| 1 | package plugin_test |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "os" |
| 7 | "strings" |
| 8 | "testing" |
| 9 | "time" |
| 10 | |
| 11 | "reasonix/internal/agent" |
| 12 | "reasonix/internal/capability" |
| 13 | "reasonix/internal/plugin" |
| 14 | "reasonix/internal/tool" |
| 15 | ) |
| 16 | |
| 17 | // mockSpec mirrors helperSpec (lazy_test.go): the test binary re-runs itself |
| 18 | // as a stdio MCP server via TestHelperProcess. |
| 19 | func mockSpec() plugin.Spec { |
| 20 | return plugin.Spec{ |
| 21 | Name: "mock", |
| 22 | Command: os.Args[0], |
| 23 | Args: []string{"-test.run=TestHelperProcess", "--"}, |
| 24 | Env: map[string]string{"GO_WANT_HELPER_PROCESS": "1"}, |
| 25 | Authorized: true, // install/boot sets this; proxy never invents trust |
| 26 | } |
| 27 | } |
| 28 | |
| 29 | // TestUseCapabilityFirstDiscoveryConnectListCall is the reviewer-demanded end |
| 30 | // to end: no schema cache, action=call on the server id resolves to a gated |
| 31 | // connect target (no process before approval), Execute connects and lists the |
| 32 | // tools, and a follow-up mcp-tool call executes against the live server. |
| 33 | func TestUseCapabilityFirstDiscoveryConnectListCall(t *testing.T) { |
| 34 | host := plugin.NewHost() |
| 35 | defer host.Close() |
| 36 | lifeCtx := t.Context() |
| 37 | ledger := capability.NewLedger() |
| 38 | audit := &capability.Audit{} |
| 39 | proxy := agent.NewUseCapabilityTool(lifeCtx, host, []plugin.Spec{mockSpec()}, tool.NewRegistry(), ledger, audit, nil) |
| 40 | |
| 41 | resolved, err := proxy.ResolveCall(context.Background(), json.RawMessage(`{"action":"call","capability_id":"mcp-server:mock"}`)) |
| 42 | if err != nil { |
| 43 | t.Fatalf("resolve server call: %v", err) |
| 44 | } |
| 45 | if resolved.SkipExecute || resolved.Target == nil { |
| 46 | t.Fatalf("expected a deferred connect target, got %+v", resolved) |
| 47 | } |
| 48 | if resolved.TargetName != plugin.MCPConnectPermissionName("mock") { |
| 49 | t.Fatalf("connect permission name = %q, want %q", resolved.TargetName, plugin.MCPConnectPermissionName("mock")) |
| 50 | } |
| 51 | if resolved.ReadOnly { |
| 52 | t.Fatal("connect must not claim read-only: it spawns a subprocess") |
| 53 | } |
| 54 | if host.HasClient("mock") { |
| 55 | t.Fatal("resolution must not start the server") |
| 56 | } |
| 57 | |
| 58 | listing, err := resolved.Target.Execute(context.Background(), resolved.Args) |
| 59 | if err != nil { |
| 60 | t.Fatalf("connect execute: %v", err) |
| 61 | } |
| 62 | if !strings.Contains(listing, "mcp-tool:mock/echo") { |
| 63 | t.Fatalf("listing missing echo tool id:\n%s", listing) |
| 64 | } |
| 65 | if !host.HasClient("mock") { |
| 66 | t.Fatal("server should be connected after approved execute") |
| 67 | } |
| 68 | if got := proxy.ConnectedProxyTools(); len(got["mock"]) == 0 { |
| 69 | t.Fatalf("proxy snapshot empty after connect: %v", got) |
| 70 | } |
| 71 | |
| 72 | // A repeated server-level call now lists side-effect-free. |
| 73 | again, err := proxy.ResolveCall(context.Background(), json.RawMessage(`{"action":"call","capability_id":"mcp-server:mock"}`)) |
| 74 | if err != nil { |
| 75 | t.Fatal(err) |
| 76 | } |
| 77 | if !again.SkipExecute || !strings.Contains(again.Result, "mcp-tool:mock/echo") { |
| 78 | t.Fatalf("connected server call should list immediately, got %+v", again) |
| 79 | } |
| 80 | if _, err := proxy.Execute(context.Background(), json.RawMessage(`{"action":"call","capability_id":"mcp-server:mock"}`)); err != nil { |
| 81 | t.Fatalf("connected server call execute: %v", err) |
| 82 | } |
| 83 | if entry, ok := ledger.Get("mcp-server:mock"); !ok || entry.Outcome != capability.OutcomeSucceeded { |
| 84 | t.Fatalf("connected server call ledger = %+v, found=%v", entry, ok) |
| 85 | } |
| 86 | if snap := audit.Snapshot(); snap.MCPCall != 1 || snap.MCPCallFailures != 0 { |
| 87 | t.Fatalf("connected server call audit = %d/%d, want 1/0", snap.MCPCall, snap.MCPCallFailures) |
| 88 | } |
| 89 | |
| 90 | // Then the discovered tool is callable end to end. |
| 91 | call, err := proxy.ResolveCall(context.Background(), json.RawMessage(`{"action":"call","capability_id":"mcp-tool:mock/echo","arguments":{"msg":"hi"}}`)) |
| 92 | if err != nil { |
| 93 | t.Fatal(err) |
| 94 | } |
| 95 | callCtx, cancelCall := context.WithTimeout(context.Background(), 5*time.Second) |
| 96 | defer cancelCall() |
| 97 | out, err := call.Target.Execute(callCtx, call.Args) |
| 98 | if err != nil { |
| 99 | t.Fatalf("tool execute: %v", err) |
| 100 | } |
| 101 | if out != "echo: hi" { |
| 102 | t.Fatalf("echo result = %q", out) |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | // TestUseCapabilitySharedHostSnapshot covers the cross-tab gap: a server |
| 107 | // connected by another controller on the shared host must still populate this |
| 108 | // proxy's snapshot (and thus the catalog) through resolve and inspect. |
| 109 | func TestUseCapabilitySharedHostSnapshot(t *testing.T) { |
| 110 | host := plugin.NewHost() |
| 111 | defer host.Close() |
| 112 | lifeCtx := t.Context() |
| 113 | callCtx, cancelCall := context.WithTimeout(context.Background(), 5*time.Second) |
| 114 | defer cancelCall() |
| 115 | // "Tab A" connects directly (not through any proxy). |
| 116 | if _, err := host.AddWithLifecycle(lifeCtx, callCtx, mockSpec()); err != nil { |
| 117 | t.Fatalf("tab A connect: %v", err) |
| 118 | } |
| 119 | |
| 120 | // "Tab B": empty registry (auto_start=false semantics), fresh proxy. |
| 121 | proxyB := agent.NewUseCapabilityTool(lifeCtx, host, []plugin.Spec{mockSpec()}, tool.NewRegistry(), capability.NewLedger(), nil, nil) |
| 122 | resolved, err := proxyB.ResolveCall(context.Background(), json.RawMessage(`{"action":"call","capability_id":"mcp-tool:mock/echo","arguments":{}}`)) |
| 123 | if err != nil { |
| 124 | t.Fatal(err) |
| 125 | } |
| 126 | if resolved.Target == nil { |
| 127 | t.Fatalf("expected live target via shared host, got %+v", resolved) |
| 128 | } |
| 129 | if got := proxyB.ConnectedProxyTools(); len(got["mock"]) == 0 { |
| 130 | t.Fatalf("resolve on a shared-host server must build the snapshot, got %v", got) |
| 131 | } |
| 132 | |
| 133 | // A third proxy sees it via inspect too. |
| 134 | proxyC := agent.NewUseCapabilityTool(lifeCtx, host, []plugin.Spec{mockSpec()}, tool.NewRegistry(), capability.NewLedger(), nil, func() capability.Catalog { |
| 135 | return capability.Catalog{Entries: []capability.Entry{{ |
| 136 | ID: "mcp-server:mock", Kind: capability.KindMCPServer, Name: "mock", Source: "mock", Status: capability.StatusReady, |
| 137 | }}} |
| 138 | }) |
| 139 | if _, err := proxyC.Execute(context.Background(), json.RawMessage(`{"action":"inspect","capability_id":"mcp-server:mock"}`)); err != nil { |
| 140 | t.Fatal(err) |
| 141 | } |
| 142 | if got := proxyC.ConnectedProxyTools(); len(got["mock"]) == 0 { |
| 143 | t.Fatalf("inspect on a shared-host server must build the snapshot, got %v", got) |
| 144 | } |
| 145 | } |
| 146 |