| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "context" |
| 6 | "io" |
| 7 | "net/http" |
| 8 | "net/http/httptest" |
| 9 | "sync" |
| 10 | "testing" |
| 11 | |
| 12 | "reasonix/internal/event" |
| 13 | "reasonix/internal/provider" |
| 14 | "reasonix/internal/provider/responses" |
| 15 | ) |
| 16 | |
| 17 | const responsesToolWithoutReasoningSSE = `data: {"type":"response.output_item.added","item":{"id":"fc_1","type":"function_call","call_id":"call_1","name":"echo"}} |
| 18 | |
| 19 | data: {"type":"response.output_item.done","item":{"id":"fc_1","type":"function_call","call_id":"call_1","name":"echo","arguments":"{\"text\":\"hi\"}"}} |
| 20 | |
| 21 | data: {"type":"response.completed","response":{"id":"resp_tool","usage":{"input_tokens":10,"output_tokens":2,"total_tokens":12}}} |
| 22 | |
| 23 | ` |
| 24 | |
| 25 | const responsesFinalAnswerSSE = `data: {"type":"response.output_text.delta","item_id":"msg_1","content_index":0,"delta":"done"} |
| 26 | |
| 27 | data: {"type":"response.completed","response":{"id":"resp_final","usage":{"input_tokens":12,"output_tokens":1,"total_tokens":13}}} |
| 28 | |
| 29 | ` |
| 30 | |
| 31 | func TestResponsesToolTurnWithoutReasoningContinuesSafely(t *testing.T) { |
| 32 | tests := []struct { |
| 33 | name, baseURL, model, effort string |
| 34 | wantRequests int |
| 35 | wantReasoningRetries int |
| 36 | }{ |
| 37 | {name: "deepseek flash", baseURL: "https://api.deepseek.com", model: "deepseek-v4-flash", effort: "high", wantRequests: 2}, |
| 38 | {name: "deepseek pro", baseURL: "https://api.deepseek.com", model: "deepseek-v4-pro", effort: "high", wantRequests: 2}, |
| 39 | {name: "deepseek pro reasoning disabled", baseURL: "https://api.deepseek.com", model: "deepseek-v4-pro", effort: "none", wantRequests: 2}, |
| 40 | {name: "mimo", baseURL: "https://api.xiaomimimo.com/v1", model: "mimo-v2.5-pro", effort: "high", wantRequests: 2}, |
| 41 | } |
| 42 | |
| 43 | for _, tc := range tests { |
| 44 | t.Run(tc.name, func(t *testing.T) { |
| 45 | var mu sync.Mutex |
| 46 | var bodies [][]byte |
| 47 | srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { |
| 48 | body, err := io.ReadAll(r.Body) |
| 49 | if err != nil { |
| 50 | t.Errorf("read request: %v", err) |
| 51 | w.WriteHeader(http.StatusInternalServerError) |
| 52 | return |
| 53 | } |
| 54 | mu.Lock() |
| 55 | bodies = append(bodies, append([]byte(nil), body...)) |
| 56 | mu.Unlock() |
| 57 | |
| 58 | w.Header().Set("Content-Type", "text/event-stream") |
| 59 | if bytes.Contains(body, []byte(`"type":"function_call_output"`)) { |
| 60 | _, _ = io.WriteString(w, responsesFinalAnswerSSE) |
| 61 | return |
| 62 | } |
| 63 | _, _ = io.WriteString(w, responsesToolWithoutReasoningSSE) |
| 64 | })) |
| 65 | defer srv.Close() |
| 66 | |
| 67 | prov := responses.New(responses.Config{ |
| 68 | Name: "responses-replay", APIKey: "test-key", BaseURL: tc.baseURL, |
| 69 | RequestURL: srv.URL, Model: tc.model, Effort: tc.effort, |
| 70 | }) |
| 71 | sink := &recordSink{} |
| 72 | a := New(prov, echoRegistry(), NewSession(""), Options{MissingReasoningWarnStateDir: t.TempDir()}, sink) |
| 73 | if err := a.Run(withNoClosedLoop(context.Background()), "go"); err != nil { |
| 74 | t.Fatalf("Run: %v", err) |
| 75 | } |
| 76 | |
| 77 | mu.Lock() |
| 78 | gotBodies := append([][]byte(nil), bodies...) |
| 79 | mu.Unlock() |
| 80 | if len(gotBodies) != tc.wantRequests { |
| 81 | t.Fatalf("HTTP requests = %d, want %d", len(gotBodies), tc.wantRequests) |
| 82 | } |
| 83 | if tc.wantReasoningRetries == 1 && !bytes.Equal(gotBodies[0], gotBodies[1]) { |
| 84 | t.Fatal("missing-reasoning probe did not replay the exact frozen request") |
| 85 | } |
| 86 | if got := sink.recoveryCount(event.ProtocolRecoveryMissingReasoningRetryAttempted); got != tc.wantReasoningRetries { |
| 87 | t.Fatalf("reasoning retries = %d, want %d", got, tc.wantReasoningRetries) |
| 88 | } |
| 89 | if got := len(sink.kinds(event.ToolResult)); got != 1 { |
| 90 | t.Fatalf("tool results = %d, want exactly one execution", got) |
| 91 | } |
| 92 | |
| 93 | var toolTurns int |
| 94 | for _, message := range a.Session().Snapshot() { |
| 95 | if message.Role == provider.RoleAssistant && len(message.ToolCalls) > 0 { |
| 96 | toolTurns++ |
| 97 | if message.ReasoningContent != "" { |
| 98 | t.Fatalf("missing provider reasoning was fabricated as %q", message.ReasoningContent) |
| 99 | } |
| 100 | } |
| 101 | } |
| 102 | if toolTurns != 1 { |
| 103 | t.Fatalf("saved tool turns = %d, want 1", toolTurns) |
| 104 | } |
| 105 | }) |
| 106 | } |
| 107 | } |
| 108 |