| 1 | package plugin |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "fmt" |
| 7 | "os" |
| 8 | "path/filepath" |
| 9 | "slices" |
| 10 | "strings" |
| 11 | "sync" |
| 12 | "testing" |
| 13 | "time" |
| 14 | |
| 15 | "reasonix/internal/mcplaunch" |
| 16 | "reasonix/internal/tool" |
| 17 | ) |
| 18 | |
| 19 | type destructiveLazyTarget struct { |
| 20 | name string |
| 21 | calls int |
| 22 | } |
| 23 | |
| 24 | type mutableLazyTarget struct { |
| 25 | name string |
| 26 | calls int |
| 27 | } |
| 28 | |
| 29 | func (t *mutableLazyTarget) Name() string { return t.name } |
| 30 | func (t *mutableLazyTarget) Description() string { return "writer test target" } |
| 31 | func (t *mutableLazyTarget) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 32 | func (t *mutableLazyTarget) ReadOnly() bool { return false } |
| 33 | func (t *mutableLazyTarget) Execute(context.Context, json.RawMessage) (string, error) { |
| 34 | t.calls++ |
| 35 | return "executed", nil |
| 36 | } |
| 37 | |
| 38 | func (t *destructiveLazyTarget) Name() string { return t.name } |
| 39 | func (t *destructiveLazyTarget) Description() string { return "destructive test target" } |
| 40 | func (t *destructiveLazyTarget) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 41 | func (t *destructiveLazyTarget) ReadOnly() bool { return true } |
| 42 | func (t *destructiveLazyTarget) MCPDestructiveHint() bool { return true } |
| 43 | func (t *destructiveLazyTarget) Execute(context.Context, json.RawMessage) (string, error) { |
| 44 | t.calls++ |
| 45 | return "executed", nil |
| 46 | } |
| 47 | |
| 48 | // helperSpec returns a Spec that re-invokes this test binary as a minimal MCP |
| 49 | // stdio server (see TestHelperProcess in plugin_test.go). Reused across every |
| 50 | // lazy_test case so the helper-process contract — "echo: <msg>" responder with |
| 51 | // tools/list exposing echo and zed — stays the single source of truth. |
| 52 | func helperSpec() Spec { |
| 53 | return Spec{ |
| 54 | Name: "mock", |
| 55 | Command: os.Args[0], |
| 56 | Args: []string{"-test.run=TestHelperProcess", "--"}, |
| 57 | Env: map[string]string{"GO_WANT_HELPER_PROCESS": "1"}, |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | // writeMockCache primes the on-disk cache for spec with the two tools the |
| 62 | // helper subprocess exposes (echo, zed). We mirror the real schemas so a |
| 63 | // cache-hit lazyTool surfaces the same Schema() bytes that a freshly handshaked |
| 64 | // remoteTool would — the test for "model sees real schema before any Execute" |
| 65 | // depends on this equivalence. |
| 66 | func writeMockCache(t *testing.T, spec Spec) { |
| 67 | t.Helper() |
| 68 | cs := CachedSchema{ |
| 69 | CacheKey: SchemaCacheKey(spec), |
| 70 | Capabilities: map[string]bool{"prompts": false, "resources": false}, |
| 71 | Tools: []CachedTool{ |
| 72 | { |
| 73 | Name: "echo", |
| 74 | Description: "Echo back the message.", |
| 75 | Schema: json.RawMessage(`{"type":"object","properties":{"msg":{"type":"string"}},"required":["z","msg"]}`), |
| 76 | }, |
| 77 | { |
| 78 | Name: "zed", |
| 79 | Description: "Sorted after echo.", |
| 80 | Schema: json.RawMessage(`{"type":"object"}`), |
| 81 | }, |
| 82 | }, |
| 83 | } |
| 84 | if err := SaveCachedSchema(spec.Name, cs); err != nil { |
| 85 | t.Fatalf("SaveCachedSchema: %v", err) |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | // waitForServer polls host.ServerNames() until name appears or timeout |
| 90 | // elapses. The lazy path spawns via a goroutine, so tests need a bounded poll |
| 91 | // rather than a fixed sleep — five seconds covers a slow CI subprocess fork |
| 92 | // while still aborting clearly on a real hang. |
| 93 | func waitForServer(t *testing.T, host *Host, name string, timeout time.Duration) { |
| 94 | t.Helper() |
| 95 | deadline := time.Now().Add(timeout) |
| 96 | for time.Now().Before(deadline) { |
| 97 | if slices.Contains(host.ServerNames(), name) { |
| 98 | return |
| 99 | } |
| 100 | time.Sleep(10 * time.Millisecond) |
| 101 | } |
| 102 | t.Fatalf("server %q never appeared in host.ServerNames() within %v (got %v)", name, timeout, host.ServerNames()) |
| 103 | } |
| 104 | |
| 105 | func waitForCachedSchema(t *testing.T, spec Spec, timeout time.Duration) *CachedSchema { |
| 106 | t.Helper() |
| 107 | deadline := time.Now().Add(timeout) |
| 108 | for time.Now().Before(deadline) { |
| 109 | if cs, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)); ok { |
| 110 | return cs |
| 111 | } |
| 112 | time.Sleep(10 * time.Millisecond) |
| 113 | } |
| 114 | t.Fatalf("cached schema for %q never appeared within %v", spec.Name, timeout) |
| 115 | return nil |
| 116 | } |
| 117 | |
| 118 | func TestHostCloseWaitsForLazyBackgroundWrite(t *testing.T) { |
| 119 | host := NewHost() |
| 120 | started := make(chan struct{}) |
| 121 | release := make(chan struct{}) |
| 122 | host.queueBackgroundWrite(func() { |
| 123 | close(started) |
| 124 | <-release |
| 125 | }) |
| 126 | <-started |
| 127 | |
| 128 | closed := make(chan struct{}) |
| 129 | go func() { |
| 130 | host.Close() |
| 131 | close(closed) |
| 132 | }() |
| 133 | select { |
| 134 | case <-closed: |
| 135 | t.Fatal("Host.Close returned before the lazy background write finished") |
| 136 | case <-time.After(50 * time.Millisecond): |
| 137 | } |
| 138 | |
| 139 | close(release) |
| 140 | select { |
| 141 | case <-closed: |
| 142 | case <-time.After(2 * time.Second): |
| 143 | t.Fatal("Host.Close did not return after the lazy background write finished") |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | // TestLazyCacheHitSyncSpawn drives the cache-hit branch end-to-end: cache is |
| 148 | // pre-populated, the model can see real schemas before any spawn, and the |
| 149 | // first Execute synchronously handshakes, swaps the placeholder for the real |
| 150 | // *remoteTool, and forwards through in one turn. This is the "warm start" |
| 151 | // payoff — lazy plugins should be indistinguishable from eager once they have |
| 152 | // a cache. |
| 153 | func TestLazyCacheHitSyncSpawn(t *testing.T) { |
| 154 | redirectCache(t) |
| 155 | spec := helperSpec() |
| 156 | writeMockCache(t, spec) |
| 157 | |
| 158 | cs, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 159 | if !ok { |
| 160 | t.Fatal("LoadCachedSchema: miss right after save (sanity)") |
| 161 | } |
| 162 | |
| 163 | host := NewHost() |
| 164 | defer host.Close() |
| 165 | reg := tool.NewRegistry() |
| 166 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 167 | defer cancel() |
| 168 | |
| 169 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 170 | if len(tools) != 2 { |
| 171 | t.Fatalf("LazyToolset returned %d tools, want 2 (echo + zed)", len(tools)) |
| 172 | } |
| 173 | for _, lt := range tools { |
| 174 | reg.Add(lt) |
| 175 | } |
| 176 | |
| 177 | // Before any Execute: registry exposes real cached schemas (not the empty |
| 178 | // {"type":"object"} stub). The model relies on this to call the tool with |
| 179 | // real args on the very first turn. |
| 180 | echoBefore, ok := reg.Get("mcp__mock__echo") |
| 181 | if !ok { |
| 182 | t.Fatal("registry missing mcp__mock__echo after LazyToolset") |
| 183 | } |
| 184 | if _, isLazy := echoBefore.(*lazyTool); !isLazy { |
| 185 | t.Fatalf("pre-Execute echo should be a *lazyTool, got %T", echoBefore) |
| 186 | } |
| 187 | gotSchema := string(echoBefore.Schema()) |
| 188 | if !strings.Contains(gotSchema, `"msg"`) || !strings.Contains(gotSchema, `"required"`) { |
| 189 | t.Fatalf("cached schema not surfaced through lazyTool.Schema(): %s", gotSchema) |
| 190 | } |
| 191 | |
| 192 | // First Execute: cache-hit path runs the handshake synchronously and |
| 193 | // forwards to the real tool — the user sees "echo: hi" in this same turn. |
| 194 | out, err := echoBefore.Execute(ctx, json.RawMessage(`{"msg":"hi"}`)) |
| 195 | if err != nil { |
| 196 | t.Fatalf("Execute: %v", err) |
| 197 | } |
| 198 | if out != "echo: hi" { |
| 199 | t.Fatalf("Execute result = %q, want %q", out, "echo: hi") |
| 200 | } |
| 201 | |
| 202 | // The spawn actually happened — host now lists the mock server. |
| 203 | names := host.ServerNames() |
| 204 | if len(names) != 1 || names[0] != "mock" { |
| 205 | t.Fatalf("host.ServerNames() = %v, want [mock]", names) |
| 206 | } |
| 207 | |
| 208 | // After Execute, the registry entry must STILL be the placeholder: cache-hit |
| 209 | // placeholders are pinned for the whole session so the request's tools |
| 210 | // array stays byte-identical even when the live handshake differs from the |
| 211 | // cache (see trySwap). Execution keeps forwarding to the real tool through |
| 212 | // the shared spawn state. |
| 213 | echoAfter, _ := reg.Get("mcp__mock__echo") |
| 214 | if _, isLazy := echoAfter.(*lazyTool); !isLazy { |
| 215 | t.Fatalf("post-Execute echo should remain the pinned *lazyTool, got %T", echoAfter) |
| 216 | } |
| 217 | if got := string(echoAfter.Schema()); got != gotSchema { |
| 218 | t.Fatalf("registry schema bytes changed across the handshake:\nbefore: %s\nafter: %s", gotSchema, got) |
| 219 | } |
| 220 | // Second call goes straight through the ready state to the real tool. |
| 221 | out2, err := echoAfter.Execute(ctx, json.RawMessage(`{"msg":"again"}`)) |
| 222 | if err != nil { |
| 223 | t.Fatalf("second Execute: %v", err) |
| 224 | } |
| 225 | if out2 != "echo: again" { |
| 226 | t.Fatalf("second Execute result = %q, want %q", out2, "echo: again") |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | func TestLazyCacheHitReusesExistingSharedHostClient(t *testing.T) { |
| 231 | redirectCache(t) |
| 232 | spec := helperSpec() |
| 233 | writeMockCache(t, spec) |
| 234 | |
| 235 | cs, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 236 | if !ok { |
| 237 | t.Fatal("LoadCachedSchema: miss right after save (sanity)") |
| 238 | } |
| 239 | |
| 240 | host := NewHost() |
| 241 | defer host.Close() |
| 242 | reg := tool.NewRegistry() |
| 243 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 244 | defer cancel() |
| 245 | |
| 246 | if _, err := host.Add(ctx, spec); err != nil { |
| 247 | t.Fatalf("preconnect shared host: %v", err) |
| 248 | } |
| 249 | |
| 250 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 251 | for _, lt := range tools { |
| 252 | reg.Add(lt) |
| 253 | } |
| 254 | echoBefore, ok := reg.Get("mcp__mock__echo") |
| 255 | if !ok { |
| 256 | t.Fatal("registry missing mcp__mock__echo after LazyToolset") |
| 257 | } |
| 258 | |
| 259 | out, err := echoBefore.Execute(ctx, json.RawMessage(`{"msg":"hi"}`)) |
| 260 | if err != nil { |
| 261 | t.Fatalf("Execute against existing shared host client: %v", err) |
| 262 | } |
| 263 | if out != "echo: hi" { |
| 264 | t.Fatalf("Execute result = %q, want %q", out, "echo: hi") |
| 265 | } |
| 266 | if got := host.ServerNames(); len(got) != 1 || got[0] != "mock" { |
| 267 | t.Fatalf("shared host should still have exactly one mock server, got %v", got) |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | func TestLazyRemoveCancelsInFlightGenerationWithoutResurrection(t *testing.T) { |
| 272 | redirectCache(t) |
| 273 | spec := helperSpec() |
| 274 | spec.Env["GO_WANT_HELPER_INIT_MS"] = "500" |
| 275 | host := NewHost() |
| 276 | defer host.Close() |
| 277 | reg := tool.NewRegistry() |
| 278 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 279 | defer cancel() |
| 280 | |
| 281 | for _, placeholder := range LazyToolset(spec, nil, host, reg, ctx, true) { |
| 282 | reg.Add(placeholder) |
| 283 | } |
| 284 | deadline := time.Now().Add(3 * time.Second) |
| 285 | for { |
| 286 | host.spawningMu.Lock() |
| 287 | spawning := len(host.spawning) > 0 |
| 288 | host.spawningMu.Unlock() |
| 289 | if spawning { |
| 290 | break |
| 291 | } |
| 292 | if time.Now().After(deadline) { |
| 293 | t.Fatal("lazy spawn never entered the in-flight state") |
| 294 | } |
| 295 | time.Sleep(5 * time.Millisecond) |
| 296 | } |
| 297 | |
| 298 | prefix, found := host.Remove(spec.Name) |
| 299 | if !found { |
| 300 | t.Fatal("Host.Remove did not cancel the in-flight lazy generation") |
| 301 | } |
| 302 | reg.RemovePrefix(prefix) |
| 303 | done := make(chan struct{}) |
| 304 | go func() { |
| 305 | host.deferredWG.Wait() |
| 306 | close(done) |
| 307 | }() |
| 308 | select { |
| 309 | case <-done: |
| 310 | case <-time.After(5 * time.Second): |
| 311 | t.Fatal("cancelled lazy generation did not finish") |
| 312 | } |
| 313 | if host.HasClient(spec.Name) || len(host.ServerNames()) != 0 { |
| 314 | t.Fatalf("removed lazy server was resurrected: %v", host.ServerNames()) |
| 315 | } |
| 316 | if _, ok := reg.Get(ToolPrefix(spec.Name) + "connect"); ok { |
| 317 | t.Fatal("removed lazy placeholder was re-registered") |
| 318 | } |
| 319 | if _, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)); ok { |
| 320 | t.Fatal("cancelled lazy generation wrote a new schema cache") |
| 321 | } |
| 322 | tools, err := host.Add(ctx, spec) |
| 323 | if err != nil { |
| 324 | t.Fatalf("re-add after cancelled generation: %v", err) |
| 325 | } |
| 326 | if len(tools) == 0 || !host.HasClient(spec.Name) { |
| 327 | t.Fatalf("new generation did not connect after removal: tools=%d clients=%v", len(tools), host.ServerNames()) |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | func TestAddWithLifecycleCoalescesConcurrentSameServer(t *testing.T) { |
| 332 | spec := helperSpec() |
| 333 | spec.Env["GO_WANT_HELPER_INIT_MS"] = "200" |
| 334 | |
| 335 | host := NewHost() |
| 336 | defer host.Close() |
| 337 | lifeCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 338 | defer cancel() |
| 339 | |
| 340 | start := make(chan struct{}) |
| 341 | errs := make([]error, 2) |
| 342 | toolCounts := make([]int, 2) |
| 343 | var wg sync.WaitGroup |
| 344 | for i := range errs { |
| 345 | wg.Add(1) |
| 346 | go func(i int) { |
| 347 | defer wg.Done() |
| 348 | <-start |
| 349 | callCtx, cancelCall := context.WithTimeout(lifeCtx, 5*time.Second) |
| 350 | defer cancelCall() |
| 351 | tools, err := host.AddWithLifecycle(lifeCtx, callCtx, spec) |
| 352 | errs[i] = err |
| 353 | toolCounts[i] = len(tools) |
| 354 | }(i) |
| 355 | } |
| 356 | close(start) |
| 357 | wg.Wait() |
| 358 | |
| 359 | for i, err := range errs { |
| 360 | if err != nil { |
| 361 | t.Fatalf("AddWithLifecycle call %d failed: %v (all errors: %v)", i, err, errs) |
| 362 | } |
| 363 | if toolCounts[i] != 2 { |
| 364 | t.Fatalf("AddWithLifecycle call %d returned %d tools, want 2", i, toolCounts[i]) |
| 365 | } |
| 366 | } |
| 367 | if got := host.ServerNames(); len(got) != 1 || got[0] != "mock" { |
| 368 | t.Fatalf("host should contain exactly one connected server, got %v", got) |
| 369 | } |
| 370 | } |
| 371 | |
| 372 | func TestLazyCacheHitSlowStartupContinuesInBackground(t *testing.T) { |
| 373 | redirectCache(t) |
| 374 | spec := helperSpec() |
| 375 | spec.StartupTimeout = 2 * time.Second |
| 376 | spec.Env["GO_WANT_HELPER_INIT_MS"] = "200" |
| 377 | writeMockCache(t, spec) |
| 378 | |
| 379 | cs, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 380 | if !ok { |
| 381 | t.Fatal("LoadCachedSchema: miss right after save (sanity)") |
| 382 | } |
| 383 | |
| 384 | host := NewHost() |
| 385 | defer host.Close() |
| 386 | reg := tool.NewRegistry() |
| 387 | ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second) |
| 388 | defer cancel() |
| 389 | |
| 390 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 391 | for _, lt := range tools { |
| 392 | reg.Add(lt) |
| 393 | } |
| 394 | echo, ok := reg.Get("mcp__mock__echo") |
| 395 | if !ok { |
| 396 | t.Fatal("registry missing mcp__mock__echo after LazyToolset") |
| 397 | } |
| 398 | lazyEcho, ok := echo.(*lazyTool) |
| 399 | if !ok { |
| 400 | t.Fatalf("pre-Execute echo should be a *lazyTool, got %T", echo) |
| 401 | } |
| 402 | lazyEcho.shared.waitBudget = 25 * time.Millisecond |
| 403 | beforeName := echo.Name() |
| 404 | beforeDescription := echo.Description() |
| 405 | beforeSchema := string(echo.Schema()) |
| 406 | |
| 407 | if _, err := echo.Execute(ctx, json.RawMessage(`{"msg":"slow"}`)); err == nil || !strings.Contains(err.Error(), "continues in background") { |
| 408 | t.Fatalf("first Execute error = %v, want background startup notice", err) |
| 409 | } |
| 410 | waitForServer(t, host, spec.Name, 2*time.Second) |
| 411 | out, err := echo.Execute(ctx, json.RawMessage(`{"msg":"retry"}`)) |
| 412 | if err != nil { |
| 413 | t.Fatalf("second Execute after background startup should succeed: %v", err) |
| 414 | } |
| 415 | if out != "echo: retry" { |
| 416 | t.Fatalf("Execute result = %q, want %q", out, "echo: retry") |
| 417 | } |
| 418 | if echo.Name() != beforeName || echo.Description() != beforeDescription || string(echo.Schema()) != beforeSchema { |
| 419 | t.Fatalf("provider-visible cached tool changed across background startup") |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | func TestLazyToolsetInheritsInstalledServerReaderAuthorization(t *testing.T) { |
| 424 | redirectCache(t) |
| 425 | spec := helperSpec() |
| 426 | spec.LaunchManager = mcplaunch.NewManager(filepath.Join(t.TempDir(), mcplaunch.StateFilename), t.TempDir()) |
| 427 | spec.Authorized = true |
| 428 | if err := SaveCachedSchema(spec.Name, CachedSchema{ |
| 429 | CacheKey: SchemaCacheKey(spec), |
| 430 | Tools: []CachedTool{{ |
| 431 | Name: "echo", Description: "Echo back the message.", |
| 432 | Schema: json.RawMessage(`{"type":"object","properties":{"msg":{"type":"string"}}}`), ReadOnly: true, |
| 433 | }}, |
| 434 | }); err != nil { |
| 435 | t.Fatal(err) |
| 436 | } |
| 437 | cs, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 438 | if !ok { |
| 439 | t.Fatal("LoadCachedSchema: miss right after save") |
| 440 | } |
| 441 | |
| 442 | host := NewHost() |
| 443 | defer host.Close() |
| 444 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 445 | defer cancel() |
| 446 | tools := LazyToolset(spec, cs, host, tool.NewRegistry(), ctx, false) |
| 447 | var echo tool.Tool |
| 448 | for _, candidate := range tools { |
| 449 | if candidate.Name() == "mcp__mock__echo" { |
| 450 | echo = candidate |
| 451 | break |
| 452 | } |
| 453 | } |
| 454 | if echo == nil || !echo.ReadOnly() { |
| 455 | t.Fatalf("installed cached reader missing or not read-only: %T", echo) |
| 456 | } |
| 457 | if authority, ok := echo.(tool.MCPServerAuthorization); !ok || !authority.MCPServerAuthorized() { |
| 458 | t.Fatalf("lazy installed reader did not inherit authorization: %T", echo) |
| 459 | } |
| 460 | } |
| 461 | |
| 462 | // TestLazyCacheMissAsyncSpawn drives the cache-miss branch: with no cache, a |
| 463 | // single "connect" placeholder shows up; first Execute returns a retry hint and |
| 464 | // kicks the spawn async; once that spawn finishes, the registry swaps to the |
| 465 | // real tools under their real names, and the connect stub is dropped. This is |
| 466 | // the "model warm-up" contract — the model must not see stale schemas, so we |
| 467 | // refuse to forward the first call and instead ask for one more turn. |
| 468 | func TestLazyCacheMissAsyncSpawn(t *testing.T) { |
| 469 | redirectCache(t) |
| 470 | spec := helperSpec() |
| 471 | |
| 472 | host := NewHost() |
| 473 | defer host.Close() |
| 474 | reg := tool.NewRegistry() |
| 475 | ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second) |
| 476 | defer cancel() |
| 477 | |
| 478 | tools := LazyToolset(spec, nil, host, reg, ctx, false) |
| 479 | if len(tools) != 1 { |
| 480 | t.Fatalf("cache-miss LazyToolset must return 1 connect stub, got %d", len(tools)) |
| 481 | } |
| 482 | for _, lt := range tools { |
| 483 | reg.Add(lt) |
| 484 | } |
| 485 | |
| 486 | connect, ok := reg.Get("mcp__mock__connect") |
| 487 | if !ok { |
| 488 | t.Fatalf("registry missing mcp__mock__connect; names=%v", reg.Names()) |
| 489 | } |
| 490 | |
| 491 | // First Execute must NOT forward — schema is unknown, so the model would |
| 492 | // be feeding garbage. It returns a retry hint and triggers spawn async. |
| 493 | _, err := connect.Execute(ctx, json.RawMessage(`{}`)) |
| 494 | if err == nil { |
| 495 | t.Fatal("first Execute on cache-miss placeholder should error with a retry hint") |
| 496 | } |
| 497 | msg := err.Error() |
| 498 | if !strings.Contains(msg, "initializing") && !strings.Contains(msg, "next turn") { |
| 499 | t.Fatalf("first-Execute error %q should mention 'initializing' or 'next turn'", msg) |
| 500 | } |
| 501 | |
| 502 | // A retry waits on the lazy state machine's completion signal, so its return |
| 503 | // proves the registry swap has published the real tools. |
| 504 | if _, err := connect.Execute(ctx, json.RawMessage(`{}`)); err != nil { |
| 505 | t.Fatalf("second Execute after cache-miss spawn: %v", err) |
| 506 | } |
| 507 | |
| 508 | if _, found := reg.Get("mcp__mock__connect"); found { |
| 509 | t.Errorf("connect stub should be removed after swap, names=%v", reg.Names()) |
| 510 | } |
| 511 | if _, found := reg.Get("mcp__mock__echo"); !found { |
| 512 | t.Errorf("real mcp__mock__echo missing after swap, names=%v", reg.Names()) |
| 513 | } |
| 514 | if _, found := reg.Get("mcp__mock__zed"); !found { |
| 515 | t.Errorf("real mcp__mock__zed missing after swap, names=%v", reg.Names()) |
| 516 | } |
| 517 | } |
| 518 | |
| 519 | func TestLazySwapDoesNotRaceRegistrySchemas(t *testing.T) { |
| 520 | redirectCache(t) |
| 521 | spec := helperSpec() |
| 522 | spec.Env["GO_WANT_HELPER_INIT_MS"] = "50" |
| 523 | writeMockCache(t, spec) |
| 524 | cs, _ := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 525 | |
| 526 | host := NewHost() |
| 527 | defer host.Close() |
| 528 | reg := tool.NewRegistry() |
| 529 | ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second) |
| 530 | defer cancel() |
| 531 | |
| 532 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 533 | for _, lt := range tools { |
| 534 | reg.Add(lt) |
| 535 | } |
| 536 | echo, _ := reg.Get("mcp__mock__echo") |
| 537 | if echo == nil { |
| 538 | t.Fatal("missing mcp__mock__echo placeholder") |
| 539 | } |
| 540 | |
| 541 | done := make(chan struct{}) |
| 542 | var wg sync.WaitGroup |
| 543 | wg.Go(func() { |
| 544 | for { |
| 545 | select { |
| 546 | case <-done: |
| 547 | return |
| 548 | default: |
| 549 | _ = reg.Schemas() |
| 550 | } |
| 551 | } |
| 552 | }) |
| 553 | |
| 554 | out, err := echo.Execute(ctx, json.RawMessage(`{"msg":"race"}`)) |
| 555 | close(done) |
| 556 | wg.Wait() |
| 557 | if err != nil { |
| 558 | t.Fatalf("Execute: %v", err) |
| 559 | } |
| 560 | if out != "echo: race" { |
| 561 | t.Fatalf("Execute result = %q, want %q", out, "echo: race") |
| 562 | } |
| 563 | } |
| 564 | |
| 565 | // TestLazyBackgroundKick covers the background-tier path: kick=true plus a |
| 566 | // cache hit means the spawn races boot, finishes before the model calls, and |
| 567 | // the first Execute hits the "already-ready, swap on the way through" branch. |
| 568 | // The model never sees a placeholder schema-wise either, since the cache |
| 569 | // fed Schema() before kick even started. |
| 570 | func TestLazyBackgroundKick(t *testing.T) { |
| 571 | redirectCache(t) |
| 572 | spec := helperSpec() |
| 573 | writeMockCache(t, spec) |
| 574 | cs, _ := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 575 | |
| 576 | host := NewHost() |
| 577 | defer host.Close() |
| 578 | reg := tool.NewRegistry() |
| 579 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 580 | defer cancel() |
| 581 | |
| 582 | tools := LazyToolset(spec, cs, host, reg, ctx, true) // kick=true |
| 583 | if len(tools) != 2 { |
| 584 | t.Fatalf("LazyToolset(kick=true) returned %d tools, want 2", len(tools)) |
| 585 | } |
| 586 | for _, lt := range tools { |
| 587 | reg.Add(lt) |
| 588 | } |
| 589 | |
| 590 | // Wait for the background spawn to complete — proof that kick fired off |
| 591 | // the handshake without us calling Execute. |
| 592 | waitForServer(t, host, "mock", 5*time.Second) |
| 593 | |
| 594 | // Now Execute: the state is already spawnReady, so this should swap + |
| 595 | // forward in one shot without a second Add call. The result must still be |
| 596 | // correct. |
| 597 | echo, _ := reg.Get("mcp__mock__echo") |
| 598 | out, err := echo.Execute(ctx, json.RawMessage(`{"msg":"bg"}`)) |
| 599 | if err != nil { |
| 600 | t.Fatalf("Execute after background ready: %v", err) |
| 601 | } |
| 602 | if out != "echo: bg" { |
| 603 | t.Fatalf("Execute result = %q, want %q", out, "echo: bg") |
| 604 | } |
| 605 | |
| 606 | // One spawn, not two — kick + Execute must collapse onto the same run. |
| 607 | if names := host.ServerNames(); len(names) != 1 { |
| 608 | t.Fatalf("host.ServerNames() = %v, want exactly one 'mock'", names) |
| 609 | } |
| 610 | } |
| 611 | |
| 612 | func TestLazyBackgroundCacheMissPersistsSchemaAndCompletesAdvertisedConnect(t *testing.T) { |
| 613 | redirectCache(t) |
| 614 | spec := helperSpec() |
| 615 | |
| 616 | host := NewHost() |
| 617 | defer host.Close() |
| 618 | reg := tool.NewRegistry() |
| 619 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 620 | defer cancel() |
| 621 | |
| 622 | tools := LazyToolset(spec, nil, host, reg, ctx, true) // cache miss + background kick |
| 623 | if len(tools) != 1 { |
| 624 | t.Fatalf("cache-miss LazyToolset returned %d tools, want one connect placeholder", len(tools)) |
| 625 | } |
| 626 | connect, ok := tools[0].(*lazyTool) |
| 627 | if !ok { |
| 628 | t.Fatalf("cache-miss placeholder type = %T, want *lazyTool", tools[0]) |
| 629 | } |
| 630 | for _, lt := range tools { |
| 631 | reg.Add(lt) |
| 632 | } |
| 633 | |
| 634 | waitForServer(t, host, "mock", 5*time.Second) |
| 635 | cs := waitForCachedSchema(t, spec, 5*time.Second) |
| 636 | if len(cs.Tools) != 2 { |
| 637 | t.Fatalf("cached schema has %d tools, want 2", len(cs.Tools)) |
| 638 | } |
| 639 | got := map[string]bool{} |
| 640 | for _, ct := range cs.Tools { |
| 641 | got[ct.Name] = true |
| 642 | } |
| 643 | if !got["echo"] || !got["zed"] { |
| 644 | t.Fatalf("cached tools = %v, want echo and zed", got) |
| 645 | } |
| 646 | if _, found := reg.Get(connect.Name()); found { |
| 647 | t.Fatalf("connect placeholder remained provider-visible after discovery; names=%v", reg.Names()) |
| 648 | } |
| 649 | if out, err := connect.Execute(ctx, json.RawMessage(`{}`)); err != nil || !strings.Contains(out, "real tools are now available") { |
| 650 | t.Fatalf("already-advertised connect after discovery = (%q, %v), want controlled connected result", out, err) |
| 651 | } |
| 652 | } |
| 653 | |
| 654 | func TestLazyBackgroundCloseCancelsInFlightKick(t *testing.T) { |
| 655 | redirectCache(t) |
| 656 | spec := helperSpec() |
| 657 | spec.Name = "slow" |
| 658 | spec.Env["GO_WANT_HELPER_INIT_MS"] = "5000" |
| 659 | writeMockCache(t, spec) |
| 660 | cs, _ := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 661 | |
| 662 | host := NewHost() |
| 663 | reg := tool.NewRegistry() |
| 664 | |
| 665 | tools := LazyToolset(spec, cs, host, reg, context.Background(), true) |
| 666 | for _, lt := range tools { |
| 667 | reg.Add(lt) |
| 668 | } |
| 669 | |
| 670 | done := make(chan struct{}) |
| 671 | go func() { |
| 672 | host.Close() |
| 673 | close(done) |
| 674 | }() |
| 675 | select { |
| 676 | case <-done: |
| 677 | case <-time.After(2 * time.Second): |
| 678 | t.Fatal("Host.Close did not cancel the in-flight background lazy spawn") |
| 679 | } |
| 680 | |
| 681 | if names := host.ServerNames(); len(names) != 0 { |
| 682 | t.Fatalf("closed host retained connected servers: %v", names) |
| 683 | } |
| 684 | } |
| 685 | |
| 686 | // TestLazyConcurrentExecuteOnlyOneSpawn pins the de-duplication contract: 10 |
| 687 | // goroutines racing through Execute on the same lazyTool may only trigger ONE |
| 688 | // spawn (and therefore one connected mock server on the host). The state |
| 689 | // machine's mu+state gate is what makes this true; this test would catch a |
| 690 | // regression where someone moved the state transition outside the lock or |
| 691 | // swapped to a TOCTOU check. |
| 692 | // |
| 693 | // Note: by design (see lazy.go), only the winner of the race forwards |
| 694 | // synchronously; the losers observe spawnInFlight and return a "retry next |
| 695 | // turn" hint rather than blocking. We assert that contract too: at least one |
| 696 | // goroutine got "echo: r<i>", and the racers that didn't win got the |
| 697 | // initializing hint — never a spurious error and never a stale or partial |
| 698 | // result. After all goroutines complete, a fresh Execute hits spawnReady and |
| 699 | // forwards normally. |
| 700 | func TestLazyConcurrentExecuteOnlyOneSpawn(t *testing.T) { |
| 701 | redirectCache(t) |
| 702 | spec := helperSpec() |
| 703 | writeMockCache(t, spec) |
| 704 | cs, _ := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 705 | |
| 706 | host := NewHost() |
| 707 | defer host.Close() |
| 708 | reg := tool.NewRegistry() |
| 709 | ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second) |
| 710 | defer cancel() |
| 711 | |
| 712 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 713 | for _, lt := range tools { |
| 714 | reg.Add(lt) |
| 715 | } |
| 716 | echo, _ := reg.Get("mcp__mock__echo") |
| 717 | |
| 718 | const goroutines = 10 |
| 719 | var wg sync.WaitGroup |
| 720 | results := make([]string, goroutines) |
| 721 | errs := make([]error, goroutines) |
| 722 | wg.Add(goroutines) |
| 723 | for i := range goroutines { |
| 724 | go func(i int) { |
| 725 | defer wg.Done() |
| 726 | out, err := echo.Execute(ctx, json.RawMessage(fmt.Sprintf(`{"msg":"r%d"}`, i))) |
| 727 | results[i], errs[i] = out, err |
| 728 | }(i) |
| 729 | } |
| 730 | wg.Wait() |
| 731 | |
| 732 | // Every result must be either the real "echo: rN" output or the explicit |
| 733 | // initializing hint — nothing else. At least one goroutine (the racing |
| 734 | // winner) must succeed, otherwise the state machine deadlocked the win. |
| 735 | winners := 0 |
| 736 | for i, err := range errs { |
| 737 | want := fmt.Sprintf("echo: r%d", i) |
| 738 | switch { |
| 739 | case err == nil && results[i] == want: |
| 740 | winners++ |
| 741 | case err != nil && strings.Contains(err.Error(), "initializing"): |
| 742 | // expected loser |
| 743 | default: |
| 744 | t.Errorf("goroutine %d: result=%q err=%v — must be either %q or an 'initializing' hint", i, results[i], err, want) |
| 745 | } |
| 746 | } |
| 747 | if winners == 0 { |
| 748 | t.Fatal("no goroutine succeeded — at least the race winner must forward through") |
| 749 | } |
| 750 | |
| 751 | // Exactly one Client landed on the host: the mu+state gate kept the 9 |
| 752 | // losers off the spawn path. This is the headline invariant of the lazy |
| 753 | // design — racing the first call must not fork-bomb the subprocess. |
| 754 | mockCount := 0 |
| 755 | for _, n := range host.ServerNames() { |
| 756 | if n == "mock" { |
| 757 | mockCount++ |
| 758 | } |
| 759 | } |
| 760 | if mockCount != 1 { |
| 761 | t.Fatalf("expected 1 'mock' server after concurrent Execute, got %d (names=%v)", mockCount, host.ServerNames()) |
| 762 | } |
| 763 | |
| 764 | // A follow-up Execute (now in spawnReady) goes through cleanly: the |
| 765 | // "retry on next turn" hint was honest, not a permanent error. |
| 766 | out, err := echo.Execute(ctx, json.RawMessage(`{"msg":"after"}`)) |
| 767 | if err != nil { |
| 768 | t.Fatalf("post-race Execute: %v", err) |
| 769 | } |
| 770 | if out != "echo: after" { |
| 771 | t.Fatalf("post-race Execute = %q, want %q", out, "echo: after") |
| 772 | } |
| 773 | } |
| 774 | |
| 775 | // TestLazyHandshakeFailureSurfaced covers the spawnFailed sticky branch: a |
| 776 | // bogus command can't start, the first Execute returns an error that mentions |
| 777 | // "failed to start", and a second Execute returns the SAME error (the state |
| 778 | // machine doesn't retry — we don't want to fork a doomed subprocess every |
| 779 | // turn until the user fixes config). |
| 780 | func TestLazyHandshakeFailureSurfaced(t *testing.T) { |
| 781 | redirectCache(t) |
| 782 | // Bogus command: process exec will fail outright. |
| 783 | spec := Spec{Name: "missing", Command: "reasonix-nonexistent-binary-for-lazy-test"} |
| 784 | |
| 785 | // Hand-craft a cache so the cache-HIT branch runs (synchronous spawn, |
| 786 | // failure surfaces directly to the first caller rather than via a retry |
| 787 | // hint). The CacheKey must match — otherwise LoadCachedSchema would miss |
| 788 | // and we'd be exercising the async path. |
| 789 | cs := &CachedSchema{ |
| 790 | CacheKey: SchemaCacheKey(spec), |
| 791 | Capabilities: map[string]bool{}, |
| 792 | Tools: []CachedTool{{ |
| 793 | Name: "doit", |
| 794 | Description: "noop", |
| 795 | Schema: json.RawMessage(`{"type":"object"}`), |
| 796 | }}, |
| 797 | } |
| 798 | |
| 799 | host := NewHost() |
| 800 | defer host.Close() |
| 801 | reg := tool.NewRegistry() |
| 802 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 803 | defer cancel() |
| 804 | |
| 805 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 806 | if len(tools) != 1 { |
| 807 | t.Fatalf("LazyToolset returned %d tools, want 1 (doit)", len(tools)) |
| 808 | } |
| 809 | for _, lt := range tools { |
| 810 | reg.Add(lt) |
| 811 | } |
| 812 | doit, _ := reg.Get("mcp__missing__doit") |
| 813 | |
| 814 | _, err1 := doit.Execute(ctx, json.RawMessage(`{}`)) |
| 815 | if err1 == nil { |
| 816 | t.Fatal("Execute on a bogus command should error") |
| 817 | } |
| 818 | if !strings.Contains(err1.Error(), "failed to start") { |
| 819 | t.Fatalf("error %q should mention 'failed to start'", err1.Error()) |
| 820 | } |
| 821 | |
| 822 | // Second call: same error, no retry. spawnFailed is sticky on purpose — |
| 823 | // the operator must fix config and restart, not have us fork-bomb on |
| 824 | // every turn. |
| 825 | _, err2 := doit.Execute(ctx, json.RawMessage(`{}`)) |
| 826 | if err2 == nil { |
| 827 | t.Fatal("second Execute after spawnFailed should still error") |
| 828 | } |
| 829 | if !strings.Contains(err2.Error(), "failed to start") { |
| 830 | t.Fatalf("second error %q should still mention 'failed to start' (state machine must stay in spawnFailed)", err2.Error()) |
| 831 | } |
| 832 | } |
| 833 | |
| 834 | // TestLazyToolsetCacheHitSchemaVisible is the model-facing visibility test: |
| 835 | // immediately after LazyToolset returns and BEFORE any Execute, lazyTool.Schema() |
| 836 | // must equal the canonicalized cached schema. The whole point of the cache is |
| 837 | // that the model sees real schemas at turn-start; if Schema() returned the |
| 838 | // "{}" stub here, the model would call with empty args and the cache-hit |
| 839 | // path would never get a useful first call. |
| 840 | func TestLazyToolsetCacheHitSchemaVisible(t *testing.T) { |
| 841 | redirectCache(t) |
| 842 | spec := helperSpec() |
| 843 | |
| 844 | rawSchema := json.RawMessage(`{"properties":{"msg":{"type":"string"}},"type":"object","required":["msg"]}`) |
| 845 | cs := &CachedSchema{ |
| 846 | CacheKey: SchemaCacheKey(spec), |
| 847 | Capabilities: map[string]bool{}, |
| 848 | Tools: []CachedTool{{ |
| 849 | Name: "echo", |
| 850 | Description: "Echo back.", |
| 851 | Schema: rawSchema, |
| 852 | }}, |
| 853 | } |
| 854 | |
| 855 | host := NewHost() |
| 856 | defer host.Close() |
| 857 | reg := tool.NewRegistry() |
| 858 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 859 | defer cancel() |
| 860 | |
| 861 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 862 | if len(tools) != 1 { |
| 863 | t.Fatalf("LazyToolset returned %d tools, want 1", len(tools)) |
| 864 | } |
| 865 | got := string(tools[0].Schema()) |
| 866 | want := string(canonicalizeSchema(rawSchema)) |
| 867 | if got != want { |
| 868 | t.Fatalf("lazyTool.Schema() = %s,\nwant canonicalized cached schema = %s", got, want) |
| 869 | } |
| 870 | |
| 871 | // And we never spawned: Schema() must be free, otherwise the cache |
| 872 | // optimisation is moot. |
| 873 | if names := host.ServerNames(); len(names) != 0 { |
| 874 | t.Fatalf("Schema() must not spawn; host.ServerNames() = %v", names) |
| 875 | } |
| 876 | } |
| 877 | |
| 878 | // registrySchemaBytes marshals the registry's full tool schemas — the exact |
| 879 | // surface that feeds the provider request's tools array. |
| 880 | func registrySchemaBytes(t *testing.T, reg *tool.Registry) string { |
| 881 | t.Helper() |
| 882 | b, err := json.Marshal(reg.Schemas()) |
| 883 | if err != nil { |
| 884 | t.Fatalf("marshal schemas: %v", err) |
| 885 | } |
| 886 | return string(b) |
| 887 | } |
| 888 | |
| 889 | // TestLazyCacheHitPinsToolBytesAcrossDivergentHandshake is the session |
| 890 | // byte-stability guard: the cached snapshot deliberately DIFFERS from what the |
| 891 | // live handshake will report (stale description/schema, and it omits one tool |
| 892 | // the live server exposes). After the background spawn completes, the |
| 893 | // registry's schema bytes must be identical to what the model saw at boot — |
| 894 | // the divergence surfaces in the refreshed disk cache (next session), never |
| 895 | // mid-session in the tools array. |
| 896 | func TestLazyCacheHitPinsToolBytesAcrossDivergentHandshake(t *testing.T) { |
| 897 | redirectCache(t) |
| 898 | spec := helperSpec() |
| 899 | stale := CachedSchema{ |
| 900 | CacheKey: SchemaCacheKey(spec), |
| 901 | Capabilities: map[string]bool{}, |
| 902 | Tools: []CachedTool{{ |
| 903 | Name: "echo", |
| 904 | Description: "STALE description from a previous session.", |
| 905 | Schema: json.RawMessage(`{"type":"object","properties":{"msg":{"type":"string"}}}`), |
| 906 | // live handshake also exposes "zed" — absent here on purpose. |
| 907 | }}, |
| 908 | } |
| 909 | if err := SaveCachedSchema(spec.Name, stale); err != nil { |
| 910 | t.Fatalf("SaveCachedSchema: %v", err) |
| 911 | } |
| 912 | cs, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 913 | if !ok { |
| 914 | t.Fatal("LoadCachedSchema miss after save") |
| 915 | } |
| 916 | |
| 917 | host := NewHost() |
| 918 | defer host.Close() |
| 919 | reg := tool.NewRegistry() |
| 920 | ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second) |
| 921 | defer cancel() |
| 922 | |
| 923 | for _, lt := range LazyToolset(spec, cs, host, reg, ctx, true) { |
| 924 | reg.Add(lt) |
| 925 | } |
| 926 | bootBytes := registrySchemaBytes(t, reg) |
| 927 | |
| 928 | // Let the background handshake finish and give trySwap every chance to run. |
| 929 | waitForServer(t, host, "mock", 5*time.Second) |
| 930 | echo, _ := reg.Get("mcp__mock__echo") |
| 931 | if out, err := echo.Execute(ctx, json.RawMessage(`{"msg":"pin"}`)); err != nil || out != "echo: pin" { |
| 932 | t.Fatalf("Execute after schema drift = %q, %v; want live execution", out, err) |
| 933 | } |
| 934 | |
| 935 | if got := registrySchemaBytes(t, reg); got != bootBytes { |
| 936 | t.Fatalf("tools array bytes changed mid-session after a divergent handshake:\nboot: %s\nnow: %s", bootBytes, got) |
| 937 | } |
| 938 | if _, found := reg.Get("mcp__mock__zed"); found { |
| 939 | t.Fatal("live-only tool joined the registry mid-session; it must wait for the next session") |
| 940 | } |
| 941 | |
| 942 | // The refreshed cache carries the live truth for the NEXT session. The |
| 943 | // stale cache this test wrote is itself loadable, so poll until the |
| 944 | // refresh actually lands (the background save races Execute's return on |
| 945 | // slow machines) rather than accepting the first loadable snapshot. |
| 946 | deadline := time.Now().Add(5 * time.Second) |
| 947 | for { |
| 948 | refreshed, ok := LoadCachedSchema(spec.Name, SchemaCacheKey(spec)) |
| 949 | if ok { |
| 950 | names := map[string]bool{} |
| 951 | for _, ct := range refreshed.Tools { |
| 952 | names[ct.Name] = true |
| 953 | } |
| 954 | if names["echo"] && names["zed"] { |
| 955 | break |
| 956 | } |
| 957 | if time.Now().After(deadline) { |
| 958 | t.Fatalf("refreshed cache tools = %v, want live set {echo, zed}", refreshed.Tools) |
| 959 | } |
| 960 | } else if time.Now().After(deadline) { |
| 961 | t.Fatal("cached schema never became loadable") |
| 962 | } |
| 963 | time.Sleep(10 * time.Millisecond) |
| 964 | } |
| 965 | } |
| 966 | |
| 967 | func TestLazyToolPromotesLiveDestructiveHintBeforeExecution(t *testing.T) { |
| 968 | const name = "mcp__srv__wipe" |
| 969 | target := &destructiveLazyTarget{name: name} |
| 970 | shared := &lazySpawn{ |
| 971 | spec: Spec{Name: "srv"}, |
| 972 | state: spawnReady, |
| 973 | real: map[string]tool.Tool{name: target}, |
| 974 | swapped: true, |
| 975 | } |
| 976 | lazy := &lazyTool{ |
| 977 | shared: shared, |
| 978 | name: name, |
| 979 | rawName: "wipe", |
| 980 | readOnly: true, |
| 981 | hasCache: true, |
| 982 | } |
| 983 | |
| 984 | if out, err := lazy.Execute(context.Background(), nil); err == nil || !strings.Contains(err.Error(), "retry") || out != "" { |
| 985 | t.Fatalf("first Execute = (%q,%v), want retry before destructive execution", out, err) |
| 986 | } |
| 987 | if target.calls != 0 || !lazy.MCPDestructiveHint() { |
| 988 | t.Fatalf("after promotion calls=%d destructive=%v, want 0/true", target.calls, lazy.MCPDestructiveHint()) |
| 989 | } |
| 990 | |
| 991 | out, err := lazy.Execute(context.Background(), nil) |
| 992 | if err != nil || out != "executed" || target.calls != 1 { |
| 993 | t.Fatalf("second Execute = (%q,%v), calls=%d, want execution after metadata refresh retry", out, err, target.calls) |
| 994 | } |
| 995 | } |
| 996 | |
| 997 | func TestLazyToolDemotesStaleReaderBeforeExecution(t *testing.T) { |
| 998 | const name = "mcp__srv__mutate" |
| 999 | target := &mutableLazyTarget{name: name} |
| 1000 | shared := &lazySpawn{ |
| 1001 | spec: Spec{Name: "srv"}, |
| 1002 | state: spawnReady, |
| 1003 | real: map[string]tool.Tool{name: target}, |
| 1004 | swapped: true, |
| 1005 | } |
| 1006 | lazy := &lazyTool{ |
| 1007 | shared: shared, name: name, rawName: "mutate", readOnly: true, hasCache: true, |
| 1008 | } |
| 1009 | |
| 1010 | if out, err := lazy.Execute(context.Background(), nil); err == nil || !strings.Contains(err.Error(), "Plan/read-only safety boundary") || out != "" { |
| 1011 | t.Fatalf("first Execute = (%q,%v), want retry before writer execution", out, err) |
| 1012 | } |
| 1013 | if target.calls != 0 || lazy.ReadOnly() { |
| 1014 | t.Fatalf("after demotion calls=%d readOnly=%v, want 0/false", target.calls, lazy.ReadOnly()) |
| 1015 | } |
| 1016 | |
| 1017 | out, err := lazy.Execute(context.Background(), nil) |
| 1018 | if err != nil || out != "executed" || target.calls != 1 { |
| 1019 | t.Fatalf("second Execute = (%q,%v), calls=%d", out, err, target.calls) |
| 1020 | } |
| 1021 | } |
| 1022 | |
| 1023 | // TestLazyEmptyCachedToolsFallsBackToConnectStub: a snapshot with zero tools |
| 1024 | // presents nothing the model could call, so it must take the cache-miss stub |
| 1025 | // path instead of letting live tools join the registry mid-session unnamed. |
| 1026 | func TestLazyEmptyCachedToolsFallsBackToConnectStub(t *testing.T) { |
| 1027 | redirectCache(t) |
| 1028 | spec := helperSpec() |
| 1029 | cs := &CachedSchema{CacheKey: SchemaCacheKey(spec), Tools: nil} |
| 1030 | |
| 1031 | host := NewHost() |
| 1032 | defer host.Close() |
| 1033 | reg := tool.NewRegistry() |
| 1034 | ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second) |
| 1035 | defer cancel() |
| 1036 | |
| 1037 | tools := LazyToolset(spec, cs, host, reg, ctx, false) |
| 1038 | if len(tools) != 1 || tools[0].Name() != "mcp__mock__connect" { |
| 1039 | t.Fatalf("empty-cache toolset = %v, want single connect stub", tools) |
| 1040 | } |
| 1041 | } |
| 1042 | |
| 1043 | // TestAddWithLifecycleSurvivesHandshakeCtxCancel proves the on-demand proxy |
| 1044 | // pattern: connect with a short handshake budget, cancel it immediately after |
| 1045 | // connect, and the stdio child must stay alive (its lifetime is lifeCtx) so |
| 1046 | // the tool call that triggered the connect can still execute. |
| 1047 | func TestAddWithLifecycleSurvivesHandshakeCtxCancel(t *testing.T) { |
| 1048 | spec := helperSpec() |
| 1049 | host := NewHost() |
| 1050 | defer host.Close() |
| 1051 | |
| 1052 | lifeCtx := t.Context() |
| 1053 | handshakeCtx, cancelHandshake := context.WithTimeout(context.Background(), 5*time.Second) |
| 1054 | tools, err := host.AddWithLifecycle(lifeCtx, handshakeCtx, spec) |
| 1055 | cancelHandshake() // the proxy's deferred cancel fires right after connect |
| 1056 | if err != nil { |
| 1057 | t.Fatalf("AddWithLifecycle: %v", err) |
| 1058 | } |
| 1059 | var echo tool.Tool |
| 1060 | for _, tl := range tools { |
| 1061 | if strings.HasSuffix(tl.Name(), "__echo") { |
| 1062 | echo = tl |
| 1063 | } |
| 1064 | } |
| 1065 | if echo == nil { |
| 1066 | t.Fatalf("no echo tool in %d tools", len(tools)) |
| 1067 | } |
| 1068 | out, err := echo.Execute(context.Background(), json.RawMessage(`{"msg":"hi"}`)) |
| 1069 | if err != nil { |
| 1070 | t.Fatalf("Execute after handshake ctx cancel: %v — the child died with the handshake context", err) |
| 1071 | } |
| 1072 | if out != "echo: hi" { |
| 1073 | t.Fatalf("Execute result = %q, want %q", out, "echo: hi") |
| 1074 | } |
| 1075 | } |
| 1076 |