AutoZero/agent/sideeffect_test.go

package agent

import (
	"context"
	"errors"
	"testing"

	"git.gopher.ink/mirrors2/AutoZero/llm"
	"git.gopher.ink/mirrors2/AutoZero/tools"
)

// TestToolEventMarksSideEffects：副作用标记要随工具事件传给宿主，前端才能在
// 回合失败时禁用自动重试。
func TestToolEventMarksSideEffects(t *testing.T) {
	prov := &scriptedProvider{fn: func(call int, _ string, _ []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
		if call == 1 {
			return llm.Message{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{
				{ID: "r1", Name: "rotool", Input: "{}"},
				{ID: "s1", Name: "sidetool", Input: "{}"},
			}}, nil
		}
		return llm.Message{Role: llm.RoleAssistant, Content: "done"}, nil
	}}
	ro := &tools.Tool{Name: "rotool", ReadOnly: true, ConcurrencySafe: true, Run: func(context.Context, string) (tools.Result, error) {
		return tools.TextResult("r"), nil
	}}
	side := &tools.Tool{Name: "sidetool", SideEffect: true, Kind: tools.KindExec, Run: func(context.Context, string) (tools.Result, error) {
		return tools.TextResult("s"), nil
	}}
	a, _ := newTestAgent(t, ro, side)
	a.Provider = prov
	started := map[string]bool{}
	a.Opt.OnTool = func(ev ToolEvent) {
		if !ev.Done {
			started[ev.Name] = ev.SideEffect
		}
	}
	if err := a.Run(context.Background(), "go"); err != nil {
		t.Fatal(err)
	}
	if !started["sidetool"] {
		t.Error("side-effect tool must be marked in ToolEvent")
	}
	if started["rotool"] {
		t.Error("read-only tool must not be marked side-effect")
	}
}

// TestSideEffectsDisableWholeTurnRetry：副作用已经执行后，回合失败要带
// ErrAfterSideEffects，宿主据此只给手动重试；工具本身只跑一次。
func TestSideEffectsDisableWholeTurnRetry(t *testing.T) {
	runs := 0
	calls := 0
	prov := &scriptedProvider{fn: func(call int, _ string, _ []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
		calls++
		if calls == 1 {
			return llm.Message{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{
				{ID: "s1", Name: "sidetool", Input: "{}"},
			}}, nil
		}
		return llm.Message{}, errors.New("boom")
	}}
	side := &tools.Tool{Name: "sidetool", SideEffect: true, Kind: tools.KindExec, Run: func(context.Context, string) (tools.Result, error) {
		runs++
		return tools.TextResult("done"), nil
	}}
	a, _ := newTestAgent(t, side)
	a.Provider = prov

	err := a.Run(context.Background(), "go")
	if !errors.Is(err, ErrAfterSideEffects) {
		t.Fatalf("err = %v, want ErrAfterSideEffects", err)
	}
	if runs != 1 {
		t.Fatalf("side effect ran %d times, want 1", runs)
	}
}

// TestSideEffectResultIsDurableBeforeNextTool：同一批串行工具里，前一个副作用
// 工具的结果必须在后一个开始前就进会话——崩溃/重试时才能知道它已经跑过。
func TestSideEffectResultIsDurableBeforeNextTool(t *testing.T) {
	var sessHasPrior bool
	var sessionMessages func() []llm.Message
	prov := &scriptedProvider{fn: func(call int, _ string, _ []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
		if call == 1 {
			return llm.Message{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{
				{ID: "t1", Name: "first", Input: "{}"},
				{ID: "t2", Name: "second", Input: "{}"},
			}}, nil
		}
		return llm.Message{Role: llm.RoleAssistant, Content: "done"}, nil
	}}
	first := &tools.Tool{Name: "first", SideEffect: true, Kind: tools.KindExec, Run: func(context.Context, string) (tools.Result, error) {
		return tools.TextResult("first-result"), nil
	}}
	second := &tools.Tool{Name: "second", SideEffect: true, Kind: tools.KindExec, Run: func(context.Context, string) (tools.Result, error) {
		for _, m := range sessionMessages() {
			if m.Role == llm.RoleTool && m.ToolCallID == "t1" {
				sessHasPrior = true
			}
		}
		return tools.TextResult("second-result"), nil
	}}
	a, s := newTestAgent(t, first, second)
	a.Provider = prov
	sessionMessages = func() []llm.Message { return s.Messages }

	if err := a.Run(context.Background(), "go"); err != nil {
		t.Fatal(err)
	}
	if !sessHasPrior {
		t.Fatal("second tool ran before the first tool's result was persisted")
	}
}

// TestResumeKeepsSideEffectFence：续跑一个已经执行过副作用工具的回合时，
// 失败仍然不能自动整轮重试——历史里的 tool 结果就是证据。
func TestResumeKeepsSideEffectFence(t *testing.T) {
	prov := &scriptedProvider{fn: func(int, string, []llm.Message, []llm.ToolDef) (llm.Message, error) {
		return llm.Message{}, errors.New("boom")
	}}
	side := &tools.Tool{Name: "sidetool", SideEffect: true, Kind: tools.KindExec, Run: func(context.Context, string) (tools.Result, error) {
		t.Error("resume must not re-run the side-effect tool")
		return tools.TextResult(""), nil
	}}
	a, sess := newTestAgent(t, side)
	a.Provider = prov
	sess.Append(llm.Message{Role: llm.RoleUser, Content: "go"})
	sess.Append(llm.Message{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{{ID: "t1", Name: "sidetool"}}})
	sess.Append(llm.Message{Role: llm.RoleTool, ToolCallID: "t1", Tool: "sidetool", Content: "ok"})

	err := a.Run(context.Background(), "")
	if !errors.Is(err, ErrAfterSideEffects) {
		t.Fatalf("err = %v, want ErrAfterSideEffects", err)
	}
}

// TestSideEffectsDoNotWrapCancellation：用户主动停止必须保持 context.Canceled，
// 不能被副作用标记包成普通错误——界面靠它区分"停止"和"出错"。
func TestSideEffectsDoNotWrapCancellation(t *testing.T) {
	calls := 0
	prov := &scriptedProvider{fn: func(int, string, []llm.Message, []llm.ToolDef) (llm.Message, error) {
		calls++
		if calls == 1 {
			return llm.Message{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{
				{ID: "s1", Name: "sidetool", Input: "{}"},
			}}, nil
		}
		return llm.Message{}, context.Canceled
	}}
	side := &tools.Tool{Name: "sidetool", SideEffect: true, Kind: tools.KindExec, Run: func(context.Context, string) (tools.Result, error) {
		return tools.TextResult("done"), nil
	}}
	a, _ := newTestAgent(t, side)
	a.Provider = prov

	err := a.Run(context.Background(), "go")
	if !errors.Is(err, context.Canceled) {
		t.Fatalf("err = %v, want context.Canceled", err)
	}
	if errors.Is(err, ErrAfterSideEffects) {
		t.Fatalf("cancellation must not be wrapped as a retryable failure: %v", err)
	}
}