三层回退恢复机制(L0写前留档/L1恢复梯子/L2离线回滚)+ guard 父守护

- L0: files 插件写受保护系统路径(/etc 等)前自动留档,AbstractBeforeWrite 到 data/file_baseline
- L1: failback 受限 worker 执行恢复梯子 probe→还原DNS/proxy→还原LLM配置+ReloadFromConfig→probe,N轮有界
- L2: tracker changeset 持久化原文 blob,guard 离线 RollbackFromDisk 回滚 agentfs;SystemSnapshot 支撑
- guard 父守护: 心跳 IPC(PING/ACK unix socket, 文件心跳回退)、失败计数、退出码协议(42/43/44)、最后手段
- 发行版路径适配: system.protected_paths/network_paths 可注入,默认面向主流 Linux
- Windows 兼容: guard.go/failback.go 加 //go:build linux, guard_windows.go 提供 no-op 桩
- 修复: guard.yaml last_resort 键冲突、changeset Content 不落盘导致离线回滚丢原文

Build 全绿, vet 干净, system/recovery/ipc/tracker 单元测试全过
This commit is contained in:
root
2026-08-05 16:00:08 +08:00
parent fc614477ed
commit d08ed3312e
28 changed files with 3675 additions and 167 deletions

View File

@ -2,7 +2,10 @@ package config
import (
"database/sql"
"encoding/json"
"fmt"
"log"
"os"
"path/filepath"
"sort"
"strconv"
@ -28,10 +31,11 @@ type ConfigDef struct {
}
type ConfigRegistry struct {
mu sync.RWMutex
db *sql.DB
dbPath string
defs map[string]*ConfigDef
mu sync.RWMutex
db *sql.DB
dbPath string
defs map[string]*ConfigDef
llmSnapFile string
}
func NewConfigRegistry(dbPath string) *ConfigRegistry {
@ -265,6 +269,9 @@ func (r *ConfigRegistry) Get(key string) (interface{}, error) {
func (r *ConfigRegistry) Set(key string, value interface{}) error {
r.mu.Lock()
defer r.mu.Unlock()
if strings.HasPrefix(key, "core.llm.") {
r.writeLLMSnapshotLocked()
}
_, err := r.db.Exec(`INSERT OR REPLACE INTO config (key, value) VALUES (?, ?)`, key, fmt.Sprint(value))
if err == nil && strings.HasPrefix(key, "core.llm.sources.") {
rest := strings.TrimPrefix(key, "core.llm.sources.")
@ -303,6 +310,107 @@ func (r *ConfigRegistry) Delete(key string) error {
return err
}
// SnapshotCoreLLM 捕获全部 core.llm.* 键值(LLM 源密度快照),供写前留档。
// 返回值是 key→value 的不可变拷贝;写入 guard 配置恢复的基线。
func (r *ConfigRegistry) SnapshotCoreLLM() map[string]string {
r.mu.RLock()
defer r.mu.RUnlock()
return r.listPrefixLocked("core.llm.")
}
// SetLLMSnapshotFile 设定写前留档文件:此后任意写入 core.llm.* 键时,
// 先把当前 llm 配置整体快照到该文件(guard 恢复的外部基线)。
func (r *ConfigRegistry) SetLLMSnapshotFile(path string) {
r.mu.Lock()
defer r.mu.Unlock()
if path != "" {
r.writeLLMSnapshotToFileLocked(path)
}
r.llmSnapFile = path
}
// writeLLMSnapshotLocked 调用方须持有写锁;若已配置快照文件则写入当前 llm 快照。
func (r *ConfigRegistry) writeLLMSnapshotLocked() {
if r.llmSnapFile == "" {
return
}
r.writeLLMSnapshotToFileLocked(r.llmSnapFile)
}
func (r *ConfigRegistry) writeLLMSnapshotToFileLocked(path string) {
snap := r.listPrefixLocked("core.llm.")
if err := SaveLLMSnapshot(path, snap); err != nil {
log.Printf("[config] save llm snapshot %s: %v", path, err)
}
}
// RestoreCoreLLM 精确还原到快照状态:快照里有的键回写旧值,
// 当前存在但快照里没有的 core.llm.* 键删除(保持与快照一致)。
func (r *ConfigRegistry) RestoreCoreLLM(snap map[string]string) error {
r.mu.Lock()
defer r.mu.Unlock()
current := r.listPrefixLocked("core.llm.")
seen := make(map[string]bool, len(snap))
for k, v := range snap {
seen[k] = true
if _, err := r.db.Exec(`INSERT OR REPLACE INTO config (key, value) VALUES (?, ?)`, k, v); err != nil {
return err
}
}
for k := range current {
if seen[k] {
continue
}
if _, err := r.db.Exec(`DELETE FROM config WHERE key = ?`, k); err != nil {
return err
}
}
return nil
}
// SaveLLMSnapshot 把 LLM 快照持久化到文件(guard 恢复的外部基线)。
func SaveLLMSnapshot(path string, snap map[string]string) error {
if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
return err
}
data, err := json.MarshalIndent(snap, "", " ")
if err != nil {
return err
}
return os.WriteFile(path, data, 0600)
}
// LoadLLMSnapshot 从文件读回 LLM 快照。
func LoadLLMSnapshot(path string) (map[string]string, error) {
data, err := os.ReadFile(path)
if err != nil {
return nil, err
}
snap := make(map[string]string)
if err := json.Unmarshal(data, &snap); err != nil {
return nil, err
}
return snap, nil
}
// listPrefixLocked 调用方须持有锁;返回 prefix 开头的全部键值。
func (r *ConfigRegistry) listPrefixLocked(prefix string) map[string]string {
out := make(map[string]string)
rows, err := r.db.Query(`SELECT key, value FROM config WHERE key LIKE ? ORDER BY key`, prefix+"%")
if err != nil {
return out
}
defer rows.Close()
for rows.Next() {
var k, v string
if err := rows.Scan(&k, &v); err == nil {
out[k] = v
}
}
return out
}
func (r *ConfigRegistry) Dump() map[string]interface{} {
r.mu.RLock()
defer r.mu.RUnlock()
@ -464,6 +572,7 @@ func (r *ConfigRegistry) seedCoreDefs(dataDir string) {
reg(ConfigDef{Key: "core.daemon.heartbeat_interval", Default: "15s", Type: "duration", DisplayName: "心跳间隔", Description: "Agent 心跳检查间隔", Category: "daemon"})
reg(ConfigDef{Key: "core.daemon.check_interval", Default: "30s", Type: "duration", DisplayName: "检查间隔", Description: "网络状态检查间隔", Category: "daemon"})
reg(ConfigDef{Key: "core.daemon.log_level", Default: "info", Type: "select", DisplayName: "日志级别", Description: "日志输出级别", Options: []string{"debug", "info", "warn", "error"}, Category: "daemon"})
reg(ConfigDef{Key: "core.defaults.llm_endpoints", Default: "", Type: "string", DisplayName: "探活端点", Description: "健康检查的 LLM 探活端点,逗号分隔;留空自动取 LLM 源 base_url", Category: "daemon"})
reg(ConfigDef{Key: "core.llm.provider", Default: "deepseek", Type: "string", DisplayName: "默认提供商", Description: "默认 LLM 提供商名称,需匹配 sources 中的定义", Category: "llm"})
reg(ConfigDef{Key: "core.llm.model", Default: "deepseek-v4-flash", Type: "string", DisplayName: "默认模型", Description: "默认 LLM 模型名称", Category: "llm"})
@ -698,6 +807,23 @@ func (r *ConfigRegistry) ToConfig() *types.Config {
cfg.Plugin.Dir = read("core.plugin.dir", cfg.Plugin.Dir)
// 探活端点:优先显式配置,缺省取 LLM 源 base_url(去重),保证健康检查有实际目标
if eps := read("core.defaults.llm_endpoints", ""); eps != "" {
for _, ep := range strings.Split(eps, ",") {
if ep = strings.TrimSpace(ep); ep != "" {
cfg.Defaults.LLMEndpoints = append(cfg.Defaults.LLMEndpoints, ep)
}
}
} else {
seen := make(map[string]bool)
for _, src := range cfg.LLM.Sources {
if src.BaseURL != "" && !seen[src.BaseURL] {
seen[src.BaseURL] = true
cfg.Defaults.LLMEndpoints = append(cfg.Defaults.LLMEndpoints, src.BaseURL)
}
}
}
cfg.InputProcessing.Image.FallbackProvider = read("core.input_processing.image.fallback_provider", cfg.InputProcessing.Image.FallbackProvider)
cfg.InputProcessing.Image.FallbackModel = read("core.input_processing.image.fallback_model", cfg.InputProcessing.Image.FallbackModel)
cfg.InputProcessing.Image.DescribePrompt = read("core.input_processing.image.describe_prompt", cfg.InputProcessing.Image.DescribePrompt)

View File

@ -220,3 +220,73 @@ func TestGetHelpers(t *testing.T) {
t.Fatalf("GetBool fallback: expected true, got %v", got)
}
}
func TestSnapshotRestoreCoreLLM(t *testing.T) {
r := NewConfigRegistry("")
defer r.Close()
r.Set("core.llm.sources.main.base_url", "https://a")
r.Set("core.llm.sources.main.model", "m1")
r.Set("core.llm.sources.main.api_key", "k1")
snap := r.SnapshotCoreLLM()
if len(snap) != 3 {
t.Fatalf("expected 3 keys, got %d: %v", len(snap), snap)
}
// 模拟写坏
r.Set("core.llm.sources.main.base_url", "https://broken")
r.Set("core.llm.sources.main.api_key", "hacked")
r.Set("core.llm.sources.extra.model", "intruder")
if err := r.RestoreCoreLLM(snap); err != nil {
t.Fatalf("RestoreCoreLLM: %v", err)
}
got := r.SnapshotCoreLLM()
if len(got) != 3 {
t.Fatalf("after restore expected 3 keys, got %d: %v", len(got), got)
}
for k, v := range snap {
if got[k] != v {
t.Errorf("key %s: want %q got %q", k, v, got[k])
}
}
}
func TestLLMSnapshotFile(t *testing.T) {
path := filepath.Join(t.TempDir(), "llm_snapshot.json")
snap := map[string]string{"core.llm.sources.main.base_url": "https://a", "core.llm.sources.main.model": "m1"}
if err := SaveLLMSnapshot(path, snap); err != nil {
t.Fatalf("Save: %v", err)
}
got, err := LoadLLMSnapshot(path)
if err != nil {
t.Fatalf("Load: %v", err)
}
if got["core.llm.sources.main.base_url"] != "https://a" || got["core.llm.sources.main.model"] != "m1" {
t.Fatalf("round-trip mismatch: %v", got)
}
}
func TestSetLLMSnapshotFile(t *testing.T) {
r := NewConfigRegistry("")
defer r.Close()
r.Set("core.llm.sources.main.base_url", "https://orig")
r.Set("core.llm.sources.main.model", "m0")
path := filepath.Join(t.TempDir(), "llm_pre.json")
r.SetLLMSnapshotFile(path)
// 再次写入:写前自动留档应记录当前值 orig/m0,随后才被覆盖
r.Set("core.llm.sources.main.base_url", "https://broken")
got, err := LoadLLMSnapshot(path)
if err != nil {
t.Fatalf("Load snapshot: %v", err)
}
if got["core.llm.sources.main.base_url"] != "https://orig" {
t.Fatalf("write-ahead snapshot should record pre-write value, got %q", got["core.llm.sources.main.base_url"])
}
if got["core.llm.sources.main.model"] != "m0" {
t.Fatalf("snapshot missing untouched key model: %v", got)
}
}