| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "strings" |
| 10 | "sync/atomic" |
| 11 | "testing" |
| 12 | |
| 13 | "reasonix/internal/diff" |
| 14 | "reasonix/internal/event" |
| 15 | "reasonix/internal/provider" |
| 16 | "reasonix/internal/tool" |
| 17 | "reasonix/internal/tool/builtin" |
| 18 | ) |
| 19 | |
| 20 | type failingWriterTool struct { |
| 21 | name string |
| 22 | calls *int32 |
| 23 | } |
| 24 | |
| 25 | func (f failingWriterTool) Name() string { return f.name } |
| 26 | func (f failingWriterTool) Description() string { return "always fails to write" } |
| 27 | func (f failingWriterTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 28 | func (f failingWriterTool) ReadOnly() bool { return false } |
| 29 | func (f failingWriterTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 30 | if f.calls != nil { |
| 31 | atomic.AddInt32(f.calls, 1) |
| 32 | } |
| 33 | return "", errors.New("old_string not found in prompt.txt") |
| 34 | } |
| 35 | |
| 36 | type stateAwareWriterTool struct { |
| 37 | name string |
| 38 | calls *int32 |
| 39 | valid *atomic.Bool |
| 40 | } |
| 41 | |
| 42 | type previewSuccessFailWriterTool struct { |
| 43 | name string |
| 44 | calls *int32 |
| 45 | } |
| 46 | |
| 47 | func (f previewSuccessFailWriterTool) Name() string { return f.name } |
| 48 | func (f previewSuccessFailWriterTool) Description() string { |
| 49 | return "previews successfully but cannot write" |
| 50 | } |
| 51 | func (f previewSuccessFailWriterTool) Schema() json.RawMessage { |
| 52 | return json.RawMessage(`{"type":"object"}`) |
| 53 | } |
| 54 | func (f previewSuccessFailWriterTool) ReadOnly() bool { return false } |
| 55 | func (f previewSuccessFailWriterTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 56 | if f.calls != nil { |
| 57 | atomic.AddInt32(f.calls, 1) |
| 58 | } |
| 59 | return "", errors.New("write prompt.txt: permission denied") |
| 60 | } |
| 61 | func (f previewSuccessFailWriterTool) Preview(json.RawMessage) (diff.Change, error) { |
| 62 | return diff.Change{Path: "prompt.txt"}, nil |
| 63 | } |
| 64 | |
| 65 | func (f stateAwareWriterTool) Name() string { return f.name } |
| 66 | func (f stateAwareWriterTool) Description() string { return "fails until target state changes" } |
| 67 | func (f stateAwareWriterTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 68 | func (f stateAwareWriterTool) ReadOnly() bool { return false } |
| 69 | func (f stateAwareWriterTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 70 | if f.calls != nil { |
| 71 | atomic.AddInt32(f.calls, 1) |
| 72 | } |
| 73 | if f.valid != nil && f.valid.Load() { |
| 74 | return "edited prompt.txt", nil |
| 75 | } |
| 76 | return "", errors.New("old_string not found in prompt.txt") |
| 77 | } |
| 78 | func (f stateAwareWriterTool) Preview(json.RawMessage) (diff.Change, error) { |
| 79 | if f.valid != nil && f.valid.Load() { |
| 80 | return diff.Change{Path: "prompt.txt"}, nil |
| 81 | } |
| 82 | return diff.Change{}, errors.New("old_string not found in prompt.txt") |
| 83 | } |
| 84 | |
| 85 | func TestRepeatGuardBlocksRepeatedSuccessfulBashFileWrite(t *testing.T) { |
| 86 | var calls int32 |
| 87 | reg := tool.NewRegistry() |
| 88 | reg.Add(fakeTool{name: "bash", readOnly: false, calls: &calls}) |
| 89 | args := `{"command":"python -c \"with open('prompt.txt', 'w') as f: f.write('hello')\""}` |
| 90 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 91 | {toolCallChunk("c1", "bash", args), {Type: provider.ChunkDone}}, |
| 92 | {toolCallChunk("c2", "bash", args), {Type: provider.ChunkDone}}, |
| 93 | {toolCallChunk("c3", "bash", args), {Type: provider.ChunkDone}}, |
| 94 | {{Type: provider.ChunkText, Text: "done"}, {Type: provider.ChunkDone}}, |
| 95 | }} |
| 96 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 97 | |
| 98 | if err := a.Run(context.Background(), "update the prompt file"); err != nil { |
| 99 | t.Fatalf("Run: %v", err) |
| 100 | } |
| 101 | if got := atomic.LoadInt32(&calls); got != 2 { |
| 102 | t.Fatalf("bash executed %d times, want 2 before the repeat guard blocks", got) |
| 103 | } |
| 104 | results := toolResults(a.session, "bash") |
| 105 | if len(results) != 3 { |
| 106 | t.Fatalf("tool results = %d, want 3", len(results)) |
| 107 | } |
| 108 | last := results[len(results)-1] |
| 109 | if !strings.Contains(last, "[loop guard]") || !strings.Contains(last, "edit_file") { |
| 110 | t.Fatalf("third repeated write should nudge the model to change tools, got %q", last) |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | func TestRepeatGuardAllowsRepeatedNonWritingBashCommand(t *testing.T) { |
| 115 | var calls int32 |
| 116 | reg := tool.NewRegistry() |
| 117 | reg.Add(fakeTool{name: "bash", readOnly: false, calls: &calls}) |
| 118 | args := `{"command":"go test ./internal/agent"}` |
| 119 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 120 | {toolCallChunk("c1", "bash", args), {Type: provider.ChunkDone}}, |
| 121 | {toolCallChunk("c2", "bash", args), {Type: provider.ChunkDone}}, |
| 122 | {toolCallChunk("c3", "bash", args), {Type: provider.ChunkDone}}, |
| 123 | {{Type: provider.ChunkText, Text: "done"}, {Type: provider.ChunkDone}}, |
| 124 | }} |
| 125 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 126 | |
| 127 | if err := a.Run(context.Background(), "verify repeatedly"); err != nil { |
| 128 | t.Fatalf("Run: %v", err) |
| 129 | } |
| 130 | if got := atomic.LoadInt32(&calls); got != 3 { |
| 131 | t.Fatalf("bash executed %d times, want 3 for non-writing commands", got) |
| 132 | } |
| 133 | if last := lastToolResult(a.session, "bash"); strings.Contains(last, "[loop guard]") { |
| 134 | t.Fatalf("non-writing bash should not trip the repeat guard, got %q", last) |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | func TestRepeatGuardBashWriteRedirectDetectionUsesAST(t *testing.T) { |
| 139 | tests := []struct { |
| 140 | name string |
| 141 | cmd string |
| 142 | want bool |
| 143 | }{ |
| 144 | { |
| 145 | name: "stdout file redirect", |
| 146 | cmd: "printf hi > prompt.txt", |
| 147 | want: true, |
| 148 | }, |
| 149 | { |
| 150 | name: "stderr file redirect", |
| 151 | cmd: "printf hi 2>err.log", |
| 152 | want: true, |
| 153 | }, |
| 154 | { |
| 155 | name: "append redirect", |
| 156 | cmd: "printf hi >> prompt.txt", |
| 157 | want: true, |
| 158 | }, |
| 159 | { |
| 160 | name: "combined redirect", |
| 161 | cmd: "printf hi &> prompt.txt", |
| 162 | want: true, |
| 163 | }, |
| 164 | { |
| 165 | name: "read write redirect", |
| 166 | cmd: "cat <> prompt.txt", |
| 167 | want: true, |
| 168 | }, |
| 169 | { |
| 170 | name: "null sink redirect", |
| 171 | cmd: "printf hi >/dev/null", |
| 172 | want: false, |
| 173 | }, |
| 174 | { |
| 175 | name: "powershell null sink spelling", |
| 176 | cmd: "printf hi >$null", |
| 177 | want: false, |
| 178 | }, |
| 179 | { |
| 180 | name: "windows nul sink spelling", |
| 181 | cmd: "printf hi >NUL", |
| 182 | want: false, |
| 183 | }, |
| 184 | { |
| 185 | name: "fd duplication", |
| 186 | cmd: "printf hi 2>&1", |
| 187 | want: false, |
| 188 | }, |
| 189 | { |
| 190 | name: "quoted redirect text", |
| 191 | cmd: `printf '%s\n' 'a > b'`, |
| 192 | want: false, |
| 193 | }, |
| 194 | { |
| 195 | name: "heredoc body redirect text", |
| 196 | cmd: "cat <<'EOF'\n> prompt.txt\nEOF", |
| 197 | want: false, |
| 198 | }, |
| 199 | { |
| 200 | name: "heredoc with file redirect", |
| 201 | cmd: "cat <<'EOF' > prompt.txt\nbody\nEOF", |
| 202 | want: true, |
| 203 | }, |
| 204 | } |
| 205 | |
| 206 | for _, tt := range tests { |
| 207 | t.Run(tt.name, func(t *testing.T) { |
| 208 | if got := isShellFileWriteCommand(tt.cmd); got != tt.want { |
| 209 | t.Fatalf("isShellFileWriteCommand(%q) = %v, want %v", tt.cmd, got, tt.want) |
| 210 | } |
| 211 | }) |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | func TestRepeatGuardNormalizesStaticBashFields(t *testing.T) { |
| 216 | singleQuoted := normalizeShellCommand(`printf '%s\n' 'hello world'`) |
| 217 | doubleQuoted := normalizeShellCommand(`printf "%s\n" "hello world"`) |
| 218 | if singleQuoted != doubleQuoted { |
| 219 | t.Fatalf("normalized quote styles differ:\n single: %q\n double: %q", singleQuoted, doubleQuoted) |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | func TestRepeatGuardAllowsTwoRepeatedWriterSuccesses(t *testing.T) { |
| 224 | var calls int32 |
| 225 | reg := tool.NewRegistry() |
| 226 | reg.Add(fakeTool{name: "write_file", readOnly: false, calls: &calls}) |
| 227 | args := `{"path":"prompt.txt","content":"hello"}` |
| 228 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 229 | {toolCallChunk("c1", "write_file", args), {Type: provider.ChunkDone}}, |
| 230 | {toolCallChunk("c2", "write_file", args), {Type: provider.ChunkDone}}, |
| 231 | {{Type: provider.ChunkText, Text: "done"}, {Type: provider.ChunkDone}}, |
| 232 | }} |
| 233 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 234 | |
| 235 | if err := a.Run(context.Background(), "write twice"); err != nil { |
| 236 | t.Fatalf("Run: %v", err) |
| 237 | } |
| 238 | if got := atomic.LoadInt32(&calls); got != 2 { |
| 239 | t.Fatalf("writer executed %d times, want 2 before the guard threshold", got) |
| 240 | } |
| 241 | if last := lastToolResult(a.session, "write_file"); strings.Contains(last, "[loop guard]") { |
| 242 | t.Fatalf("second repeated writer call should still be allowed, got %q", last) |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | func TestRepeatGuardBlocksStaleEditLoopAcrossSuccessfulReads(t *testing.T) { |
| 247 | var editCalls int32 |
| 248 | var readCalls int32 |
| 249 | reg := tool.NewRegistry() |
| 250 | reg.Add(failingWriterTool{name: "edit_file", calls: &editCalls}) |
| 251 | reg.Add(fakeTool{name: "read_file", readOnly: true, calls: &readCalls}) |
| 252 | editArgs1 := `{"path":"prompt.txt","old_string":"stale","new_string":"ready-v1"}` |
| 253 | editArgs2 := `{"path":"prompt.txt","old_string":"stale","new_string":"ready-v2"}` |
| 254 | editArgs3 := `{"path":"prompt.txt","old_string":"stale","new_string":"ready-v3"}` |
| 255 | readArgs := `{"path":"prompt.txt"}` |
| 256 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 257 | {toolCallChunk("e1", "edit_file", editArgs1), {Type: provider.ChunkDone}}, |
| 258 | {toolCallChunk("r1", "read_file", readArgs), {Type: provider.ChunkDone}}, |
| 259 | {toolCallChunk("e2", "edit_file", editArgs2), {Type: provider.ChunkDone}}, |
| 260 | {toolCallChunk("r2", "read_file", readArgs), {Type: provider.ChunkDone}}, |
| 261 | {toolCallChunk("e3", "edit_file", editArgs3), {Type: provider.ChunkDone}}, |
| 262 | {{Type: provider.ChunkText, Text: "blocked"}, {Type: provider.ChunkDone}}, |
| 263 | }} |
| 264 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 265 | |
| 266 | if err := a.Run(context.Background(), "fix prompt.txt"); err != nil { |
| 267 | t.Fatalf("Run: %v", err) |
| 268 | } |
| 269 | if got := atomic.LoadInt32(&editCalls); got != 2 { |
| 270 | t.Fatalf("edit_file executed %d times, want 2 before the repeat guard blocks", got) |
| 271 | } |
| 272 | if got := atomic.LoadInt32(&readCalls); got != 2 { |
| 273 | t.Fatalf("read_file executed %d times, want 2", got) |
| 274 | } |
| 275 | last := lastToolResult(a.session, "edit_file") |
| 276 | for _, want := range []string{"[loop guard]", "already failed 2 times", "Re-reading alone"} { |
| 277 | if !strings.Contains(last, want) { |
| 278 | t.Fatalf("blocked stale edit result should mention %q, got %q", want, last) |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | |
| 283 | func TestRepeatGuardRetainsStaleEditFailuresAcrossGoalScope(t *testing.T) { |
| 284 | var editCalls int32 |
| 285 | reg := tool.NewRegistry() |
| 286 | reg.Add(failingWriterTool{name: "edit_file", calls: &editCalls}) |
| 287 | editArgs := `{"path":"prompt.txt","old_string":"stale","new_string":"ready"}` |
| 288 | readArgs := `{"path":"prompt.txt"}` |
| 289 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 290 | {toolCallChunk("e1", "edit_file", editArgs), {Type: provider.ChunkDone}}, |
| 291 | {toolCallChunk("r1", "read_file", readArgs), {Type: provider.ChunkDone}}, |
| 292 | {{Type: provider.ChunkText, Text: "[goal:continue]"}, {Type: provider.ChunkDone}}, |
| 293 | {toolCallChunk("e2", "edit_file", editArgs), {Type: provider.ChunkDone}}, |
| 294 | {toolCallChunk("r2", "read_file", readArgs), {Type: provider.ChunkDone}}, |
| 295 | {{Type: provider.ChunkText, Text: "[goal:continue]"}, {Type: provider.ChunkDone}}, |
| 296 | {toolCallChunk("e3", "edit_file", editArgs), {Type: provider.ChunkDone}}, |
| 297 | {{Type: provider.ChunkText, Text: "[goal:blocked:stale edit]"}, {Type: provider.ChunkDone}}, |
| 298 | }} |
| 299 | reg.Add(fakeTool{name: "read_file", readOnly: true}) |
| 300 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 301 | ctx := WithDeliveryExecutionScope(context.Background(), DeliveryExecutionScope{ |
| 302 | ID: "goal-scope-1", |
| 303 | TaskText: "fix prompt.txt", |
| 304 | }) |
| 305 | |
| 306 | for i := 0; i < 3; i++ { |
| 307 | if err := a.Run(ctx, "continue goal"); err != nil { |
| 308 | t.Fatalf("Run %d: %v", i+1, err) |
| 309 | } |
| 310 | } |
| 311 | if got := atomic.LoadInt32(&editCalls); got != 2 { |
| 312 | t.Fatalf("edit_file executed %d times across one goal scope, want 2", got) |
| 313 | } |
| 314 | if last := lastToolResult(a.session, "edit_file"); !strings.Contains(last, "[loop guard]") { |
| 315 | t.Fatalf("third goal-scope edit should be blocked, got %q", last) |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | func TestRepeatGuardClearsOrdinaryWriteFailureAcrossGoalRuns(t *testing.T) { |
| 320 | var editCalls int32 |
| 321 | reg := tool.NewRegistry() |
| 322 | reg.Add(previewSuccessFailWriterTool{name: "edit_file", calls: &editCalls}) |
| 323 | editArgs := `{"path":"prompt.txt","old_string":"current","new_string":"ready"}` |
| 324 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 325 | {toolCallChunk("e1", "edit_file", editArgs), {Type: provider.ChunkDone}}, |
| 326 | {{Type: provider.ChunkText, Text: "[goal:continue]"}, {Type: provider.ChunkDone}}, |
| 327 | {toolCallChunk("e2", "edit_file", editArgs), {Type: provider.ChunkDone}}, |
| 328 | {{Type: provider.ChunkText, Text: "[goal:continue]"}, {Type: provider.ChunkDone}}, |
| 329 | {toolCallChunk("e3", "edit_file", editArgs), {Type: provider.ChunkDone}}, |
| 330 | {{Type: provider.ChunkText, Text: "[goal:blocked:permission denied]"}, {Type: provider.ChunkDone}}, |
| 331 | }} |
| 332 | a := New(prov, reg, NewSession(""), Options{}, event.Discard) |
| 333 | ctx := WithDeliveryExecutionScope(context.Background(), DeliveryExecutionScope{ |
| 334 | ID: "goal-scope-1", |
| 335 | TaskText: "fix prompt.txt", |
| 336 | }) |
| 337 | |
| 338 | for i := 0; i < 3; i++ { |
| 339 | if err := a.Run(ctx, "continue goal"); err != nil { |
| 340 | t.Fatalf("Run %d: %v", i+1, err) |
| 341 | } |
| 342 | } |
| 343 | if got := atomic.LoadInt32(&editCalls); got != 3 { |
| 344 | t.Fatalf("edit_file executed %d times across Goal Runs, want ordinary write failure retried each Run", got) |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | func TestRepeatGuardDoesNotUsePreviewToClearWriteFailure(t *testing.T) { |
| 349 | var editCalls int32 |
| 350 | reg := tool.NewRegistry() |
| 351 | reg.Add(previewSuccessFailWriterTool{name: "edit_file", calls: &editCalls}) |
| 352 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 353 | ctx := context.Background() |
| 354 | edit := provider.ToolCall{Name: "edit_file", Arguments: `{"path":"prompt.txt","old_string":"current","new_string":"ready"}`} |
| 355 | |
| 356 | executeBatchOutputs(a, ctx, []provider.ToolCall{edit}) |
| 357 | executeBatchOutputs(a, ctx, []provider.ToolCall{edit}) |
| 358 | last := executeBatchOutputs(a, ctx, []provider.ToolCall{edit})[0] |
| 359 | |
| 360 | if !strings.Contains(last, "[loop guard]") { |
| 361 | t.Fatalf("successful preview must not clear a repeated write failure, got %q", last) |
| 362 | } |
| 363 | for _, staleHint := range []string{"stale anchor", "Re-reading alone", "new old_string"} { |
| 364 | if strings.Contains(last, staleHint) { |
| 365 | t.Fatalf("ordinary write failure should not use stale-anchor guidance %q, got %q", staleHint, last) |
| 366 | } |
| 367 | } |
| 368 | if got := atomic.LoadInt32(&editCalls); got != 2 { |
| 369 | t.Fatalf("edit_file executed %d times, want write failure blocked before third execution", got) |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | func TestRepeatGuardKeepsStaleFailureAfterUnrelatedMutation(t *testing.T) { |
| 374 | var editCalls int32 |
| 375 | reg := tool.NewRegistry() |
| 376 | reg.Add(failingWriterTool{name: "edit_file", calls: &editCalls}) |
| 377 | reg.Add(fakeTool{name: "write_file", readOnly: false}) |
| 378 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 379 | ctx := context.Background() |
| 380 | edit := provider.ToolCall{Name: "edit_file", Arguments: `{"path":"prompt.txt","old_string":"stale","new_string":"ready"}`} |
| 381 | |
| 382 | executeBatchOutputs(a, ctx, []provider.ToolCall{edit}) |
| 383 | executeBatchOutputs(a, ctx, []provider.ToolCall{edit}) |
| 384 | executeBatchOutputs(a, ctx, []provider.ToolCall{{ |
| 385 | Name: "write_file", Arguments: `{"path":"other.txt","content":"unrelated"}`, |
| 386 | }}) |
| 387 | last := executeBatchOutputs(a, ctx, []provider.ToolCall{edit})[0] |
| 388 | |
| 389 | if !strings.Contains(last, "[loop guard]") { |
| 390 | t.Fatalf("unrelated mutation should not clear stale failure history, got %q", last) |
| 391 | } |
| 392 | if got := atomic.LoadInt32(&editCalls); got != 2 { |
| 393 | t.Fatalf("edit_file executed %d times, want unrelated mutation to preserve the guard", got) |
| 394 | } |
| 395 | } |
| 396 | |
| 397 | func TestRepeatGuardKeepsStaleFailureAfterSameFileMutation(t *testing.T) { |
| 398 | dir := t.TempDir() |
| 399 | path := filepath.Join(dir, "prompt.txt") |
| 400 | if err := os.WriteFile(path, []byte("status=ready\n"), 0o600); err != nil { |
| 401 | t.Fatal(err) |
| 402 | } |
| 403 | reg := tool.NewRegistry() |
| 404 | for _, tl := range (builtin.Workspace{Dir: dir}).Tools("edit_file") { |
| 405 | reg.Add(tl) |
| 406 | } |
| 407 | a := New(nil, reg, NewSession(""), Options{WriteWorkspaceRoot: dir}, event.Discard) |
| 408 | ctx := context.Background() |
| 409 | stale := provider.ToolCall{ |
| 410 | Name: "edit_file", |
| 411 | Arguments: `{"path":"prompt.txt","old_string":"status=stale","new_string":"status=fixed"}`, |
| 412 | } |
| 413 | |
| 414 | executeBatchOutputs(a, ctx, []provider.ToolCall{stale}) |
| 415 | executeBatchOutputs(a, ctx, []provider.ToolCall{stale}) |
| 416 | executeBatchOutputs(a, ctx, []provider.ToolCall{{ |
| 417 | Name: "edit_file", |
| 418 | Arguments: `{"path":"prompt.txt","old_string":"status=ready","new_string":"status=done"}`, |
| 419 | }}) |
| 420 | last := executeBatchOutputs(a, ctx, []provider.ToolCall{stale})[0] |
| 421 | |
| 422 | if !strings.Contains(last, "[loop guard]") { |
| 423 | t.Fatalf("same-file mutation must not renew a still-stale anchor budget, got %q", last) |
| 424 | } |
| 425 | data, err := os.ReadFile(path) |
| 426 | if err != nil { |
| 427 | t.Fatal(err) |
| 428 | } |
| 429 | if got := string(data); got != "status=done\n" { |
| 430 | t.Fatalf("file content = %q, want successful unrelated edit preserved", got) |
| 431 | } |
| 432 | } |
| 433 | |
| 434 | func TestRepeatGuardNormalizesFailureTargetPaths(t *testing.T) { |
| 435 | var editCalls int32 |
| 436 | dir := t.TempDir() |
| 437 | reg := tool.NewRegistry() |
| 438 | reg.Add(failingWriterTool{name: "edit_file", calls: &editCalls}) |
| 439 | a := New(nil, reg, NewSession(""), Options{WriteWorkspaceRoot: dir}, event.Discard) |
| 440 | ctx := context.Background() |
| 441 | args := []string{ |
| 442 | `{"path":"prompt.txt","old_string":"stale","new_string":"ready-v1"}`, |
| 443 | `{"path":"./prompt.txt","old_string":"stale","new_string":"ready-v2"}`, |
| 444 | `{"path":` + string(mustJSON(t, filepath.Join(dir, "prompt.txt"))) + `,"old_string":"stale","new_string":"ready-v3"}`, |
| 445 | } |
| 446 | |
| 447 | executeBatchOutputs(a, ctx, []provider.ToolCall{{Name: "edit_file", Arguments: args[0]}}) |
| 448 | executeBatchOutputs(a, ctx, []provider.ToolCall{{Name: "edit_file", Arguments: args[1]}}) |
| 449 | last := executeBatchOutputs(a, ctx, []provider.ToolCall{{Name: "edit_file", Arguments: args[2]}})[0] |
| 450 | |
| 451 | if !strings.Contains(last, "[loop guard]") { |
| 452 | t.Fatalf("path aliases should share one repeated-failure signature, got %q", last) |
| 453 | } |
| 454 | if got := atomic.LoadInt32(&editCalls); got != 2 { |
| 455 | t.Fatalf("edit_file executed %d times, want absolute-path retry blocked", got) |
| 456 | } |
| 457 | } |
| 458 | |
| 459 | func TestRepeatGuardAllowsRetryAfterExternalTargetChange(t *testing.T) { |
| 460 | var editCalls int32 |
| 461 | var valid atomic.Bool |
| 462 | reg := tool.NewRegistry() |
| 463 | reg.Add(stateAwareWriterTool{name: "edit_file", calls: &editCalls, valid: &valid}) |
| 464 | a := New(nil, reg, NewSession(""), Options{}, event.Discard) |
| 465 | ctx := context.Background() |
| 466 | edit := provider.ToolCall{Name: "edit_file", Arguments: `{"path":"prompt.txt","old_string":"stale","new_string":"ready"}`} |
| 467 | |
| 468 | executeBatchOutputs(a, ctx, []provider.ToolCall{edit}) |
| 469 | executeBatchOutputs(a, ctx, []provider.ToolCall{edit}) |
| 470 | valid.Store(true) |
| 471 | last := executeBatchOutputs(a, ctx, []provider.ToolCall{edit})[0] |
| 472 | |
| 473 | if strings.Contains(last, "[loop guard]") { |
| 474 | t.Fatalf("changed target state should allow the retry, got %q", last) |
| 475 | } |
| 476 | if got := atomic.LoadInt32(&editCalls); got != 3 { |
| 477 | t.Fatalf("edit_file executed %d times, want retry after external target change", got) |
| 478 | } |
| 479 | } |
| 480 |