| 1 | package plugin |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "strings" |
| 7 | "testing" |
| 8 | "time" |
| 9 | |
| 10 | mcpsdk "github.com/modelcontextprotocol/go-sdk/mcp" |
| 11 | "reasonix/internal/tool" |
| 12 | ) |
| 13 | |
| 14 | // appsFixtureServer advertises the Apps extension and serves tools with |
| 15 | // visibility/_meta.ui metadata, capturing tools/list as the client sees it. |
| 16 | type appsFixtureServer struct { |
| 17 | } |
| 18 | |
| 19 | func (f *appsFixtureServer) server(t *testing.T, advertiseApps bool) *mcpsdk.Server { |
| 20 | t.Helper() |
| 21 | capabilities := &mcpsdk.ServerCapabilities{} |
| 22 | if advertiseApps { |
| 23 | capabilities.AddExtension(AppsUIExtensionID, map[string]any{"mimeTypes": []any{AppsMimeType}}) |
| 24 | } |
| 25 | server := mcpsdk.NewServer(&mcpsdk.Implementation{Name: "apps-fixture", Version: "1"}, &mcpsdk.ServerOptions{ |
| 26 | Capabilities: capabilities, |
| 27 | }) |
| 28 | addAppTool := func(name string, meta map[string]any) { |
| 29 | t := &mcpsdk.Tool{ |
| 30 | Name: name, |
| 31 | Description: name + " tool", |
| 32 | InputSchema: map[string]any{"type": "object", "properties": map[string]any{}}, |
| 33 | Meta: mcpsdk.Meta(meta), |
| 34 | } |
| 35 | server.AddTool(t, func(ctx context.Context, req *mcpsdk.CallToolRequest) (*mcpsdk.CallToolResult, error) { |
| 36 | return &mcpsdk.CallToolResult{ |
| 37 | Content: []mcpsdk.Content{&mcpsdk.TextContent{Text: name + " done"}}, |
| 38 | StructuredContent: map[string]any{"ok": true, "tool": name}, |
| 39 | }, nil |
| 40 | }) |
| 41 | } |
| 42 | addAppTool("both_tool", map[string]any{}) |
| 43 | // Top-level visibility is retained for pre-stable servers. |
| 44 | addAppTool("model_only", map[string]any{"visibility": []string{"model"}}) |
| 45 | addAppTool("app_only", map[string]any{ |
| 46 | "ui": map[string]any{ |
| 47 | "visibility": []string{"app"}, |
| 48 | "resourceUri": "ui://stable/index.html", |
| 49 | }, |
| 50 | }) |
| 51 | addAppTool("app_rich", map[string]any{ |
| 52 | "ui": map[string]any{ |
| 53 | "resourceUri": "ui://app/rich.html", |
| 54 | "csp": map[string]any{"connect-src": []string{"https://api.example.com"}}, |
| 55 | "visibility": []string{"model", "app"}, |
| 56 | }, |
| 57 | }) |
| 58 | addAppTool("nested_wins", map[string]any{ |
| 59 | "visibility": []string{"model"}, |
| 60 | "ui": map[string]any{ |
| 61 | "visibility": []string{"app"}, |
| 62 | }, |
| 63 | }) |
| 64 | server.AddTool(&mcpsdk.Tool{ |
| 65 | Name: "complete_result", |
| 66 | Description: "complete CallToolResult fixture", |
| 67 | InputSchema: map[string]any{"type": "object", "properties": map[string]any{}}, |
| 68 | Meta: mcpsdk.Meta{"ui": map[string]any{"visibility": []string{"app"}}}, |
| 69 | }, func(context.Context, *mcpsdk.CallToolRequest) (*mcpsdk.CallToolResult, error) { |
| 70 | size := int64(12) |
| 71 | return &mcpsdk.CallToolResult{ |
| 72 | Meta: mcpsdk.Meta{"trace": "app-call"}, |
| 73 | Content: []mcpsdk.Content{ |
| 74 | &mcpsdk.TextContent{Text: "complete result"}, |
| 75 | &mcpsdk.ResourceLink{URI: "https://example.test/result", Name: "result", MIMEType: "application/json", Size: &size, Meta: mcpsdk.Meta{"resource": "metadata"}}, |
| 76 | }, |
| 77 | StructuredContent: map[string]any{"ok": false, "reason": "fixture"}, |
| 78 | IsError: true, |
| 79 | }, nil |
| 80 | }) |
| 81 | server.AddResource( |
| 82 | &mcpsdk.Resource{URI: "ui://app/rich.html", Name: "rich-app", MIMEType: "text/html;profile=mcp-app"}, |
| 83 | func(context.Context, *mcpsdk.ReadResourceRequest) (*mcpsdk.ReadResourceResult, error) { |
| 84 | return &mcpsdk.ReadResourceResult{Contents: []*mcpsdk.ResourceContents{{ |
| 85 | URI: "ui://app/rich.html", MIMEType: "text/html;profile=mcp-app", Text: "<html>rich</html>", |
| 86 | Meta: mcpsdk.Meta{"ui": map[string]any{"csp": map[string]any{ |
| 87 | "connectDomains": []string{"https://api.resource.example"}, |
| 88 | "resourceDomains": []string{"https://cdn.resource.example"}, |
| 89 | }}}, |
| 90 | }}}, nil |
| 91 | }, |
| 92 | ) |
| 93 | return server |
| 94 | } |
| 95 | |
| 96 | // startAppsClient connects a desktop-profile client through the real build |
| 97 | // path and returns the started Client plus the fixture for assertions. |
| 98 | func startAppsClientWithAgreement(t *testing.T, advertiseApps bool) (*Host, *Client, toolCatalogSnapshot, *appsFixtureServer) { |
| 99 | t.Helper() |
| 100 | fixture := &appsFixtureServer{} |
| 101 | host := NewHostWithProfile(HostProfileDesktopApps) |
| 102 | lifeCtx, cancel := context.WithCancel(context.Background()) |
| 103 | transport := &sdkSessionTransport{ |
| 104 | name: "apps-fixture", |
| 105 | spec: Spec{Name: "apps-fixture", Type: "http", StartupTimeout: 2 * time.Second}, |
| 106 | profile: HostProfileDesktopApps, |
| 107 | lifeCtx: lifeCtx, |
| 108 | cancel: cancel, |
| 109 | state: SessionStateConnecting, |
| 110 | reconnectDelays: []time.Duration{time.Millisecond}, |
| 111 | } |
| 112 | transport.endpointFactory = func(ctx context.Context) (sdkEndpoint, error) { |
| 113 | clientSide, serverSide := mcpsdk.NewInMemoryTransports() |
| 114 | go func() { _ = fixture.server(t, advertiseApps).Run(ctx, serverSide) }() |
| 115 | return sdkEndpoint{transport: clientSide}, nil |
| 116 | } |
| 117 | t.Cleanup(transport.close) |
| 118 | client := &Client{name: "apps-fixture", t: transport, spec: Spec{Name: "apps-fixture"}, profile: HostProfileDesktopApps, transport: "http"} |
| 119 | if err := client.initialize(t.Context()); err != nil { |
| 120 | t.Fatalf("initialize: %v", err) |
| 121 | } |
| 122 | if _, err := client.listTools(t.Context()); err != nil { |
| 123 | t.Fatalf("tools/list: %v", err) |
| 124 | } |
| 125 | client.toolsMu.RLock() |
| 126 | snapshot := client.toolCatalog |
| 127 | snapshot.infos = append([]ToolInfo(nil), snapshot.infos...) |
| 128 | snapshot.adapters = append([]tool.Tool(nil), snapshot.adapters...) |
| 129 | snapshot.appAdapters = append([]tool.Tool(nil), snapshot.appAdapters...) |
| 130 | client.toolsMu.RUnlock() |
| 131 | return host, client, snapshot, fixture |
| 132 | } |
| 133 | |
| 134 | func startAppsClient(t *testing.T) (*Host, *Client, toolCatalogSnapshot, *appsFixtureServer) { |
| 135 | t.Helper() |
| 136 | return startAppsClientWithAgreement(t, true) |
| 137 | } |
| 138 | |
| 139 | func TestMetaVisibilitySplitsCatalogs(t *testing.T) { |
| 140 | _, _, catalog, _ := startAppsClient(t) |
| 141 | |
| 142 | var modelNames, appNames []string |
| 143 | for _, tl := range catalog.adapters { |
| 144 | modelNames = append(modelNames, tl.Name()) |
| 145 | } |
| 146 | for _, tl := range catalog.appAdapters { |
| 147 | appNames = append(appNames, tl.Name()) |
| 148 | } |
| 149 | joined := strings.Join(modelNames, ",") |
| 150 | for _, banned := range []string{"app_only", "nested_wins", "complete_result"} { |
| 151 | if strings.Contains(joined, banned) { |
| 152 | t.Fatalf("model catalog contains %s: %v", banned, modelNames) |
| 153 | } |
| 154 | } |
| 155 | for _, want := range []string{"both_tool", "model_only", "app_rich"} { |
| 156 | if !strings.Contains(joined, want) { |
| 157 | t.Fatalf("model catalog missing %s: %v", want, modelNames) |
| 158 | } |
| 159 | } |
| 160 | appJoined := strings.Join(appNames, ",") |
| 161 | for _, want := range []string{"both_tool", "app_only", "app_rich", "nested_wins", "complete_result"} { |
| 162 | if !strings.Contains(appJoined, want) { |
| 163 | t.Fatalf("app catalog missing %s: %v", want, appNames) |
| 164 | } |
| 165 | } |
| 166 | if strings.Contains(appJoined, "model_only") { |
| 167 | t.Fatalf("app catalog contains model-only tool: %v", appNames) |
| 168 | } |
| 169 | // ToolInfo (use_capability list source) must also exclude app-only. |
| 170 | for _, info := range catalog.infos { |
| 171 | if info.Name == "app_only" { |
| 172 | t.Fatal("app-only tool visible in ToolInfo list") |
| 173 | } |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | func TestMetaUIResourceNestedAndFlat(t *testing.T) { |
| 178 | _, _, catalog, _ := startAppsClient(t) |
| 179 | byName := map[string]tool.Tool{} |
| 180 | for _, tl := range catalog.appAdapters { |
| 181 | byName[tl.Name()] = tl |
| 182 | } |
| 183 | rich, ok := byName[toolName("apps-fixture", "app_rich")].(*remoteTool) |
| 184 | if !ok || rich.UIResourceURI() != "ui://app/rich.html" { |
| 185 | t.Fatalf("nested ui.resourceUri not parsed: %+v", rich) |
| 186 | } |
| 187 | if len(rich.UICSP()["connect-src"]) != 1 || rich.UICSP()["connect-src"][0] != "https://api.example.com" { |
| 188 | t.Fatalf("csp not parsed: %v", rich.UICSP()) |
| 189 | } |
| 190 | legacy, ok := byName[toolName("apps-fixture", "app_only")].(*remoteTool) |
| 191 | if !ok || legacy.UIResourceURI() != "ui://stable/index.html" { |
| 192 | t.Fatalf("stable nested resourceUri not parsed: %+v", legacy) |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | func TestAppsRequireTwoWayExtensionAgreement(t *testing.T) { |
| 197 | host, client, catalog, _ := startAppsClientWithAgreement(t, false) |
| 198 | if client.appsNegotiated() { |
| 199 | t.Fatal("Apps negotiated without the server extension") |
| 200 | } |
| 201 | if len(catalog.appAdapters) != 0 { |
| 202 | t.Fatalf("app catalog populated without agreement: %v", catalog.appAdapters) |
| 203 | } |
| 204 | for _, info := range catalog.infos { |
| 205 | if info.Name == "app_only" || info.Name == "nested_wins" || info.Name == "complete_result" { |
| 206 | t.Fatalf("stable app-only tool leaked to the model catalog: %q", info.Name) |
| 207 | } |
| 208 | } |
| 209 | host.mu.Lock() |
| 210 | host.clients = append(host.clients, client) |
| 211 | host.mu.Unlock() |
| 212 | inst := host.RegisterAppInstance("apps-fixture", "app_rich", catalog.generation, "call", "ui://app/rich.html") |
| 213 | if _, ok := host.AppInstanceResourceDescriptor(inst.Token); ok { |
| 214 | t.Fatal("App resource opened without two-way extension agreement") |
| 215 | } |
| 216 | var rich *remoteTool |
| 217 | for _, candidate := range catalog.adapters { |
| 218 | if rt, ok := candidate.(*remoteTool); ok && rt.rawName == "app_rich" { |
| 219 | rich = rt |
| 220 | } |
| 221 | } |
| 222 | if rich == nil { |
| 223 | t.Fatal("model-visible app_rich tool not found") |
| 224 | } |
| 225 | ctx, collector := tool.WithMCPAppCollector(t.Context()) |
| 226 | if _, _, err := rich.ExecuteWithImages(ctx, json.RawMessage(`{}`)); err != nil { |
| 227 | t.Fatal(err) |
| 228 | } |
| 229 | if collector.Server != "" { |
| 230 | t.Fatal("rich presentation stamped without two-way extension agreement") |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | func TestAppCallResultPreservesStandardFields(t *testing.T) { |
| 235 | _, _, catalog, _ := startAppsClient(t) |
| 236 | var complete *remoteTool |
| 237 | for _, candidate := range catalog.appAdapters { |
| 238 | if rt, ok := candidate.(*remoteTool); ok && rt.rawName == "complete_result" { |
| 239 | complete = rt |
| 240 | } |
| 241 | } |
| 242 | if complete == nil { |
| 243 | t.Fatal("complete_result App tool not found") |
| 244 | } |
| 245 | raw, text, reportedError, err := complete.ExecuteForApp(t.Context(), json.RawMessage(`{}`)) |
| 246 | if err != nil { |
| 247 | t.Fatal(err) |
| 248 | } |
| 249 | if text != "complete result" || !reportedError { |
| 250 | t.Fatalf("host projection = %q, isError=%v", text, reportedError) |
| 251 | } |
| 252 | var result map[string]any |
| 253 | if err := json.Unmarshal(raw, &result); err != nil { |
| 254 | t.Fatal(err) |
| 255 | } |
| 256 | content, _ := result["content"].([]any) |
| 257 | resource, _ := content[1].(map[string]any) |
| 258 | meta, _ := resource["_meta"].(map[string]any) |
| 259 | if result["isError"] != true || result["structuredContent"] == nil || result["_meta"] == nil || meta["resource"] != "metadata" { |
| 260 | t.Fatalf("complete CallToolResult was not preserved: %s", raw) |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | func TestRichResultStampedOnCallContext(t *testing.T) { |
| 265 | _, _, catalog, _ := startAppsClient(t) |
| 266 | var rich *remoteTool |
| 267 | for _, tl := range catalog.adapters { |
| 268 | if rt, ok := tl.(*remoteTool); ok && rt.rawName == "app_rich" { |
| 269 | rich = rt |
| 270 | break |
| 271 | } |
| 272 | } |
| 273 | if rich == nil { |
| 274 | t.Fatal("app_rich not found") |
| 275 | } |
| 276 | ctx, collector := tool.WithMCPAppCollector(t.Context()) |
| 277 | out, _, err := rich.ExecuteWithImages(ctx, json.RawMessage(`{}`)) |
| 278 | if err != nil { |
| 279 | t.Fatalf("execute: %v", err) |
| 280 | } |
| 281 | if !strings.Contains(out, "app_rich done") { |
| 282 | t.Fatalf("text form lost: %q", out) |
| 283 | } |
| 284 | stamped := collector.Sanitized() |
| 285 | if stamped == nil || stamped.Server == "" { |
| 286 | t.Fatal("Apps presentation not collected") |
| 287 | } |
| 288 | if stamped.Server != "apps-fixture" || stamped.Tool != "app_rich" || stamped.ResourceURI != "ui://app/rich.html" { |
| 289 | t.Fatalf("stamped identity = %+v", stamped) |
| 290 | } |
| 291 | if len(stamped.Structured) == 0 || !strings.Contains(string(stamped.Structured), `"tool":"app_rich"`) { |
| 292 | t.Fatalf("structured content not captured: %s", stamped.Structured) |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | func TestAppInstanceRegistryBoundAndReclaimed(t *testing.T) { |
| 297 | host := NewHostWithProfile(HostProfileDesktopApps) |
| 298 | reg := host.appInstances |
| 299 | inst := host.RegisterAppInstance("srv", "tool", 3, "call-1", "ui://x/a.html") |
| 300 | if len(inst.Token) != 48 { |
| 301 | t.Fatalf("token length = %d, want 48 hex chars", len(inst.Token)) |
| 302 | } |
| 303 | if got, ok := host.LookupAppInstance(inst.Token); !ok || got.Server != "srv" { |
| 304 | t.Fatalf("lookup failed: %+v %v", got, ok) |
| 305 | } |
| 306 | for range maxAppInstances + 4 { |
| 307 | host.RegisterAppInstance("srv", "tool", 3, "call", "ui://x/b.html") |
| 308 | } |
| 309 | if reg.Len() > maxAppInstances { |
| 310 | t.Fatalf("registry exceeded bound: %d", reg.Len()) |
| 311 | } |
| 312 | if _, ok := host.LookupAppInstance(inst.Token); ok { |
| 313 | t.Fatal("oldest instance not evicted at capacity") |
| 314 | } |
| 315 | other := host.RegisterAppInstance("other", "t", 1, "c", "ui://y/a.html") |
| 316 | host.appInstances.ReleaseServer("other") |
| 317 | if _, ok := reg.Lookup(other.Token); ok { |
| 318 | t.Fatal("release-server did not reclaim the server's instances") |
| 319 | } |
| 320 | } |
| 321 | |
| 322 | func TestAppInstanceReleaseCancelsNestedCalls(t *testing.T) { |
| 323 | host := NewHostWithProfile(HostProfileDesktopApps) |
| 324 | inst := host.RegisterAppInstance("srv", "tool", 3, "call", "ui://x/a.html") |
| 325 | ctx, ok := host.AppInstanceContext(inst.Token) |
| 326 | if !ok || ctx.Err() != nil { |
| 327 | t.Fatal("live App instance has no call context") |
| 328 | } |
| 329 | host.ReleaseAppInstance(inst.Token) |
| 330 | select { |
| 331 | case <-ctx.Done(): |
| 332 | case <-time.After(time.Second): |
| 333 | t.Fatal("releasing App instance did not cancel nested calls") |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | func TestReadResourceForAppReturnsResourceLevelCSP(t *testing.T) { |
| 338 | host, client, snapshot, _ := startAppsClient(t) |
| 339 | host.mu.Lock() |
| 340 | host.clients = append(host.clients, client) |
| 341 | host.mu.Unlock() |
| 342 | inst := host.RegisterAppInstance("apps-fixture", "app_rich", snapshot.generation, "call", "ui://app/rich.html") |
| 343 | if csp, ok := host.AppInstanceResourceDescriptor(inst.Token); !ok || len(csp["connect-src"]) != 1 { |
| 344 | t.Fatalf("tool resource descriptor = %#v, %v", csp, ok) |
| 345 | } |
| 346 | wrong := host.RegisterAppInstance("apps-fixture", "app_rich", snapshot.generation, "call", "ui://app/other.html") |
| 347 | if _, ok := host.AppInstanceResourceDescriptor(wrong.Token); ok { |
| 348 | t.Fatal("mismatched resource URI was accepted") |
| 349 | } |
| 350 | content, mime, csp, err := client.readResourceWithMime(t.Context(), "ui://app/rich.html") |
| 351 | if err != nil { |
| 352 | t.Fatal(err) |
| 353 | } |
| 354 | if content != "<html>rich</html>" || mime != "text/html;profile=mcp-app" { |
| 355 | t.Fatalf("content/mime = %q %q", content, mime) |
| 356 | } |
| 357 | if got := csp["connectDomains"]; len(got) != 1 || got[0] != "https://api.resource.example" { |
| 358 | t.Fatalf("resource CSP = %#v", csp) |
| 359 | } |
| 360 | } |
| 361 | |
| 362 | func TestAppInstanceRegistryFreezesResourceAndBoundsSnapshotMemory(t *testing.T) { |
| 363 | host := NewHostWithProfile(HostProfileDesktopApps) |
| 364 | csp := map[string][]string{"connect-src": {"https://api.example.test"}} |
| 365 | first := host.RegisterAppInstance("srv", "tool", 3, "call-1", "ui://x/a.html") |
| 366 | if !host.BindAppResource(first.Token, "<html>first</html>", "text/html", "digest-1", csp) { |
| 367 | t.Fatal("bind first resource failed") |
| 368 | } |
| 369 | csp["connect-src"][0] = "https://mutated.example.test" |
| 370 | snapshot, ok := host.AppResource(first.Token) |
| 371 | if !ok || snapshot.Digest != "digest-1" || snapshot.Content != "<html>first</html>" { |
| 372 | t.Fatalf("snapshot = %+v, %v", snapshot, ok) |
| 373 | } |
| 374 | if got := snapshot.CSP["connect-src"][0]; got != "https://api.example.test" { |
| 375 | t.Fatalf("snapshot CSP was not copied: %q", got) |
| 376 | } |
| 377 | |
| 378 | chunk := strings.Repeat("x", 1<<20) |
| 379 | for i := range 17 { |
| 380 | inst := host.RegisterAppInstance("srv", "tool", 3, "call", "ui://x/b.html") |
| 381 | if !host.BindAppResource(inst.Token, chunk, "text/html", "digest", nil) { |
| 382 | t.Fatalf("bind resource %d failed", i) |
| 383 | } |
| 384 | } |
| 385 | if host.appInstances.bytes > maxAppResourceRegistryBytes { |
| 386 | t.Fatalf("snapshot bytes = %d, limit = %d", host.appInstances.bytes, maxAppResourceRegistryBytes) |
| 387 | } |
| 388 | if _, ok := host.AppResource(first.Token); ok { |
| 389 | t.Fatal("oldest snapshot was not evicted by the aggregate memory budget") |
| 390 | } |
| 391 | tooLarge := host.RegisterAppInstance("srv", "tool", 3, "call", "ui://x/large.html") |
| 392 | if host.BindAppResource(tooLarge.Token, strings.Repeat("z", maxAppResourceSnapshotBytes+1), "text/html", "digest", nil) { |
| 393 | t.Fatal("oversized resource snapshot was accepted") |
| 394 | } |
| 395 | metadataBomb := host.RegisterAppInstance("srv", "tool", 3, "call", "ui://x/csp.html") |
| 396 | if host.BindAppResource(metadataBomb.Token, "<html></html>", "text/html", "digest", map[string][]string{ |
| 397 | "resourceDomains": {strings.Repeat("z", maxAppResourceSnapshotBytes)}, |
| 398 | }) { |
| 399 | t.Fatal("oversized resource CSP was accepted") |
| 400 | } |
| 401 | } |
| 402 |