mirror of
https://gitcode.com/JianFeeeee/HomeAgent.git
synced 2026-11-01 14:57:38 +00:00
问题(WSL2 真机诊断报告问题 4,影响最大):开终端注入「已启动」、 超时注入「已超时关闭」,每条都唤醒一整轮 LLM 推理。实测后果是 agent 反复向用户解释"终端怎么又没了",真正要做的活一件没做。 这属于本插件**自身逻辑矛盾**,不是内核缺陷: 1. 「已启动」是与 terminal_create 返回值的重复告知。返回值已含 id/command/timeout/rows/cols/notify_mode,信息完整。 2. 「已超时关闭」在创建时已由 timeout 字段告知,agent 本可推导; 且注入后紧接着 removeSession,agent 去 terminal_list 查必然是 0 个, 只能盲目排查——而这是插件自己删的。 3. 同一函数里 IsExpired() 语义相反:r.err 分支把它列为**必须静默**, 而超时分支却在注入。 4. 本插件自己的 notify=exit/interval/buffer 参数本就定义了"什么该通知"。 两处无条件注入是不经该配置的旁路,与自身设计相矛盾。 顺带:r.err 不再以 %v 原文进 prompt(原来注入 `read /dev/ptmx: input/output error`,把 agent 引去查 devpts 挂载, 好几轮推理花在这上面)。改为给**结论**「读取结束」,原始错误只进日志。 判据(plugin_test.go): · 新增 2 条**反向判据** TestNoStartupNotificationInjected / TestNoExpiredNotificationInjected —— 没有它们,后人把注入加回来时 其它测试依然全绿,缺陷会静默复发 · 原 7 处用「等已启动通知」做时序同步的测试改为 waitReadLoopReady (轮询 mockTerm.readCalls(),不靠固定 sleep 也不依赖业务通知) ★ 保留不动:进程自己结束/读取错误的通知(agent 无法从工具返回推导), 由 watch 配置控制。判据 TestProcessExit_StillNotifies 钉住这一点。
1076 lines
31 KiB
Go
1076 lines
31 KiB
Go
//go:build linux || windows
|
||
|
||
package agentcli
|
||
|
||
import (
|
||
"encoding/json"
|
||
"fmt"
|
||
"strings"
|
||
"sync"
|
||
"testing"
|
||
"time"
|
||
|
||
sdk "github.com/JianFeeeee/HologramAgent/internal/sdk"
|
||
sdkpub "github.com/JianFeeeee/hologramagentsdk/sdk"
|
||
)
|
||
|
||
type toolCapture struct {
|
||
handlers map[string]sdk.ToolHandler
|
||
defs map[string]sdk.ToolDef
|
||
}
|
||
|
||
func newToolCapture() *toolCapture {
|
||
return &toolCapture{
|
||
handlers: make(map[string]sdk.ToolHandler),
|
||
defs: make(map[string]sdk.ToolDef),
|
||
}
|
||
}
|
||
|
||
func (tc *toolCapture) RegisterTool(name string, def sdk.ToolDef, handler sdk.ToolHandler) error {
|
||
tc.handlers[name] = handler
|
||
tc.defs[name] = def
|
||
return nil
|
||
}
|
||
func (tc *toolCapture) RegisterStage(stage sdk.Stage, handler sdk.StageHandler) {}
|
||
func (tc *toolCapture) RegisterAPI(name string) error { return nil }
|
||
|
||
func setupPlugin() (*Plugin, *toolCapture, error) {
|
||
p := New("agentcli")
|
||
tc := newToolCapture()
|
||
sdk := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: tc.RegisterTool,
|
||
RegStage: tc.RegisterStage,
|
||
RegAPI: tc.RegisterAPI,
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
if err := p.Start(sdk); err != nil {
|
||
return nil, nil, err
|
||
}
|
||
return p, tc, nil
|
||
}
|
||
|
||
func TestKeyMapping(t *testing.T) {
|
||
tests := []struct {
|
||
key string
|
||
expected []byte
|
||
}{
|
||
{"enter", []byte{0x0D}},
|
||
{"tab", []byte{0x09}},
|
||
{"escape", []byte{0x1B}},
|
||
{"esc", []byte{0x1B}},
|
||
{"backspace", []byte{0x7F}},
|
||
{"delete", []byte{0x1B, 0x5B, 0x33, 0x7E}},
|
||
{"home", []byte{0x1B, 0x5B, 0x48}},
|
||
{"end", []byte{0x1B, 0x5B, 0x46}},
|
||
{"up", []byte{0x1B, 0x5B, 0x41}},
|
||
{"down", []byte{0x1B, 0x5B, 0x42}},
|
||
{"left", []byte{0x1B, 0x5B, 0x44}},
|
||
{"right", []byte{0x1B, 0x5B, 0x43}},
|
||
{"page_up", []byte{0x1B, 0x5B, 0x35, 0x7E}},
|
||
{"page_down", []byte{0x1B, 0x5B, 0x36, 0x7E}},
|
||
{"ctrl_a", []byte{0x01}},
|
||
{"ctrl_z", []byte{0x1A}},
|
||
{"alt_a", []byte{0x1B, 'a'}},
|
||
{"alt_z", []byte{0x1B, 'z'}},
|
||
{"f1", []byte{0x1B, 0x5B, 0x50}},
|
||
{"f4", []byte{0x1B, 0x5B, 0x53}},
|
||
{"f12", []byte{0x1B, 0x5B, 0x32, 0x34, 0x7E}},
|
||
}
|
||
|
||
for _, tt := range tests {
|
||
t.Run(tt.key, func(t *testing.T) {
|
||
result, err := mapKey(tt.key)
|
||
if err != nil {
|
||
t.Fatalf("unexpected error: %v", err)
|
||
}
|
||
if !bytesEqual(result, tt.expected) {
|
||
t.Fatalf("expected %v, got %v", tt.expected, result)
|
||
}
|
||
})
|
||
}
|
||
}
|
||
|
||
func TestKeyMappingInvalid(t *testing.T) {
|
||
_, err := mapKey("unknown_key")
|
||
if err == nil {
|
||
t.Fatal("expected error for unknown key")
|
||
}
|
||
}
|
||
|
||
func bytesEqual(a, b []byte) bool {
|
||
if len(a) != len(b) {
|
||
return false
|
||
}
|
||
for i := range a {
|
||
if a[i] != b[i] {
|
||
return false
|
||
}
|
||
}
|
||
return true
|
||
}
|
||
|
||
func TestListEmpty(t *testing.T) {
|
||
p, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
defer p.Stop()
|
||
|
||
handler := tc.handlers["terminal_list"]
|
||
result, err := handler(map[string]interface{}{})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if resp["status"] != "ok" {
|
||
t.Fatalf("expected status ok, got %v", resp["status"])
|
||
}
|
||
if resp["count"].(float64) != 0 {
|
||
t.Fatalf("expected count 0, got %v", resp["count"])
|
||
}
|
||
}
|
||
|
||
func TestCreateAndCloseTerminal(t *testing.T) {
|
||
p, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
defer p.Stop()
|
||
|
||
createHandler := tc.handlers["terminal_create"]
|
||
result, err := createHandler(map[string]interface{}{
|
||
"command": "echo hello",
|
||
"timeout": "10s",
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var createResp map[string]interface{}
|
||
json.Unmarshal(data, &createResp)
|
||
|
||
if v, ok := createResp["error"]; ok {
|
||
t.Skipf("PTY not available in this environment: %v", v)
|
||
}
|
||
if createResp["status"] != "created" {
|
||
t.Fatalf("expected status created, got %v", createResp["status"])
|
||
}
|
||
|
||
id := createResp["id"].(string)
|
||
if id == "" {
|
||
t.Fatal("expected non-empty terminal id")
|
||
}
|
||
|
||
// Give the terminal a moment to output
|
||
time.Sleep(200 * time.Millisecond)
|
||
|
||
// Read output
|
||
readHandler := tc.handlers["terminal_read"]
|
||
result, err = readHandler(map[string]interface{}{
|
||
"id": id,
|
||
"clear": true,
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ = json.Marshal(result)
|
||
var readResp map[string]interface{}
|
||
json.Unmarshal(data, &readResp)
|
||
if readResp["status"] != "ok" {
|
||
t.Fatalf("expected status ok, got %v", readResp["status"])
|
||
}
|
||
|
||
// Close
|
||
closeHandler := tc.handlers["terminal_close"]
|
||
result, err = closeHandler(map[string]interface{}{
|
||
"id": id,
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ = json.Marshal(result)
|
||
var closeResp map[string]interface{}
|
||
json.Unmarshal(data, &closeResp)
|
||
if closeResp["status"] != "closed" {
|
||
t.Fatalf("expected status closed, got %v", closeResp["status"])
|
||
}
|
||
}
|
||
|
||
func TestCreateTerminalMissingArgs(t *testing.T) {
|
||
p, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
defer p.Stop()
|
||
|
||
handler := tc.handlers["terminal_create"]
|
||
result, err := handler(map[string]interface{}{})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if v, ok := resp["error"]; ok {
|
||
t.Skipf("PTY not available in this environment: %v", v)
|
||
}
|
||
if resp["status"] != "created" {
|
||
t.Fatalf("expected status created, got %v", resp["status"])
|
||
}
|
||
id := resp["id"].(string)
|
||
// 经注册的 handler 关闭(handler 内部会带上 sdk);
|
||
// 直接调 p.handleClose 需自备 sdk 参数。
|
||
closeHandler := tc.handlers["terminal_close"]
|
||
if _, err := closeHandler(map[string]interface{}{"id": id}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
}
|
||
|
||
func TestWriteToNonexistentTerminal(t *testing.T) {
|
||
_, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
handler := tc.handlers["terminal_write"]
|
||
result, err := handler(map[string]interface{}{
|
||
"id": "nonexistent",
|
||
"input": "test",
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if _, ok := resp["error"]; !ok {
|
||
t.Fatal("expected error for nonexistent terminal")
|
||
}
|
||
}
|
||
|
||
func TestReadNonexistentTerminal(t *testing.T) {
|
||
_, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
handler := tc.handlers["terminal_read"]
|
||
result, err := handler(map[string]interface{}{
|
||
"id": "nonexistent",
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if _, ok := resp["error"]; !ok {
|
||
t.Fatal("expected error for nonexistent terminal")
|
||
}
|
||
}
|
||
|
||
func TestResizeNonexistentTerminal(t *testing.T) {
|
||
_, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
handler := tc.handlers["terminal_resize"]
|
||
result, err := handler(map[string]interface{}{
|
||
"id": "nonexistent",
|
||
"rows": float64(40),
|
||
"cols": float64(120),
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if _, ok := resp["error"]; !ok {
|
||
t.Fatal("expected error for nonexistent terminal")
|
||
}
|
||
}
|
||
|
||
func TestCloseNonexistentTerminal(t *testing.T) {
|
||
_, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
handler := tc.handlers["terminal_close"]
|
||
result, err := handler(map[string]interface{}{
|
||
"id": "nonexistent",
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if _, ok := resp["error"]; !ok {
|
||
t.Fatal("expected error for nonexistent terminal")
|
||
}
|
||
}
|
||
|
||
func TestTerminalWriteRequiresId(t *testing.T) {
|
||
_, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
handler := tc.handlers["terminal_write"]
|
||
result, err := handler(map[string]interface{}{
|
||
"input": "test",
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if _, ok := resp["error"]; !ok {
|
||
t.Fatal("expected error when id is missing")
|
||
}
|
||
}
|
||
|
||
func TestTerminalWriteRequiresContent(t *testing.T) {
|
||
_, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
handler := tc.handlers["terminal_write"]
|
||
result, err := handler(map[string]interface{}{
|
||
"id": "test",
|
||
})
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
data, _ := json.Marshal(result)
|
||
var resp map[string]interface{}
|
||
json.Unmarshal(data, &resp)
|
||
|
||
if _, ok := resp["error"]; !ok {
|
||
t.Fatal("expected error when both input and key are missing")
|
||
}
|
||
}
|
||
|
||
func TestToolsRegistered(t *testing.T) {
|
||
p, tc, err := setupPlugin()
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
defer p.Stop()
|
||
|
||
expectedTools := []string{
|
||
"terminal_create",
|
||
"terminal_write",
|
||
"terminal_read",
|
||
"terminal_resize",
|
||
"terminal_close",
|
||
"terminal_list",
|
||
}
|
||
|
||
for _, name := range expectedTools {
|
||
if _, ok := tc.handlers[name]; !ok {
|
||
t.Errorf("tool %s not registered", name)
|
||
}
|
||
}
|
||
}
|
||
|
||
// ——— Phase 6: 通知节流测试(mock 终端 + 捕获注入) ———
|
||
|
||
type injectCapture struct {
|
||
mu sync.Mutex
|
||
texts []string
|
||
}
|
||
|
||
func (c *injectCapture) InjectInterruptText(source, channel, text string) {
|
||
c.mu.Lock()
|
||
c.texts = append(c.texts, text)
|
||
c.mu.Unlock()
|
||
}
|
||
func (c *injectCapture) InjectText(source, channel, text string) {
|
||
c.mu.Lock()
|
||
c.texts = append(c.texts, text)
|
||
c.mu.Unlock()
|
||
}
|
||
func (c *injectCapture) InjectTextNoMemory(source, channel, text string) {
|
||
c.mu.Lock()
|
||
c.texts = append(c.texts, text)
|
||
c.mu.Unlock()
|
||
}
|
||
|
||
func (c *injectCapture) SetToolBlocks(blocks []sdkpub.ContentBlock) {
|
||
// 测试桩:忽略多模态块
|
||
}
|
||
func (c *injectCapture) InjectInputSync(source, channel, text string) string { return "" }
|
||
|
||
// 三个带媒体的注入方法同样记录文本:本测试只关心「注入了什么话」,
|
||
// 媒体块的转发在 core 的 injectedBlocks 测试里覆盖。
|
||
func (c *injectCapture) InjectInputMedia(source, channel, text string, blocks []sdkpub.ContentBlock) {
|
||
c.mu.Lock()
|
||
c.texts = append(c.texts, text)
|
||
c.mu.Unlock()
|
||
}
|
||
|
||
func (c *injectCapture) InjectInputMediaSync(source, channel, text string, blocks []sdkpub.ContentBlock) string {
|
||
return ""
|
||
}
|
||
|
||
func (c *injectCapture) InjectInterruptMedia(source, channel, text string, blocks []sdkpub.ContentBlock) {
|
||
c.mu.Lock()
|
||
c.texts = append(c.texts, text)
|
||
c.mu.Unlock()
|
||
}
|
||
|
||
// ---- 带 InjectOptions 的注入(1.2.0):同样只记文本 ----
|
||
|
||
func (c *injectCapture) InjectTextOpts(source, channel, text string, opts sdkpub.InjectOptions) {
|
||
c.InjectText(source, channel, text)
|
||
}
|
||
func (c *injectCapture) InjectInterruptTextOpts(source, channel, text string, opts sdkpub.InjectOptions) {
|
||
c.InjectInterruptText(source, channel, text)
|
||
}
|
||
func (c *injectCapture) InjectInputSyncOpts(source, channel, text string, opts sdkpub.InjectOptions) string {
|
||
return ""
|
||
}
|
||
func (c *injectCapture) InjectInputMediaOpts(source, channel, text string, blocks []sdkpub.ContentBlock, opts sdkpub.InjectOptions) {
|
||
c.InjectInputMedia(source, channel, text, blocks)
|
||
}
|
||
func (c *injectCapture) InjectInputMediaSyncOpts(source, channel, text string, blocks []sdkpub.ContentBlock, opts sdkpub.InjectOptions) string {
|
||
return ""
|
||
}
|
||
func (c *injectCapture) InjectInterruptMediaOpts(source, channel, text string, blocks []sdkpub.ContentBlock, opts sdkpub.InjectOptions) {
|
||
c.InjectInterruptMedia(source, channel, text, blocks)
|
||
}
|
||
|
||
func (c *injectCapture) snapshot() []string {
|
||
c.mu.Lock()
|
||
defer c.mu.Unlock()
|
||
out := make([]string, len(c.texts))
|
||
copy(out, c.texts)
|
||
return out
|
||
}
|
||
|
||
// mockTerm 可控输出流的假终端:Read 从 data chan 取数据,可模拟进程退出/读取错误
|
||
type mockTerm struct {
|
||
mu sync.Mutex
|
||
data chan []byte
|
||
running bool
|
||
err error
|
||
// reads 记录 Read 被调用次数,供 waitReadLoopReady 判断 readLoop 已起来。
|
||
// 不用固定 sleep:机器慢时 sleep 不够、快时白等。
|
||
reads int
|
||
}
|
||
|
||
// readCalls 返回 Read 被调用次数(并发安全)。
|
||
func (m *mockTerm) readCalls() int {
|
||
m.mu.Lock()
|
||
defer m.mu.Unlock()
|
||
return m.reads
|
||
}
|
||
|
||
func newMockTerm() *mockTerm {
|
||
return &mockTerm{data: make(chan []byte, 16), running: true}
|
||
}
|
||
|
||
func (m *mockTerm) Read(buf []byte) (int, error) {
|
||
m.mu.Lock()
|
||
m.reads++
|
||
m.mu.Unlock()
|
||
for {
|
||
m.mu.Lock()
|
||
err := m.err
|
||
running := m.running
|
||
m.mu.Unlock()
|
||
if err != nil {
|
||
return 0, err
|
||
}
|
||
if !running {
|
||
return 0, fmt.Errorf("process exited")
|
||
}
|
||
select {
|
||
case data, ok := <-m.data:
|
||
if !ok {
|
||
return 0, fmt.Errorf("closed")
|
||
}
|
||
n := copy(buf, data)
|
||
return n, nil
|
||
case <-time.After(20 * time.Millisecond):
|
||
}
|
||
}
|
||
}
|
||
|
||
func (m *mockTerm) WriteString(s string) (int, error) { return len(s), nil }
|
||
func (m *mockTerm) Resize(rows, cols uint16) error { return nil }
|
||
func (m *mockTerm) Running() bool {
|
||
m.mu.Lock()
|
||
defer m.mu.Unlock()
|
||
return m.running
|
||
}
|
||
|
||
// Kill 模拟真实 PTY 的行为:把一份 **buffer 里的输出 + EIO** 交给 Read。
|
||
//
|
||
// 两点必须还原,否则测不出线上的回声:
|
||
//
|
||
// 1. Read 返回错误(真实 PTY 被 Kill 后读端立刻 `read /dev/ptmx:
|
||
// input/output error`)。旧版这里是空操作,于是"主动关闭"在测试里
|
||
// 从不产生回声,判据永远绿(假绿)。
|
||
// 2. **已产出的数据先于错误到达**。真实内核里 PTY 缓冲中残留的输出
|
||
// 会先被读走,EIO 随后才到——这一点决定了 readLoop 的 select
|
||
// 在"stopCh 已关"与"readCh 有数据/错误"之间如何竞争。
|
||
func (m *mockTerm) Kill() error {
|
||
m.mu.Lock()
|
||
m.running = false
|
||
if m.err == nil {
|
||
m.err = fmt.Errorf("read /dev/ptmx: input/output error")
|
||
}
|
||
m.mu.Unlock()
|
||
return nil
|
||
}
|
||
func (m *mockTerm) Close() error { return nil }
|
||
|
||
func (m *mockTerm) push(data []byte) {
|
||
m.data <- data
|
||
}
|
||
|
||
func (m *mockTerm) setRunning(v bool) {
|
||
m.mu.Lock()
|
||
m.running = v
|
||
m.mu.Unlock()
|
||
}
|
||
|
||
func (m *mockTerm) setErr(err error) {
|
||
m.mu.Lock()
|
||
m.err = err
|
||
m.mu.Unlock()
|
||
}
|
||
|
||
func newTestSession(term ptyTerm) *TerminalSession {
|
||
return &TerminalSession{
|
||
id: "t1",
|
||
session: term,
|
||
createdAt: time.Now(),
|
||
timeout: 10 * time.Minute,
|
||
stopCh: make(chan struct{}),
|
||
done: make(chan struct{}),
|
||
}
|
||
}
|
||
|
||
func startReadLoop(p *Plugin, s *sdk.PluginSDK, t *TerminalSession) {
|
||
p.wg.Add(1)
|
||
go p.readLoop(t, s)
|
||
}
|
||
|
||
// waitReadLoopReady 等 readLoop 真正进入工作状态。
|
||
//
|
||
// ★ 为什么需要它:原测试用「已启动」注入当同步点(等那条通知出现再继续)。
|
||
// 但 2026-10-25 起该注入已移除(它是"同一条信息告知两次",每条都会唤醒
|
||
// 一整轮推理)。同步点因此需要换成不依赖注入的方式。
|
||
//
|
||
// 这里用「mockTerm 的 Read 至少被调用过一次」作为信号:readLoop 的第一件事
|
||
// 就是 Read,比固定 sleep 可靠,也不引入对业务通知的依赖。
|
||
func waitReadLoopReady(t *testing.T, term *mockTerm) {
|
||
t.Helper()
|
||
deadline := time.Now().Add(2 * time.Second)
|
||
for time.Now().Before(deadline) {
|
||
if term.readCalls() > 0 {
|
||
return
|
||
}
|
||
time.Sleep(2 * time.Millisecond)
|
||
}
|
||
t.Fatal("readLoop 未在 2s 内进入读取状态")
|
||
}
|
||
|
||
func waitInjected(c *injectCapture, substr string, timeout time.Duration) bool {
|
||
deadline := time.Now().Add(timeout)
|
||
for time.Now().Before(deadline) {
|
||
for _, text := range c.snapshot() {
|
||
if strings.Contains(text, substr) {
|
||
return true
|
||
}
|
||
}
|
||
time.Sleep(20 * time.Millisecond)
|
||
}
|
||
return false
|
||
}
|
||
|
||
// Phase 6: 持续吐进度时,通知频率显著低于 500ms/条(节流生效)
|
||
func TestReadLoopNotifyThrottle(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 2048
|
||
p.notifyInterval = 2 * time.Second
|
||
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
startReadLoop(p, sdkInst, ts)
|
||
|
||
// 等 readLoop 真正起来(不再依赖"已启动"注入:那已被移除,
|
||
// 理由见 plugin.go 中"不注入终端已启动通知"的说明)
|
||
waitReadLoopReady(t, term)
|
||
|
||
// 持续以 100B/50ms(=2KB/s) 吐进度 3 秒
|
||
stop := make(chan struct{})
|
||
go func() {
|
||
ticker := time.NewTicker(50 * time.Millisecond)
|
||
defer ticker.Stop()
|
||
chunk := make([]byte, 100)
|
||
for i := range chunk {
|
||
chunk[i] = 'x'
|
||
}
|
||
for {
|
||
select {
|
||
case <-stop:
|
||
return
|
||
case <-ticker.C:
|
||
term.push(chunk)
|
||
}
|
||
}
|
||
}()
|
||
|
||
time.Sleep(3 * time.Second)
|
||
close(stop)
|
||
|
||
notifies := 0
|
||
for _, text := range capture.snapshot() {
|
||
if strings.Contains(text, "有新输出") {
|
||
notifies++
|
||
}
|
||
}
|
||
// 3 秒持续输出,500ms/条 的旧行为应有 6 条;节流后 ≤3 条
|
||
if notifies > 3 {
|
||
t.Errorf("notify throttle ineffective: %d notifies in 3s (expected <=3)", notifies)
|
||
}
|
||
if notifies == 0 {
|
||
t.Error("expected at least one output notification")
|
||
}
|
||
|
||
close(ts.stopCh)
|
||
<-ts.done
|
||
}
|
||
|
||
// Phase 6: 进程退出 → 立即通知(两条路径:PTY Read 返回 EOF 走"读取结束",
|
||
// 或 reader 阻塞时顶部 terminalRunning 检测走"进程已退出")
|
||
func TestReadLoopNotifyOnExit(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 2048
|
||
p.notifyInterval = 2 * time.Second
|
||
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
startReadLoop(p, sdkInst, ts)
|
||
|
||
// 等 readLoop 真正起来(不再依赖"已启动"注入:那已被移除,
|
||
// 理由见 plugin.go 中"不注入终端已启动通知"的说明)
|
||
waitReadLoopReady(t, term)
|
||
|
||
term.setRunning(false)
|
||
gotExit := waitInjected(capture, "进程已退出", 2*time.Second)
|
||
gotReadEnd := waitInjected(capture, "读取结束", time.Second)
|
||
if !gotExit && !gotReadEnd {
|
||
t.Error("expected immediate notification on process exit (either 进程已退出 or 读取结束)")
|
||
}
|
||
close(ts.stopCh)
|
||
}
|
||
|
||
// Phase 6: 读取错误/EOF → 立即通知
|
||
func TestReadLoopNotifyOnReadError(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 2048
|
||
p.notifyInterval = 2 * time.Second
|
||
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
startReadLoop(p, sdkInst, ts)
|
||
|
||
// 等 readLoop 真正起来(不再依赖"已启动"注入:那已被移除,
|
||
// 理由见 plugin.go 中"不注入终端已启动通知"的说明)
|
||
waitReadLoopReady(t, term)
|
||
|
||
term.setErr(fmt.Errorf("read timeout"))
|
||
if !waitInjected(capture, "读取结束", 3*time.Second) {
|
||
t.Error("expected immediate notification on read error")
|
||
}
|
||
close(ts.stopCh)
|
||
<-ts.done
|
||
}
|
||
|
||
// 主动关闭终端后**不得**再注入任何收尾状态。
|
||
//
|
||
// 这是用户报的线上问题(2026-10-25):
|
||
//
|
||
// 关闭终端 → agent 收到「[终端 term_6 读取结束: read /dev/ptmx: input/output error]」
|
||
// → 又收到「[终端 term_6 中的进程已退出]」
|
||
// → 而 terminal_list 此时已是 0 个会话
|
||
//
|
||
// 后果不只是刷屏:每条注入都会唤醒一轮推理,agent 只能一遍遍解释
|
||
// "这是旧终端的回声,不用管它"。而且它**永远无法自证清白**——
|
||
// 会话列表已经是空的,它拿不到任何证据说明这条消息是历史遗留。
|
||
//
|
||
// 根因:主动 Close() 会 ①关 stopCh ②Kill 进程,而 ③PTY 读端随即返回
|
||
// 「input/output error」、④terminalRunning 变 false。旧实现在这两处
|
||
// **无条件注入**,把"我们自己关掉的"当成"进程自己结束了"上报。
|
||
//
|
||
// 判据:Close 之后只允许**关闭前**已发生的注入存在,不得新增。
|
||
// 变异:删掉 isClosing 判断 ⇒ 本判据变红。
|
||
func TestClose_NoEchoNotificationsAfterClose(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20 // 抬高阈值,避免缓冲通知混进来干扰判据
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
mterm := newMockTerm()
|
||
ts := newTestSession(mterm)
|
||
startReadLoop(p, sdkInst, ts)
|
||
|
||
// 等 readLoop 真正起来(不再依赖"已启动"注入:那已被移除,
|
||
// 理由见 plugin.go 中"不注入终端已启动通知"的说明)
|
||
waitReadLoopReady(t, mterm)
|
||
before := len(capture.snapshot())
|
||
|
||
// 主动关闭:模拟 close 工具路径(关 stopCh + 杀进程 + 释放 PTY)。
|
||
//
|
||
// 注意:Close() 内部会 <-t.done 等 readLoop 退出,所以它返回时
|
||
// readLoop 已经走完——那正是回声注入发生的时机,判据必须等到它之后。
|
||
ts.Close()
|
||
|
||
// 给 readLoop 充分时间走完所有收尾分支(含 r.err 与 terminalRunning)。
|
||
// 用固定等待而不是"等到某条消息"——我们要断言的是**什么都没有**。
|
||
time.Sleep(600 * time.Millisecond)
|
||
|
||
after := capture.snapshot()
|
||
if len(after) != before {
|
||
t.Fatalf("主动关闭后不应新增注入,实际多了 %d 条:\n%s",
|
||
len(after)-before, strings.Join(after[before:], "\n"))
|
||
}
|
||
t.Logf("关闭前后注入数:%d(无回声)", len(after))
|
||
// 明确点出两种最典型的回声,便于失败时一眼看出原因。
|
||
for _, s := range after[before:] {
|
||
if strings.Contains(s, "读取结束") || strings.Contains(s, "进程已退出") ||
|
||
strings.Contains(s, "命令已执行结束") {
|
||
t.Fatalf("关闭后出现了收尾回声:%s", s)
|
||
}
|
||
}
|
||
}
|
||
|
||
// 反向判据:进程**自己**结束时必须有通知。
|
||
//
|
||
// 这条守住修复不过度——不能因为要消除回声,把正常的上报也一起关掉。
|
||
// 变异:把 isClosing 判断写成无条件 return ⇒ 本判据变红。
|
||
func TestProcessExit_StillNotifies(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
startReadLoop(p, sdkInst, ts)
|
||
|
||
// 等 readLoop 真正起来(不再依赖"已启动"注入:那已被移除,
|
||
// 理由见 plugin.go 中"不注入终端已启动通知"的说明)
|
||
waitReadLoopReady(t, term)
|
||
|
||
// 进程自己退出(**不是** Close):mockTerm 返回 process exited。
|
||
term.mu.Lock()
|
||
term.running = false
|
||
term.mu.Unlock()
|
||
|
||
got := waitInjected(capture, "进程已退出", 3*time.Second) ||
|
||
waitInjected(capture, "命令已执行结束", 3*time.Second) ||
|
||
waitInjected(capture, "读取结束", 3*time.Second)
|
||
if !got {
|
||
t.Fatal("进程自己结束必须有通知(修复不能把正常上报一起关掉)")
|
||
}
|
||
}
|
||
|
||
// 竞态版回声判据:Close() 会让 stopCh 与 Read 错误**同时**就绪,
|
||
// 而 readLoop 的 select 在两者都就绪时**随机**选一个分支。
|
||
//
|
||
// 所以「关闭后不注入回声」不能只在"select 恰好选中 stopCh"时成立——
|
||
// 必须在**任何**顺序下都成立。上面那条判据只跑一次,抓不到这个随机性
|
||
// (实测:把 isClosing 改成恒 false 仍然绿)。
|
||
//
|
||
// 这里连开关闭 20 次,只要有一次出现回声就算失败。
|
||
func TestClose_NoEcho_Repeated(t *testing.T) {
|
||
for i := 0; i < 20; i++ {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
startReadLoop(p, sdkInst, ts)
|
||
waitReadLoopReady(t, term)
|
||
before := len(capture.snapshot())
|
||
|
||
// ★ 关键:先在 PTY 缓冲里**留下未读数据**,再关闭。
|
||
//
|
||
// 线上关闭终端时 PTY 缓冲里通常还有未读输出(提示符、命令尾部)。
|
||
// 这时 readLoop 的 select 在 `<-t.stopCh` 与 `<-readCh` 之间是
|
||
// **随机**选择——选到 readCh 就会先处理数据,随后紧跟着拿到
|
||
// EIO(因为 Kill 后读端立刻报错),于是走进 r.err 分支。
|
||
// 若这里不预置数据,select 几乎总是选中 stopCh,回声永远复现不出来
|
||
// (这就是我第一版判据显示"全绿"的原因)。
|
||
term.push([]byte("leftover output" + "\r\n"))
|
||
time.Sleep(30 * time.Millisecond)
|
||
|
||
ts.Close()
|
||
time.Sleep(120 * time.Millisecond)
|
||
|
||
after := capture.snapshot()
|
||
for _, msg := range after[before:] {
|
||
if strings.Contains(msg, "读取结束") || strings.Contains(msg, "进程已退出") ||
|
||
strings.Contains(msg, "命令已执行结束") {
|
||
t.Fatalf("第 %d 次关闭后出现回声(select 随机选中了 r.err 分支): %s", i, msg)
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// 关闭终端后,session 必须**立刻**从表里消失。
|
||
//
|
||
// 这是"关闭终端后反复注入过期状态"的**直接**成因(2026-10-25 线上):
|
||
//
|
||
// close 工具 → session.Close() → readLoop 从 <-t.stopCh 静默 return
|
||
// → 但**没删 sessions 条目**
|
||
// → cleanupLoop 每 30s 扫到它,发现 !terminalRunning
|
||
// → 又注入一条「[终端 term_6 中的进程已退出]」
|
||
// → 而 terminal_list 已经是 0 个会话
|
||
//
|
||
// 所以用户看到的不是"一次回声",而是**每 30 秒一次、永不停止**的
|
||
// 过期状态注入——每条都会唤醒一轮推理,agent 只能反复解释"这是旧回声"。
|
||
//
|
||
// 判据:Close() 返回后 sessions 里不得再有它。
|
||
// 变异:把 removeSession 调用删掉 ⇒ 本判据变红。
|
||
func TestClose_RemovesSessionFromTable(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
mterm2 := newMockTerm()
|
||
ts := newTestSession(mterm2)
|
||
ts.id = "term_leak"
|
||
p.mu.Lock()
|
||
p.sessions = map[string]*TerminalSession{"term_leak": ts}
|
||
p.mu.Unlock()
|
||
|
||
startReadLoop(p, sdkInst, ts)
|
||
// 等 readLoop 真正起来(不再依赖"已启动"注入:那已被移除,
|
||
// 理由见 plugin.go 中"不注入终端已启动通知"的说明)
|
||
waitReadLoopReady(t, mterm2)
|
||
|
||
ts.Close() // 内部会 <-t.done,等 readLoop 走完
|
||
|
||
p.mu.Lock()
|
||
_, still := p.sessions["term_leak"]
|
||
remaining := len(p.sessions)
|
||
p.mu.Unlock()
|
||
|
||
if still {
|
||
t.Fatalf("关闭后 session 仍留在表里(%d 个)——cleanupLoop 会每 30s "+
|
||
"把它当成「进程已退出」再注入一次,形成持续回声", remaining)
|
||
}
|
||
}
|
||
|
||
// cleanupLoop 对**已关闭**的终端不得再注入通知。
|
||
//
|
||
// 这是上一条的补强:即使将来某条路径又漏删了表,清理路径也必须靠
|
||
// isClosing 兜住,不能把"我们自己关掉的"当成"进程自己结束了"上报。
|
||
func TestCleanup_DoesNotNotifyForClosedTerminal(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20
|
||
capture := &injectCapture{}
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
ts.id = "term_closed"
|
||
ts.closed = true // 已被 close 工具关闭(但故意留在表里)
|
||
|
||
if !p.isClosing(ts) {
|
||
t.Fatal("isClosing 应识别 closed 标记")
|
||
}
|
||
|
||
// 模拟「漏删表」的坏状态:关闭了但还在 sessions 里。
|
||
p.sessions = map[string]*TerminalSession{"term_closed": ts}
|
||
term.setRunning(false)
|
||
|
||
// 走一遍 cleanupLoop 的判定逻辑(与生产同一段代码路径的判据部分)。
|
||
p.mu.Lock()
|
||
var wouldNotify []string
|
||
for id, t := range p.sessions {
|
||
if !p.isClosing(t) && !terminalRunning(t) {
|
||
wouldNotify = append(wouldNotify, id)
|
||
}
|
||
}
|
||
p.mu.Unlock()
|
||
|
||
if len(wouldNotify) > 0 {
|
||
t.Fatalf("已关闭的终端不应被 cleanupLoop 当作「进程已退出」上报:%v", wouldNotify)
|
||
}
|
||
_ = capture
|
||
}
|
||
|
||
// ── 2026-10-25 新增:禁止「自造状态回声」的判据 ──────────────────
|
||
//
|
||
// 背景(真机事故):agent 通过 terminal_create 建终端后,插件注入
|
||
// 「[终端 X 已启动]」;超时后又注入「[终端 X 已超时关闭]」。
|
||
// 这两条都是**插件自己造成、且 agent 已能从工具返回值推导**的状态,
|
||
// 而每条注入都会唤醒一整轮 LLM 推理。实测后果:agent 反复向用户
|
||
// 解释"终端怎么又没了",真正要做的活一件没做。
|
||
//
|
||
// 下面两条是**反向判据**:断言这些注入**不再出现**。没有它们,
|
||
// 后人(或我自己)很容易为了"让 agent 感知"把注入加回来,而那时
|
||
// 上面那些 waitReadLoopReady 的测试依然全绿 —— 缺陷会静默复发。
|
||
|
||
// 判据一:建终端不得产生「已启动」注入(terminal_create 返回值已足够)。
|
||
func TestNoStartupNotificationInjected(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20 // 不触发缓冲通知,隔离变量
|
||
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
startReadLoop(p, sdkInst, ts)
|
||
waitReadLoopReady(t, term)
|
||
|
||
// 给足时间让任何潜在注入发生
|
||
time.Sleep(150 * time.Millisecond)
|
||
|
||
for _, text := range capture.snapshot() {
|
||
if strings.Contains(text, "已启动") {
|
||
t.Fatalf("不得注入「已启动」(terminal_create 返回值已告知): %q", text)
|
||
}
|
||
}
|
||
|
||
close(ts.stopCh)
|
||
<-ts.done
|
||
}
|
||
|
||
// 判据二:超时关闭不得注入(创建时 timeout 已告知)。
|
||
func TestNoExpiredNotificationInjected(t *testing.T) {
|
||
p := New("agentcli")
|
||
p.notifyBytes = 1 << 20
|
||
|
||
capture := &injectCapture{}
|
||
sdkInst := sdk.New("agentcli", sdk.SDKConfig{
|
||
RegTool: newToolCapture().RegisterTool,
|
||
RegStage: func(sdk.Stage, sdk.StageHandler) {},
|
||
RegAPI: func(string) error { return nil },
|
||
Settings: sdk.NewSettings("agentcli", nil),
|
||
})
|
||
sdkInst.SetIOInjector(capture)
|
||
|
||
term := newMockTerm()
|
||
ts := newTestSession(term)
|
||
// 设为已超时:readLoop 下一轮循环就会走 IsExpired 分支
|
||
ts.mu.Lock()
|
||
ts.timeout = time.Nanosecond
|
||
ts.createdAt = time.Now().Add(-time.Hour)
|
||
ts.mu.Unlock()
|
||
|
||
startReadLoop(p, sdkInst, ts)
|
||
waitReadLoopReady(t, term)
|
||
|
||
// 等 removeSession 真的发生(说明走了超时分支)
|
||
deadline := time.Now().Add(2 * time.Second)
|
||
for time.Now().Before(deadline) {
|
||
p.mu.Lock()
|
||
_, still := p.sessions[ts.id]
|
||
p.mu.Unlock()
|
||
if !still {
|
||
break
|
||
}
|
||
time.Sleep(2 * time.Millisecond)
|
||
}
|
||
|
||
for _, text := range capture.snapshot() {
|
||
if strings.Contains(text, "已超时关闭") {
|
||
t.Fatalf("不得注入「已超时关闭」(创建时 timeout 已告知): %q", text)
|
||
}
|
||
}
|
||
}
|