| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "reflect" |
| 8 | "strings" |
| 9 | "sync" |
| 10 | "sync/atomic" |
| 11 | "testing" |
| 12 | "time" |
| 13 | "unicode/utf8" |
| 14 | |
| 15 | "reasonix/internal/event" |
| 16 | "reasonix/internal/evidence" |
| 17 | "reasonix/internal/provider" |
| 18 | "reasonix/internal/tool" |
| 19 | _ "reasonix/internal/tool/builtin" |
| 20 | ) |
| 21 | |
| 22 | // TestTruncateToolOutputUnderCap leaves small payloads alone — the cap should |
| 23 | // never rewrite content that already fits. |
| 24 | func TestTruncateToolOutputUnderCap(t *testing.T) { |
| 25 | in := strings.Repeat("a", maxToolOutputBytes) |
| 26 | got, notice := truncateToolOutput(in) |
| 27 | if got != in { |
| 28 | t.Errorf("payload at exactly the cap was rewritten") |
| 29 | } |
| 30 | if notice != "" { |
| 31 | t.Errorf("at-cap payload should not emit a notice, got %q", notice) |
| 32 | } |
| 33 | } |
| 34 | |
| 35 | // TestTruncateToolOutputHeadTail keeps head+tail of an oversize payload and |
| 36 | // inserts a marker; the notice must report the elided byte count truthfully. |
| 37 | func TestTruncateToolOutputHeadTail(t *testing.T) { |
| 38 | head := strings.Repeat("H", maxToolOutputBytes) |
| 39 | tail := strings.Repeat("T", maxToolOutputBytes) |
| 40 | in := head + tail |
| 41 | out, notice := truncateToolOutput(in) |
| 42 | if !strings.HasPrefix(out, "H") || !strings.HasSuffix(out, "T") { |
| 43 | t.Errorf("head/tail not preserved at the edges: %q…%q", out[:20], out[len(out)-20:]) |
| 44 | } |
| 45 | if !strings.Contains(out, "truncated") { |
| 46 | t.Errorf("truncation marker missing: %q", out) |
| 47 | } |
| 48 | if len(out) >= len(in) { |
| 49 | t.Errorf("output not shorter than input: in=%d out=%d", len(in), len(out)) |
| 50 | } |
| 51 | if !strings.Contains(notice, "truncated") { |
| 52 | t.Errorf("notice missing: %q", notice) |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | // TestTruncateToolOutputRuneBoundaries puts multibyte runes exactly across the |
| 57 | // head and tail cut points; the result must still be valid UTF-8. |
| 58 | func TestTruncateToolOutputRuneBoundaries(t *testing.T) { |
| 59 | in := strings.Repeat("中", maxToolOutputBytes) // 3 bytes each — guarantees a cut inside a rune |
| 60 | out, _ := truncateToolOutput(in) |
| 61 | if !utf8.ValidString(out) { |
| 62 | t.Errorf("truncated output is not valid UTF-8") |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | // TestFinishReasonMessage only yields a warning for abnormal terminations. |
| 67 | // Normal stops are silent (ok=false) so the per-turn line stays clean. |
| 68 | func TestFinishReasonMessage(t *testing.T) { |
| 69 | silent := []string{"", "stop", "tool_calls"} |
| 70 | for _, r := range silent { |
| 71 | if msg, ok := finishReasonMessage(&provider.Usage{FinishReason: r}); ok { |
| 72 | t.Errorf("finish_reason=%q should be silent, got %q", r, msg) |
| 73 | } |
| 74 | } |
| 75 | loud := map[string]string{ |
| 76 | "length": "max output", |
| 77 | "content_filter": "content filter", |
| 78 | "repetition_truncation": "repetition", |
| 79 | } |
| 80 | for reason, fragment := range loud { |
| 81 | msg, ok := finishReasonMessage(&provider.Usage{FinishReason: reason}) |
| 82 | if !ok || !strings.Contains(msg, fragment) { |
| 83 | t.Errorf("finish_reason=%q: got (%q, %v), want fragment %q", reason, msg, ok, fragment) |
| 84 | } |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | // TestEmptyFinalNotice keeps the user-facing line short while preserving the |
| 89 | // diagnostics that tell empty-answer causes apart in expandable details. |
| 90 | func TestEmptyFinalNotice(t *testing.T) { |
| 91 | msg := emptyFinalNotice() |
| 92 | for _, hidden := range []string{"blocked", "finish=", "reasoning="} { |
| 93 | if strings.Contains(msg, hidden) { |
| 94 | t.Errorf("notice %q should not expose internal diagnostic %q", msg, hidden) |
| 95 | } |
| 96 | } |
| 97 | detail := emptyFinalNoticeDetail("deepseek-flash", &provider.Usage{FinishReason: "stop"}, 512) |
| 98 | for _, want := range []string{"deepseek-flash", "finish=stop", "reasoning=512"} { |
| 99 | if !strings.Contains(detail, want) { |
| 100 | t.Errorf("notice detail %q missing %q", detail, want) |
| 101 | } |
| 102 | } |
| 103 | if got := emptyFinalNoticeDetail("p", nil, 0); !strings.Contains(got, "finish=unknown") { |
| 104 | t.Errorf("nil usage should report finish=unknown, got %q", got) |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | // parallel-dispatch tests |
| 109 | |
| 110 | // fakeTool is a minimal Tool stand-in for dispatch tests; ReadOnly is |
| 111 | // configurable and Execute sleeps a fixed duration so we can measure |
| 112 | // serial vs parallel behaviour by wall-clock. |
| 113 | type fakeTool struct { |
| 114 | name string |
| 115 | readOnly bool |
| 116 | delay time.Duration |
| 117 | err error |
| 118 | calls *int32 // shared counter to assert all dispatched |
| 119 | } |
| 120 | |
| 121 | type blockingTool struct { |
| 122 | name string |
| 123 | started chan struct{} |
| 124 | release chan struct{} |
| 125 | } |
| 126 | |
| 127 | func (b blockingTool) Name() string { return b.name } |
| 128 | func (b blockingTool) Description() string { return "" } |
| 129 | func (b blockingTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 130 | func (b blockingTool) ReadOnly() bool { return true } |
| 131 | func (b blockingTool) Execute(ctx context.Context, _ json.RawMessage) (string, error) { |
| 132 | close(b.started) |
| 133 | select { |
| 134 | case <-b.release: |
| 135 | return b.name + " done", nil |
| 136 | case <-ctx.Done(): |
| 137 | return "", ctx.Err() |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | type workspaceSignalSink struct { |
| 142 | mu sync.Mutex |
| 143 | events []event.Event |
| 144 | mutations chan event.WorkspaceMutation |
| 145 | } |
| 146 | |
| 147 | type failingToolDispatchSink struct{ err error } |
| 148 | |
| 149 | func (s failingToolDispatchSink) Emit(event.Event) {} |
| 150 | func (s failingToolDispatchSink) EmitChecked(e event.Event) error { |
| 151 | if e.Kind == event.ToolDispatch { |
| 152 | return s.err |
| 153 | } |
| 154 | return nil |
| 155 | } |
| 156 | |
| 157 | func newWorkspaceSignalSink() *workspaceSignalSink { |
| 158 | return &workspaceSignalSink{mutations: make(chan event.WorkspaceMutation, 8)} |
| 159 | } |
| 160 | |
| 161 | func (s *workspaceSignalSink) Emit(e event.Event) { |
| 162 | s.mu.Lock() |
| 163 | s.events = append(s.events, e) |
| 164 | s.mu.Unlock() |
| 165 | } |
| 166 | |
| 167 | func (s *workspaceSignalSink) RecordWorkspaceMutation(m event.WorkspaceMutation) { |
| 168 | s.mutations <- m |
| 169 | } |
| 170 | |
| 171 | func (s *workspaceSignalSink) kinds(kind event.Kind) []event.Event { |
| 172 | s.mu.Lock() |
| 173 | defer s.mu.Unlock() |
| 174 | var out []event.Event |
| 175 | for _, e := range s.events { |
| 176 | if e.Kind == kind { |
| 177 | out = append(out, e) |
| 178 | } |
| 179 | } |
| 180 | return out |
| 181 | } |
| 182 | |
| 183 | func (f fakeTool) Name() string { return f.name } |
| 184 | func (f fakeTool) Description() string { return "" } |
| 185 | func (f fakeTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 186 | func (f fakeTool) ReadOnly() bool { return f.readOnly } |
| 187 | func (f fakeTool) Execute(ctx context.Context, _ json.RawMessage) (string, error) { |
| 188 | if f.calls != nil { |
| 189 | atomic.AddInt32(f.calls, 1) |
| 190 | } |
| 191 | select { |
| 192 | case <-time.After(f.delay): |
| 193 | case <-ctx.Done(): |
| 194 | return "", ctx.Err() |
| 195 | } |
| 196 | if f.err != nil { |
| 197 | return "", f.err |
| 198 | } |
| 199 | return f.name + " done", nil |
| 200 | } |
| 201 | |
| 202 | func TestPartitionToolCallsAllReadOnly(t *testing.T) { |
| 203 | reg := tool.NewRegistry() |
| 204 | reg.Add(fakeTool{name: "ro1", readOnly: true}) |
| 205 | reg.Add(fakeTool{name: "ro2", readOnly: true}) |
| 206 | calls := []provider.ToolCall{{Name: "ro1"}, {Name: "ro2"}} |
| 207 | got := partitionToolCalls(reg, calls) |
| 208 | want := []toolCallBatch{{start: 0, end: 2, parallel: true}} |
| 209 | if !reflect.DeepEqual(got, want) { |
| 210 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | // TestPartitionToolCallsSegmentsAroundWriters verifies a writer only serializes |
| 215 | // its own provider-order position; read-only runs on either side stay batchable. |
| 216 | func TestPartitionToolCallsSegmentsAroundWriters(t *testing.T) { |
| 217 | reg := tool.NewRegistry() |
| 218 | reg.Add(fakeTool{name: "ro", readOnly: true}) |
| 219 | reg.Add(fakeTool{name: "rw", readOnly: false}) |
| 220 | calls := []provider.ToolCall{{Name: "ro"}, {Name: "rw"}, {Name: "ro"}} |
| 221 | got := partitionToolCalls(reg, calls) |
| 222 | want := []toolCallBatch{ |
| 223 | {start: 0, end: 1, parallel: true}, |
| 224 | {start: 1, end: 2}, |
| 225 | {start: 2, end: 3, parallel: true}, |
| 226 | } |
| 227 | if !reflect.DeepEqual(got, want) { |
| 228 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 229 | } |
| 230 | } |
| 231 | |
| 232 | // TestPartitionToolCallsUnknownToolSerial keeps unknown-tool errors |
| 233 | // deterministic by forcing unknown calls into single-call serial batches. |
| 234 | func TestPartitionToolCallsUnknownToolSerial(t *testing.T) { |
| 235 | reg := tool.NewRegistry() |
| 236 | reg.Add(fakeTool{name: "ro", readOnly: true}) |
| 237 | calls := []provider.ToolCall{{Name: "ro"}, {Name: "vanished"}, {Name: "ro"}} |
| 238 | got := partitionToolCalls(reg, calls) |
| 239 | want := []toolCallBatch{ |
| 240 | {start: 0, end: 1, parallel: true}, |
| 241 | {start: 1, end: 2}, |
| 242 | {start: 2, end: 3, parallel: true}, |
| 243 | } |
| 244 | if !reflect.DeepEqual(got, want) { |
| 245 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 246 | } |
| 247 | } |
| 248 | |
| 249 | // A retired tool accidentally present in an embedding registry carries no |
| 250 | // active receipt semantics and does not change read-only partitioning. |
| 251 | func TestPartitionToolCallsRetiredToolHasNoBarrier(t *testing.T) { |
| 252 | reg := tool.NewRegistry() |
| 253 | reg.Add(fakeTool{name: "read_file", readOnly: true}) |
| 254 | reg.Add(fakeTool{name: "complete_step", readOnly: true}) |
| 255 | |
| 256 | calls := []provider.ToolCall{{Name: "read_file"}, {Name: "complete_step"}} |
| 257 | got := partitionToolCalls(reg, calls) |
| 258 | want := []toolCallBatch{{start: 0, end: 2, parallel: true}} |
| 259 | if !reflect.DeepEqual(got, want) { |
| 260 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | func TestPartitionToolCallsCompressSerial(t *testing.T) { |
| 265 | reg := tool.NewRegistry() |
| 266 | reg.Add(fakeTool{name: "read_file", readOnly: true}) |
| 267 | reg.Add(fakeTool{name: "compress", readOnly: true}) |
| 268 | |
| 269 | calls := []provider.ToolCall{{Name: "read_file"}, {Name: "compress"}, {Name: "read_file"}} |
| 270 | got := partitionToolCalls(reg, calls) |
| 271 | want := []toolCallBatch{ |
| 272 | {start: 0, end: 1, parallel: true}, |
| 273 | {start: 1, end: 2}, |
| 274 | {start: 2, end: 3, parallel: true}, |
| 275 | } |
| 276 | if !reflect.DeepEqual(got, want) { |
| 277 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | func TestPartitionToolCallsTodoWriteSerial(t *testing.T) { |
| 282 | reg := tool.NewRegistry() |
| 283 | reg.Add(fakeTool{name: "read_file", readOnly: true}) |
| 284 | reg.Add(fakeTool{name: "todo_write", readOnly: true}) |
| 285 | |
| 286 | calls := []provider.ToolCall{{Name: "read_file"}, {Name: "todo_write"}, {Name: "read_file"}} |
| 287 | got := partitionToolCalls(reg, calls) |
| 288 | want := []toolCallBatch{ |
| 289 | {start: 0, end: 1, parallel: true}, |
| 290 | {start: 1, end: 2}, |
| 291 | {start: 2, end: 3, parallel: true}, |
| 292 | } |
| 293 | if !reflect.DeepEqual(got, want) { |
| 294 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | func TestPartitionToolCallsBackgroundCollectorsSerial(t *testing.T) { |
| 299 | reg := tool.NewRegistry() |
| 300 | reg.Add(fakeTool{name: "read_file", readOnly: true}) |
| 301 | reg.Add(fakeTool{name: "wait", readOnly: true}) |
| 302 | reg.Add(fakeTool{name: "bash_output", readOnly: true}) |
| 303 | |
| 304 | calls := []provider.ToolCall{{Name: "read_file"}, {Name: "wait"}, {Name: "bash_output"}, {Name: "read_file"}} |
| 305 | got := partitionToolCalls(reg, calls) |
| 306 | want := []toolCallBatch{ |
| 307 | {start: 0, end: 1, parallel: true}, |
| 308 | {start: 1, end: 2}, |
| 309 | {start: 2, end: 3}, |
| 310 | {start: 3, end: 4, parallel: true}, |
| 311 | } |
| 312 | if !reflect.DeepEqual(got, want) { |
| 313 | t.Fatalf("partitionToolCalls = %+v, want %+v", got, want) |
| 314 | } |
| 315 | } |
| 316 | |
| 317 | // TestExecuteBatchParallelReadOnly checks that three 80ms read-only calls |
| 318 | // complete in well under 3×80ms — the wall-clock proof of true parallelism. |
| 319 | func TestExecuteBatchParallelReadOnly(t *testing.T) { |
| 320 | const delay = 80 * time.Millisecond |
| 321 | calls := int32(0) |
| 322 | reg := tool.NewRegistry() |
| 323 | reg.Add(fakeTool{name: "a", readOnly: true, delay: delay, calls: &calls}) |
| 324 | reg.Add(fakeTool{name: "b", readOnly: true, delay: delay, calls: &calls}) |
| 325 | reg.Add(fakeTool{name: "c", readOnly: true, delay: delay, calls: &calls}) |
| 326 | |
| 327 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 328 | |
| 329 | start := time.Now() |
| 330 | batch := a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{{Name: "a"}, {Name: "b"}, {Name: "c"}}) |
| 331 | results := batch.results |
| 332 | elapsed := time.Since(start) |
| 333 | |
| 334 | if calls != 3 { |
| 335 | t.Errorf("dispatched %d calls, want 3", calls) |
| 336 | } |
| 337 | if len(results) != 3 || results[0] != "a done" || results[1] != "b done" || results[2] != "c done" { |
| 338 | t.Errorf("results out of order or wrong: %v", results) |
| 339 | } |
| 340 | // Allow generous slack for CI; even 2x serial would prove we got parallelism. |
| 341 | if elapsed >= 2*delay { |
| 342 | t.Errorf("read-only batch took %v (>= %v) — not parallel", elapsed, 2*delay) |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | func TestExecuteBatchDoesNotRunAnyToolWhenDispatchPersistenceFails(t *testing.T) { |
| 347 | wantErr := errors.New("ledger unavailable") |
| 348 | var calls int32 |
| 349 | reg := tool.NewRegistry() |
| 350 | reg.Add(fakeTool{name: "read_file", readOnly: true, calls: &calls}) |
| 351 | reg.Add(fakeTool{name: "write_file", calls: &calls}) |
| 352 | a := New(nil, reg, NewSession(""), Options{}, failingToolDispatchSink{err: wantErr}) |
| 353 | batch := a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{ |
| 354 | {ID: "read", Name: "read_file"}, |
| 355 | {ID: "write", Name: "write_file"}, |
| 356 | }) |
| 357 | if !errors.Is(batch.err, wantErr) { |
| 358 | t.Fatalf("batch error = %v, want %v", batch.err, wantErr) |
| 359 | } |
| 360 | if got := atomic.LoadInt32(&calls); got != 0 { |
| 361 | t.Fatalf("executed %d tools after dispatch persistence failed", got) |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | func TestExecuteBatchStampsToolResultTimestamps(t *testing.T) { |
| 366 | const delay = 30 * time.Millisecond |
| 367 | reg := tool.NewRegistry() |
| 368 | reg.Add(fakeTool{name: "a", readOnly: true, delay: delay}) |
| 369 | |
| 370 | sink := &recordSink{} |
| 371 | a := New(nil, reg, NewSession(""), Options{}, sink) |
| 372 | |
| 373 | before := time.Now().UnixMilli() |
| 374 | a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{{Name: "a"}}) |
| 375 | after := time.Now().UnixMilli() |
| 376 | |
| 377 | results := sink.kinds(event.ToolResult) |
| 378 | if len(results) != 1 { |
| 379 | t.Fatalf("got %d tool results, want 1", len(results)) |
| 380 | } |
| 381 | tr := results[0].Tool |
| 382 | if tr.StartedAt < before || tr.StartedAt > after { |
| 383 | t.Errorf("StartedAt = %d, want within [%d, %d]", tr.StartedAt, before, after) |
| 384 | } |
| 385 | if tr.EndedAt != tr.StartedAt+tr.DurationMs { |
| 386 | t.Errorf("EndedAt = %d, want StartedAt+DurationMs = %d", tr.EndedAt, tr.StartedAt+tr.DurationMs) |
| 387 | } |
| 388 | if tr.DurationMs < delay.Milliseconds() { |
| 389 | t.Errorf("DurationMs = %d, want >= %d", tr.DurationMs, delay.Milliseconds()) |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | func TestExecuteBatchMarksOnlyExecutedWritersForWorkspaceRefresh(t *testing.T) { |
| 394 | reg := tool.NewRegistry() |
| 395 | reg.Add(fakeTool{name: "write_file"}) |
| 396 | reg.Add(fakeTool{name: "read_file", readOnly: true}) |
| 397 | sink := &recordSink{} |
| 398 | a := New(nil, reg, NewSession(""), Options{}, sink) |
| 399 | a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{ |
| 400 | {Name: "write_file", Arguments: `{"path":"pkg/main.go","content":"x"}`}, |
| 401 | {Name: "read_file", Arguments: `{"path":"pkg/main.go"}`}, |
| 402 | }) |
| 403 | results := sink.kinds(event.ToolResult) |
| 404 | if len(results) != 2 { |
| 405 | t.Fatalf("got %d results, want 2", len(results)) |
| 406 | } |
| 407 | if !results[0].Tool.WorkspaceMutation || len(results[0].Tool.WorkspacePaths) != 1 || results[0].Tool.WorkspacePaths[0] != "pkg/main.go" { |
| 408 | t.Fatalf("writer metadata = %+v", results[0].Tool) |
| 409 | } |
| 410 | if results[1].Tool.WorkspaceMutation { |
| 411 | t.Fatalf("read-only call was marked as mutation: %+v", results[1].Tool) |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | func TestExecuteBatchMarksFailedWriterForWorkspaceRefresh(t *testing.T) { |
| 416 | reg := tool.NewRegistry() |
| 417 | reg.Add(fakeTool{name: "write_file", err: errors.New("partial write")}) |
| 418 | sink := &recordSink{} |
| 419 | a := New(nil, reg, NewSession(""), Options{}, sink) |
| 420 | a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{{Name: "write_file", Arguments: `{"path":"partial.go"}`}}) |
| 421 | results := sink.kinds(event.ToolResult) |
| 422 | if len(results) != 1 || !results[0].Tool.WorkspaceMutation { |
| 423 | t.Fatalf("failed writer did not invalidate workspace: %+v", results) |
| 424 | } |
| 425 | } |
| 426 | |
| 427 | func TestExecuteBatchPublishesWorkspaceMutationBeforeLaterToolCompletes(t *testing.T) { |
| 428 | started := make(chan struct{}) |
| 429 | release := make(chan struct{}) |
| 430 | reg := tool.NewRegistry() |
| 431 | reg.Add(fakeTool{name: "write_file"}) |
| 432 | reg.Add(blockingTool{name: "slow_read", started: started, release: release}) |
| 433 | sink := newWorkspaceSignalSink() |
| 434 | a := New(nil, reg, NewSession(""), Options{}, sink) |
| 435 | done := make(chan struct{}) |
| 436 | go func() { |
| 437 | defer close(done) |
| 438 | a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{ |
| 439 | {Name: "write_file", Arguments: `{"path":"ready.go","content":"x"}`}, |
| 440 | {Name: "slow_read", Arguments: `{}`}, |
| 441 | }) |
| 442 | }() |
| 443 | |
| 444 | select { |
| 445 | case <-started: |
| 446 | case <-time.After(2 * time.Second): |
| 447 | t.Fatal("later tool did not start") |
| 448 | } |
| 449 | select { |
| 450 | case mutation := <-sink.mutations: |
| 451 | if mutation.ToolName != "write_file" || len(mutation.Paths) != 1 || mutation.Paths[0] != "ready.go" { |
| 452 | t.Fatalf("workspace mutation = %+v", mutation) |
| 453 | } |
| 454 | case <-time.After(2 * time.Second): |
| 455 | t.Fatal("writer completion did not publish before the later tool completed") |
| 456 | } |
| 457 | if results := sink.kinds(event.ToolResult); len(results) != 1 || results[0].Tool.Name != "write_file" { |
| 458 | t.Fatalf("completed writer must be checkpointed before the next tool: %+v", results) |
| 459 | } |
| 460 | close(release) |
| 461 | select { |
| 462 | case <-done: |
| 463 | case <-time.After(2 * time.Second): |
| 464 | t.Fatal("batch did not finish after releasing the later tool") |
| 465 | } |
| 466 | } |
| 467 | |
| 468 | func TestWorkspaceMutationClassifierTreatsGitCommitAsGitMetadata(t *testing.T) { |
| 469 | mutation, ok := workspaceMutationForCall("call", "bash", json.RawMessage(`{"command":"git commit -m test"}`), false) |
| 470 | if !ok || !mutation.GitMeta || mutation.Content || mutation.WorkingTree || mutation.Tree || !mutation.AllPaths { |
| 471 | t.Fatalf("git commit workspace invalidation = %+v, ok=%v", mutation, ok) |
| 472 | } |
| 473 | mutation, ok = workspaceMutationForCall("call", "bash", json.RawMessage(`{"command":"git commit -am test"}`), false) |
| 474 | if !ok || !mutation.GitMeta || !mutation.Content || !mutation.WorkingTree || !mutation.Tree { |
| 475 | t.Fatalf("content-writing git commit invalidation = %+v, ok=%v", mutation, ok) |
| 476 | } |
| 477 | if mutation, ok = workspaceMutationForCall("call", "bash", json.RawMessage(`{"command":"go test ./..."}`), false); ok { |
| 478 | t.Fatalf("ordinary verifier invalidated durable workspace state: %+v", mutation) |
| 479 | } |
| 480 | if mutation, ok = workspaceMutationForCall("call", "bash", json.RawMessage(`{"command":"date --set tomorrow"}`), false); ok { |
| 481 | t.Fatalf("host-only state write invalidated the workspace: %+v", mutation) |
| 482 | } |
| 483 | if mutation, ok = workspaceMutationForCall("call", "remember", json.RawMessage(`{"name":"preference"}`), false); ok { |
| 484 | t.Fatalf("host-only memory write invalidated the workspace: %+v", mutation) |
| 485 | } |
| 486 | } |
| 487 | |
| 488 | func TestExecuteBatchCancelledCallsCarryNoTimestamps(t *testing.T) { |
| 489 | reg := tool.NewRegistry() |
| 490 | reg.Add(fakeTool{name: "a", readOnly: true}) |
| 491 | |
| 492 | sink := &recordSink{} |
| 493 | a := New(nil, reg, NewSession(""), Options{}, sink) |
| 494 | |
| 495 | ctx, cancel := context.WithCancel(context.Background()) |
| 496 | cancel() |
| 497 | a.executeBatch(ctx, &a.turn, []provider.ToolCall{{Name: "a"}}) |
| 498 | |
| 499 | results := sink.kinds(event.ToolResult) |
| 500 | if len(results) != 1 { |
| 501 | t.Fatalf("got %d tool results, want 1", len(results)) |
| 502 | } |
| 503 | tr := results[0].Tool |
| 504 | if tr.StartedAt != 0 || tr.EndedAt != 0 { |
| 505 | t.Errorf("never-ran call has StartedAt=%d EndedAt=%d, want both zero", tr.StartedAt, tr.EndedAt) |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | // TestExecuteBatchSegmentsAroundWrites ensures a write call only serializes its |
| 510 | // own position in the provider-ordered batch: read-only runs before and after it |
| 511 | // may still parallelise within their contiguous segments. |
| 512 | func TestExecuteBatchSegmentsAroundWrites(t *testing.T) { |
| 513 | // A larger per-call delay keeps fixed scheduler jitter on loaded CI a small |
| 514 | // fraction of the segment time, so the tight relative bound below stays |
| 515 | // reliable instead of being widened toward the serial floor. |
| 516 | const delay = 150 * time.Millisecond |
| 517 | reg := tool.NewRegistry() |
| 518 | reg.Add(fakeTool{name: "ro1", readOnly: true, delay: delay}) |
| 519 | reg.Add(fakeTool{name: "ro2", readOnly: true, delay: delay}) |
| 520 | reg.Add(fakeTool{name: "ro3", readOnly: true, delay: delay}) |
| 521 | reg.Add(fakeTool{name: "ro4", readOnly: true, delay: delay}) |
| 522 | reg.Add(fakeTool{name: "rw", readOnly: false, delay: delay}) |
| 523 | |
| 524 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 525 | |
| 526 | start := time.Now() |
| 527 | batch := a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{ |
| 528 | {Name: "ro1"}, |
| 529 | {Name: "ro2"}, |
| 530 | {Name: "rw"}, |
| 531 | {Name: "ro3"}, |
| 532 | {Name: "ro4"}, |
| 533 | }) |
| 534 | results := batch.results |
| 535 | elapsed := time.Since(start) |
| 536 | |
| 537 | want := []string{"ro1 done", "ro2 done", "rw done", "ro3 done", "ro4 done"} |
| 538 | if len(results) != len(want) { |
| 539 | t.Fatalf("got %d results, want %d: %v", len(results), len(want), results) |
| 540 | } |
| 541 | for i := range want { |
| 542 | if stripReceiptCitation(results[i]) != want[i] { |
| 543 | t.Fatalf("results out of order or wrong: got %v want %v", results, want) |
| 544 | } |
| 545 | } |
| 546 | // Desired shape is roughly 3*delay: (ro1|ro2), then rw, then (ro3|ro4). |
| 547 | // Old all-serial behaviour is roughly 5*delay and should fail this bound. |
| 548 | if elapsed >= 4*delay { |
| 549 | t.Errorf("mixed batch took %v (>= %v) — read-only segments did not parallelise", elapsed, 4*delay) |
| 550 | } |
| 551 | if elapsed < 2*delay { |
| 552 | t.Errorf("mixed batch took only %v — write call appears to have overlapped a read-only segment", elapsed) |
| 553 | } |
| 554 | } |
| 555 | |
| 556 | func TestExecuteBatchPairsRetiredCompleteStepWithoutBlockingPeers(t *testing.T) { |
| 557 | reg := tool.NewRegistry() |
| 558 | reg.Add(fakeTool{name: "bash", readOnly: false}) |
| 559 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 560 | |
| 561 | a.SeedTodoState([]evidence.TodoItem{{Content: "Run checks", Status: "pending"}}) |
| 562 | batch := a.executeBatch(context.Background(), &a.turn, []provider.ToolCall{ |
| 563 | {Name: "bash", Arguments: `{"command":"go test ./internal/..."}`}, |
| 564 | {Name: "complete_step", Arguments: `{ |
| 565 | "step":"Run checks", |
| 566 | "result":"checks passed", |
| 567 | "evidence":[{"kind":"verification","summary":"tests passed","command":"go test ./internal/..."}] |
| 568 | }`}, |
| 569 | }) |
| 570 | results := batch.results |
| 571 | |
| 572 | if len(results) != 2 { |
| 573 | t.Fatalf("got %d results, want 2", len(results)) |
| 574 | } |
| 575 | if !strings.Contains(results[1], "tool_retired") || !a.task.ledger.HasSuccessfulCommand("go test ./internal/...") { |
| 576 | t.Fatalf("retired result=%q bash receipt missing=%v", results[1], !a.task.ledger.HasSuccessfulCommand("go test ./internal/...")) |
| 577 | } |
| 578 | } |
| 579 | |
| 580 | func TestExecuteOneFailedReceiptDoesNotVerify(t *testing.T) { |
| 581 | reg := tool.NewRegistry() |
| 582 | reg.Add(fakeTool{name: "bash", readOnly: false, err: errors.New("boom")}) |
| 583 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 584 | |
| 585 | out := a.executeOne(context.Background(), &a.turn, provider.ToolCall{Name: "bash", Arguments: `{"command":"go test ./..."}`}) |
| 586 | if out.errMsg == "" { |
| 587 | t.Fatal("failing fake tool should return an error outcome") |
| 588 | } |
| 589 | if a.task.ledger.HasSuccessfulCommand("go test ./...") { |
| 590 | t.Fatal("failed bash receipt must not verify") |
| 591 | } |
| 592 | } |
| 593 | |
| 594 | func TestRunResetsEvidenceLedger(t *testing.T) { |
| 595 | reg := tool.NewRegistry() |
| 596 | reg.Add(fakeTool{name: "bash", readOnly: false}) |
| 597 | prov := &mockProvider{name: "p", chunks: []provider.Chunk{{Type: provider.ChunkText, Text: "done"}}} |
| 598 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 599 | |
| 600 | a.executeOne(context.Background(), &a.turn, provider.ToolCall{Name: "bash", Arguments: `{"command":"go test ./..."}`}) |
| 601 | if !a.task.ledger.HasSuccessfulCommand("go test ./...") { |
| 602 | t.Fatal("setup failed to record evidence") |
| 603 | } |
| 604 | |
| 605 | if err := a.Run(context.Background(), "next turn"); err != nil { |
| 606 | t.Fatalf("Run: %v", err) |
| 607 | } |
| 608 | if a.task.ledger.HasSuccessfulCommand("go test ./...") { |
| 609 | t.Fatal("new user turn should not inherit previous receipts") |
| 610 | } |
| 611 | } |
| 612 |