| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "strings" |
| 9 | "sync" |
| 10 | "testing" |
| 11 | "time" |
| 12 | |
| 13 | "reasonix/internal/agent/testutil" |
| 14 | "reasonix/internal/event" |
| 15 | "reasonix/internal/provider" |
| 16 | "reasonix/internal/tool" |
| 17 | ) |
| 18 | |
| 19 | // slowTool is a tool that takes a noticeable amount of time to execute, |
| 20 | // simulating a long-running bash command or other blocking operation. |
| 21 | type slowTool struct{} |
| 22 | |
| 23 | func (slowTool) Name() string { return "slow_tool" } |
| 24 | |
| 25 | func (slowTool) Description() string { return "A tool that executes slowly" } |
| 26 | |
| 27 | func (slowTool) Schema() json.RawMessage { |
| 28 | return json.RawMessage(`{"type":"object","properties":{"duration_ms":{"type":"number","description":"How long to sleep in milliseconds"}},"required":["duration_ms"]}`) |
| 29 | } |
| 30 | |
| 31 | func (slowTool) ReadOnly() bool { return false } |
| 32 | |
| 33 | func (slowTool) Execute(ctx context.Context, args json.RawMessage) (string, error) { |
| 34 | var p struct { |
| 35 | DurationMs int `json:"duration_ms"` |
| 36 | } |
| 37 | if err := json.Unmarshal(args, &p); err != nil { |
| 38 | return "", err |
| 39 | } |
| 40 | if p.DurationMs <= 0 { |
| 41 | p.DurationMs = 500 |
| 42 | } |
| 43 | |
| 44 | // Simulate work that respects context cancellation |
| 45 | select { |
| 46 | case <-time.After(time.Duration(p.DurationMs) * time.Millisecond): |
| 47 | return "done", nil |
| 48 | case <-ctx.Done(): |
| 49 | return "", ctx.Err() |
| 50 | } |
| 51 | } |
| 52 | |
| 53 | // trackingTool is a tool that records when it was executed and can simulate delays. |
| 54 | type trackingTool struct { |
| 55 | name string |
| 56 | readOnly bool |
| 57 | } |
| 58 | |
| 59 | func (t trackingTool) Name() string { |
| 60 | if t.name != "" { |
| 61 | return t.name |
| 62 | } |
| 63 | return "tracking" |
| 64 | } |
| 65 | func (trackingTool) Description() string { return "Tracks execution" } |
| 66 | func (trackingTool) Schema() json.RawMessage { |
| 67 | return json.RawMessage(`{"type":"object","properties":{"name":{"type":"string"},"delay_ms":{"type":"number"},"should_fail":{"type":"boolean"}},"required":["name"]}`) |
| 68 | } |
| 69 | func (t trackingTool) ReadOnly() bool { return t.readOnly } |
| 70 | func (trackingTool) Execute(ctx context.Context, args json.RawMessage) (string, error) { |
| 71 | var p struct { |
| 72 | Name string `json:"name"` |
| 73 | DelayMs int `json:"delay_ms"` |
| 74 | ShouldFail bool `json:"should_fail"` |
| 75 | } |
| 76 | if err := json.Unmarshal(args, &p); err != nil { |
| 77 | return "", err |
| 78 | } |
| 79 | |
| 80 | executedMu.Lock() |
| 81 | executed = append(executed, p.Name+"_start") |
| 82 | executedMu.Unlock() |
| 83 | |
| 84 | if p.ShouldFail { |
| 85 | return "", context.Canceled |
| 86 | } |
| 87 | |
| 88 | // Simulate work that respects context cancellation |
| 89 | if p.DelayMs > 0 { |
| 90 | select { |
| 91 | case <-time.After(time.Duration(p.DelayMs) * time.Millisecond): |
| 92 | // Completed the delay successfully |
| 93 | case <-ctx.Done(): |
| 94 | executedMu.Lock() |
| 95 | executed = append(executed, p.Name+"_cancelled") |
| 96 | executedMu.Unlock() |
| 97 | return "", ctx.Err() |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | executedMu.Lock() |
| 102 | executed = append(executed, p.Name+"_done") |
| 103 | executedMu.Unlock() |
| 104 | |
| 105 | return p.Name + " done", nil |
| 106 | } |
| 107 | |
| 108 | // Global variables for tracking across tests |
| 109 | var ( |
| 110 | executedMu sync.Mutex |
| 111 | executed []string |
| 112 | ) |
| 113 | |
| 114 | type stuckStreamProvider struct{} |
| 115 | |
| 116 | func (stuckStreamProvider) Name() string { return "stuck-stream" } |
| 117 | |
| 118 | func (stuckStreamProvider) Stream(context.Context, provider.Request) (<-chan provider.Chunk, error) { |
| 119 | return make(chan provider.Chunk), nil |
| 120 | } |
| 121 | |
| 122 | type closedStreamProvider struct{} |
| 123 | |
| 124 | func (closedStreamProvider) Name() string { return "closed-stream" } |
| 125 | |
| 126 | func (closedStreamProvider) Stream(context.Context, provider.Request) (<-chan provider.Chunk, error) { |
| 127 | ch := make(chan provider.Chunk) |
| 128 | close(ch) |
| 129 | return ch, nil |
| 130 | } |
| 131 | |
| 132 | func TestCanceledContextClosedProviderStreamReturnsCancel(t *testing.T) { |
| 133 | for i := 0; i < 100; i++ { |
| 134 | ctx, cancel := context.WithCancel(context.Background()) |
| 135 | cancel() |
| 136 | |
| 137 | a := New(closedStreamProvider{}, tool.NewRegistry(), NewSession(""), Options{}, &recordSink{}) |
| 138 | err := a.Run(ctx, "already cancelled") |
| 139 | if !errors.Is(err, context.Canceled) { |
| 140 | t.Fatalf("Run error on iteration %d = %v, want context cancellation", i, err) |
| 141 | } |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | func TestCancelDuringStuckProviderStreamReturnsPromptly(t *testing.T) { |
| 146 | a := New(stuckStreamProvider{}, tool.NewRegistry(), NewSession(""), Options{}, &recordSink{}) |
| 147 | |
| 148 | ctx, cancel := context.WithCancel(context.Background()) |
| 149 | done := make(chan error, 1) |
| 150 | go func() { |
| 151 | done <- a.Run(ctx, "wait on provider") |
| 152 | }() |
| 153 | |
| 154 | time.Sleep(50 * time.Millisecond) |
| 155 | cancel() |
| 156 | |
| 157 | select { |
| 158 | case err := <-done: |
| 159 | if err == nil { |
| 160 | t.Fatal("Run returned nil after context cancellation") |
| 161 | } |
| 162 | if !errors.Is(err, context.Canceled) { |
| 163 | t.Fatalf("Run error = %v, want context cancellation", err) |
| 164 | } |
| 165 | case <-time.After(500 * time.Millisecond): |
| 166 | t.Fatal("Run did not return promptly after provider stream context cancellation") |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | type activeReasoningUntilCancelProvider struct{} |
| 171 | |
| 172 | func (activeReasoningUntilCancelProvider) Name() string { return "active-reasoning" } |
| 173 | |
| 174 | func (p activeReasoningUntilCancelProvider) Stream(ctx context.Context, _ provider.Request) (<-chan provider.Chunk, error) { |
| 175 | ch := make(chan provider.Chunk) |
| 176 | go func() { |
| 177 | defer close(ch) |
| 178 | for offset := 224; ; offset += 4 { |
| 179 | select { |
| 180 | case <-ctx.Done(): |
| 181 | return |
| 182 | case ch <- provider.Chunk{Type: provider.ChunkReasoning, Text: fmt.Sprintf("%d unknown\n", offset)}: |
| 183 | } |
| 184 | } |
| 185 | }() |
| 186 | return ch, nil |
| 187 | } |
| 188 | |
| 189 | type reasoningGuardCancelProvider struct { |
| 190 | canceled chan struct{} |
| 191 | } |
| 192 | |
| 193 | func (reasoningGuardCancelProvider) Name() string { return "reasoning-guard-cancel" } |
| 194 | |
| 195 | func (p reasoningGuardCancelProvider) Stream(ctx context.Context, _ provider.Request) (<-chan provider.Chunk, error) { |
| 196 | ch := make(chan provider.Chunk) |
| 197 | go func() { |
| 198 | defer close(ch) |
| 199 | defer close(p.canceled) |
| 200 | for { |
| 201 | select { |
| 202 | case <-ctx.Done(): |
| 203 | return |
| 204 | case ch <- provider.Chunk{Type: provider.ChunkReasoning, Text: "0123456789abcdef"}: |
| 205 | } |
| 206 | } |
| 207 | }() |
| 208 | return ch, nil |
| 209 | } |
| 210 | |
| 211 | func TestRunawayReasoningStopsAtAgentSideByteGuard(t *testing.T) { |
| 212 | sink := &recordSink{} |
| 213 | a := New(activeReasoningUntilCancelProvider{}, tool.NewRegistry(), NewSession(""), Options{ReasoningByteLimit: 64}, sink) |
| 214 | |
| 215 | ctx, cancel := context.WithTimeout(context.Background(), 500*time.Millisecond) |
| 216 | defer cancel() |
| 217 | |
| 218 | err := a.Run(ctx, "parse this binary by offset") |
| 219 | if !errors.Is(err, errReasoningByteLimitExceeded) { |
| 220 | t.Fatalf("Run error = %v, want reasoning limit guard", err) |
| 221 | } |
| 222 | if got := len(sink.kinds(event.Reasoning)); got == 0 { |
| 223 | t.Fatal("no reasoning chunks emitted; repro did not exercise the active-output path") |
| 224 | } |
| 225 | usages := sink.kinds(event.Usage) |
| 226 | if len(usages) != 1 { |
| 227 | t.Fatalf("usage events = %d, want one best-effort usage event", len(usages)) |
| 228 | } |
| 229 | if u := usages[0].Usage; u == nil || u.FinishReason != "client_reasoning_limit" || !u.Estimated || u.TotalTokens <= 0 || u.ReasoningTokens <= 0 { |
| 230 | t.Fatalf("usage = %+v, want client reasoning limit with estimated reasoning tokens", u) |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | func TestReasoningByteGuardCancelsProviderStream(t *testing.T) { |
| 235 | canceled := make(chan struct{}) |
| 236 | a := New(reasoningGuardCancelProvider{canceled: canceled}, tool.NewRegistry(), NewSession(""), Options{ReasoningByteLimit: 32}, event.Discard) |
| 237 | |
| 238 | if err := a.Run(context.Background(), "trigger the reasoning guard"); !errors.Is(err, errReasoningByteLimitExceeded) { |
| 239 | t.Fatalf("Run error = %v, want reasoning limit guard", err) |
| 240 | } |
| 241 | select { |
| 242 | case <-canceled: |
| 243 | case <-time.After(time.Second): |
| 244 | t.Fatal("provider context remained live after the reasoning guard returned") |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | func TestInterruptedReasoningEmitsBestEffortUsage(t *testing.T) { |
| 249 | sink := &recordSink{} |
| 250 | a := New(activeReasoningUntilCancelProvider{}, tool.NewRegistry(), NewSession(""), Options{}, sink) |
| 251 | |
| 252 | ctx, cancel := context.WithCancel(context.Background()) |
| 253 | done := make(chan error, 1) |
| 254 | go func() { |
| 255 | done <- a.Run(ctx, "parse this binary by offset") |
| 256 | }() |
| 257 | |
| 258 | deadline := time.After(500 * time.Millisecond) |
| 259 | for len(sink.kinds(event.Reasoning)) == 0 { |
| 260 | select { |
| 261 | case <-deadline: |
| 262 | t.Fatal("timed out waiting for streamed reasoning") |
| 263 | default: |
| 264 | time.Sleep(time.Millisecond) |
| 265 | } |
| 266 | } |
| 267 | cancel() |
| 268 | |
| 269 | select { |
| 270 | case err := <-done: |
| 271 | if !errors.Is(err, context.Canceled) { |
| 272 | t.Fatalf("Run error = %v, want context cancellation", err) |
| 273 | } |
| 274 | case <-time.After(500 * time.Millisecond): |
| 275 | t.Fatal("Run did not return after cancellation") |
| 276 | } |
| 277 | |
| 278 | usages := sink.kinds(event.Usage) |
| 279 | if len(usages) != 1 { |
| 280 | t.Fatalf("usage events = %d, want one best-effort usage event", len(usages)) |
| 281 | } |
| 282 | if u := usages[0].Usage; u == nil || u.FinishReason != "interrupted" || !u.Estimated || u.TotalTokens <= 0 || u.ReasoningTokens <= 0 { |
| 283 | t.Fatalf("usage = %+v, want interrupted finish with estimated reasoning tokens", u) |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | func TestReasoningByteGuardDoesNotSetProviderOutputBudget(t *testing.T) { |
| 288 | tests := []struct { |
| 289 | name string |
| 290 | limit int |
| 291 | }{ |
| 292 | {name: "default"}, |
| 293 | {name: "custom", limit: 65}, |
| 294 | {name: "disabled", limit: -1}, |
| 295 | } |
| 296 | for _, tt := range tests { |
| 297 | t.Run(tt.name, func(t *testing.T) { |
| 298 | prov := testutil.NewMock("m", testutil.Turn{Text: "done"}) |
| 299 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{ReasoningByteLimit: tt.limit}, event.Discard) |
| 300 | if err := a.Run(context.Background(), "go"); err != nil { |
| 301 | t.Fatal(err) |
| 302 | } |
| 303 | req := prov.LastRequest() |
| 304 | if req == nil || req.MaxTokens != 0 { |
| 305 | t.Fatalf("request = %+v, reasoning bytes must not become a total output budget", req) |
| 306 | } |
| 307 | }) |
| 308 | } |
| 309 | |
| 310 | t.Run("stable across tool loop", func(t *testing.T) { |
| 311 | prov := testutil.NewMock("m", |
| 312 | testutil.Turn{ToolCalls: []provider.ToolCall{{ID: "call-1", Name: "read", Arguments: `{}`}}}, |
| 313 | testutil.Turn{Text: "done"}, |
| 314 | ) |
| 315 | registry := tool.NewRegistry() |
| 316 | registry.Add(fakeTool{name: "read", readOnly: true}) |
| 317 | a := New(prov, registry, NewSession(""), Options{MaxOutputTokens: 8192}, event.Discard) |
| 318 | if err := a.Run(context.Background(), "go"); err != nil { |
| 319 | t.Fatal(err) |
| 320 | } |
| 321 | requests := prov.Requests() |
| 322 | if len(requests) != 2 { |
| 323 | t.Fatalf("requests = %d, want two provider turns", len(requests)) |
| 324 | } |
| 325 | for i, req := range requests { |
| 326 | if req.MaxTokens != 8192 { |
| 327 | t.Fatalf("request %d max_tokens = %d, want stable 8192", i+1, req.MaxTokens) |
| 328 | } |
| 329 | } |
| 330 | }) |
| 331 | } |
| 332 | |
| 333 | func TestBestEffortStreamUsageMarksOnlySyntheticCountsEstimated(t *testing.T) { |
| 334 | exact := &provider.Usage{PromptTokens: 10, CompletionTokens: 20, TotalTokens: 30, ReasoningTokens: 15} |
| 335 | got := bestEffortStreamUsage(exact, 4, 4, "interrupted") |
| 336 | if got.Estimated { |
| 337 | t.Fatalf("usage = %+v, exact counts should remain exact", got) |
| 338 | } |
| 339 | if got.FinishReason != "interrupted" { |
| 340 | t.Fatalf("finish reason = %q, want interrupted", got.FinishReason) |
| 341 | } |
| 342 | |
| 343 | got = bestEffortStreamUsage(exact, 200, 400, "interrupted") |
| 344 | if !got.Estimated || got.CompletionTokens != 150 || got.ReasoningTokens != 100 || got.TotalTokens != 160 { |
| 345 | t.Fatalf("usage = %+v, want byte-derived estimates", got) |
| 346 | } |
| 347 | } |
| 348 | |
| 349 | // TestCancelDuringToolExecutionBreaksOutPromptly verifies that when the context |
| 350 | // is cancelled while tools are executing, the agent loop breaks out immediately |
| 351 | // rather than continuing to execute remaining tools. |
| 352 | func TestCancelDuringToolExecutionBreaksOutPromptly(t *testing.T) { |
| 353 | reg := tool.NewRegistry() |
| 354 | reg.Add(slowTool{}) |
| 355 | |
| 356 | // Script: first turn calls two slow tools, but we'll cancel after the first starts |
| 357 | mp := testutil.NewMock("m", |
| 358 | testutil.Turn{ |
| 359 | Text: "", |
| 360 | ToolCalls: []provider.ToolCall{ |
| 361 | {ID: "call-1", Name: "slow_tool", Arguments: `{"duration_ms": 2000}`}, // 2 second tool |
| 362 | {ID: "call-2", Name: "slow_tool", Arguments: `{"duration_ms": 2000}`}, // another 2 second tool |
| 363 | }, |
| 364 | }, |
| 365 | ) |
| 366 | |
| 367 | sink := &recordSink{} |
| 368 | a := New(mp, reg, NewSession(""), Options{}, sink) |
| 369 | |
| 370 | // Create a cancellable context and cancel it shortly after starting |
| 371 | ctx, cancel := context.WithCancel(context.Background()) |
| 372 | |
| 373 | start := time.Now() |
| 374 | done := make(chan error, 1) |
| 375 | go func() { |
| 376 | done <- a.Run(ctx, "test cancel during tool execution") |
| 377 | }() |
| 378 | |
| 379 | // Cancel after a short delay to simulate user pressing Esc mid-execution |
| 380 | go func() { |
| 381 | time.Sleep(300 * time.Millisecond) |
| 382 | cancel() |
| 383 | }() |
| 384 | |
| 385 | // Wait for the run to complete (should be fast due to cancel, not 4+ seconds) |
| 386 | var err error |
| 387 | select { |
| 388 | case err = <-done: |
| 389 | case <-time.After(5 * time.Second): |
| 390 | t.Fatal("Run did not complete within 5s after cancel — context cancellation did not interrupt tool execution") |
| 391 | } |
| 392 | |
| 393 | elapsed := time.Since(start) |
| 394 | |
| 395 | // Should have run until the cancel (~300ms) but not completed both tools (4s+) |
| 396 | if elapsed < 250*time.Millisecond { |
| 397 | t.Fatalf("command exited too fast (%v) — cancel didn't actually interrupt execution; err=%v", elapsed, err) |
| 398 | } |
| 399 | if elapsed > 2*time.Second { |
| 400 | t.Fatalf("cancel took too long (%v) — should have broken out after first tool, not waited for all tools", elapsed) |
| 401 | } |
| 402 | |
| 403 | // The error should be related to context cancellation |
| 404 | if err == nil { |
| 405 | t.Log("Run returned nil error after cancel (acceptable if tools detected ctx.Done)") |
| 406 | } else { |
| 407 | t.Logf("Run returned error after cancel: %v (elapsed: %v)", err, elapsed) |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | // TestCancelDuringBatchStopsRemainingTools verifies that when context is |
| 412 | // cancelled during a batch of tool executions, remaining tools are not executed. |
| 413 | func TestCancelDuringBatchStopsRemainingTools(t *testing.T) { |
| 414 | // Reset tracking |
| 415 | executedMu.Lock() |
| 416 | executed = nil |
| 417 | executedMu.Unlock() |
| 418 | |
| 419 | reg := tool.NewRegistry() |
| 420 | reg.Add(trackingTool{}) |
| 421 | |
| 422 | // Script: model wants to execute three tools in sequence |
| 423 | mp := testutil.NewMock("m", |
| 424 | testutil.Turn{ |
| 425 | Text: "", |
| 426 | ToolCalls: []provider.ToolCall{ |
| 427 | {ID: "call-1", Name: "tracking", Arguments: `{"name": "tool1", "delay_ms": 50}`}, |
| 428 | {ID: "call-2", Name: "tracking", Arguments: `{"name": "tool2", "delay_ms": 5000}`}, // Long-running tool |
| 429 | {ID: "call-3", Name: "tracking", Arguments: `{"name": "tool3", "delay_ms": 50}`}, |
| 430 | }, |
| 431 | }, |
| 432 | ) |
| 433 | |
| 434 | sink := &recordSink{} |
| 435 | a := New(mp, reg, NewSession(""), Options{}, sink) |
| 436 | |
| 437 | ctx, cancel := context.WithCancel(context.Background()) |
| 438 | defer cancel() |
| 439 | |
| 440 | done := make(chan error, 1) |
| 441 | go func() { |
| 442 | done <- a.Run(ctx, "test batch cancel") |
| 443 | }() |
| 444 | |
| 445 | // Cancel while tool2 is still running (after tool1 completes but during tool2) |
| 446 | go func() { |
| 447 | time.Sleep(300 * time.Millisecond) |
| 448 | cancel() |
| 449 | }() |
| 450 | |
| 451 | var err error |
| 452 | select { |
| 453 | case err = <-done: |
| 454 | case <-time.After(10 * time.Second): |
| 455 | t.Fatal("Run did not complete within 10s") |
| 456 | } |
| 457 | |
| 458 | executedMu.Lock() |
| 459 | executedCopy := make([]string, len(executed)) |
| 460 | copy(executedCopy, executed) |
| 461 | executedMu.Unlock() |
| 462 | |
| 463 | t.Logf("Executed tools: %v (err=%v)", executedCopy, err) |
| 464 | |
| 465 | // We expect tool1 to have completed, tool2 to have been cancelled mid-execution, |
| 466 | // and tool3 to NOT have started at all due to our ctx.Err() check after each tool. |
| 467 | if len(executedCopy) < 2 { // At least tool1_start should be there |
| 468 | t.Error("Expected at least one tool to start execution") |
| 469 | } |
| 470 | |
| 471 | // Check that tool3 never started |
| 472 | for _, name := range executedCopy { |
| 473 | if strings.HasPrefix(name, "tool3") { |
| 474 | t.Error("tool3 should not have executed after cancel interrupted the batch") |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | // Verify tool2 was cancelled |
| 479 | foundTool2Cancelled := false |
| 480 | for _, name := range executedCopy { |
| 481 | if name == "tool2_cancelled" { |
| 482 | foundTool2Cancelled = true |
| 483 | } |
| 484 | } |
| 485 | if !foundTool2Cancelled { |
| 486 | t.Log("Note: tool2 may have completed or been cancelled - check timing") |
| 487 | } |
| 488 | |
| 489 | toolsByID := toolMessagesByID(a.Session().Messages) |
| 490 | if got := toolsByID["call-1"]; !strings.Contains(got, "tool1 done") { |
| 491 | t.Fatalf("completed tool result was not persisted before cancellation: %q", got) |
| 492 | } |
| 493 | if got := toolsByID["call-3"]; !strings.Contains(got, "cancelled") { |
| 494 | t.Fatalf("skipped tool result was not persisted as cancelled: %q", got) |
| 495 | } |
| 496 | } |
| 497 | |
| 498 | // TestCancelBeforeParallelBatchSkipsTheWholeRemainingBatch verifies that a |
| 499 | // cancellation in a serial writer segment prevents the next read-only parallel |
| 500 | // segment from starting. |
| 501 | func TestCancelBeforeParallelBatchSkipsTheWholeRemainingBatch(t *testing.T) { |
| 502 | executedMu.Lock() |
| 503 | executed = nil |
| 504 | executedMu.Unlock() |
| 505 | |
| 506 | reg := tool.NewRegistry() |
| 507 | reg.Add(trackingTool{}) |
| 508 | reg.Add(trackingTool{name: "readonly_tracking", readOnly: true}) |
| 509 | |
| 510 | mp := testutil.NewMock("m", |
| 511 | testutil.Turn{ |
| 512 | Text: "", |
| 513 | ToolCalls: []provider.ToolCall{ |
| 514 | {ID: "call-1", Name: "tracking", Arguments: `{"name": "writer", "delay_ms": 5000}`}, |
| 515 | {ID: "call-2", Name: "readonly_tracking", Arguments: `{"name": "read1", "delay_ms": 50}`}, |
| 516 | {ID: "call-3", Name: "readonly_tracking", Arguments: `{"name": "read2", "delay_ms": 50}`}, |
| 517 | }, |
| 518 | }, |
| 519 | ) |
| 520 | |
| 521 | sink := &recordSink{} |
| 522 | a := New(mp, reg, NewSession(""), Options{}, sink) |
| 523 | |
| 524 | ctx, cancel := context.WithCancel(context.Background()) |
| 525 | defer cancel() |
| 526 | |
| 527 | done := make(chan error, 1) |
| 528 | go func() { |
| 529 | done <- a.Run(ctx, "test cancel before parallel batch") |
| 530 | }() |
| 531 | go func() { |
| 532 | time.Sleep(300 * time.Millisecond) |
| 533 | cancel() |
| 534 | }() |
| 535 | |
| 536 | select { |
| 537 | case err := <-done: |
| 538 | if err == nil { |
| 539 | t.Fatal("Run returned nil, want context cancellation") |
| 540 | } |
| 541 | case <-time.After(5 * time.Second): |
| 542 | t.Fatal("Run did not complete within 5s") |
| 543 | } |
| 544 | |
| 545 | executedMu.Lock() |
| 546 | executedCopy := append([]string(nil), executed...) |
| 547 | executedMu.Unlock() |
| 548 | for _, name := range executedCopy { |
| 549 | if strings.HasPrefix(name, "read") { |
| 550 | t.Fatalf("read-only parallel batch should not start after cancel, executed: %v", executedCopy) |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | results := sink.kinds(event.ToolResult) |
| 555 | if len(results) != 3 { |
| 556 | t.Fatalf("ToolResult events = %d, want 3", len(results)) |
| 557 | } |
| 558 | for _, e := range results[1:] { |
| 559 | if e.Tool.Err == "" { |
| 560 | t.Fatalf("cancelled unstarted tool result should carry an error: %+v", e.Tool) |
| 561 | } |
| 562 | if !strings.Contains(e.Tool.Output, "cancelled") { |
| 563 | t.Fatalf("cancelled unstarted tool result should explain cancellation: %+v", e.Tool) |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | |
| 568 | func TestCancelInsideLargeParallelBatchStopsSchedulingNewTools(t *testing.T) { |
| 569 | executedMu.Lock() |
| 570 | executed = nil |
| 571 | executedMu.Unlock() |
| 572 | |
| 573 | reg := tool.NewRegistry() |
| 574 | reg.Add(trackingTool{name: "readonly_tracking", readOnly: true}) |
| 575 | |
| 576 | var calls []provider.ToolCall |
| 577 | for i := 0; i < 12; i++ { |
| 578 | calls = append(calls, provider.ToolCall{ |
| 579 | ID: fmt.Sprintf("call-%02d", i), |
| 580 | Name: "readonly_tracking", |
| 581 | Arguments: fmt.Sprintf(`{"name": "read%02d", "delay_ms": 5000}`, i), |
| 582 | }) |
| 583 | } |
| 584 | |
| 585 | mp := testutil.NewMock("m", testutil.Turn{ToolCalls: calls}) |
| 586 | a := New(mp, reg, NewSession(""), Options{}, &recordSink{}) |
| 587 | |
| 588 | ctx, cancel := context.WithCancel(context.Background()) |
| 589 | defer cancel() |
| 590 | |
| 591 | done := make(chan error, 1) |
| 592 | go func() { |
| 593 | done <- a.Run(ctx, "test cancel inside parallel batch") |
| 594 | }() |
| 595 | go func() { |
| 596 | time.Sleep(300 * time.Millisecond) |
| 597 | cancel() |
| 598 | }() |
| 599 | |
| 600 | select { |
| 601 | case err := <-done: |
| 602 | if err == nil { |
| 603 | t.Fatal("Run returned nil, want context cancellation") |
| 604 | } |
| 605 | case <-time.After(5 * time.Second): |
| 606 | t.Fatal("Run did not complete within 5s") |
| 607 | } |
| 608 | |
| 609 | executedMu.Lock() |
| 610 | executedCopy := append([]string(nil), executed...) |
| 611 | executedMu.Unlock() |
| 612 | for _, name := range executedCopy { |
| 613 | for i := 8; i < 12; i++ { |
| 614 | if strings.HasPrefix(name, fmt.Sprintf("read%02d", i)) { |
| 615 | t.Fatalf("parallel scheduler started a tool after cancellation: %v", executedCopy) |
| 616 | } |
| 617 | } |
| 618 | } |
| 619 | |
| 620 | toolsByID := toolMessagesByID(a.Session().Messages) |
| 621 | if len(toolsByID) != len(calls) { |
| 622 | t.Fatalf("persisted tool messages = %d, want %d: %#v", len(toolsByID), len(calls), toolsByID) |
| 623 | } |
| 624 | if got := toolsByID["call-08"]; !strings.Contains(got, "cancelled") { |
| 625 | t.Fatalf("unstarted parallel tool result was not persisted as cancelled: %q", got) |
| 626 | } |
| 627 | } |
| 628 | |
| 629 | func toolMessagesByID(msgs []provider.Message) map[string]string { |
| 630 | out := make(map[string]string) |
| 631 | for _, m := range msgs { |
| 632 | if m.Role == provider.RoleTool && !m.LocalOnly { |
| 633 | out[m.ToolCallID] = m.Content |
| 634 | } |
| 635 | } |
| 636 | return out |
| 637 | } |
| 638 |