| 1 | package control |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "strings" |
| 7 | "testing" |
| 8 | "time" |
| 9 | |
| 10 | "reasonix/internal/agent" |
| 11 | "reasonix/internal/event" |
| 12 | "reasonix/internal/permission" |
| 13 | "reasonix/internal/provider" |
| 14 | "reasonix/internal/sandbox" |
| 15 | "reasonix/internal/tool" |
| 16 | ) |
| 17 | |
| 18 | // TestAutoApproveToolsStillRequiresExplicitPlanApproval proves that YOLO/full |
| 19 | // tool access does not bypass the separate Plan Mode collaboration gate. |
| 20 | func TestAutoApproveToolsStillRequiresExplicitPlanApproval(t *testing.T) { |
| 21 | prov := &scriptedTurns{turns: planThenExecuteTurns( |
| 22 | "Plan:\n1. Add the config field\n2. Wire it into boot\n3. Add tests", |
| 23 | "Done — implemented the approved plan.", |
| 24 | )} |
| 25 | ag := newPlanTestAgent(prov) |
| 26 | |
| 27 | approvalRequests := make(chan event.Approval, 1) |
| 28 | c := newOwnedTestController(t, Options{ |
| 29 | Runner: ag, |
| 30 | Executor: ag, |
| 31 | Sink: event.FuncSink(func(e event.Event) { |
| 32 | switch e.Kind { |
| 33 | case event.ApprovalRequest: |
| 34 | approvalRequests <- e.Approval |
| 35 | } |
| 36 | }), |
| 37 | }) |
| 38 | c.SetToolApprovalMode(ToolApprovalDangerFullAccess) |
| 39 | c.SetPlanMode(true) |
| 40 | |
| 41 | input := "实现 issue #2395：新增配置项、自动判断复杂任务、补测试和文档" |
| 42 | done := make(chan error, 1) |
| 43 | go func() { done <- c.runTurnWithRaw(context.Background(), input, input) }() |
| 44 | |
| 45 | var approval event.Approval |
| 46 | select { |
| 47 | case approval = <-approvalRequests: |
| 48 | case <-time.After(30 * time.Second): |
| 49 | t.Fatal("tool auto-approval must not suppress plan approval") |
| 50 | } |
| 51 | if approval.Tool != planApprovalTool { |
| 52 | t.Fatalf("approval tool = %q, want %q", approval.Tool, planApprovalTool) |
| 53 | } |
| 54 | |
| 55 | if !c.PlanMode() { |
| 56 | t.Fatal("controller should stay in plan mode while waiting for approval") |
| 57 | } |
| 58 | c.Approve(approval.ID, true, false, false) |
| 59 | |
| 60 | select { |
| 61 | case err := <-done: |
| 62 | if err != nil { |
| 63 | t.Fatalf("runTurnWithRaw: %v", err) |
| 64 | } |
| 65 | case <-time.After(30 * time.Second): |
| 66 | t.Fatal("approved plan did not continue into execution") |
| 67 | } |
| 68 | if got := agent.StripTransientUserBlocks(firstUserMessage(ag.Session().Messages)); !strings.HasPrefix(got, PlanModeMarker) { |
| 69 | t.Fatalf("first model input = %q, want the plan marker prefixed", got) |
| 70 | } |
| 71 | if c.PlanMode() { |
| 72 | t.Fatal("plan mode should be off after approval") |
| 73 | } |
| 74 | if !c.AutoApproveTools() { |
| 75 | t.Fatal("tool auto-approval should remain on after plan approval") |
| 76 | } |
| 77 | if got := c.Todos(); len(got) != 0 { |
| 78 | t.Fatalf("approved plan seeded todo state: %+v", got) |
| 79 | } |
| 80 | if prov.call != 3 { |
| 81 | t.Fatalf("provider called %d times, want 3 (plan + read + answer)", prov.call) |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | // TestRequestApprovalHonorsAutoApproveTools guards the underlying gate: ordinary |
| 86 | // tool approvals must return allow immediately without emitting anything under |
| 87 | // tool auto-approval. |
| 88 | func TestRequestApprovalHonorsAutoApproveTools(t *testing.T) { |
| 89 | var approvalRequested bool |
| 90 | c := newOwnedTestController(t, Options{ |
| 91 | Sink: event.FuncSink(func(e event.Event) { |
| 92 | if e.Kind == event.ApprovalRequest { |
| 93 | approvalRequested = true |
| 94 | } |
| 95 | }), |
| 96 | }) |
| 97 | c.SetToolApprovalMode(ToolApprovalDangerFullAccess) |
| 98 | |
| 99 | done := make(chan bool, 1) |
| 100 | go func() { |
| 101 | allow, _, err := c.requestApproval(context.Background(), "multi_edit", "/tmp/file", nil) |
| 102 | if err != nil { |
| 103 | t.Errorf("requestApproval: %v", err) |
| 104 | } |
| 105 | done <- allow |
| 106 | }() |
| 107 | |
| 108 | select { |
| 109 | case allow := <-done: |
| 110 | if !allow { |
| 111 | t.Fatal("tool auto-approval should allow the approval") |
| 112 | } |
| 113 | case <-time.After(30 * time.Second): |
| 114 | t.Fatal("requestApproval blocked under tool auto-approval") |
| 115 | } |
| 116 | |
| 117 | if approvalRequested { |
| 118 | t.Fatal("tool auto-approval must not emit an ApprovalRequest event") |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | func TestToolApprovalModeAutoKeepsAskRules(t *testing.T) { |
| 123 | c := newOwnedTestController(t, Options{ |
| 124 | Policy: permission.New("ask", nil, []string{"bash(git commit*)"}, []string{"bash(rm*)"}), |
| 125 | }) |
| 126 | c.SetToolApprovalMode(ToolApprovalAuto) |
| 127 | |
| 128 | gate := c.newInteractiveGate() |
| 129 | if got := gate.Policy.Decide("bash", false, json.RawMessage(`{"command":"go test ./..."}`)); got != permission.Allow { |
| 130 | t.Fatalf("auto mode fallback = %v, want allow", got) |
| 131 | } |
| 132 | if got := gate.Policy.Decide("bash", false, json.RawMessage(`{"command":"git commit -m x"}`)); got != permission.Ask { |
| 133 | t.Fatalf("explicit ask rule = %v, want ask", got) |
| 134 | } |
| 135 | if got := gate.Policy.Decide("bash", false, json.RawMessage(`{"command":"rm -rf build"}`)); got != permission.Deny { |
| 136 | t.Fatalf("deny rule = %v, want deny", got) |
| 137 | } |
| 138 | if c.AutoApproveTools() { |
| 139 | t.Fatal("auto approval must not report as YOLO") |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | func TestToolApprovalModeDontAskDeniesWithoutPrompt(t *testing.T) { |
| 144 | requests := 0 |
| 145 | c := newOwnedTestController(t, Options{ |
| 146 | Policy: permission.New("ask", nil, []string{"bash(git commit*)"}, nil). |
| 147 | WithSessionAllow([]string{"bash(go test*)"}), |
| 148 | Sink: event.FuncSink(func(e event.Event) { |
| 149 | if e.Kind == event.ApprovalRequest { |
| 150 | requests++ |
| 151 | } |
| 152 | }), |
| 153 | }) |
| 154 | c.SetToolApprovalMode(ToolApprovalDontAsk) |
| 155 | gate := c.newInteractiveGate() |
| 156 | |
| 157 | allow, _, err := gate.Check(context.Background(), "bash", json.RawMessage(`{"command":"go test ./..."}`), false) |
| 158 | if err != nil || !allow { |
| 159 | t.Fatalf("session-allowed call = (%v, %v), want allow", allow, err) |
| 160 | } |
| 161 | allow, _, err = gate.Check(context.Background(), "bash", json.RawMessage(`{"command":"git commit -m x"}`), false) |
| 162 | if err != nil || allow { |
| 163 | t.Fatalf("explicit ask under dontAsk = (%v, %v), want deny", allow, err) |
| 164 | } |
| 165 | allow, _, err = gate.Check(context.Background(), "write_file", json.RawMessage(`{"path":"x.txt"}`), false) |
| 166 | if err != nil || allow { |
| 167 | t.Fatalf("fallback under dontAsk = (%v, %v), want deny", allow, err) |
| 168 | } |
| 169 | if requests != 0 { |
| 170 | t.Fatalf("dontAsk emitted %d approval requests, want 0", requests) |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | func TestLegacyAutoMigrationDoesNotApprovePendingFallback(t *testing.T) { |
| 175 | approvalRequests := make(chan event.Approval, 1) |
| 176 | c := newOwnedTestController(t, Options{ |
| 177 | Policy: permission.New("ask", nil, nil, nil), |
| 178 | Sink: event.FuncSink(func(e event.Event) { |
| 179 | if e.Kind == event.ApprovalRequest { |
| 180 | approvalRequests <- e.Approval |
| 181 | } |
| 182 | }), |
| 183 | }) |
| 184 | |
| 185 | done := make(chan bool, 1) |
| 186 | errs := make(chan error, 1) |
| 187 | go func() { |
| 188 | allow, _, err := c.requestApproval(context.Background(), "multi_edit", "/tmp/file", nil) |
| 189 | if err != nil { |
| 190 | errs <- err |
| 191 | return |
| 192 | } |
| 193 | done <- allow |
| 194 | }() |
| 195 | |
| 196 | var approval event.Approval |
| 197 | select { |
| 198 | case approval = <-approvalRequests: |
| 199 | case <-time.After(30 * time.Second): |
| 200 | t.Fatal("approval request was not emitted") |
| 201 | } |
| 202 | |
| 203 | c.SetToolApprovalMode(ToolApprovalAuto) |
| 204 | |
| 205 | select { |
| 206 | case err := <-errs: |
| 207 | t.Fatalf("requestApproval returned unexpectedly: %v", err) |
| 208 | case allow := <-done: |
| 209 | t.Fatalf("legacy auto migration resolved pending approval unexpectedly: allow=%v", allow) |
| 210 | case <-time.After(50 * time.Millisecond): |
| 211 | } |
| 212 | c.Approve(approval.ID, true, true, false) |
| 213 | select { |
| 214 | case err := <-errs: |
| 215 | t.Fatalf("requestApproval: %v", err) |
| 216 | case allow := <-done: |
| 217 | if !allow { |
| 218 | t.Fatal("manual session approval returned deny") |
| 219 | } |
| 220 | case <-time.After(30 * time.Second): |
| 221 | t.Fatal("pending fallback approval stayed blocked after manual approval") |
| 222 | } |
| 223 | if c.AutoApproveTools() { |
| 224 | t.Fatal("auto mode must not report as YOLO") |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | func TestToolApprovalModeAutoDoesNotDrainPendingExplicitAsk(t *testing.T) { |
| 229 | approvalRequests := make(chan event.Approval, 1) |
| 230 | c := newOwnedTestController(t, Options{ |
| 231 | Policy: permission.New("ask", nil, []string{"bash(git commit*)"}, nil), |
| 232 | Sink: event.FuncSink(func(e event.Event) { |
| 233 | if e.Kind == event.ApprovalRequest { |
| 234 | approvalRequests <- e.Approval |
| 235 | } |
| 236 | }), |
| 237 | }) |
| 238 | |
| 239 | done := make(chan bool, 1) |
| 240 | errs := make(chan error, 1) |
| 241 | go func() { |
| 242 | allow, _, err := c.requestApproval(context.Background(), "bash", "git commit -m x", nil) |
| 243 | if err != nil { |
| 244 | errs <- err |
| 245 | return |
| 246 | } |
| 247 | done <- allow |
| 248 | }() |
| 249 | |
| 250 | var approval event.Approval |
| 251 | select { |
| 252 | case approval = <-approvalRequests: |
| 253 | case <-time.After(30 * time.Second): |
| 254 | t.Fatal("approval request was not emitted") |
| 255 | } |
| 256 | |
| 257 | c.SetToolApprovalMode(ToolApprovalAuto) |
| 258 | |
| 259 | select { |
| 260 | case err := <-errs: |
| 261 | t.Fatalf("requestApproval: %v", err) |
| 262 | case allow := <-done: |
| 263 | t.Fatalf("auto mode must not answer explicit ask rules; got allow=%v", allow) |
| 264 | case <-time.After(50 * time.Millisecond): |
| 265 | } |
| 266 | |
| 267 | c.Approve(approval.ID, true, false, false) |
| 268 | |
| 269 | select { |
| 270 | case err := <-errs: |
| 271 | t.Fatalf("requestApproval: %v", err) |
| 272 | case allow := <-done: |
| 273 | if !allow { |
| 274 | t.Fatal("manual approval should allow the explicit ask request") |
| 275 | } |
| 276 | case <-time.After(30 * time.Second): |
| 277 | t.Fatal("explicit ask approval stayed blocked after manual Approve") |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | func TestToolApprovalModeYoloBypassesApprovalPrompts(t *testing.T) { |
| 282 | c := newOwnedTestController(t, Options{}) |
| 283 | c.SetToolApprovalMode(ToolApprovalYolo) |
| 284 | if !c.AutoApproveTools() { |
| 285 | t.Fatal("YOLO mode should satisfy legacy AutoApproveTools") |
| 286 | } |
| 287 | allow, remember, err := c.requestApproval(context.Background(), "bash", "go test ./...", nil) |
| 288 | if err != nil || !allow || remember { |
| 289 | t.Fatalf("requestApproval in YOLO = (%v,%v,%v), want allow without remember", allow, remember, err) |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | func TestPlanApprovalIgnoresAutoApproveTools(t *testing.T) { |
| 294 | approvalRequests := make(chan event.Approval, 1) |
| 295 | c := newOwnedTestController(t, Options{ |
| 296 | Sink: event.FuncSink(func(e event.Event) { |
| 297 | if e.Kind == event.ApprovalRequest { |
| 298 | approvalRequests <- e.Approval |
| 299 | } |
| 300 | }), |
| 301 | }) |
| 302 | c.SetToolApprovalMode(ToolApprovalDangerFullAccess) |
| 303 | |
| 304 | done := make(chan bool, 1) |
| 305 | errs := make(chan error, 1) |
| 306 | go func() { |
| 307 | allow, _, err := c.requestApproval(context.Background(), planApprovalTool, "", nil) |
| 308 | if err != nil { |
| 309 | errs <- err |
| 310 | return |
| 311 | } |
| 312 | done <- allow |
| 313 | }() |
| 314 | |
| 315 | var approval event.Approval |
| 316 | select { |
| 317 | case approval = <-approvalRequests: |
| 318 | case <-time.After(30 * time.Second): |
| 319 | t.Fatal("plan approval must still prompt under tool auto-approval") |
| 320 | } |
| 321 | if approval.Tool != planApprovalTool { |
| 322 | t.Fatalf("approval tool = %q, want %q", approval.Tool, planApprovalTool) |
| 323 | } |
| 324 | select { |
| 325 | case allow := <-done: |
| 326 | t.Fatalf("plan approval must wait for the user under tool auto-approval; got allow=%v", allow) |
| 327 | case err := <-errs: |
| 328 | t.Fatalf("requestApproval: %v", err) |
| 329 | default: |
| 330 | } |
| 331 | |
| 332 | c.Approve(approval.ID, true, false, false) |
| 333 | |
| 334 | select { |
| 335 | case err := <-errs: |
| 336 | t.Fatalf("requestApproval: %v", err) |
| 337 | case allow := <-done: |
| 338 | if !allow { |
| 339 | t.Fatal("manual plan approval should allow") |
| 340 | } |
| 341 | case <-time.After(30 * time.Second): |
| 342 | t.Fatal("plan approval stayed blocked after Approve") |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | // Legacy SetAutoApproveTools conservatively maps to workspace access and must |
| 347 | // not answer an approval created under an older permission revision. |
| 348 | func TestSetAutoApproveToolsDoesNotResolvePendingApproval(t *testing.T) { |
| 349 | c, ids, _ := approvalIDs(t) |
| 350 | |
| 351 | done := make(chan bool, 1) |
| 352 | errs := make(chan error, 1) |
| 353 | go func() { |
| 354 | allow, _, err := c.requestApproval(context.Background(), "multi_edit", "/tmp/file", nil) |
| 355 | if err != nil { |
| 356 | errs <- err |
| 357 | return |
| 358 | } |
| 359 | done <- allow |
| 360 | }() |
| 361 | |
| 362 | var approvalID string |
| 363 | select { |
| 364 | case approvalID = <-ids: |
| 365 | case <-time.After(30 * time.Second): |
| 366 | t.Fatal("approval request was not emitted") |
| 367 | } |
| 368 | |
| 369 | c.SetAutoApproveTools(true) |
| 370 | |
| 371 | select { |
| 372 | case err := <-errs: |
| 373 | t.Fatalf("requestApproval: %v", err) |
| 374 | case allow := <-done: |
| 375 | t.Fatalf("legacy mode change answered pending approval: allow=%v", allow) |
| 376 | case <-time.After(50 * time.Millisecond): |
| 377 | } |
| 378 | if c.AutoApproveTools() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 379 | t.Fatalf("legacy mode = %q full=%v, want workspace-write without full access", c.ToolApprovalMode(), c.AutoApproveTools()) |
| 380 | } |
| 381 | c.Approve(approvalID, true, false, false) |
| 382 | if allow := <-done; !allow { |
| 383 | t.Fatal("manual approval should allow") |
| 384 | } |
| 385 | } |
| 386 | |
| 387 | func TestSandboxEscapeApprovalIgnoresAutoApproveTools(t *testing.T) { |
| 388 | approvalRequests := make(chan event.Approval, 1) |
| 389 | c := newOwnedTestController(t, Options{ |
| 390 | Sink: event.FuncSink(func(e event.Event) { |
| 391 | if e.Kind == event.ApprovalRequest { |
| 392 | approvalRequests <- e.Approval |
| 393 | } |
| 394 | }), |
| 395 | }) |
| 396 | c.SetToolApprovalMode(ToolApprovalDangerFullAccess) |
| 397 | |
| 398 | type escapeResult struct { |
| 399 | allow bool |
| 400 | reason string |
| 401 | err error |
| 402 | } |
| 403 | done := make(chan escapeResult, 1) |
| 404 | go func() { |
| 405 | allow, reason, err := sandboxEscapeApprover{c}.ApproveSandboxEscape(context.Background(), sandbox.EscapeRequest{ |
| 406 | Command: "go test ./...", |
| 407 | Reason: "Windows sandbox failed. Run this command unconfined once?", |
| 408 | }) |
| 409 | done <- escapeResult{allow: allow, reason: reason, err: err} |
| 410 | }() |
| 411 | |
| 412 | var approval event.Approval |
| 413 | select { |
| 414 | case approval = <-approvalRequests: |
| 415 | case <-time.After(30 * time.Second): |
| 416 | t.Fatal("sandbox escape approval request was not emitted") |
| 417 | } |
| 418 | if approval.Tool != SandboxEscapeApprovalTool { |
| 419 | t.Fatalf("approval tool = %q, want %q", approval.Tool, SandboxEscapeApprovalTool) |
| 420 | } |
| 421 | |
| 422 | c.SetAutoApproveTools(true) |
| 423 | select { |
| 424 | case got := <-done: |
| 425 | t.Fatalf("tool auto-approval must not answer sandbox escape; got %+v", got) |
| 426 | case <-time.After(50 * time.Millisecond): |
| 427 | } |
| 428 | |
| 429 | c.Approve(approval.ID, true, true, false) |
| 430 | select { |
| 431 | case got := <-done: |
| 432 | if got.err != nil || !got.allow || got.reason != "" { |
| 433 | t.Fatalf("sandbox escape result = %+v, want allowed without reason/error", got) |
| 434 | } |
| 435 | case <-time.After(30 * time.Second): |
| 436 | t.Fatal("sandbox escape approval stayed blocked after Approve") |
| 437 | } |
| 438 | |
| 439 | if !(sandboxEscapeApprover{c}).SandboxEscapeSessionAllowed(context.Background(), sandbox.EscapeRequest{Command: "npm test"}) { |
| 440 | t.Fatal("sandbox escape session checker = false, want true after session grant") |
| 441 | } |
| 442 | allow, reason, err := sandboxEscapeApprover{c}.ApproveSandboxEscape(context.Background(), sandbox.EscapeRequest{ |
| 443 | Command: "npm test", |
| 444 | Reason: "Windows sandbox failed. Run this command unconfined once?", |
| 445 | }) |
| 446 | if err != nil || !allow || reason != "" { |
| 447 | t.Fatalf("sandbox escape session grant result = (%v,%q,%v), want allow", allow, reason, err) |
| 448 | } |
| 449 | select { |
| 450 | case approval := <-approvalRequests: |
| 451 | t.Fatalf("sandbox escape session grant emitted another approval: %+v", approval) |
| 452 | default: |
| 453 | } |
| 454 | } |
| 455 | |
| 456 | func TestSetAutoApproveToolsDoesNotDrainPendingPlanApproval(t *testing.T) { |
| 457 | approvalRequests := make(chan event.Approval, 1) |
| 458 | c := newOwnedTestController(t, Options{ |
| 459 | Sink: event.FuncSink(func(e event.Event) { |
| 460 | if e.Kind == event.ApprovalRequest { |
| 461 | approvalRequests <- e.Approval |
| 462 | } |
| 463 | }), |
| 464 | }) |
| 465 | |
| 466 | done := make(chan bool, 1) |
| 467 | errs := make(chan error, 1) |
| 468 | go func() { |
| 469 | allow, _, err := c.requestApproval(context.Background(), planApprovalTool, "", nil) |
| 470 | if err != nil { |
| 471 | errs <- err |
| 472 | return |
| 473 | } |
| 474 | done <- allow |
| 475 | }() |
| 476 | |
| 477 | var approval event.Approval |
| 478 | select { |
| 479 | case approval = <-approvalRequests: |
| 480 | case <-time.After(30 * time.Second): |
| 481 | t.Fatal("plan approval request was not emitted") |
| 482 | } |
| 483 | |
| 484 | c.SetAutoApproveTools(true) |
| 485 | |
| 486 | select { |
| 487 | case err := <-errs: |
| 488 | t.Fatalf("requestApproval: %v", err) |
| 489 | case allow := <-done: |
| 490 | t.Fatalf("SetAutoApproveTools must not auto-answer pending plan approval; got allow=%v", allow) |
| 491 | case <-time.After(50 * time.Millisecond): |
| 492 | } |
| 493 | if c.AutoApproveTools() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 494 | t.Fatalf("legacy mode = %q full=%v, want workspace-write", c.ToolApprovalMode(), c.AutoApproveTools()) |
| 495 | } |
| 496 | |
| 497 | c.Approve(approval.ID, true, false, false) |
| 498 | |
| 499 | select { |
| 500 | case err := <-errs: |
| 501 | t.Fatalf("requestApproval: %v", err) |
| 502 | case allow := <-done: |
| 503 | if !allow { |
| 504 | t.Fatal("manual plan approval should allow") |
| 505 | } |
| 506 | case <-time.After(30 * time.Second): |
| 507 | t.Fatal("plan approval stayed blocked after Approve") |
| 508 | } |
| 509 | } |
| 510 | |
| 511 | func TestSetAutoApproveToolsDoesNotDrainPendingPlanModeReadOnlyCommandTrust(t *testing.T) { |
| 512 | approvalRequests := make(chan event.Approval, 1) |
| 513 | c := newOwnedTestController(t, Options{ |
| 514 | Sink: event.FuncSink(func(e event.Event) { |
| 515 | if e.Kind == event.ApprovalRequest { |
| 516 | approvalRequests <- e.Approval |
| 517 | } |
| 518 | }), |
| 519 | }) |
| 520 | |
| 521 | type trustResult struct { |
| 522 | allow bool |
| 523 | reason string |
| 524 | err error |
| 525 | } |
| 526 | done := make(chan trustResult, 1) |
| 527 | req := agent.PlanModeReadOnlyTrustRequest{ |
| 528 | ToolName: agent.PlanModeReadOnlyCommandApprovalTool, |
| 529 | Command: "gh issue view 5867", |
| 530 | Prefix: "gh issue view", |
| 531 | } |
| 532 | go func() { |
| 533 | allow, reason, err := planModeReadOnlyTrustApprover{c}.CheckPlanModeReadOnlyTrust(context.Background(), req) |
| 534 | done <- trustResult{allow: allow, reason: reason, err: err} |
| 535 | }() |
| 536 | |
| 537 | var approval event.Approval |
| 538 | select { |
| 539 | case approval = <-approvalRequests: |
| 540 | case <-time.After(30 * time.Second): |
| 541 | t.Fatal("plan-mode bash read-only command trust approval request was not emitted") |
| 542 | } |
| 543 | if approval.Tool != agent.PlanModeReadOnlyCommandApprovalTool { |
| 544 | t.Fatalf("approval tool = %q, want %q", approval.Tool, agent.PlanModeReadOnlyCommandApprovalTool) |
| 545 | } |
| 546 | |
| 547 | c.SetAutoApproveTools(true) |
| 548 | |
| 549 | select { |
| 550 | case got := <-done: |
| 551 | t.Fatalf("SetAutoApproveTools must not auto-answer plan-mode bash read-only command trust; got %+v", got) |
| 552 | case <-time.After(50 * time.Millisecond): |
| 553 | } |
| 554 | if c.AutoApproveTools() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 555 | t.Fatalf("legacy mode = %q full=%v, want workspace-write", c.ToolApprovalMode(), c.AutoApproveTools()) |
| 556 | } |
| 557 | |
| 558 | c.Approve(approval.ID, true, false, false) |
| 559 | select { |
| 560 | case got := <-done: |
| 561 | if got.err != nil || !got.allow || got.reason != "" { |
| 562 | t.Fatalf("manual plan-mode bash read-only command trust approval = %+v, want allow", got) |
| 563 | } |
| 564 | case <-time.After(30 * time.Second): |
| 565 | t.Fatal("plan-mode bash read-only command trust approval stayed blocked after Approve") |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | // The legacy combined mode API no longer grants full access and cannot answer |
| 570 | // an approval already waiting under another snapshot. |
| 571 | func TestSetModeLegacyPermissionDoesNotResolvePendingApproval(t *testing.T) { |
| 572 | c, ids, _ := approvalIDs(t) |
| 573 | |
| 574 | done := make(chan bool, 1) |
| 575 | go func() { |
| 576 | allow, _, _ := c.requestApproval(context.Background(), "multi_edit", "/tmp/file", nil) |
| 577 | done <- allow |
| 578 | }() |
| 579 | |
| 580 | var approvalID string |
| 581 | select { |
| 582 | case approvalID = <-ids: |
| 583 | case <-time.After(30 * time.Second): |
| 584 | t.Fatal("approval request was not emitted") |
| 585 | } |
| 586 | |
| 587 | c.SetMode(false, true) |
| 588 | |
| 589 | select { |
| 590 | case allow := <-done: |
| 591 | t.Fatalf("legacy SetMode answered pending approval: allow=%v", allow) |
| 592 | case <-time.After(50 * time.Millisecond): |
| 593 | } |
| 594 | c.Approve(approvalID, true, false, false) |
| 595 | if allow := <-done; !allow { |
| 596 | t.Fatal("manual approval should allow") |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | func TestSetModeLegacyAppliesPlanAndWorkspacePermission(t *testing.T) { |
| 601 | c, _, _ := approvalIDs(t) |
| 602 | |
| 603 | c.SetMode(true, false) |
| 604 | if !c.PlanMode() || c.AutoApproveTools() { |
| 605 | t.Fatalf("plan mode: plan=%v autoApproveTools=%v, want true/false", c.PlanMode(), c.AutoApproveTools()) |
| 606 | } |
| 607 | |
| 608 | c.SetMode(false, true) |
| 609 | if c.PlanMode() || c.AutoApproveTools() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 610 | t.Fatalf("legacy write mode: plan=%v permission=%q", c.PlanMode(), c.ToolApprovalMode()) |
| 611 | } |
| 612 | |
| 613 | c.SetMode(true, true) |
| 614 | if !c.PlanMode() || c.AutoApproveTools() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 615 | t.Fatalf("legacy plan mode: plan=%v permission=%q", c.PlanMode(), c.ToolApprovalMode()) |
| 616 | } |
| 617 | |
| 618 | c.SetMode(false, false) |
| 619 | if c.PlanMode() || c.AutoApproveTools() { |
| 620 | t.Fatalf("normal mode: plan=%v autoApproveTools=%v, want false/false", c.PlanMode(), c.AutoApproveTools()) |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | type planModeCountingRunner struct { |
| 625 | calls int |
| 626 | last bool |
| 627 | } |
| 628 | |
| 629 | func (*planModeCountingRunner) Run(context.Context, string) error { return nil } |
| 630 | func (r *planModeCountingRunner) SetPlanMode(v bool) { |
| 631 | r.calls++ |
| 632 | r.last = v |
| 633 | } |
| 634 | |
| 635 | func TestApplyModeUsesRunnerPlanPropagationOnce(t *testing.T) { |
| 636 | runner := &planModeCountingRunner{} |
| 637 | c := newOwnedTestController(t, Options{Runner: runner}) |
| 638 | c.ApplyMode(true, true) |
| 639 | if runner.calls != 1 || !runner.last { |
| 640 | t.Fatalf("runner SetPlanMode calls=%d last=%v, want 1/true", runner.calls, runner.last) |
| 641 | } |
| 642 | if !c.PlanMode() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 643 | t.Fatalf("controller plan=%v approval=%q, want true/workspace-write after legacy migration", c.PlanMode(), c.ToolApprovalMode()) |
| 644 | } |
| 645 | c.SetPlanMode(false) |
| 646 | if runner.calls != 2 || runner.last { |
| 647 | t.Fatalf("SetPlanMode runner calls=%d last=%v, want 2/false", runner.calls, runner.last) |
| 648 | } |
| 649 | } |
| 650 | |
| 651 | func TestApplyModePlanPropagationRunnerFallbacks(t *testing.T) { |
| 652 | for _, tc := range []struct { |
| 653 | name string |
| 654 | runner func(*agent.Agent) agent.Runner |
| 655 | }{ |
| 656 | {name: "single agent", runner: func(executor *agent.Agent) agent.Runner { return executor }}, |
| 657 | {name: "runner without setter", runner: func(*agent.Agent) agent.Runner { |
| 658 | return appendingRunner{session: agent.NewSession("runner")} |
| 659 | }}, |
| 660 | {name: "nil runner", runner: func(*agent.Agent) agent.Runner { return nil }}, |
| 661 | } { |
| 662 | t.Run(tc.name, func(t *testing.T) { |
| 663 | phaseCalls := 0 |
| 664 | reg := tool.NewRegistry() |
| 665 | reg.Add(plannerUnsafeReadTool{calls: &phaseCalls}) |
| 666 | prov := &scriptedTurns{turns: [][]provider.Chunk{ |
| 667 | toolCallTurn("phase-1", "planner_phase_only", `{}`), |
| 668 | textTurn("done"), |
| 669 | }} |
| 670 | executor := agent.New(prov, reg, agent.NewSession("executor"), agent.Options{}, event.Discard) |
| 671 | c := newOwnedTestController(t, Options{Runner: tc.runner(executor), Executor: executor}) |
| 672 | c.ApplyMode(true, true) |
| 673 | if err := executor.Run(context.Background(), "try the execution-phase tool"); err != nil { |
| 674 | t.Fatalf("executor Run: %v", err) |
| 675 | } |
| 676 | if phaseCalls != 0 { |
| 677 | t.Fatalf("phase-opted-out tool executed %d times, want 0", phaseCalls) |
| 678 | } |
| 679 | }) |
| 680 | } |
| 681 | } |
| 682 | |
| 683 | type plannerUnsafeReadTool struct { |
| 684 | calls *int |
| 685 | } |
| 686 | |
| 687 | func (plannerUnsafeReadTool) Name() string { return "planner_phase_only" } |
| 688 | func (plannerUnsafeReadTool) Description() string { return "planner phase test tool" } |
| 689 | func (plannerUnsafeReadTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 690 | func (plannerUnsafeReadTool) ReadOnly() bool { return true } |
| 691 | func (plannerUnsafeReadTool) PlanModeSafe() bool { return false } |
| 692 | func (t plannerUnsafeReadTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 693 | (*t.calls)++ |
| 694 | return "executed", nil |
| 695 | } |
| 696 | |
| 697 | func TestApplyModePropagatesPlanToCoordinatorPlannerAndMigratesLegacyYolo(t *testing.T) { |
| 698 | plannerCalls := 0 |
| 699 | plannerTools := agent.PlannerToolRegistry(tool.NewRegistry()) |
| 700 | plannerTools.Add(plannerUnsafeReadTool{calls: &plannerCalls}) |
| 701 | planner := &scriptedTurns{turns: [][]provider.Chunk{ |
| 702 | toolCallTurn("planner-tool", "planner_phase_only", `{}`), |
| 703 | planTurn("1. inspect the current behavior\n2. implement the fix"), |
| 704 | }} |
| 705 | execProvider := &scriptedTurns{turns: [][]provider.Chunk{textTurn("executor done")}} |
| 706 | executor := agent.New(execProvider, tool.NewRegistry(), agent.NewSession("exec"), agent.Options{}, event.Discard) |
| 707 | coordinator := agent.NewCoordinator(planner, agent.NewSession("planner"), nil, plannerTools, agent.Options{}, executor, 0, event.Discard, nil) |
| 708 | c := newOwnedTestController(t, Options{Runner: coordinator, Executor: executor}) |
| 709 | |
| 710 | c.ApplyMode(true, true) |
| 711 | if err := c.Run(context.Background(), "prepare the change"); err != nil { |
| 712 | t.Fatalf("Run: %v", err) |
| 713 | } |
| 714 | if plannerCalls != 0 { |
| 715 | t.Fatalf("planner phase-only tool executed %d times, want 0 while Plan is active", plannerCalls) |
| 716 | } |
| 717 | if !c.PlanMode() || c.ToolApprovalMode() != ToolApprovalWorkspaceWrite { |
| 718 | t.Fatalf("after run plan=%v approval=%q, want true/workspace-write", c.PlanMode(), c.ToolApprovalMode()) |
| 719 | } |
| 720 | } |
| 721 | |
| 722 | type askCallResult struct { |
| 723 | answers []event.AskAnswer |
| 724 | err error |
| 725 | } |
| 726 | |
| 727 | func sampleAskQuestions() []event.AskQuestion { |
| 728 | return []event.AskQuestion{ |
| 729 | { |
| 730 | ID: "approach", |
| 731 | Header: "Approach", |
| 732 | Prompt: "Which path?", |
| 733 | Options: []event.AskOption{ |
| 734 | {Label: "Recommended path"}, |
| 735 | {Label: "Alternative path"}, |
| 736 | }, |
| 737 | }, |
| 738 | { |
| 739 | ID: "scope", |
| 740 | Header: "Scope", |
| 741 | Prompt: "How broad?", |
| 742 | Options: []event.AskOption{ |
| 743 | {Label: "Minimal"}, |
| 744 | {Label: "Broad"}, |
| 745 | }, |
| 746 | Multi: true, |
| 747 | }, |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | func askController(t *testing.T, c *Controller, questions []event.AskQuestion) <-chan askCallResult { |
| 752 | t.Helper() |
| 753 | done := make(chan askCallResult, 1) |
| 754 | go func() { |
| 755 | answers, err := c.Ask(context.Background(), questions) |
| 756 | done <- askCallResult{answers: answers, err: err} |
| 757 | }() |
| 758 | return done |
| 759 | } |
| 760 | |
| 761 | func waitAskRequest(t *testing.T, askCh <-chan event.Ask) event.Ask { |
| 762 | t.Helper() |
| 763 | select { |
| 764 | case ask := <-askCh: |
| 765 | return ask |
| 766 | case <-time.After(30 * time.Second): |
| 767 | t.Fatal("Ask did not emit AskRequest") |
| 768 | } |
| 769 | return event.Ask{} |
| 770 | } |
| 771 | |
| 772 | func waitAskResult(t *testing.T, done <-chan askCallResult) askCallResult { |
| 773 | t.Helper() |
| 774 | select { |
| 775 | case result := <-done: |
| 776 | if result.err != nil { |
| 777 | t.Fatalf("Ask: %v", result.err) |
| 778 | } |
| 779 | return result |
| 780 | case <-time.After(30 * time.Second): |
| 781 | t.Fatal("Ask stayed blocked") |
| 782 | } |
| 783 | return askCallResult{} |
| 784 | } |
| 785 | |
| 786 | func assertAskAnswers(t *testing.T, got, want []event.AskAnswer) { |
| 787 | t.Helper() |
| 788 | if len(got) != len(want) { |
| 789 | t.Fatalf("answers len = %d, want %d: %#v", len(got), len(want), got) |
| 790 | } |
| 791 | for i := range want { |
| 792 | if got[i].QuestionID != want[i].QuestionID || len(got[i].Selected) != len(want[i].Selected) { |
| 793 | t.Fatalf("answers[%d] = %#v, want %#v", i, got[i], want[i]) |
| 794 | } |
| 795 | for j := range want[i].Selected { |
| 796 | if got[i].Selected[j] != want[i].Selected[j] { |
| 797 | t.Fatalf("answers[%d] = %#v, want %#v", i, got[i], want[i]) |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | } |
| 802 | |
| 803 | func TestBypassDoesNotAutoAnswerAsk(t *testing.T) { |
| 804 | userAnswers := []event.AskAnswer{ |
| 805 | {QuestionID: "approach", Selected: []string{"Alternative path"}}, |
| 806 | {QuestionID: "scope", Selected: []string{"Broad"}}, |
| 807 | } |
| 808 | askCh := make(chan event.Ask, 1) |
| 809 | c := newOwnedTestController(t, Options{ |
| 810 | Sink: event.FuncSink(func(e event.Event) { |
| 811 | if e.Kind == event.AskRequest { |
| 812 | askCh <- e.Ask |
| 813 | } |
| 814 | }), |
| 815 | }) |
| 816 | c.SetBypass(true) |
| 817 | |
| 818 | done := askController(t, c, sampleAskQuestions()) |
| 819 | ask := waitAskRequest(t, askCh) |
| 820 | |
| 821 | // Even with bypass/YOLO on, Ask must wait for the user's non-default choice. |
| 822 | c.AnswerQuestion(ask.ID, userAnswers) |
| 823 | result := waitAskResult(t, done) |
| 824 | assertAskAnswers(t, result.answers, userAnswers) |
| 825 | } |
| 826 | |
| 827 | func TestAskPromptsAcrossInteractiveModes(t *testing.T) { |
| 828 | userAnswers := []event.AskAnswer{ |
| 829 | {QuestionID: "approach", Selected: []string{"Alternative path"}}, |
| 830 | {QuestionID: "scope", Selected: []string{"Broad"}}, |
| 831 | } |
| 832 | tests := []struct { |
| 833 | name string |
| 834 | setup func(*Controller) |
| 835 | }{ |
| 836 | {name: "normal"}, |
| 837 | {name: "plan", setup: func(c *Controller) { c.SetMode(true, false) }}, |
| 838 | {name: "yolo", setup: func(c *Controller) { c.SetMode(false, true) }}, |
| 839 | } |
| 840 | |
| 841 | for _, tt := range tests { |
| 842 | t.Run(tt.name, func(t *testing.T) { |
| 843 | askCh := make(chan event.Ask, 1) |
| 844 | c := newOwnedTestController(t, Options{ |
| 845 | Sink: event.FuncSink(func(e event.Event) { |
| 846 | if e.Kind == event.AskRequest { |
| 847 | askCh <- e.Ask |
| 848 | } |
| 849 | }), |
| 850 | }) |
| 851 | if tt.setup != nil { |
| 852 | tt.setup(c) |
| 853 | } |
| 854 | |
| 855 | done := askController(t, c, sampleAskQuestions()) |
| 856 | ask := waitAskRequest(t, askCh) |
| 857 | |
| 858 | // Answer with non-recommended options to prove this is the user's |
| 859 | // selection, not an automatic recommended-option fallback. |
| 860 | c.AnswerQuestion(ask.ID, userAnswers) |
| 861 | result := waitAskResult(t, done) |
| 862 | assertAskAnswers(t, result.answers, userAnswers) |
| 863 | }) |
| 864 | } |
| 865 | } |
| 866 | |
| 867 | func TestSetAutoApproveToolsDoesNotDrainPendingAsk(t *testing.T) { |
| 868 | askCh := make(chan event.Ask, 1) |
| 869 | c := newOwnedTestController(t, Options{ |
| 870 | Sink: event.FuncSink(func(e event.Event) { |
| 871 | if e.Kind == event.AskRequest { |
| 872 | askCh <- e.Ask |
| 873 | } |
| 874 | }), |
| 875 | }) |
| 876 | |
| 877 | done := askController(t, c, sampleAskQuestions()) |
| 878 | ask := waitAskRequest(t, askCh) |
| 879 | |
| 880 | c.SetToolApprovalMode(ToolApprovalDangerFullAccess) |
| 881 | |
| 882 | select { |
| 883 | case result := <-done: |
| 884 | t.Fatalf("SetAutoApproveTools must not answer pending AskRequest; got %#v", result.answers) |
| 885 | case <-time.After(50 * time.Millisecond): |
| 886 | } |
| 887 | |
| 888 | userAnswers := []event.AskAnswer{ |
| 889 | {QuestionID: "approach", Selected: []string{"Alternative path"}}, |
| 890 | {QuestionID: "scope", Selected: []string{"Broad"}}, |
| 891 | } |
| 892 | c.AnswerQuestion(ask.ID, userAnswers) |
| 893 | result := waitAskResult(t, done) |
| 894 | assertAskAnswers(t, result.answers, userAnswers) |
| 895 | } |
| 896 | |
| 897 | func TestDismissedAskCancelsTurnWithoutModelContinuation(t *testing.T) { |
| 898 | askCh := make(chan event.Ask, 1) |
| 899 | turnDone := make(chan event.Event, 1) |
| 900 | c := newOwnedTestController(t, Options{ |
| 901 | Sink: event.FuncSink(func(e event.Event) { |
| 902 | switch e.Kind { |
| 903 | case event.AskRequest: |
| 904 | askCh <- e.Ask |
| 905 | case event.TurnDone: |
| 906 | turnDone <- e |
| 907 | } |
| 908 | }), |
| 909 | }) |
| 910 | |
| 911 | continued := false |
| 912 | if got := c.runGuarded(func(ctx context.Context) error { |
| 913 | _, err := c.Ask(ctx, sampleAskQuestions()) |
| 914 | if err == nil { |
| 915 | continued = true |
| 916 | } |
| 917 | return err |
| 918 | }); got != turnStarted { |
| 919 | t.Fatalf("runGuarded = %v, want turnStarted", got) |
| 920 | } |
| 921 | ask := waitAskRequest(t, askCh) |
| 922 | c.AnswerQuestion(ask.ID, nil) |
| 923 | |
| 924 | select { |
| 925 | case done := <-turnDone: |
| 926 | if !done.Cancelled { |
| 927 | t.Fatalf("dismissed Ask TurnDone = %+v, want Cancelled", done) |
| 928 | } |
| 929 | case <-time.After(30 * time.Second): |
| 930 | t.Fatal("dismissed Ask did not finish the turn") |
| 931 | } |
| 932 | if continued { |
| 933 | t.Fatal("dismissed Ask returned a model-facing result instead of stopping the turn") |
| 934 | } |
| 935 | } |
| 936 | |
| 937 | // TestApplyToolApprovalModeDoesNotAuthorizePendingApprovals pins the preset |
| 938 | // revision contract: changing the boundary never turns an older prompt into |
| 939 | // an authorization. New calls evaluate the new preset from a fresh snapshot. |
| 940 | func TestApplyToolApprovalModeDoesNotAuthorizePendingApprovals(t *testing.T) { |
| 941 | c := newOwnedTestController(t, Options{ |
| 942 | Policy: permission.New("ask", nil, []string{"bash(git commit*)"}, nil), |
| 943 | }) |
| 944 | |
| 945 | autoOKID, autoOKReply := c.approval.register("bash", "go test ./...", "") |
| 946 | askRuleID, askRuleReply := c.approval.register("bash", "git commit -m x", "") |
| 947 | planID, planReply := c.approval.registerDecision(planApprovalTool, "", "", true, false) |
| 948 | |
| 949 | drained := c.ApplyToolApprovalMode(ToolApprovalAuto) |
| 950 | if len(drained) != 0 { |
| 951 | t.Fatalf("workspace preset resolved old approvals: %v", drained) |
| 952 | } |
| 953 | select { |
| 954 | case r := <-autoOKReply: |
| 955 | t.Fatalf("workspace preset resolved old approval: %+v", r) |
| 956 | default: |
| 957 | } |
| 958 | select { |
| 959 | case <-askRuleReply: |
| 960 | t.Fatal("explicit ask-rule approval must stay pending under auto") |
| 961 | default: |
| 962 | } |
| 963 | |
| 964 | drained = c.ApplyToolApprovalMode(ToolApprovalYolo) |
| 965 | if len(drained) != 0 { |
| 966 | t.Fatalf("full-access preset resolved old approvals: %v", drained) |
| 967 | } |
| 968 | select { |
| 969 | case r := <-askRuleReply: |
| 970 | t.Fatalf("full-access preset resolved old approval: %+v", r) |
| 971 | default: |
| 972 | } |
| 973 | |
| 974 | // The fresh plan decision survives both switches and stays pending. |
| 975 | select { |
| 976 | case <-planReply: |
| 977 | t.Fatal("fresh plan approval must never drain on a posture switch") |
| 978 | default: |
| 979 | } |
| 980 | if !c.approval.hasPending() { |
| 981 | t.Fatalf("plan approval %s should still be pending", planID) |
| 982 | } |
| 983 | // Clean up the synthetic pending approvals without granting them. |
| 984 | c.Approve(autoOKID, false, false, false) |
| 985 | c.Approve(askRuleID, false, false, false) |
| 986 | c.Approve(planID, false, false, false) |
| 987 | } |
| 988 |