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