majordomo/agent/finalize_test.go

package agent

import (
	"context"
	"encoding/json"
	"strings"
	"testing"

	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/fake"
)

func TestIsWeakFinal(t *testing.T) {
	long := strings.Repeat("As I said, this is the full answer. ", 6) // >120, contains "as i said"
	cases := []struct {
		name string
		in   string
		want bool
	}{
		{"empty", "", true},
		{"whitespace", "   \n\t ", true},
		{"already-answered", "(Already answered above.)", true},
		{"see-above", "see above", true},
		{"as-i-said-short", "As I said, it's 60 minutes.", true},
		{"crisp-number", "42", false},
		{"crisp-yes", "Yes.", false},
		{"crisp-status", "It's down, restarting now.", false},
		{"long-with-as-i-said", long, false}, // >120 chars: not weak despite the phrase
	}
	for _, c := range cases {
		t.Run(c.name, func(t *testing.T) {
			if got := isWeakFinal(c.in); got != c.want {
				t.Errorf("isWeakFinal(%q) = %v, want %v", c.in, got, c.want)
			}
		})
	}
}

func asst(text string, tools ...llm.ToolCall) llm.Message {
	m := llm.Message{Role: llm.RoleAssistant}
	if text != "" {
		m.Parts = []llm.Part{llm.Text(text)}
	}
	m.ToolCalls = tools
	return m
}

func TestFinalOutput(t *testing.T) {
	cite := []llm.ToolCall{{ID: "c1", Name: "cite", Arguments: json.RawMessage(`{}`)}}
	longAnswer := strings.TrimSpace(strings.Repeat("Free group calls are capped at sixty minutes. ", 6)) // >200

	tests := []struct {
		name     string
		msgs     []llm.Message
		terminal string
		want     string
	}{
		{
			name: "front-loaded answer recovered over back-ref closer",
			msgs: []llm.Message{
				llm.UserText("q?"),
				asst(longAnswer, cite...),
				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
				asst("(Already answered above.)"),
			},
			terminal: "(Already answered above.)",
			want:     longAnswer,
		},
		{
			name: "empty terminal recovers prior substantive answer",
			msgs: []llm.Message{
				llm.UserText("q?"),
				asst(longAnswer, cite...),
				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
				asst(""),
			},
			terminal: "",
			want:     longAnswer,
		},
		{
			name: "healthy terminal answer is unchanged",
			msgs: []llm.Message{
				llm.UserText("q?"),
				asst("Let me check.", cite...),
				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
				asst(longAnswer),
			},
			terminal: longAnswer,
			want:     longAnswer,
		},
		{
			name: "short crisp answer not overridden by a short preamble prior",
			msgs: []llm.Message{
				llm.UserText("is it up?"),
				asst("Let me check the server status.", cite...),
				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
				asst("It's down, restarting now."),
			},
			terminal: "It's down, restarting now.", // not weak → returned as-is
			want:     "It's down, restarting now.",
		},
		{
			name: "weak terminal but only a preamble prior: no recovery",
			msgs: []llm.Message{
				llm.UserText("q?"),
				asst("Let me look that up for you.", cite...),
				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
				asst("(see above)"),
			},
			terminal: "(see above)",
			want:     "(see above)", // preamble excluded; falls back to terminal
		},
	}
	for _, tc := range tests {
		t.Run(tc.name, func(t *testing.T) {
			if got := finalOutput(tc.msgs, tc.terminal); got != tc.want {
				t.Errorf("finalOutput = %q, want %q", got, tc.want)
			}
		})
	}
}

func citeToolbox(t *testing.T) *llm.Toolbox {
	t.Helper()
	return llm.NewToolbox("sources", llm.Tool{
		Name:        "cite",
		Description: "Record a citation.",
		Parameters:  json.RawMessage(`{"type":"object","properties":{}}`),
		Handler: func(_ context.Context, _ json.RawMessage) (any, error) {
			return map[string]bool{"ok": true}, nil
		},
	})
}

// TestRun_RecoversFrontLoadedAnswer reproduces the glm-5.2 shape end-to-end: a
// turn carrying the full answer text AND a tool call, then a degenerate
// terminal turn. The recovered answer must be delivered with no extra model
// call (zero-cost recovery from the transcript).
func TestRun_RecoversFrontLoadedAnswer(t *testing.T) {
	longAnswer := strings.TrimSpace(strings.Repeat("Free group calls are capped at sixty minutes. ", 6))
	fp := fake.New("fp")
	fp.Enqueue("test-model",
		fake.ReplyWith(llm.Response{
			Parts:        []llm.Part{llm.Text(longAnswer)},
			ToolCalls:    []llm.ToolCall{{ID: "c1", Name: "cite", Arguments: json.RawMessage(`{}`)}},
			FinishReason: llm.FinishToolCalls,
			Usage:        llm.Usage{InputTokens: 10, OutputTokens: 5},
		}),
		fake.Reply("(Already answered above.)"),
	)

	a := New(newModel(t, fp), "sys", WithToolbox(citeToolbox(t)))
	res, err := a.Run(context.Background(), "is there a meet time limit?")
	if err != nil {
		t.Fatalf("Run: %v", err)
	}
	if res.Output != longAnswer {
		t.Errorf("Output = %q, want recovered front-loaded answer", res.Output)
	}
	if n := len(fp.Calls()); n != 2 {
		t.Errorf("model calls = %d, want 2 (no extra nudge turn)", n)
	}
}

// TestRun_HealthyTerminalUnchanged guards against regressing the normal case:
// a deferred answer in the terminal turn is delivered verbatim.
func TestRun_HealthyTerminalUnchanged(t *testing.T) {
	fp := fake.New("fp")
	fp.Enqueue("test-model",
		toolCallReply("c1", "cite", `{}`),
		fake.Reply("The limit is 60 minutes for free group calls."),
	)
	a := New(newModel(t, fp), "sys", WithToolbox(citeToolbox(t)))
	res, err := a.Run(context.Background(), "q?")
	if err != nil {
		t.Fatalf("Run: %v", err)
	}
	if res.Output != "The limit is 60 minutes for free group calls." {
		t.Errorf("Output = %q, want terminal answer unchanged", res.Output)
	}
}