| 1 | package main |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "strings" |
| 7 | "sync/atomic" |
| 8 | "testing" |
| 9 | "time" |
| 10 | |
| 11 | "reasonix/internal/agent" |
| 12 | agenttest "reasonix/internal/agent/testutil" |
| 13 | "reasonix/internal/control" |
| 14 | "reasonix/internal/event" |
| 15 | "reasonix/internal/provider" |
| 16 | "reasonix/internal/tool" |
| 17 | ) |
| 18 | |
| 19 | type desktopCountingTool struct { |
| 20 | name string |
| 21 | readOnly bool |
| 22 | calls *int32 |
| 23 | } |
| 24 | |
| 25 | func (t desktopCountingTool) Name() string { return t.name } |
| 26 | func (t desktopCountingTool) Description() string { return "test tool" } |
| 27 | func (t desktopCountingTool) Schema() json.RawMessage { |
| 28 | return json.RawMessage(`{"type":"object","properties":{"command":{"type":"string"}}}`) |
| 29 | } |
| 30 | func (t desktopCountingTool) ReadOnly() bool { return t.readOnly } |
| 31 | func (t desktopCountingTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 32 | atomic.AddInt32(t.calls, 1) |
| 33 | return "ok", nil |
| 34 | } |
| 35 | |
| 36 | func TestDesktopE2EBlocksRepeatedSuccessfulBashFileWrite(t *testing.T) { |
| 37 | if testing.Short() { |
| 38 | t.Skip("skipping desktop E2E repeat-guard test in short mode") |
| 39 | } |
| 40 | |
| 41 | var calls int32 |
| 42 | reg := tool.NewRegistry() |
| 43 | reg.Add(desktopCountingTool{name: "bash", calls: &calls}) |
| 44 | args := `{"command":"python -c \"with open('prompt.txt', 'w') as f: f.write('hello')\""}` |
| 45 | prov := agenttest.NewMock("scripted-desktop", |
| 46 | agenttest.Turn{ToolCalls: []provider.ToolCall{{ID: "c1", Name: "bash", Arguments: args}}}, |
| 47 | agenttest.Turn{ToolCalls: []provider.ToolCall{{ID: "c2", Name: "bash", Arguments: args}}}, |
| 48 | agenttest.Turn{ToolCalls: []provider.ToolCall{{ID: "c3", Name: "bash", Arguments: args}}}, |
| 49 | agenttest.Turn{Text: "done"}, |
| 50 | ) |
| 51 | events := make(chan event.Event, 32) |
| 52 | sink := event.FuncSink(func(e event.Event) { events <- e }) |
| 53 | ag := agent.New(prov, reg, agent.NewSession(""), agent.Options{}, sink) |
| 54 | ctrl := control.New(control.Options{Runner: ag, Executor: ag, Sink: sink}) |
| 55 | app := NewApp() |
| 56 | app.setTestCtrl(ctrl, "scripted-desktop") |
| 57 | |
| 58 | app.SubmitToTab("test", "update the prompt file") |
| 59 | |
| 60 | var results []event.Event |
| 61 | deadline := time.After(5 * time.Second) |
| 62 | for { |
| 63 | select { |
| 64 | case e := <-events: |
| 65 | if e.Kind == event.ToolResult { |
| 66 | results = append(results, e) |
| 67 | } |
| 68 | if e.Kind == event.TurnDone { |
| 69 | if e.Err != nil { |
| 70 | t.Fatalf("turn failed: %v", e.Err) |
| 71 | } |
| 72 | if got := atomic.LoadInt32(&calls); got != 2 { |
| 73 | t.Fatalf("bash executed %d times, want 2 before the repeat guard blocks", got) |
| 74 | } |
| 75 | if len(results) != 3 { |
| 76 | t.Fatalf("tool results = %d, want 3", len(results)) |
| 77 | } |
| 78 | last := results[len(results)-1].Tool.Output |
| 79 | if !strings.Contains(last, "[loop guard]") || !strings.Contains(last, "edit_file") { |
| 80 | t.Fatalf("third repeated write should nudge the model to change tools, got %q", last) |
| 81 | } |
| 82 | return |
| 83 | } |
| 84 | case <-deadline: |
| 85 | t.Fatal("timed out waiting for desktop turn to finish") |
| 86 | } |
| 87 | } |
| 88 | } |
| 89 |