
| 1036 | worker_id: "worker-1".to_string(), |
| 1037 | attempt: 1, |
| 1038 | exit_code: Some(0), |
| 1039 | artifacts: vec![], |
| 1040 | final_answer: None, |
| 1041 | saved_session_id: None, |
| 1042 | resolved_route: None, |
| 1043 | effective_permissions: None, |
| 1044 | }; |
| 1045 | |
| 1046 | let pass = verify_task_result( |
| 1047 | tmp.path(), |
| 1048 | &task("exit", Some(FleetScorerSpec::ExitCode)), |
| 1049 | &input, |
| 1050 | ); |
| 1051 | assert_eq!(pass.result, FleetTaskResult::Pass); |
| 1052 | assert_eq!(pass.failure_kind, None); |
| 1053 | |
| 1054 | let regex = verify_task_result( |
| 1055 | tmp.path(), |
| 1056 | &task( |
| 1057 | "regex", |
| 1058 | Some(FleetScorerSpec::RegexMatch { |
| 1059 | path: PathBuf::from("result.txt"), |
| 1060 | pattern: "status=ok".to_string(), |
| 1061 | }), |
| 1062 | ), |
| 1063 | &input, |
| 1064 | ); |
| 1065 | assert_eq!(regex.result, FleetTaskResult::Pass); |
| 1066 | |
| 1067 | let json_path = verify_task_result( |
| 1068 | tmp.path(), |
| 1069 | &task( |
| 1070 | "json", |
| 1071 | Some(FleetScorerSpec::JsonPath { |
| 1072 | path: PathBuf::from("result.json"), |
| 1073 | expression: "$.status".to_string(), |
| 1074 | }), |
| 1075 | ), |
| 1076 | &input, |
| 1077 | ); |
| 1078 | assert_eq!(json_path.result, FleetTaskResult::Pass); |
| 1079 | |
| 1080 | let manual = verify_task_result( |
| 1081 | tmp.path(), |
| 1082 | &task("manual", Some(FleetScorerSpec::Manual)), |
| 1083 | &input, |
| 1084 | ); |
| 1085 | assert_eq!(manual.result, FleetTaskResult::Partial); |
| 1086 | |
| 1087 | let no_scorer_empty = verify_task_result(tmp.path(), &task("unscored", None), &input); |
| 1088 | assert_eq!(no_scorer_empty.result, FleetTaskResult::Partial); |
| 1089 | assert!( |
| 1090 | no_scorer_empty |
| 1091 | .score |
| 1092 | .notes |
| 1093 | .as_deref() |
| 1094 | .unwrap_or_default() |
| 1095 | .contains("no verifiable output") |
| 1096 | ); |
| 1097 | |
| 1098 | let failed = verify_task_result( |
| 1099 | tmp.path(), |
| 1100 | &task( |
| 1101 | "missing", |
| 1102 | Some(FleetScorerSpec::FileExists { |
| 1103 | path: PathBuf::from("missing.txt"), |
| 1104 | }), |
| 1105 | ), |
| 1106 | &input, |
| 1107 | ); |
| 1108 | assert_eq!(failed.result, FleetTaskResult::Fail); |
| 1109 | assert_eq!(failed.failure_kind, Some(FleetTaskFailureKind::Task)); |
| 1110 | |
| 1111 | let verifier_failed = verify_task_result( |
| 1112 | tmp.path(), |
| 1113 | &task( |
| 1114 | "bad-regex", |
| 1115 | Some(FleetScorerSpec::RegexMatch { |
| 1116 | path: PathBuf::from("result.txt"), |
| 1117 | pattern: "[".to_string(), |
| 1118 | }), |
| 1119 | ), |
| 1120 | &input, |
| 1121 | ); |
| 1122 | assert_eq!(verifier_failed.result, FleetTaskResult::Fail); |
| 1123 | assert_eq!( |
| 1124 | verifier_failed.failure_kind, |
| 1125 | Some(FleetTaskFailureKind::Verifier) |
| 1126 | ); |
| 1127 | } |
| 1128 | |
| 1129 | #[test] |
| 1130 | fn unscored_worker_surfaces_summary_deliverable_instead_of_no_output() { |
| 1131 | let tmp = TempDir::new().unwrap(); |
| 1132 | let input = FleetTaskVerificationInput { |
| 1133 | run_id: FleetRunId::from("run-1"), |
| 1134 | task_id: "task-a".to_string(), |
| 1135 | worker_id: "worker-1".to_string(), |
| 1136 | attempt: 1, |
| 1137 | exit_code: Some(0), |
| 1138 | artifacts: vec![], |
| 1139 | final_answer: Some(FleetWorkerFinalAnswer { |
| 1140 | excerpt: "The Changelog review is complete".to_string(), |
| 1141 | chars: 32, |
| 1142 | }), |
| 1143 | saved_session_id: None, |
| 1144 | resolved_route: None, |
| 1145 | effective_permissions: None, |
| 1146 | }; |
| 1147 | let verification = verify_task_result(tmp.path(), &task("unscored", None), &input); |
| 1148 | assert_eq!(verification.result, FleetTaskResult::Partial); |
| 1149 | let notes = verification |
| 1150 | .score |
| 1151 | .notes |
| 1152 | .as_deref() |
| 1153 | .unwrap_or_default() |
| 1154 | .to_string(); |
| 1155 | assert!( |
| 1156 | notes.contains("worker produced 32 characters of deliverable"), |
| 1157 | "unexpected notes: {notes}" |
| 1158 | ); |
| 1159 | assert!(notes.contains("Changelog review is complete")); |
| 1160 | assert!(!notes.contains("no verifiable output")); |
| 1161 | } |
| 1162 | |
| 1163 | #[test] |
| 1164 | fn fleet_task_spec_receipt_records_artifacts_scores_and_failure_kind() { |
| 1165 | let tmp = TempDir::new().unwrap(); |
| 1166 | let ledger = FleetLedger::open(tmp.path()).unwrap(); |
| 1167 | let log = write_fleet_artifact_ref( |
| 1168 | tmp.path(), |
| 1169 | &FleetRunId::from("run-1"), |
| 1170 | "task-a", |
| 1171 | "worker-1", |
| 1172 | FleetArtifactKind::Log, |
| 1173 | "worker.log", |
| 1174 | b"exit_code=1", |
| 1175 | Some("text/plain"), |
| 1176 | ) |
| 1177 | .unwrap(); |
| 1178 | let input = FleetTaskVerificationInput { |
| 1179 | run_id: FleetRunId::from("run-1"), |
| 1180 | task_id: "task-a".to_string(), |
| 1181 | worker_id: "worker-1".to_string(), |
| 1182 | attempt: 3, |
| 1183 | exit_code: Some(1), |
| 1184 | artifacts: vec![log], |
| 1185 | final_answer: None, |
| 1186 | saved_session_id: None, |
| 1187 | resolved_route: None, |
| 1188 | effective_permissions: Some(FleetEffectivePermissions { |
| 1189 | write: false, |
| 1190 | network: false, |
| 1191 | shell: "read_only".to_string(), |
| 1192 | tool_scope: "explicit".to_string(), |
| 1193 | tools: vec!["read_file".to_string()], |
| 1194 | background: true, |
| 1195 | max_spawn_depth: 0, |
| 1196 | profile_id: None, |
| 1197 | profile_origin: None, |
| 1198 | source: "worker_runtime_profile".to_string(), |
| 1199 | }), |
| 1200 | }; |
| 1201 | let verification = verify_task_result( |
| 1202 | tmp.path(), |
| 1203 | &task("task-a", Some(FleetScorerSpec::ExitCode)), |
| 1204 | &input, |
| 1205 | ); |
| 1206 | |
| 1207 | let receipt = |
| 1208 | record_verification_receipt(&ledger, tmp.path(), &input, verification).unwrap(); |
| 1209 | |
| 1210 | assert_eq!(receipt.result, FleetTaskResult::Fail); |
| 1211 | assert_eq!(receipt.failure_kind, Some(FleetTaskFailureKind::Task)); |
| 1212 | assert_eq!(receipt.attempt, Some(3)); |
| 1213 | assert_eq!(receipt.terminal_seq, None); |
| 1214 | assert_eq!(receipt.effective_permissions, input.effective_permissions); |
| 1215 | assert_eq!(receipt.artifacts.len(), 2); |
| 1216 | assert!(matches!( |
| 1217 | receipt.artifacts.last().unwrap().kind, |
| 1218 | FleetArtifactKind::Receipt |
| 1219 | )); |
| 1220 | assert!( |
| 1221 | receipt |
| 1222 | .artifacts |
| 1223 | .last() |
| 1224 | .unwrap() |
| 1225 | .path |
| 1226 | .to_string_lossy() |
| 1227 | .contains("verification-receipt-attempt-0000000003-") |
| 1228 | ); |
| 1229 | let state = ledger.rebuild_state().unwrap(); |
| 1230 | assert_eq!( |
| 1231 | state.receipts["run-1:task-a"].failure_kind, |
| 1232 | Some(FleetTaskFailureKind::Task) |
| 1233 | ); |
| 1234 | } |
| 1235 | |
| 1236 | #[test] |
| 1237 | fn verification_evidence_is_attempt_and_content_addressed() { |
| 1238 | let tmp = TempDir::new().unwrap(); |
| 1239 | let mut input = FleetTaskVerificationInput { |
| 1240 | run_id: FleetRunId::from("run-1"), |
| 1241 | task_id: "task-a".to_string(), |
| 1242 | worker_id: "worker-1".to_string(), |
| 1243 | attempt: 1, |
| 1244 | exit_code: Some(1), |
| 1245 | artifacts: Vec::new(), |
| 1246 | final_answer: None, |
| 1247 | saved_session_id: None, |
| 1248 | resolved_route: None, |
| 1249 | effective_permissions: None, |
| 1250 | }; |
| 1251 | let scorer = task("task-a", Some(FleetScorerSpec::ExitCode)); |
| 1252 | let stale_verification = verify_task_result(tmp.path(), &scorer, &input); |
| 1253 | let stale = prepare_verification_receipt(tmp.path(), &input, stale_verification).unwrap(); |
| 1254 | |
| 1255 | input.attempt = 2; |
| 1256 | input.exit_code = Some(0); |
| 1257 | let winning_verification = verify_task_result(tmp.path(), &scorer, &input); |
| 1258 | let winning = |
| 1259 | prepare_verification_receipt(tmp.path(), &input, winning_verification).unwrap(); |
| 1260 | |
| 1261 | let stale_path = &stale.artifacts.last().unwrap().path; |
| 1262 | let winning_path = &winning.artifacts.last().unwrap().path; |
| 1263 | assert_ne!(stale_path, winning_path); |
| 1264 | assert!(stale_path.to_string_lossy().contains("attempt-0000000001-")); |
| 1265 | assert!( |
| 1266 | winning_path |
| 1267 | .to_string_lossy() |
| 1268 | .contains("attempt-0000000002-") |
| 1269 | ); |
| 1270 | assert!(tmp.path().join(stale_path).is_file()); |
| 1271 | assert!(tmp.path().join(winning_path).is_file()); |
| 1272 | assert_eq!(stale.result, FleetTaskResult::Fail); |
| 1273 | assert_eq!(winning.result, FleetTaskResult::Pass); |
| 1274 | } |
| 1275 | } |
| 1276 |