| 1 | package boot |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "context" |
| 6 | "errors" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "slices" |
| 10 | "strings" |
| 11 | "testing" |
| 12 | |
| 13 | "reasonix/internal/agent" |
| 14 | "reasonix/internal/control" |
| 15 | "reasonix/internal/provider" |
| 16 | "reasonix/internal/session" |
| 17 | ) |
| 18 | |
| 19 | func TestBuildRuntimeDisablesImplicitSkillInvocation(t *testing.T) { |
| 20 | isolateConfigHome(t) |
| 21 | dir := robustTempDir(t) |
| 22 | t.Chdir(dir) |
| 23 | writeRuntimeFixture(t, dir) |
| 24 | configPath := filepath.Join(dir, "reasonix.toml") |
| 25 | content, err := os.ReadFile(configPath) |
| 26 | if err != nil { |
| 27 | t.Fatalf("read fixture config: %v", err) |
| 28 | } |
| 29 | content = append(content, []byte("\n[skills]\ndisable_implicit_invocation = true\n")...) |
| 30 | if err := os.WriteFile(configPath, content, 0o644); err != nil { |
| 31 | t.Fatalf("write skills config: %v", err) |
| 32 | } |
| 33 | res := buildRuntimeFixture(t) |
| 34 | if res.Controller.ImplicitSkillInvocationEnabled() { |
| 35 | t.Fatal("controller should disable implicit skill invocation") |
| 36 | } |
| 37 | if res.Assembly == nil || res.Assembly.ImplicitSkillInvocation { |
| 38 | t.Fatal("reused assembly should record implicit skill invocation as disabled") |
| 39 | } |
| 40 | if strings.Contains(res.Snapshot.SystemPrompt(), "One-liner index") { |
| 41 | t.Fatal("skill index should not be provider-visible when implicit invocation is disabled") |
| 42 | } |
| 43 | for _, entry := range res.Controller.ToolContractEntries() { |
| 44 | switch entry.Name { |
| 45 | case "run_skill", "read_skill", "read_only_skill", "install_skill": |
| 46 | t.Fatalf("skill tool %q should not be exposed to the model", entry.Name) |
| 47 | } |
| 48 | } |
| 49 | if _, ok := res.Controller.RunSkill("/explore inspect"); !ok { |
| 50 | t.Fatal("explicit /skill invocation should remain available") |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | func TestRebuildFromForceFullRebuildRefreshesSkillPolicy(t *testing.T) { |
| 55 | isolateConfigHome(t) |
| 56 | dir := robustTempDir(t) |
| 57 | t.Chdir(dir) |
| 58 | writeRuntimeFixture(t, dir) |
| 59 | configPath := filepath.Join(dir, "reasonix.toml") |
| 60 | content, err := os.ReadFile(configPath) |
| 61 | if err != nil { |
| 62 | t.Fatalf("read fixture config: %v", err) |
| 63 | } |
| 64 | content = append(content, []byte("\n[skills]\ndisable_implicit_invocation = true\n")...) |
| 65 | if err := os.WriteFile(configPath, content, 0o644); err != nil { |
| 66 | t.Fatalf("write disabled skill policy: %v", err) |
| 67 | } |
| 68 | previous, err := BuildRuntime(context.Background(), Options{}) |
| 69 | if err != nil { |
| 70 | t.Fatalf("initial BuildRuntime: %v", err) |
| 71 | } |
| 72 | if previous.Controller.ImplicitSkillInvocationEnabled() { |
| 73 | t.Fatal("initial controller should disable implicit skill invocation") |
| 74 | } |
| 75 | content = bytes.Replace(content, []byte("disable_implicit_invocation = true"), []byte("disable_implicit_invocation = false"), 1) |
| 76 | if err := os.WriteFile(configPath, content, 0o644); err != nil { |
| 77 | t.Fatalf("write enabled skill policy: %v", err) |
| 78 | } |
| 79 | res, err := RebuildFrom(context.Background(), previous, Options{ |
| 80 | RuntimeReload: RuntimeReload{ForceFullRebuild: true}, |
| 81 | }) |
| 82 | if err != nil { |
| 83 | previous.Controller.Close() |
| 84 | t.Fatalf("forced RebuildFrom: %v", err) |
| 85 | } |
| 86 | t.Cleanup(func() { |
| 87 | res.Controller.Close() |
| 88 | }) |
| 89 | if res.Controller == previous.Controller { |
| 90 | t.Fatal("forced rebuild reused the previous controller") |
| 91 | } |
| 92 | if !res.Controller.ImplicitSkillInvocationEnabled() { |
| 93 | t.Fatal("forced rebuild did not refresh the enabled skill policy") |
| 94 | } |
| 95 | previous.Controller.Close() |
| 96 | } |
| 97 | |
| 98 | // writeRuntimeFixture writes the minimal deterministic config the runtime |
| 99 | // tests share: a resolvable model, a fixed base system prompt, and the |
| 100 | // environment probe section disabled (it embeds machine-specific data). |
| 101 | func writeRuntimeFixture(t *testing.T, dir string) { |
| 102 | t.Helper() |
| 103 | writeFile(t, dir, "reasonix.toml", ` |
| 104 | default_model = "test-model" |
| 105 | |
| 106 | [agent] |
| 107 | system_prompt = "BASE SYSTEM PROMPT" |
| 108 | |
| 109 | [environment] |
| 110 | enabled = false |
| 111 | |
| 112 | [[providers]] |
| 113 | name = "test-model" |
| 114 | kind = "openai" |
| 115 | base_url = "https://example.invalid" |
| 116 | model = "x" |
| 117 | api_key_env = "REASONIX_TEST_KEY_UNSET" |
| 118 | `) |
| 119 | approveWorkspace(t, dir) |
| 120 | } |
| 121 | |
| 122 | // buildRuntimeFixture builds one runtime against the fixture and registers |
| 123 | // its controller for cleanup. |
| 124 | func buildRuntimeFixture(t *testing.T) *BuildResult { |
| 125 | t.Helper() |
| 126 | res, err := BuildRuntime(context.Background(), Options{}) |
| 127 | if err != nil { |
| 128 | t.Fatalf("BuildRuntime: %v", err) |
| 129 | } |
| 130 | if res.Controller == nil { |
| 131 | t.Fatal("BuildRuntime returned a nil controller") |
| 132 | } |
| 133 | t.Cleanup(res.Controller.Close) |
| 134 | return res |
| 135 | } |
| 136 | |
| 137 | // TestBuildRuntimeSnapshotMatchesController pins the stage-3a contract: the |
| 138 | // kernel snapshot mirrors exactly what the build wired — same system prompt, |
| 139 | // same provider-visible tool contract — its cache fingerprint is stable |
| 140 | // across identical builds, and generations increase monotonically. |
| 141 | func TestBuildRuntimeSnapshotMatchesController(t *testing.T) { |
| 142 | isolateConfigHome(t) |
| 143 | dir := robustTempDir(t) |
| 144 | t.Chdir(dir) |
| 145 | writeRuntimeFixture(t, dir) |
| 146 | |
| 147 | first := buildRuntimeFixture(t) |
| 148 | if first.Snapshot == nil { |
| 149 | t.Fatal("BuildRuntime returned a nil snapshot") |
| 150 | } |
| 151 | if first.Runtime == nil { |
| 152 | t.Fatal("BuildRuntime returned a nil runtime set") |
| 153 | } |
| 154 | if first.Runtime.Len() != 0 { |
| 155 | t.Fatalf("stage-3a runtime set holds %d closers, want 0 (sidecars arrive in stage 5)", first.Runtime.Len()) |
| 156 | } |
| 157 | if first.Snapshot.Generation() == 0 { |
| 158 | t.Fatal("snapshot generation = 0, want the counter to start at 1") |
| 159 | } |
| 160 | |
| 161 | // The snapshot's system prompt is exactly the controller's system |
| 162 | // message. |
| 163 | if got, want := first.Snapshot.SystemPrompt(), systemMessage(first.Controller.History()); got != want { |
| 164 | t.Fatalf("snapshot system prompt != controller system message\n got: %q\nwant: %q", got, want) |
| 165 | } |
| 166 | |
| 167 | // The snapshot's tool schemas are exactly the controller's tool contract, |
| 168 | // entry for entry. |
| 169 | entries := first.Controller.ToolContractEntries() |
| 170 | schemas := first.Snapshot.ToolSchemas() |
| 171 | if len(entries) == 0 { |
| 172 | t.Fatal("BuildRuntime registered no tools") |
| 173 | } |
| 174 | if len(schemas) != len(entries) { |
| 175 | t.Fatalf("snapshot holds %d tool schemas, controller contract has %d", len(schemas), len(entries)) |
| 176 | } |
| 177 | for i, e := range entries { |
| 178 | s := schemas[i] |
| 179 | if s.Name != e.Name || s.Description != e.Description || string(s.Parameters) != string(e.Schema) { |
| 180 | t.Fatalf("tool schema %d = (%q, %.40q, %.40q), want (%q, %.40q, %.40q)", |
| 181 | i, s.Name, s.Description, s.Parameters, e.Name, e.Description, e.Schema) |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | // A clean fixture records no diagnostics. |
| 186 | if diags := first.Snapshot.Diagnostics(); len(diags) != 0 { |
| 187 | t.Fatalf("snapshot diagnostics = %v, want none", diags) |
| 188 | } |
| 189 | |
| 190 | // An identical second build reproduces the same CacheHash (the |
| 191 | // provider-cache fingerprint) at a higher generation. |
| 192 | second := buildRuntimeFixture(t) |
| 193 | if second.Snapshot == nil { |
| 194 | t.Fatal("second BuildRuntime returned a nil snapshot") |
| 195 | } |
| 196 | if got, want := second.Snapshot.CacheHash(), first.Snapshot.CacheHash(); got != want { |
| 197 | t.Fatalf("CacheHash drifted across identical builds: %s vs %s", got, want) |
| 198 | } |
| 199 | if got, before := second.Snapshot.Generation(), first.Snapshot.Generation(); got <= before { |
| 200 | t.Fatalf("generation did not increase across builds: %d then %d", before, got) |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | // TestRebuildMigratesSessionState drives the success path: an old controller |
| 205 | // with a conversation, session grants, and the session axes set rebuilds into |
| 206 | // a replacement that continues the same session file with everything carried |
| 207 | // — while the old controller stays fully usable. |
| 208 | func TestRebuildMigratesSessionState(t *testing.T) { |
| 209 | isolateConfigHome(t) |
| 210 | dir := robustTempDir(t) |
| 211 | t.Chdir(dir) |
| 212 | writeRuntimeFixture(t, dir) |
| 213 | |
| 214 | old, err := BuildRuntime(context.Background(), withTestSession(t, Options{})) |
| 215 | if err != nil { |
| 216 | t.Fatalf("BuildRuntime v3: %v", err) |
| 217 | } |
| 218 | t.Cleanup(old.Controller.Close) |
| 219 | oldCtrl := old.Controller |
| 220 | |
| 221 | // Pin a v3 session and seed a conversation plus the session axes the |
| 222 | // rebuild must carry. |
| 223 | oldCtrl.EnsureSessionPath() |
| 224 | prevRef, ok := oldCtrl.SessionRef() |
| 225 | if !ok { |
| 226 | t.Fatal("old controller pinned no v3 session") |
| 227 | } |
| 228 | oldCtrl.AdoptHistory([]provider.Message{ |
| 229 | {Role: provider.RoleSystem, Content: systemMessage(oldCtrl.History())}, |
| 230 | {Role: provider.RoleUser, Content: "hello"}, |
| 231 | {Role: provider.RoleAssistant, Content: "hi there"}, |
| 232 | }, "") |
| 233 | oldCtrl.SetToolApprovalMode(control.ToolApprovalYolo) |
| 234 | oldCtrl.SetPlanMode(true) |
| 235 | oldCtrl.SetGoal("ship the kernel") |
| 236 | oldCtrl.RestoreSessionAuthorizations(control.SessionAuthorizations{ |
| 237 | Grants: []string{"bash(go test ./...)"}, |
| 238 | PlanModeReadOnlyCommands: []string{"git status"}, |
| 239 | }) |
| 240 | oldHistory := oldCtrl.History() |
| 241 | |
| 242 | res, err := Rebuild(context.Background(), oldCtrl, Options{}) |
| 243 | if err != nil { |
| 244 | t.Fatalf("Rebuild: %v", err) |
| 245 | } |
| 246 | if res.Snapshot == nil { |
| 247 | t.Fatal("Rebuild returned a nil snapshot") |
| 248 | } |
| 249 | if res.Snapshot.Generation() <= old.Snapshot.Generation() { |
| 250 | t.Fatalf("generation did not increase: old %d, new %d", old.Snapshot.Generation(), res.Snapshot.Generation()) |
| 251 | } |
| 252 | defer res.Controller.Close() |
| 253 | |
| 254 | // The conversation continues on the same immutable v3 identity with identical |
| 255 | // messages (the fixture rebuild produces the same system prompt, so the |
| 256 | // splice is invisible here). |
| 257 | gotRef, ok := res.Controller.SessionRef() |
| 258 | if !ok || gotRef != prevRef { |
| 259 | t.Fatalf("session ref = %+v, want continued %+v", gotRef, prevRef) |
| 260 | } |
| 261 | if got := res.Controller.SessionPath(); got != "" { |
| 262 | t.Fatalf("rebuilt v3 controller retained legacy path %q", got) |
| 263 | } |
| 264 | newHistory := res.Controller.History() |
| 265 | if len(newHistory) != len(oldHistory) { |
| 266 | t.Fatalf("new history has %d messages, want %d", len(newHistory), len(oldHistory)) |
| 267 | } |
| 268 | for i := range oldHistory { |
| 269 | if newHistory[i].Role != oldHistory[i].Role || newHistory[i].Content != oldHistory[i].Content { |
| 270 | t.Fatalf("history[%d] = (%s, %q), want (%s, %q)", |
| 271 | i, newHistory[i].Role, newHistory[i].Content, oldHistory[i].Role, oldHistory[i].Content) |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | // Session axes migrated. |
| 276 | if got := res.Controller.ToolApprovalMode(); got != control.ToolApprovalYolo { |
| 277 | t.Fatalf("tool approval mode = %q, want %q", got, control.ToolApprovalYolo) |
| 278 | } |
| 279 | if !res.Controller.PlanMode() { |
| 280 | t.Fatal("plan mode did not migrate") |
| 281 | } |
| 282 | if got := res.Controller.Goal(); got != "ship the kernel" { |
| 283 | t.Fatalf("goal = %q, want migrated %q", got, "ship the kernel") |
| 284 | } |
| 285 | auth := res.Controller.SessionAuthorizations() |
| 286 | if !slices.Contains(auth.Grants, "bash(go test ./...)") { |
| 287 | t.Fatalf("session grants = %v, want the migrated grant", auth.Grants) |
| 288 | } |
| 289 | if !slices.Contains(auth.PlanModeReadOnlyCommands, "git status") { |
| 290 | t.Fatalf("plan-mode trust = %v, want the migrated prefix", auth.PlanModeReadOnlyCommands) |
| 291 | } |
| 292 | |
| 293 | // old keeps working: history intact, runtime set untouched, close clean. |
| 294 | if got := len(oldCtrl.History()); got != len(oldHistory) { |
| 295 | t.Fatalf("old controller history changed during rebuild: %d, want %d", got, len(oldHistory)) |
| 296 | } |
| 297 | if old.Runtime.Closed() { |
| 298 | t.Fatal("Rebuild closed the old runtime set") |
| 299 | } |
| 300 | oldCtrl.Close() |
| 301 | } |
| 302 | |
| 303 | func TestRebuildKeepsLiveReadOnlyAfterStoredRemotePreset(t *testing.T) { |
| 304 | restoreSandbox := control.SetPresetSandboxForTest(true) |
| 305 | t.Cleanup(restoreSandbox) |
| 306 | isolateConfigHome(t) |
| 307 | dir := robustTempDir(t) |
| 308 | t.Chdir(dir) |
| 309 | writeRuntimeFixture(t, dir) |
| 310 | old, err := BuildRuntime(t.Context(), withTestSession(t, Options{})) |
| 311 | if err != nil { |
| 312 | t.Fatal(err) |
| 313 | } |
| 314 | t.Cleanup(old.Controller.Close) |
| 315 | ctrl := old.Controller |
| 316 | ctrl.EnsureSessionPath() |
| 317 | if _, _, err := ctrl.SetSessionPermissionPreset(t.Context(), control.ToolApprovalDangerFullAccess, ctrl.PermissionSnapshot().Revision); err != nil { |
| 318 | t.Fatal(err) |
| 319 | } |
| 320 | ctrl.SetToolApprovalMode(control.ToolApprovalReadOnly) |
| 321 | next, err := Rebuild(t.Context(), ctrl, Options{}) |
| 322 | if err != nil { |
| 323 | t.Fatal(err) |
| 324 | } |
| 325 | t.Cleanup(next.Controller.Close) |
| 326 | if got := next.Controller.ToolApprovalMode(); got != control.ToolApprovalReadOnly { |
| 327 | t.Fatalf("rebuild widened live read-only mode to %q", got) |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | func TestRebuildKeepsLegacySessionNativeWithHostService(t *testing.T) { |
| 332 | isolateConfigHome(t) |
| 333 | dir := robustTempDir(t) |
| 334 | t.Chdir(dir) |
| 335 | writeRuntimeFixture(t, dir) |
| 336 | |
| 337 | old, err := BuildRuntime(context.Background(), Options{}) |
| 338 | if err != nil { |
| 339 | t.Fatalf("BuildRuntime legacy: %v", err) |
| 340 | } |
| 341 | t.Cleanup(old.Controller.Close) |
| 342 | old.Controller.EnsureSessionPath() |
| 343 | old.Controller.AdoptHistory([]provider.Message{ |
| 344 | {Role: provider.RoleSystem, Content: systemMessage(old.Controller.History())}, |
| 345 | {Role: provider.RoleUser, Content: "legacy history"}, |
| 346 | }, old.Controller.SessionPath()) |
| 347 | if err := old.Controller.Snapshot(); err != nil { |
| 348 | t.Fatalf("Snapshot legacy: %v", err) |
| 349 | } |
| 350 | |
| 351 | workspace := filepath.Join(dir, "workspace") |
| 352 | storeRoot := filepath.Join(dir, "desktop-sessions-v5", "by-id") |
| 353 | service, err := session.NewService("local", session.NewFilesystemPersistence(storeRoot)) |
| 354 | if err != nil { |
| 355 | t.Fatalf("NewService: %v", err) |
| 356 | } |
| 357 | t.Cleanup(func() { _ = service.Shutdown(context.Background()) }) |
| 358 | |
| 359 | rebuilt, err := Rebuild(context.Background(), old.Controller, Options{ |
| 360 | SessionService: service, |
| 361 | SessionCreateOptions: session.CreateOptions{ |
| 362 | CWD: workspace, Origin: session.SessionOriginLegacyImport, |
| 363 | }, |
| 364 | }) |
| 365 | if err != nil { |
| 366 | t.Fatalf("Rebuild: %v", err) |
| 367 | } |
| 368 | t.Cleanup(rebuilt.Controller.Close) |
| 369 | if _, ok := rebuilt.Controller.SessionRef(); ok || !rebuilt.Controller.NativeLegacySession() { |
| 370 | t.Fatal("rebuild converted historical storage") |
| 371 | } |
| 372 | if rebuilt.Controller.SessionPath() != old.Controller.SessionPath() || rebuilt.Controller.SessionService() != service { |
| 373 | t.Fatal("rebuild lost native path or current-store service") |
| 374 | } |
| 375 | if got := rebuilt.Controller.History(); len(got) != 2 || got[1].Content != "legacy history" { |
| 376 | t.Fatalf("rebuild lost history: %+v", got) |
| 377 | } |
| 378 | if entries, err := os.ReadDir(storeRoot); err != nil && !os.IsNotExist(err) || len(entries) != 0 { |
| 379 | t.Fatalf("rebuild created canonical history: %v %v", entries, err) |
| 380 | } |
| 381 | other, err := session.NewService("local", session.NewFilesystemPersistence(filepath.Join(dir, "other-root"))) |
| 382 | if err != nil { |
| 383 | t.Fatal(err) |
| 384 | } |
| 385 | t.Cleanup(func() { _ = other.Shutdown(context.Background()) }) |
| 386 | if err := control.ActivateControllerReplacement(old.Controller, rebuilt.Controller); err != nil { |
| 387 | t.Fatal(err) |
| 388 | } |
| 389 | next, err := Rebuild(context.Background(), rebuilt.Controller, Options{SessionService: other}) |
| 390 | if err != nil { |
| 391 | t.Fatal(err) |
| 392 | } |
| 393 | t.Cleanup(next.Controller.Close) |
| 394 | if next.Controller.SessionService() != service || next.Controller.SessionCreationService() != service || next.Controller.SessionPath() != rebuilt.Controller.SessionPath() { |
| 395 | t.Fatal("second rebuild replaced the authoritative store or creation service") |
| 396 | } |
| 397 | } |
| 398 | |
| 399 | func TestNativeRebuildPreservesSelectedDAGHead(t *testing.T) { |
| 400 | isolateConfigHome(t) |
| 401 | dir := robustTempDir(t) |
| 402 | t.Chdir(dir) |
| 403 | writeRuntimeFixture(t, dir) |
| 404 | old := buildRuntimeFixture(t) |
| 405 | t.Cleanup(old.Controller.Close) |
| 406 | path := filepath.Join(dir, "selected.jsonl") |
| 407 | s := agent.NewSession(systemMessage(old.Controller.History())) |
| 408 | s.Add(provider.Message{Role: provider.RoleUser, Content: "question"}) |
| 409 | s.Add(provider.Message{Role: provider.RoleAssistant, Content: "main answer"}) |
| 410 | if err := s.Save(path); err != nil { |
| 411 | t.Fatal(err) |
| 412 | } |
| 413 | if _, err := s.ForkHead(path, s.Snapshot()[1].ID, agent.HeadKindFork, "alternate"); err != nil { |
| 414 | t.Fatal(err) |
| 415 | } |
| 416 | s.Add(provider.Message{Role: provider.RoleAssistant, Content: "alternate answer"}) |
| 417 | if err := s.Save(path); err != nil { |
| 418 | t.Fatal(err) |
| 419 | } |
| 420 | selected, err := agent.LoadSessionHeadReadOnly(path, agent.SessionMainHead) |
| 421 | if err != nil { |
| 422 | t.Fatal(err) |
| 423 | } |
| 424 | if err := old.Controller.ResumeNativeSession(selected, path); err != nil { |
| 425 | t.Fatal(err) |
| 426 | } |
| 427 | rebuilt, err := Rebuild(t.Context(), old.Controller, Options{}) |
| 428 | if err != nil { |
| 429 | t.Fatal(err) |
| 430 | } |
| 431 | t.Cleanup(rebuilt.Controller.Close) |
| 432 | head, ok := rebuilt.Controller.SessionHead() |
| 433 | if !ok || head.HeadID != agent.SessionMainHead { |
| 434 | t.Fatalf("rebuild switched head: %+v", head) |
| 435 | } |
| 436 | if err := rebuilt.Controller.Snapshot(); err != nil { |
| 437 | t.Fatal(err) |
| 438 | } |
| 439 | reopened, err := agent.LoadSessionHeadReadOnly(path, agent.SessionMainHead) |
| 440 | if err != nil { |
| 441 | t.Fatal(err) |
| 442 | } |
| 443 | if history := reopened.Snapshot(); history[len(history)-1].Content != "main answer" { |
| 444 | t.Fatalf("rebuild overwrote selected history: %+v", history) |
| 445 | } |
| 446 | other, err := agent.LoadSession(path) |
| 447 | if err != nil { |
| 448 | t.Fatal(err) |
| 449 | } |
| 450 | if history := other.Snapshot(); history[len(history)-1].Content != "alternate answer" { |
| 451 | t.Fatalf("rebuild changed default head: %+v", history) |
| 452 | } |
| 453 | } |
| 454 | |
| 455 | // TestRebuildCarriesGoalWithoutSessionPath covers the in-memory fallback: an |
| 456 | // old controller that never pinned a session file has no Goal sidecar to |
| 457 | // restore, so the running Goal migrates from memory. |
| 458 | func TestRebuildCarriesGoalWithoutSessionPath(t *testing.T) { |
| 459 | isolateConfigHome(t) |
| 460 | dir := robustTempDir(t) |
| 461 | t.Chdir(dir) |
| 462 | writeRuntimeFixture(t, dir) |
| 463 | |
| 464 | old := buildRuntimeFixture(t) |
| 465 | oldCtrl := old.Controller |
| 466 | if got := oldCtrl.SessionPath(); got != "" { |
| 467 | t.Fatalf("fresh controller session path = %q, want empty", got) |
| 468 | } |
| 469 | oldCtrl.SetGoal("ship the kernel") |
| 470 | |
| 471 | res, err := Rebuild(context.Background(), oldCtrl, Options{}) |
| 472 | if err != nil { |
| 473 | t.Fatalf("Rebuild: %v", err) |
| 474 | } |
| 475 | defer res.Controller.Close() |
| 476 | if got := res.Controller.Goal(); got != "ship the kernel" { |
| 477 | t.Fatalf("goal = %q, want seeded %q", got, "ship the kernel") |
| 478 | } |
| 479 | } |
| 480 | |
| 481 | // TestRebuildFailureKeepsOldController drives the fail-atomic path: the |
| 482 | // replacement build fails (unknown model), the error propagates, and the old |
| 483 | // controller is untouched. |
| 484 | func TestRebuildFailureKeepsOldController(t *testing.T) { |
| 485 | isolateConfigHome(t) |
| 486 | dir := robustTempDir(t) |
| 487 | t.Chdir(dir) |
| 488 | writeRuntimeFixture(t, dir) |
| 489 | |
| 490 | old := buildRuntimeFixture(t) |
| 491 | oldCtrl := old.Controller |
| 492 | oldCtrl.EnsureSessionPath() |
| 493 | prevPath := oldCtrl.SessionPath() |
| 494 | oldCtrl.AdoptHistory([]provider.Message{ |
| 495 | {Role: provider.RoleSystem, Content: systemMessage(oldCtrl.History())}, |
| 496 | {Role: provider.RoleUser, Content: "hello"}, |
| 497 | }, prevPath) |
| 498 | |
| 499 | res, err := Rebuild(context.Background(), oldCtrl, Options{Model: "definitely-unknown-model"}) |
| 500 | if err == nil { |
| 501 | t.Fatal("Rebuild succeeded, want ErrUnknownModel") |
| 502 | } |
| 503 | if !errors.Is(err, ErrUnknownModel) { |
| 504 | t.Fatalf("Rebuild error = %v, want ErrUnknownModel", err) |
| 505 | } |
| 506 | if res != nil { |
| 507 | t.Fatalf("Rebuild returned a partial result on failure: %+v", res) |
| 508 | } |
| 509 | |
| 510 | // old is fully usable: history and path intact, runtime set untouched, |
| 511 | // close clean. |
| 512 | if got := len(oldCtrl.History()); got != 2 { |
| 513 | t.Fatalf("old history = %d messages, want 2", got) |
| 514 | } |
| 515 | if got := oldCtrl.SessionPath(); got != prevPath { |
| 516 | t.Fatalf("old session path = %q, want %q", got, prevPath) |
| 517 | } |
| 518 | if old.Runtime.Closed() { |
| 519 | t.Fatal("Rebuild closed the old runtime set on failure") |
| 520 | } |
| 521 | oldCtrl.Close() |
| 522 | } |
| 523 | |
| 524 | // TestSpliceFreshSystemPrompt pins the system-message splice used to refresh |
| 525 | // the profile contract on a continued conversation. |
| 526 | func TestSpliceFreshSystemPrompt(t *testing.T) { |
| 527 | fresh := []provider.Message{{Role: provider.RoleSystem, Content: "new prompt"}} |
| 528 | t.Run("replaces the carried system message", func(t *testing.T) { |
| 529 | carried := []provider.Message{ |
| 530 | {Role: provider.RoleSystem, Content: "old prompt"}, |
| 531 | {Role: provider.RoleUser, Content: "hi"}, |
| 532 | } |
| 533 | got := spliceFreshSystemPrompt(carried, fresh) |
| 534 | if len(got) != 2 || got[0].Content != "new prompt" || got[1].Content != "hi" { |
| 535 | t.Fatalf("splice = %+v", got) |
| 536 | } |
| 537 | }) |
| 538 | t.Run("prepends when the carried conversation has none", func(t *testing.T) { |
| 539 | carried := []provider.Message{{Role: provider.RoleUser, Content: "hi"}} |
| 540 | got := spliceFreshSystemPrompt(carried, fresh) |
| 541 | if len(got) != 2 || got[0].Role != provider.RoleSystem || got[1].Content != "hi" { |
| 542 | t.Fatalf("splice = %+v", got) |
| 543 | } |
| 544 | }) |
| 545 | t.Run("no fresh system message leaves the conversation untouched", func(t *testing.T) { |
| 546 | carried := []provider.Message{{Role: provider.RoleUser, Content: "hi"}} |
| 547 | got := spliceFreshSystemPrompt(carried, []provider.Message{{Role: provider.RoleUser, Content: "yo"}}) |
| 548 | if len(got) != 1 || got[0].Content != "hi" { |
| 549 | t.Fatalf("splice = %+v", got) |
| 550 | } |
| 551 | }) |
| 552 | t.Run("does not alias the input slice", func(t *testing.T) { |
| 553 | carried := []provider.Message{{Role: provider.RoleSystem, Content: "old prompt"}} |
| 554 | got := spliceFreshSystemPrompt(carried, fresh) |
| 555 | got[0].Content = "mutated" |
| 556 | if carried[0].Content != "old prompt" { |
| 557 | t.Fatal("splice wrote through to the caller's slice") |
| 558 | } |
| 559 | }) |
| 560 | } |
| 561 |