package subagents import ( "context" "encoding/json" "fmt" "os" "path/filepath" "strings" "testing" "time" ) // TestSubagentWorkerArgs locks in the exact flag set the subprocess // runner uses to start a subagent agent in daemon mode. Past // regressions in this area: // // - "--no-sess" instead of "--no-session" (old print-mode runner): // every spawned agent died with "unknown flag" before it could // talk to the model. // // - Forgetting --cwd: the child resolved tools against the parent // zut's working directory, defeating the whole point of the // worktree isolation. // // - Forgetting --session: a daemon-mode agent without a session // file would lose context between follow-up turns, making // "send another message" mostly useless. // // The test asserts the load-bearing pieces are present in plausible // positions. If a flag is renamed, update both the runner and this // test so we notice immediately. func TestCredentialStdinHelperProcess(t *testing.T) { if os.Getenv("ZUT_SUBAGENT_CREDENTIAL_HELPER") != "1" { return } if os.Getenv("ZUT_SUBAGENT_CREDENTIAL_STDIN") != "1" { os.Exit(2) } var credential Credential if err := json.NewDecoder(os.Stdin).Decode(&credential); err != nil { os.Exit(3) } if credential.Value != "inherited-secret" || credential.Method != "apikey" { os.Exit(4) } for _, arg := range os.Args { if strings.Contains(arg, credential.Value) { os.Exit(5) } } for _, env := range os.Environ() { if strings.Contains(env, credential.Value) { os.Exit(6) } } fmt.Println(`{"type":"agent_stopped","data":{"reason":"completed"}}`) os.Exit(0) } func TestWorkerEnvironmentRedactsProviderSecrets(t *testing.T) { t.Setenv("OPENAI_API_KEY", "openai-secret") t.Setenv("CUSTOM_PROVIDER_API_KEY", "custom-secret") t.Setenv("HF_TOKEN", "hf-secret") t.Setenv("GITHUB_TOKEN", "github-secret") t.Setenv("CUSTOM_PROVIDER_TOKEN", "token-secret") t.Setenv("CUSTOM_PROVIDER_KEY", "key-secret") t.Setenv("AWS_ACCESS_KEY_ID", "aws-access-secret") t.Setenv("GOOGLE_APPLICATION_CREDENTIALS", "/tmp/service-account.json") t.Setenv("ZUT_SUBAGENT_TEST_VALUE", "not-secret") env := strings.Join(workerEnvironment("openai"), "\n") for _, secret := range []string{ "OPENAI_API_KEY=openai-secret", "CUSTOM_PROVIDER_API_KEY=custom-secret", "HF_TOKEN=hf-secret", "GITHUB_TOKEN=github-secret", "CUSTOM_PROVIDER_TOKEN=token-secret", "CUSTOM_PROVIDER_KEY=key-secret", "AWS_ACCESS_KEY_ID=aws-access-secret", "GOOGLE_APPLICATION_CREDENTIALS=/tmp/service-account.json", } { if strings.Contains(env, secret) { t.Fatalf("worker environment leaked %s", secret) } } if !strings.Contains(env, "ZUT_SUBAGENT_TEST_VALUE=not-secret") { t.Fatal("worker environment dropped unrelated variables") } t.Setenv("AWS_BEARER_TOKEN_BEDROCK", "bedrock-secret") bedrockEnv := strings.Join(workerEnvironment("amazon-bedrock"), "\n") if !strings.Contains(bedrockEnv, "AWS_BEARER_TOKEN_BEDROCK=bedrock-secret") { t.Fatal("Bedrock ambient credential chain was unexpectedly removed") } if !strings.Contains(bedrockEnv, "AWS_ACCESS_KEY_ID=aws-access-secret") { t.Fatal("Bedrock AWS access-key credential was unexpectedly removed") } } func TestExecRunnerTransfersCredentialOnlyOnStdin(t *testing.T) { t.Setenv("ZUT_SUBAGENT_CREDENTIAL_HELPER", "1") root := t.TempDir() a := &Agent{ ID: "credential-test", Dir: root, Provider: "custom", SessionPath: filepath.Join(root, "session.json"), InboxPath: filepath.Join(root, "inbox"), EventLogPath: filepath.Join(root, "events.jsonl"), } calls := 0 r := &execRunner{ agent: a, Command: []string{os.Args[0], "-test.run=^TestCredentialStdinHelperProcess$"}, resolveCredential: func(context.Context, string) (Credential, error) { calls++ return Credential{Value: "inherited-secret", Method: "apikey"}, nil }, } if err := r.Run(context.Background(), agentSink{a: a}); err != nil { t.Fatal(err) } if calls != 1 { t.Fatalf("credential resolver called %d times, want 1", calls) } } func TestApplyEventPreservesMultilineRolesAndFinalAssistant(t *testing.T) { a := &Agent{} sink := agentSink{a: a} applyEventToSink(NewEvent("user_message", map[string]any{"content": []any{ map[string]any{"type": "text", "text": "first line\n\nlast line"}, }}), sink) applyEventToSink(NewEvent("assistant_message", map[string]any{"content": []any{ map[string]any{"type": "text", "text": "complete\nanswer"}, }}), sink) snap := a.Snapshot() wantLines := []string{"user: first line", "user: ", "user: last line", "complete", "answer"} if strings.Join(snap.Lines, "\n") != strings.Join(wantLines, "\n") { t.Fatalf("transcript = %#v, want %#v", snap.Lines, wantLines) } if snap.LastAssistant != "complete\nanswer" { t.Fatalf("last assistant = %q", snap.LastAssistant) } } func TestSubagentWorkerArgs(t *testing.T) { args := subagentWorkerArgs(subagentWorkerArgsOpts{ Exe: "/path/to/zut", Dir: "/tmp/worktree", SessionPath: "/tmp/state/session.json", InboxPath: "/tmp/state/in.sock", Task: "do the thing", Reasoning: "high", Subagent: "reviewer", }) if len(args) < 7 { t.Fatalf("argv unexpectedly short: %v", args) } if args[0] != "/path/to/zut" { t.Fatalf("argv[0] = %q; want the binary path", args[0]) } // The task must come last so anything that looks flag-like in // the task body doesn't get interpreted as a flag. if args[len(args)-1] != "do the thing" { t.Fatalf("task should be last positional; got %v", args) } mustHave := map[string]string{ "--subagent-worker": "/tmp/state/in.sock", "--session": "/tmp/state/session.json", "--cwd": "/tmp/worktree", "--reasoning": "high", "--subagent": "reviewer", } for flag, value := range mustHave { i := indexOf(args, flag) if i < 0 { t.Errorf("argv missing %q: %v", flag, args) continue } if i+1 >= len(args) || args[i+1] != value { t.Errorf("argv %q value = %q; want %q", flag, safeAt(args, i+1), value) } } // Reject prior bad flags explicitly so a future revert is caught. joined := strings.Join(args, " ") for _, bad := range []string{"--print", "--no-sess ", "--no-session"} { if strings.Contains(joined, bad) { t.Fatalf("argv contains stale/wrong flag %q: %s", bad, joined) } } } func TestResolveSubagentExecutableUsesCurrentWhenAvailable(t *testing.T) { got, err := resolveSubagentExecutable(context.Background(), "/current/bin/zut", "/old/bin/zut", func(_ context.Context, candidate string) (string, error) { if candidate == "/current/bin/zut" { return candidate, nil } return "", os.ErrNotExist }) if err != nil { t.Fatal(err) } if got != "/current/bin/zut" { t.Fatalf("executable = %q, want current executable", got) } } func TestResolveSubagentExecutableFallsBackToPath(t *testing.T) { got, err := resolveSubagentExecutable(context.Background(), "/removed/.local/bin/zut", "/removed/.local/bin/zut", func(_ context.Context, candidate string) (string, error) { if candidate == "zut" { return "/active/go/bin/zut", nil } return "", os.ErrNotExist }) if err != nil { t.Fatal(err) } if got != "/active/go/bin/zut" { t.Fatalf("executable = %q, want PATH fallback", got) } } func TestResolveSubagentExecutableCancelsOtherLookupsAfterFirstSuccess(t *testing.T) { slowStarted := make(chan struct{}) slowCanceled := make(chan struct{}) got, err := resolveSubagentExecutable(context.Background(), "/current/bin/zut", "/old/bin/zut", func(ctx context.Context, candidate string) (string, error) { switch candidate { case "/current/bin/zut": close(slowStarted) <-ctx.Done() close(slowCanceled) return "", ctx.Err() case "zut": <-slowStarted return "/active/go/bin/zut", nil default: return "", os.ErrNotExist } }) if err != nil { t.Fatal(err) } if got != "/active/go/bin/zut" { t.Fatalf("executable = %q, want first successful lookup", got) } select { case <-slowCanceled: case <-time.After(time.Second): t.Fatal("slower lookup was not canceled") } } func TestSubagentWorkerArgsPropagatesProviderConnectionSettings(t *testing.T) { args := subagentWorkerArgs(subagentWorkerArgsOpts{ Exe: "/zut", Dir: "/work", SessionPath: "/state/session.json", InboxPath: "/state/inbox.sock", BaseURL: "https://gateway.example.test/v1", InsecureTLS: true, }) if i := indexOf(args, "--base-url"); i < 0 || safeAt(args, i+1) != "https://gateway.example.test/v1" { t.Fatalf("argv = %v, want --base-url with inherited endpoint", args) } if !containsArg(args, "--insecure") { t.Fatalf("argv = %v, want inherited --insecure", args) } for _, arg := range args { if strings.Contains(arg, "api-key") || strings.Contains(arg, "secret") { t.Fatalf("argv contains credential material: %v", args) } } } func TestDefaultChildArgsPropagatesProviderConnectionSettings(t *testing.T) { a := &Agent{ Dir: "/work", BaseURL: "https://gateway.example.test/v1", InsecureTLS: true, } args := defaultChildArgs("/zut", a, "/state/session.json", "/state/inbox.sock") if i := indexOf(args, "--base-url"); i < 0 || safeAt(args, i+1) != a.BaseURL { t.Fatalf("argv = %v, want Agent.BaseURL propagated", args) } if !containsArg(args, "--insecure") { t.Fatalf("argv = %v, want Agent.InsecureTLS propagated", args) } } // TestSubagentWorkerArgsEmptyTaskOmitsPositional makes sure that when the // agent is being adopted (no fresh task) we don't pass an empty // positional which the arg parser would treat as a real prompt. func TestSubagentWorkerArgsEmptyTaskOmitsPositional(t *testing.T) { args := subagentWorkerArgs(subagentWorkerArgsOpts{ Exe: "/zut", Dir: "/wt", SessionPath: "/s.json", InboxPath: "/in.sock", }) for _, a := range args { if a == "" { t.Fatalf("argv contains an empty positional: %v", args) } } // last arg should be a real flag value, not a stray positional if a := args[len(args)-1]; strings.HasPrefix(a, "--") { t.Fatalf("argv ends on a flag with no value: %v", args) } for _, flag := range []string{"--reasoning", "--subagent"} { if indexOf(args, flag) >= 0 { t.Fatalf("empty child options should omit %s: %v", flag, args) } } } // TestDefaultChildArgsSpawnIncludesTask pins the spawn shape: a // fresh (non-resuming) Agent produces argv that ends with the // original task as a positional, so the child runs it as the // initial user turn. func TestDefaultChildArgsSpawnIncludesTask(t *testing.T) { a := &Agent{Dir: "/wt", Task: "do thing"} args := defaultChildArgs("/zut", a, "/s.json", "/in.sock") if got := args[len(args)-1]; got != "do thing" { t.Fatalf("spawn argv last = %q; want %q\n%v", got, "do thing", args) } } // TestDefaultChildArgsResumeOmitsTask is the regression for the // "agent busy; send 'cancel' first" error: when an Agent is being // resumed (Resuming==true), the child argv MUST NOT include the // original Task as a positional. Otherwise the child fires the task // as a fresh user turn on every resume, racing with whatever the // user types next via the inbox. func TestDefaultChildArgsResumeOmitsTask(t *testing.T) { a := &Agent{Dir: "/wt", Task: "do thing", Resuming: true} args := defaultChildArgs("/zut", a, "/s.json", "/in.sock") for _, v := range args { if v == "do thing" { t.Fatalf("resume argv contains the task; it would re-fire as a duplicate turn\n%v", args) } } // And no trailing positional at all: every final token must be a // recognized boolean flag or a value belonging to a preceding flag, // never a stray empty string or the original task. if got := args[len(args)-1]; got == "" { t.Fatalf("resume argv ends with an empty positional: %v", args) } } func TestUpdateAgentFromEventIgnoresProviderToolLoopTurnEnd(t *testing.T) { a := &Agent{} a.setTurnState(TurnRunning, "turn-1") persisted := 0 a.persistFn = func(*Agent) { persisted++ } updateAgentFromEvent(a, NewEvent("turn_end", map[string]any{"stop": "tool_use"})) if got := a.TurnState(); got != TurnRunning { t.Fatalf("turn state = %q, want %q", got, TurnRunning) } if got := a.CurrentTurnID(); got != "turn-1" { t.Fatalf("turn ID = %q, want turn-1", got) } if persisted != 0 { t.Fatalf("persist called %d times for provider/tool-loop turn_end, want 0", persisted) } } func TestUpdateAgentFromEventPersistsTerminalTurnEndWithStep(t *testing.T) { a := &Agent{} persisted := 0 a.persistFn = func(*Agent) { persisted++ } updateAgentFromEvent(a, NewEvent("turn_end", map[string]any{"step": float64(2), "error": "boom"})) if got := a.TurnState(); got != TurnFailed { t.Fatalf("turn state = %q, want %q", got, TurnFailed) } if persisted != 1 { t.Fatalf("persist called %d times for terminal turn_end, want 1", persisted) } } func TestNotifyPromptTurnEndIgnoresProviderToolLoopTurnEnd(t *testing.T) { called := make(chan struct{}, 1) a := &Agent{} a.SetOnTurnEnd(func(step int, errMsg string) { called <- struct{}{} }) notifyPromptTurnEnd(a, NewEvent("turn_end", map[string]any{"stop": "tool_use"})) select { case <-called: t.Fatal("OnTurnEnd fired for provider/tool-loop turn_end without step") case <-time.After(50 * time.Millisecond): } } func TestNotifyPromptTurnEndFiresForDaemonPromptCompletion(t *testing.T) { type got struct { step int err string } called := make(chan got, 1) a := &Agent{} a.SetOnTurnEnd(func(step int, errMsg string) { called <- got{step: step, err: errMsg} }) notifyPromptTurnEnd(a, NewEvent("turn_end", map[string]any{"step": float64(2), "error": "boom"})) select { case g := <-called: if g.step != 2 || g.err != "boom" { t.Fatalf("callback = (%d, %q); want (2, boom)", g.step, g.err) } case <-time.After(time.Second): t.Fatal("OnTurnEnd did not fire for prompt-level turn_end with step") } } func TestSetOnTurnEndReplaysPendingNoticesInArrivalOrder(t *testing.T) { a := &Agent{} notifyPromptTurnEnd(a, NewEvent("turn_end", map[string]any{"step": float64(1)})) notifyPromptTurnEnd(a, NewEvent("turn_end", map[string]any{"step": float64(2), "error": "boom"})) type notice struct { step int err string } var got []notice a.SetOnTurnEnd(func(step int, errMsg string) { got = append(got, notice{step: step, err: errMsg}) }) want := []notice{{step: 1}, {step: 2, err: "boom"}} if len(got) != len(want) { t.Fatalf("replayed %d notices, want %d: %+v", len(got), len(want), got) } for i := range want { if got[i] != want[i] { t.Errorf("replayed notice %d = %+v, want %+v", i, got[i], want[i]) } } } func indexOf(xs []string, want string) int { for i, x := range xs { if x == want { return i } } return -1 } func safeAt(xs []string, i int) string { if i < 0 || i >= len(xs) { return "" } return xs[i] }