| 1 | use super::*; |
| 2 | use tempfile::tempdir; |
| 3 | |
| 4 | fn git(root: &Path, args: &[&str]) { |
| 5 | let output = std::process::Command::new("git") |
| 6 | .arg("-C") |
| 7 | .arg(root) |
| 8 | .args(args) |
| 9 | .output() |
| 10 | .expect("git"); |
| 11 | assert!( |
| 12 | output.status.success(), |
| 13 | "git {args:?}: {}", |
| 14 | String::from_utf8_lossy(&output.stderr) |
| 15 | ); |
| 16 | } |
| 17 | |
| 18 | fn repository(root: &Path) { |
| 19 | git(root, &["init", "--quiet"]); |
| 20 | git(root, &["config", "user.name", "Delivery test"]); |
| 21 | git(root, &["config", "user.email", "delivery@example.invalid"]); |
| 22 | fs::create_dir_all(root.join("src")).unwrap(); |
| 23 | fs::write(root.join("src/lib.rs"), "baseline\n").unwrap(); |
| 24 | git(root, &["add", "--", "src/lib.rs"]); |
| 25 | git(root, &["commit", "--quiet", "-m", "baseline"]); |
| 26 | } |
| 27 | |
| 28 | fn worker(root: &Path, write: bool, paths: &[&str], scope: &[&str]) -> (SubAgentManager, String) { |
| 29 | let mut manager = SubAgentManager::new(root.to_path_buf(), 2); |
| 30 | let id = manager.insert_test_running_agent("delivery", root); |
| 31 | let record = manager.worker_records.get_mut(&id).unwrap(); |
| 32 | record.spec.runtime_profile.permissions.write = write; |
| 33 | record.spec.launch_manifest = Some(ChildLaunchManifest { |
| 34 | owner_session: "workspace".into(), |
| 35 | child_id: id.clone(), |
| 36 | profile: record.spec.runtime_profile.clone(), |
| 37 | prompt: "produce report".into(), |
| 38 | cwd: Some(root.display().to_string()), |
| 39 | worktree: false, |
| 40 | writable_roots: scope.iter().map(|path| (*path).into()).collect(), |
| 41 | writable_files: Vec::new(), |
| 42 | coordination_contracts: Vec::new(), |
| 43 | expected_artifact: None, |
| 44 | deliverables: paths.iter().map(|path| (*path).into()).collect(), |
| 45 | resume_identity: None, |
| 46 | generation: 1, |
| 47 | resume_from_agent_id: None, |
| 48 | }); |
| 49 | record.delivery_evidence = DeliveryEvidence::capture(&record.spec); |
| 50 | if write && !scope.is_empty() { |
| 51 | manager |
| 52 | .coordination |
| 53 | .register_claim( |
| 54 | WriteScopeClaim { |
| 55 | owner: id.clone(), |
| 56 | roots: scope.iter().map(|path| (*path).into()).collect(), |
| 57 | exact_files: Vec::new(), |
| 58 | contracts: Vec::new(), |
| 59 | }, |
| 60 | false, |
| 61 | |_| false, |
| 62 | ) |
| 63 | .unwrap(); |
| 64 | } |
| 65 | (manager, id) |
| 66 | } |
| 67 | |
| 68 | fn complete(manager: &mut SubAgentManager, id: &str, report: &str) -> AgentRunVerificationSummary { |
| 69 | let mut result = manager.get_result(id).unwrap(); |
| 70 | result.status = SubAgentStatus::Completed; |
| 71 | result.result = Some(report.into()); |
| 72 | manager.complete_worker_from_result(id, &result); |
| 73 | // Deferred verification (#6210): the commit leaves verification pending; |
| 74 | // run the same snapshot→compute→store halves `ensure` runs off the lock. |
| 75 | if !manager.worker_records[id].delivery_evidence.checked |
| 76 | && let Some(inputs) = manager.delivery_verification_inputs(id, &result) |
| 77 | { |
| 78 | let verification = delivery::compute_delivery_verification(&inputs); |
| 79 | manager.store_delivery_verification(id, verification); |
| 80 | } |
| 81 | manager.worker_records[id].verification.clone() |
| 82 | } |
| 83 | |
| 84 | #[test] |
| 85 | fn declared_deliverables_narrow_default_scope_and_reject_invalid_input() { |
| 86 | let request = parse_spawn_request(&json!({ |
| 87 | "type": "implement", "prompt": "write outputs", "deliverables": ["tmp/a/report.md", "tmp/b/report.md"] |
| 88 | })).unwrap(); |
| 89 | assert!(request.write_roots.is_empty()); |
| 90 | assert_eq!(request.exact_files, ["tmp/a/report.md", "tmp/b/report.md"]); |
| 91 | for paths in [ |
| 92 | json!([""]), |
| 93 | json!(["../report.md"]), |
| 94 | json!(["/tmp/report.md"]), |
| 95 | json!([".git/config"]), |
| 96 | json!(["."]), |
| 97 | json!([1]), |
| 98 | json!("report.md"), |
| 99 | ] { |
| 100 | assert!( |
| 101 | parse_spawn_request( |
| 102 | &json!({"type":"implement", "prompt":"report", "deliverables": paths}) |
| 103 | ) |
| 104 | .is_err() |
| 105 | ); |
| 106 | } |
| 107 | assert!( |
| 108 | parse_spawn_request( |
| 109 | &json!({"type":"implement", "prompt":"report", "deliverables": vec!["x.md"; 17]}) |
| 110 | ) |
| 111 | .is_err() |
| 112 | ); |
| 113 | assert!( |
| 114 | delivery::declared_paths(&[], Some("review findings")) |
| 115 | .unwrap() |
| 116 | .is_empty() |
| 117 | ); |
| 118 | assert_eq!( |
| 119 | delivery::declared_paths(&[], Some("report.md")).unwrap(), |
| 120 | ["report.md"] |
| 121 | ); |
| 122 | } |
| 123 | |
| 124 | #[test] |
| 125 | fn deliverable_verdict_distinguishes_present_missing_empty_directory_and_scope() { |
| 126 | let tmp = tempdir().unwrap(); |
| 127 | fs::write(tmp.path().join("present.md"), "report").unwrap(); |
| 128 | fs::write(tmp.path().join("empty.md"), "").unwrap(); |
| 129 | fs::create_dir(tmp.path().join("directory")).unwrap(); |
| 130 | for (path, status) in [ |
| 131 | ("present.md", "present"), |
| 132 | ("missing.md", "missing"), |
| 133 | ("empty.md", "empty"), |
| 134 | ("directory", "not_file"), |
| 135 | ] { |
| 136 | assert_eq!( |
| 137 | delivery::check_deliverable(tmp.path(), path, true).status, |
| 138 | status |
| 139 | ); |
| 140 | } |
| 141 | assert_eq!( |
| 142 | delivery::check_deliverable(tmp.path(), "present.md", false).status, |
| 143 | "out_of_scope" |
| 144 | ); |
| 145 | } |
| 146 | |
| 147 | #[cfg(unix)] |
| 148 | #[test] |
| 149 | fn deliverables_refuse_leaf_and_parent_symlink_escape() { |
| 150 | let tmp = tempdir().unwrap(); |
| 151 | let outside = tempdir().unwrap(); |
| 152 | fs::write(outside.path().join("secret.md"), "outside").unwrap(); |
| 153 | std::os::unix::fs::symlink(outside.path().join("secret.md"), tmp.path().join("leaf.md")) |
| 154 | .unwrap(); |
| 155 | std::os::unix::fs::symlink(outside.path(), tmp.path().join("linked")).unwrap(); |
| 156 | for path in ["leaf.md", "linked/secret.md"] { |
| 157 | assert_eq!( |
| 158 | delivery::check_deliverable(tmp.path(), path, true).status, |
| 159 | "invalid_path" |
| 160 | ); |
| 161 | assert!(delivery::safe_deliverable_path(tmp.path(), path).is_err()); |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | #[test] |
| 166 | fn missing_declared_deliverable_is_visible_in_terminal_sentinel() { |
| 167 | let tmp = tempdir().unwrap(); |
| 168 | let (mut manager, id) = worker(tmp.path(), true, &["report.md"], &["."]); |
| 169 | let verification = complete(&mut manager, &id, "Finished the research."); |
| 170 | assert_eq!(verification.status, "deliverable_missing"); |
| 171 | assert_eq!(verification.deliverables[0].status, "missing"); |
| 172 | let mut result = manager.get_result(&id).unwrap(); |
| 173 | result.status = SubAgentStatus::Completed; |
| 174 | let completion = subagent_completion_with_verification( |
| 175 | "workspace", |
| 176 | &result, |
| 177 | None, |
| 178 | Some(&verification), |
| 179 | None, |
| 180 | ); |
| 181 | assert!(completion.payload.contains("deliverable_missing")); |
| 182 | assert!(completion.payload.contains("report.md")); |
| 183 | } |
| 184 | |
| 185 | #[test] |
| 186 | fn declared_output_is_not_an_undeclared_edit_and_undeclared_worker_stays_self_reported() { |
| 187 | let tmp = tempdir().unwrap(); |
| 188 | repository(tmp.path()); |
| 189 | let (mut manager, id) = worker(tmp.path(), true, &["report.md"], &["."]); |
| 190 | fs::write(tmp.path().join("report.md"), "findings").unwrap(); |
| 191 | let verification = complete(&mut manager, &id, "Report ready."); |
| 192 | assert_eq!(verification.status, "deliverables_present"); |
| 193 | assert_eq!(verification.deliverables[0].bytes, Some(8)); |
| 194 | let (mut manager, id) = worker(tmp.path(), false, &[], &[]); |
| 195 | assert_eq!( |
| 196 | complete(&mut manager, &id, "No changes.").status, |
| 197 | "self_report_only" |
| 198 | ); |
| 199 | } |
| 200 | |
| 201 | #[test] |
| 202 | fn change_like_prose_and_line_citations_are_never_edit_claims() { |
| 203 | for report in [ |
| 204 | "Fixed behavior is documented in src/lib.rs:12-19.", |
| 205 | "CHANGES: None\nThe added guard is at src/lib.rs:12-19", |
| 206 | "CHANGES: src/lib.rs:12-19", |
| 207 | "CHANGES:\n- Reviewed src/lib.rs:12-19.", |
| 208 | "CHANGES:\n- Reviewed src/lib.rs:12-19!", |
| 209 | "CHANGES:\n- Reviewed src/lib.rs:12-19:", |
| 210 | "CHANGES: [src/lib.rs:12-19](src/lib.rs#L12-L19)", |
| 211 | "CHANGES: [source](src/lib.rs:12-19)", |
| 212 | ] { |
| 213 | assert!( |
| 214 | delivery::explicit_change_paths(report).is_empty(), |
| 215 | "{report}" |
| 216 | ); |
| 217 | } |
| 218 | assert_eq!( |
| 219 | delivery::explicit_change_paths("CHANGES:\n- src/lib.rs\n- report.md"), |
| 220 | BTreeSet::from(["src/lib.rs".into(), "report.md".into()]) |
| 221 | ); |
| 222 | let tmp = tempdir().unwrap(); |
| 223 | repository(tmp.path()); |
| 224 | let (mut manager, id) = worker(tmp.path(), false, &[], &[]); |
| 225 | assert_eq!( |
| 226 | complete(&mut manager, &id, "CHANGES: src/lib.rs").status, |
| 227 | "self_report_only" |
| 228 | ); |
| 229 | } |
| 230 | |
| 231 | #[test] |
| 232 | fn unchanged_dirty_file_does_not_satisfy_a_new_edit_claim() { |
| 233 | let tmp = tempdir().unwrap(); |
| 234 | repository(tmp.path()); |
| 235 | fs::write(tmp.path().join("src/lib.rs"), "existing dirty work\n").unwrap(); |
| 236 | let (mut manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 237 | let verification = complete(&mut manager, &id, "CHANGES: src/lib.rs"); |
| 238 | assert_eq!(verification.status, "claim_mismatch"); |
| 239 | assert!(verification.summary.contains("declared but unchanged")); |
| 240 | assert!(verification.summary.contains("src/lib.rs")); |
| 241 | } |
| 242 | |
| 243 | #[test] |
| 244 | fn five_heading_output_declares_changed_files_without_claiming_evidence_or_risks() { |
| 245 | let tmp = tempdir().unwrap(); |
| 246 | repository(tmp.path()); |
| 247 | let (mut manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 248 | fs::write(tmp.path().join("src/lib.rs"), "updated\n").unwrap(); |
| 249 | manager |
| 250 | .worker_records |
| 251 | .get_mut(&id) |
| 252 | .unwrap() |
| 253 | .delivery_evidence |
| 254 | .observed_writes |
| 255 | .insert("src/lib.rs".into()); |
| 256 | let report = "### SUMMARY\n\nUpdated the parser.\n\n\ |
| 257 | ### EVIDENCE\n\n- Reviewed src/reference.rs:12-19.\n\n\ |
| 258 | ### CHANGES\n\n- `src/lib.rs` — adjusted the parser\n\n\ |
| 259 | ### RISKS\n\n- src/consumer.rs still needs a separate review\n\n\ |
| 260 | ### BLOCKERS\n\nNone.\n"; |
| 261 | for report in [ |
| 262 | report.to_string(), |
| 263 | report.replace("### CHANGES", "### changes"), |
| 264 | ] { |
| 265 | assert_eq!( |
| 266 | delivery::explicit_change_paths(&report), |
| 267 | BTreeSet::from(["src/lib.rs".into()]) |
| 268 | ); |
| 269 | assert_ne!( |
| 270 | complete(&mut manager, &id, &report).status, |
| 271 | "claim_mismatch" |
| 272 | ); |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | #[test] |
| 277 | fn changes_bullet_descriptions_do_not_invent_file_claims() { |
| 278 | let tmp = tempdir().unwrap(); |
| 279 | repository(tmp.path()); |
| 280 | let (mut manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 281 | fs::write(tmp.path().join("src/lib.rs"), "updated\n").unwrap(); |
| 282 | manager |
| 283 | .worker_records |
| 284 | .get_mut(&id) |
| 285 | .unwrap() |
| 286 | .delivery_evidence |
| 287 | .observed_writes |
| 288 | .insert("src/lib.rs".into()); |
| 289 | for declaration in [ |
| 290 | "- `src/lib.rs` — updated parsing.", |
| 291 | "- src/lib.rs: Updated parsing.", |
| 292 | "- src/lib.rs updated parsing; reviewed notes.md.", |
| 293 | "- [src/lib.rs](src/lib.rs) - Updated parsing! See notes.md.", |
| 294 | "- 'src/lib.rs' — matches src/reference.rs:12-19.", |
| 295 | ] { |
| 296 | let report = format!( |
| 297 | "### SUMMARY\n\nUpdated the parser.\n\n\ |
| 298 | ### EVIDENCE\n\n- Reviewed src/reference.rs:12-19.\n\n\ |
| 299 | ### CHANGES\n\n{declaration}\n\n\ |
| 300 | ### RISKS\n\n- Check notes.md separately.\n\n\ |
| 301 | ### BLOCKERS\n\nNone.\n" |
| 302 | ); |
| 303 | assert_eq!( |
| 304 | delivery::explicit_change_paths(&report), |
| 305 | BTreeSet::from(["src/lib.rs".into()]), |
| 306 | "{declaration}" |
| 307 | ); |
| 308 | assert_ne!( |
| 309 | complete(&mut manager, &id, &report).status, |
| 310 | "claim_mismatch", |
| 311 | "{declaration}" |
| 312 | ); |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | #[test] |
| 317 | fn explicit_change_path_lists_preserve_quoted_spaces_and_punctuation() { |
| 318 | let expected = BTreeSet::from([ |
| 319 | "src/lib.rs".into(), |
| 320 | "src/with spaces.rs".into(), |
| 321 | "src/another file.rs".into(), |
| 322 | "src/trailing.".into(), |
| 323 | ]); |
| 324 | for label in ["CHANGES:", "Changed files:", "Files changed:"] { |
| 325 | let report = format!( |
| 326 | "{label} src/lib.rs, `src/with spaces.rs`; \"src/another file.rs\" 'src/trailing.'" |
| 327 | ); |
| 328 | assert_eq!(delivery::explicit_change_paths(&report), expected); |
| 329 | } |
| 330 | assert!(delivery::explicit_change_paths("CHANGES: Updated parsing.").is_empty()); |
| 331 | assert!( |
| 332 | delivery::explicit_change_paths("CHANGES: src/reference.rs:12-19. — reviewed only.") |
| 333 | .is_empty() |
| 334 | ); |
| 335 | } |
| 336 | |
| 337 | #[test] |
| 338 | fn modification_of_already_dirty_file_is_measured_against_spawn_content() { |
| 339 | let tmp = tempdir().unwrap(); |
| 340 | repository(tmp.path()); |
| 341 | fs::write(tmp.path().join("src/lib.rs"), "existing dirty work\n").unwrap(); |
| 342 | let (mut manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 343 | fs::write( |
| 344 | tmp.path().join("src/lib.rs"), |
| 345 | "worker changed this further\n", |
| 346 | ) |
| 347 | .unwrap(); |
| 348 | assert_eq!( |
| 349 | complete(&mut manager, &id, "CHANGES: src/lib.rs").status, |
| 350 | "self_report_only" |
| 351 | ); |
| 352 | } |
| 353 | |
| 354 | #[test] |
| 355 | fn committed_change_is_compared_with_exact_spawn_head_without_timestamp_guessing() { |
| 356 | let tmp = tempdir().unwrap(); |
| 357 | repository(tmp.path()); |
| 358 | let (mut manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 359 | fs::write(tmp.path().join("src/lib.rs"), "worker commit\n").unwrap(); |
| 360 | git(tmp.path(), &["add", "--", "src/lib.rs"]); |
| 361 | git(tmp.path(), &["commit", "--quiet", "-m", "worker"]); |
| 362 | assert_eq!( |
| 363 | complete(&mut manager, &id, "CHANGES: src/lib.rs").status, |
| 364 | "self_report_only" |
| 365 | ); |
| 366 | } |
| 367 | |
| 368 | #[test] |
| 369 | fn an_observed_write_without_a_declaration_is_flagged_but_external_changes_are_not() { |
| 370 | let tmp = tempdir().unwrap(); |
| 371 | repository(tmp.path()); |
| 372 | let (mut manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 373 | fs::write(tmp.path().join("src/lib.rs"), "new change\n").unwrap(); |
| 374 | manager |
| 375 | .worker_records |
| 376 | .get_mut(&id) |
| 377 | .unwrap() |
| 378 | .delivery_evidence |
| 379 | .observed_writes |
| 380 | .insert("src/lib.rs".into()); |
| 381 | assert_eq!( |
| 382 | complete(&mut manager, &id, "CHANGES: None").status, |
| 383 | "claim_mismatch" |
| 384 | ); |
| 385 | for scope in ["reports", "."] { |
| 386 | let (mut manager, id) = worker(tmp.path(), true, &[], &[scope]); |
| 387 | fs::write( |
| 388 | tmp.path().join("src/lib.rs"), |
| 389 | format!("external {scope} change\n"), |
| 390 | ) |
| 391 | .unwrap(); |
| 392 | assert!( |
| 393 | manager.worker_records[&id] |
| 394 | .delivery_evidence |
| 395 | .observed_writes |
| 396 | .is_empty() |
| 397 | ); |
| 398 | assert_eq!( |
| 399 | complete(&mut manager, &id, "CHANGES: None").status, |
| 400 | "self_report_only" |
| 401 | ); |
| 402 | } |
| 403 | } |
| 404 | |
| 405 | #[test] |
| 406 | fn disjoint_sibling_write_paths_admit_and_ancestor_overlap_names_actual_remedy() { |
| 407 | let mut ledger = CoordinationLedger::default(); |
| 408 | let claim = |owner: &str, path: &str| WriteScopeClaim { |
| 409 | owner: owner.into(), |
| 410 | roots: vec![path.into()], |
| 411 | exact_files: Vec::new(), |
| 412 | contracts: Vec::new(), |
| 413 | }; |
| 414 | ledger |
| 415 | .register_claim(claim("a", "tmp/scan/a"), false, |_| true) |
| 416 | .unwrap(); |
| 417 | ledger |
| 418 | .register_claim(claim("b", "tmp/scan/b"), false, |_| true) |
| 419 | .unwrap(); |
| 420 | let error = ledger |
| 421 | .register_claim(claim("broad", "tmp/scan"), false, |_| true) |
| 422 | .unwrap_err(); |
| 423 | for text in [ |
| 424 | "tmp/scan/a", |
| 425 | "tmp/scan", |
| 426 | "disjoint sibling", |
| 427 | "exact_files", |
| 428 | "write_authority=read_only", |
| 429 | ] { |
| 430 | assert!(error.contains(text), "{error}"); |
| 431 | } |
| 432 | } |
| 433 | |
| 434 | #[cfg(unix)] |
| 435 | #[tokio::test] |
| 436 | #[allow(clippy::print_stderr)] // Test receipt must distinguish refused probes from exercised isolation. |
| 437 | async fn enforced_readonly_python_queries_sqlite_under_a_live_peer_write_claim() { |
| 438 | let tmp = tempdir().unwrap(); |
| 439 | let database = rusqlite::Connection::open(tmp.path().join("fixture.sqlite")).unwrap(); |
| 440 | database |
| 441 | .execute_batch( |
| 442 | "CREATE TABLE fixture(value TEXT); INSERT INTO fixture VALUES ('peer-read-receipt');", |
| 443 | ) |
| 444 | .unwrap(); |
| 445 | drop(database); |
| 446 | fs::write(tmp.path().join("peer.txt"), "preserve peer bytes").unwrap(); |
| 447 | fs::write(tmp.path().join("own.txt"), "own bytes").unwrap(); |
| 448 | let manager = new_shared_subagent_manager(tmp.path().to_path_buf(), 4); |
| 449 | { |
| 450 | let mut guard = manager.write().await; |
| 451 | guard.insert_test_running_agent("analysis", tmp.path()); |
| 452 | guard.insert_test_running_agent("peer", tmp.path()); |
| 453 | for (owner, file) in [("agent_analysis", "own.txt"), ("agent_peer", "peer.txt")] { |
| 454 | guard |
| 455 | .coordination |
| 456 | .register_claim( |
| 457 | WriteScopeClaim { |
| 458 | owner: owner.into(), |
| 459 | roots: Vec::new(), |
| 460 | exact_files: vec![file.into()], |
| 461 | contracts: Vec::new(), |
| 462 | }, |
| 463 | false, |
| 464 | |_| true, |
| 465 | ) |
| 466 | .unwrap(); |
| 467 | } |
| 468 | assert_eq!( |
| 469 | guard.live_peer_shared_write_claim_owners("agent_analysis"), |
| 470 | ["agent_peer"] |
| 471 | ); |
| 472 | } |
| 473 | let python = [ |
| 474 | "/usr/bin/python3", |
| 475 | "/opt/homebrew/bin/python3", |
| 476 | "/usr/local/bin/python3", |
| 477 | ] |
| 478 | .into_iter() |
| 479 | .find(|path| Path::new(path).is_file()) |
| 480 | .expect("Python fixture runtime"); |
| 481 | for role in [FleetRole::Builder, FleetRole::Scout] { |
| 482 | let mut runtime = super::tests::stub_runtime(); |
| 483 | runtime.manager = Arc::clone(&manager); |
| 484 | runtime.context = ToolContext::new(tmp.path()); |
| 485 | runtime.context.auto_approve = true; |
| 486 | runtime.context.elevated_sandbox_policy = |
| 487 | Some(crate::sandbox::SandboxPolicy::DangerFullAccess); |
| 488 | #[cfg(target_os = "linux")] |
| 489 | runtime |
| 490 | .context |
| 491 | .shell_manager |
| 492 | .lock() |
| 493 | .unwrap() |
| 494 | .set_prefer_bwrap(true); |
| 495 | let available = runtime |
| 496 | .context |
| 497 | .shell_manager |
| 498 | .lock() |
| 499 | .unwrap() |
| 500 | .configured_sandbox_type() |
| 501 | .is_some(); |
| 502 | runtime.worker_profile = WorkerRuntimeProfile::for_role(role.clone()); |
| 503 | let registry = SubAgentToolRegistry::new_with_owner( |
| 504 | runtime, |
| 505 | role, |
| 506 | "agent_analysis".into(), |
| 507 | "analysis".into(), |
| 508 | Some(vec!["bash".into()]), |
| 509 | Arc::new(Mutex::new(TodoList::new())), |
| 510 | Arc::new(Mutex::new(PlanState::default())), |
| 511 | ); |
| 512 | let script = "import sqlite3; c=sqlite3.connect('file:fixture.sqlite?mode=ro', uri=True); print(c.execute('SELECT value FROM fixture').fetchone()[0])"; |
| 513 | let query = json!({"command": format!("{python} -I -B -c {}", shell_words::quote(script)), "read_only": true}); |
| 514 | let read = registry.execute("agent_analysis", "bash", query).await; |
| 515 | if !available { |
| 516 | let error = read.unwrap_err().to_string(); |
| 517 | assert!(error.contains("native read-only enforcement"), "{error}"); |
| 518 | eprintln!("UNRUN: native child+peer Python probe; unavailable sandbox was refused"); |
| 519 | continue; |
| 520 | } |
| 521 | let read = read.unwrap(); |
| 522 | assert!(read.contains("peer-read-receipt"), "{read}"); |
| 523 | let write = json!({"command": format!("{python} -I -B -c {}", shell_words::quote("open('peer.txt', 'w').write('corrupt')")), "read_only": true}); |
| 524 | let error = registry |
| 525 | .execute("agent_analysis", "bash", write) |
| 526 | .await |
| 527 | .unwrap_err() |
| 528 | .to_string(); |
| 529 | assert!( |
| 530 | !error.contains("blocking peers"), |
| 531 | "read-only execution must reach native enforcement: {error}" |
| 532 | ); |
| 533 | assert_eq!( |
| 534 | fs::read_to_string(tmp.path().join("peer.txt")).unwrap(), |
| 535 | "preserve peer bytes" |
| 536 | ); |
| 537 | assert_eq!( |
| 538 | manager |
| 539 | .read() |
| 540 | .await |
| 541 | .get_result("agent_peer") |
| 542 | .unwrap() |
| 543 | .status, |
| 544 | SubAgentStatus::Running |
| 545 | ); |
| 546 | eprintln!( |
| 547 | "NATIVE_READONLY_ENFORCED: child SQLite query passed under a live peer claim; mutation denied" |
| 548 | ); |
| 549 | } |
| 550 | } |
| 551 | |
| 552 | /// Deferred-verification contract (#6210): the terminal commit stores the |
| 553 | /// worker projection but leaves verification pending for `ensure`. |
| 554 | #[test] |
| 555 | fn terminal_commit_leaves_delivery_verification_pending() { |
| 556 | let tmp = tempdir().unwrap(); |
| 557 | repository(tmp.path()); |
| 558 | let (mut manager, id) = worker(tmp.path(), true, &["report.md"], &["."]); |
| 559 | let mut result = manager.get_result(&id).unwrap(); |
| 560 | result.status = SubAgentStatus::Completed; |
| 561 | result.result = Some("Finished the research.".into()); |
| 562 | manager.complete_worker_from_result(&id, &result); |
| 563 | let record = manager.worker_records.get(&id).unwrap(); |
| 564 | assert!(!record.delivery_evidence.checked); |
| 565 | assert_eq!(record.verification.status, "self_report_only"); |
| 566 | assert_eq!( |
| 567 | record.result_summary.as_deref(), |
| 568 | Some("Finished the research.") |
| 569 | ); |
| 570 | } |
| 571 | |
| 572 | #[tokio::test] |
| 573 | async fn ensure_worker_delivery_verified_stores_verdicts_and_is_idempotent() { |
| 574 | let tmp = tempdir().unwrap(); |
| 575 | repository(tmp.path()); |
| 576 | let (manager, id) = worker(tmp.path(), true, &["report.md"], &["."]); |
| 577 | let manager = Arc::new(RwLock::new(manager)); |
| 578 | let result = { |
| 579 | let mut guard = manager.write().await; |
| 580 | let mut result = guard.get_result(&id).unwrap(); |
| 581 | result.status = SubAgentStatus::Completed; |
| 582 | result.result = Some("Finished the research.".into()); |
| 583 | guard.complete_worker_from_result(&id, &result); |
| 584 | assert!(!guard.worker_records[&id].delivery_evidence.checked); |
| 585 | result |
| 586 | }; |
| 587 | ensure_worker_delivery_verified(&manager, &id, &result).await; |
| 588 | let first = manager.read().await.worker_records[&id] |
| 589 | .verification |
| 590 | .clone(); |
| 591 | assert_eq!(first.status, "deliverable_missing"); |
| 592 | assert!( |
| 593 | manager.read().await.worker_records[&id] |
| 594 | .delivery_evidence |
| 595 | .checked |
| 596 | ); |
| 597 | // A second call is a no-op even with a different report. |
| 598 | let mut other = result.clone(); |
| 599 | other.result = Some("CHANGES: src/lib.rs".into()); |
| 600 | ensure_worker_delivery_verified(&manager, &id, &other).await; |
| 601 | assert_eq!(manager.read().await.worker_records[&id].verification, first); |
| 602 | } |
| 603 | |
| 604 | #[tokio::test] |
| 605 | async fn ensure_worker_delivery_verified_ignores_running_missing_and_checked() { |
| 606 | let tmp = tempdir().unwrap(); |
| 607 | repository(tmp.path()); |
| 608 | let (manager, id) = worker(tmp.path(), true, &[], &["src"]); |
| 609 | let manager = Arc::new(RwLock::new(manager)); |
| 610 | let running = manager.read().await.get_result(&id).unwrap(); |
| 611 | assert_eq!(running.status, SubAgentStatus::Running); |
| 612 | ensure_worker_delivery_verified(&manager, &id, &running).await; |
| 613 | assert!( |
| 614 | !manager.read().await.worker_records[&id] |
| 615 | .delivery_evidence |
| 616 | .checked |
| 617 | ); |
| 618 | // A missing worker id is a silent no-op. |
| 619 | let mut missing = running.clone(); |
| 620 | missing.agent_id = "agent_missing".to_string(); |
| 621 | missing.status = SubAgentStatus::Completed; |
| 622 | ensure_worker_delivery_verified(&manager, "agent_missing", &missing).await; |
| 623 | // A checked record keeps its stored verdict. |
| 624 | let mut done = running.clone(); |
| 625 | done.status = SubAgentStatus::Completed; |
| 626 | done.result = Some("CHANGES: src/lib.rs".into()); |
| 627 | ensure_worker_delivery_verified(&manager, &id, &done).await; |
| 628 | assert_eq!( |
| 629 | manager.read().await.worker_records[&id].verification.status, |
| 630 | "claim_mismatch" |
| 631 | ); |
| 632 | let mut changed_mind = done.clone(); |
| 633 | changed_mind.result = Some("CHANGES: None".into()); |
| 634 | ensure_worker_delivery_verified(&manager, &id, &changed_mind).await; |
| 635 | assert_eq!( |
| 636 | manager.read().await.worker_records[&id].verification.status, |
| 637 | "claim_mismatch" |
| 638 | ); |
| 639 | } |
| 640 | |
| 641 | /// Read-side backstop (#6210): a terminal detail projection heals a |
| 642 | /// verification left pending by a Stop/interrupt/close/stale commit. |
| 643 | #[tokio::test] |
| 644 | async fn detail_projection_heals_pending_delivery_verification() { |
| 645 | let tmp = tempdir().unwrap(); |
| 646 | repository(tmp.path()); |
| 647 | let (manager, id) = worker(tmp.path(), true, &["report.md"], &["."]); |
| 648 | let manager = Arc::new(RwLock::new(manager)); |
| 649 | let result = { |
| 650 | let mut guard = manager.write().await; |
| 651 | let mut result = guard.get_result(&id).unwrap(); |
| 652 | result.status = SubAgentStatus::Completed; |
| 653 | result.result = Some("Finished the research.".into()); |
| 654 | guard.complete_worker_from_result(&id, &result); |
| 655 | result |
| 656 | }; |
| 657 | let mut context = ToolContext::new(tmp.path()); |
| 658 | context.state_namespace = "workspace".to_string(); |
| 659 | let worker_record = manager |
| 660 | .read() |
| 661 | .await |
| 662 | .get_worker_record_for_session("workspace", &id); |
| 663 | assert!( |
| 664 | worker_record |
| 665 | .as_ref() |
| 666 | .is_some_and(|record| !record.delivery_evidence.checked) |
| 667 | ); |
| 668 | let projection = |
| 669 | subagent_session_projection(&manager, result, false, &context, worker_record).await; |
| 670 | assert_eq!(projection.verification.status, "deliverable_missing"); |
| 671 | assert!( |
| 672 | projection |
| 673 | .verification |
| 674 | .deliverables |
| 675 | .iter() |
| 676 | .any(|verdict| verdict.path == "report.md") |
| 677 | ); |
| 678 | assert!( |
| 679 | manager.read().await.worker_records[&id] |
| 680 | .delivery_evidence |
| 681 | .checked |
| 682 | ); |
| 683 | } |
| 684 |