| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "strings" |
| 8 | "sync" |
| 9 | "testing" |
| 10 | "time" |
| 11 | |
| 12 | "reasonix/internal/provider" |
| 13 | "reasonix/internal/tool" |
| 14 | ) |
| 15 | |
| 16 | // stubbornTool ignores its context: it returns only when released. |
| 17 | type stubbornTool struct { |
| 18 | once *sync.Once |
| 19 | started chan struct{} |
| 20 | release chan struct{} |
| 21 | } |
| 22 | |
| 23 | func (stubbornTool) Name() string { return "stubborn" } |
| 24 | func (stubbornTool) Description() string { return "ignores cancellation" } |
| 25 | func (stubbornTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 26 | func (stubbornTool) ReadOnly() bool { return true } |
| 27 | func (s stubbornTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 28 | s.once.Do(func() { close(s.started) }) |
| 29 | <-s.release |
| 30 | return "late", nil |
| 31 | } |
| 32 | |
| 33 | // fastTool reports when it has entered execution, so the test cancels only |
| 34 | // after both tools of the batch are running. |
| 35 | type fastTool struct { |
| 36 | once *sync.Once |
| 37 | started chan struct{} |
| 38 | } |
| 39 | |
| 40 | func (fastTool) Name() string { return "fast" } |
| 41 | func (fastTool) Description() string { return "always succeeds" } |
| 42 | func (fastTool) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) } |
| 43 | func (fastTool) ReadOnly() bool { return true } |
| 44 | func (f fastTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 45 | f.once.Do(func() { close(f.started) }) |
| 46 | return "ok", nil |
| 47 | } |
| 48 | |
| 49 | // A read-only parallel segment must not keep the whole turn wedged behind one |
| 50 | // tool that ignores cancellation: after the grace the batch reports that call |
| 51 | // as an unknown effect while the calls that did finish keep their results. |
| 52 | func TestParallelBatchAbandonsToolThatIgnoresCancellation(t *testing.T) { |
| 53 | oldGrace := parallelStragglerGrace |
| 54 | parallelStragglerGrace = 200 * time.Millisecond |
| 55 | t.Cleanup(func() { parallelStragglerGrace = oldGrace }) |
| 56 | |
| 57 | stub := stubbornTool{once: &sync.Once{}, started: make(chan struct{}), release: make(chan struct{})} |
| 58 | t.Cleanup(func() { close(stub.release) }) |
| 59 | fast := fastTool{once: &sync.Once{}, started: make(chan struct{})} |
| 60 | reg := tool.NewRegistry() |
| 61 | reg.Add(stub) |
| 62 | reg.Add(fast) |
| 63 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 64 | {toolCallChunk("stubborn-1", "stubborn", `{}`), toolCallChunk("fast-1", "fast", `{}`)}, |
| 65 | {{Type: provider.ChunkText, Text: "done"}}, |
| 66 | }} |
| 67 | sess := NewSession("") |
| 68 | a := New(prov, reg, sess, Options{}, &recordSink{}) |
| 69 | |
| 70 | ctx, cancel := context.WithCancel(context.Background()) |
| 71 | defer cancel() |
| 72 | done := make(chan error, 1) |
| 73 | go func() { done <- a.Run(withNoClosedLoop(ctx), "go") }() |
| 74 | select { |
| 75 | case <-stub.started: |
| 76 | case <-time.After(5 * time.Second): |
| 77 | t.Fatal("stubborn tool never started") |
| 78 | } |
| 79 | select { |
| 80 | case <-fast.started: |
| 81 | case <-time.After(5 * time.Second): |
| 82 | t.Fatal("fast tool never started") |
| 83 | } |
| 84 | cancel() |
| 85 | select { |
| 86 | case err := <-done: |
| 87 | if !errors.Is(err, context.Canceled) { |
| 88 | t.Fatalf("Run returned %v, want context.Canceled", err) |
| 89 | } |
| 90 | case <-time.After(5 * time.Second): |
| 91 | t.Fatal("cancelled batch stayed wedged behind a tool that ignores its context") |
| 92 | } |
| 93 | if got := toolResultByID(sess, "stubborn-1"); !strings.Contains(got, "did not stop after cancellation") { |
| 94 | t.Fatalf("stubborn result = %q, want the abandoned marker", got) |
| 95 | } |
| 96 | if got := toolResultByID(sess, "fast-1"); !strings.Contains(got, "ok") { |
| 97 | t.Fatalf("fast result = %q, want the finished tool's own output", got) |
| 98 | } |
| 99 | } |
| 100 |