AskaEth
454ddb0b70
主人实测：memory_items 1191 行、候选 77 条全 pending、memory_tidy_runs 仅 1 行、两晚定时没跑。 根因两条（都在这片里修掉）： ① **以前根本没有执行面** —— 只有只读试算，候选永远 pending ② **调度被重启饿死** —— interval=24h 的 ticker 从进程启动算起，两天重启十几次 ⇒ 24h 永远走不完 - 应用路径：逐条 / 按 run 批量（默认 100、硬顶 500、超限 truncated）/ revert；**软合并** （status + merged_into/merged_from/merge_count，**绝不改 content、行不删**）+ 审计 （memory.tidy.apply / memory.tidy.revert）+ **幂等**（候选状态条件更新 + 行 status 闸两道） - 调度：改成**到点即跑**（catchUp：无历史+窗口内 / 距上次 run ≥ interval / 窗口起点晚于上次 run）， 「上次 run」以 memory_tidy_runs 为**持久事实**；注入时钟用例证明**重启不饿死** - 手动真跑端点 POST /admin/memories/tidy/run（同档鉴权 + 审计 + 回执：新增/合并/软删/耗时） - 候选表加 run_id 列 + 索引（前滚；顺序=建表→补列→建索引，否则既有库会 no such column） - 3 个变异自检（删 catchUp / 去掉两道 status 闸 / truncated 恒 false） ⚠️ 生产库零动作：1191 行真跑与 77 条候选应用**未执行**（按红线由主人确认后主会话执行） ⚠️ 契约段在报告里（candidates 列表/apply/revert/tidy.run），Docs/ 待登记
396 lines
16 KiB
Go
396 lines
16 KiB
Go
package memory
|
||
|
||
import (
|
||
"context"
|
||
"database/sql"
|
||
"encoding/json"
|
||
"errors"
|
||
"fmt"
|
||
"strings"
|
||
"time"
|
||
)
|
||
|
||
// 本文件是记忆整理的存储层：分区扫描、候选落库与查询、运行留痕。它复用
|
||
// memory 包既有的 *sql.DB 与格式化辅助（formatTime/parseTime/scanItem），
|
||
// 不新建存储机制。
|
||
//
|
||
// 第一版唯一的写入就是候选表与运行表：**没有任何一条 memory_items 会被改写
|
||
// 或删除**（design/23 §3 非目标 + §7「apply 才写记忆」）。
|
||
|
||
// tidyCandidatePending 是候选的初始状态（design/23 §10）。
|
||
const tidyCandidatePending = "pending"
|
||
|
||
// tidyPartition 是整理的隔离单元：design/23 §6 的分区键
|
||
// (persona_id, scope, project_id)。分区之间**永不比较**。
|
||
type tidyPartition struct {
|
||
personaID string
|
||
scope Scope
|
||
projectID string
|
||
}
|
||
|
||
// label 渲染分区标签，供日志与截断登记使用（不含任何记忆正文）。
|
||
func (p tidyPartition) label() string {
|
||
return fmt.Sprintf("%s/%s/%s", p.personaID, p.scope, p.projectID)
|
||
}
|
||
|
||
// tidyStatusActive 是「在册」状态。整理只扫描在册条目：已合并/已归档的行
|
||
// 不再参与比对（design §10 status 列）。
|
||
const tidyStatusActive = "active"
|
||
|
||
// listTidyPartitionItemsSQL 按分区读取**在册**（status='active'）条目，
|
||
// 最新在前；rowid 打破 created_at 并列，使同一刻写入的条目顺序稳定
|
||
// （与 ListByPersona 同一口径）。分区过滤写在 SQL 里，而不是读全表再在内存
|
||
// 里筛 —— design/23 §6 要求在分区切分层就隔断。
|
||
const listTidyPartitionItemsSQL = `SELECT id, persona_id, role_id, project_id, scope, kind,
|
||
content, tags, confidence, source_conversation, created_at, updated_at,
|
||
decay_weight FROM memory_items
|
||
WHERE persona_id = ? AND scope = ? AND project_id = ? AND status = ?
|
||
ORDER BY created_at DESC, rowid DESC LIMIT ?`
|
||
|
||
const listTidyPartitionsSQL = `SELECT DISTINCT persona_id, scope, project_id
|
||
FROM memory_items WHERE status = ?
|
||
ORDER BY persona_id, scope, project_id`
|
||
|
||
const countTidyPartitionItemsSQL = `SELECT COUNT(*) FROM memory_items
|
||
WHERE persona_id = ? AND scope = ? AND project_id = ? AND status = ?`
|
||
|
||
const insertTidyCandidateSQL = `INSERT OR IGNORE INTO memory_tidy_candidates
|
||
(id, persona_id, scope, project_id, kind, action, source_ids, target_id,
|
||
proposed_content, merged_content, reason, evidence, rule_version,
|
||
score_min, score_max, status, idempotency_key, run_id, created_at, decided_at, decided_by)
|
||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, '', '')`
|
||
|
||
const insertTidyRunSQL = `INSERT INTO memory_tidy_runs
|
||
(id, started_at, finished_at, scanned, candidates, applied, skipped, errors,
|
||
vetoed, veto_reasons, rule_version)
|
||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`
|
||
|
||
// listTidyCandidatesByRunSQL 按 **run_id + status** 取候选（最新在前）。run_id
|
||
// 传空串表示「不按轮次筛」—— 那正是「全部待定应用」这条粒度（切片 TIDY-1
|
||
// 三种粒度里最粗的一档）。
|
||
const listTidyCandidatesByRunSQL = `SELECT id, persona_id, scope, project_id, kind, action,
|
||
source_ids, target_id, proposed_content, merged_content, reason, evidence,
|
||
rule_version, score_min, score_max, status, idempotency_key, run_id, created_at
|
||
FROM memory_tidy_candidates
|
||
WHERE status = ? AND (? = '' OR run_id = ?)
|
||
ORDER BY created_at DESC, rowid DESC LIMIT ?`
|
||
|
||
const countTidyCandidatesByRunSQL = `SELECT COUNT(*) FROM memory_tidy_candidates
|
||
WHERE status = ? AND (? = '' OR run_id = ?)`
|
||
|
||
const selectTidyCandidateSQL = `SELECT id, persona_id, scope, project_id, kind, action,
|
||
source_ids, target_id, proposed_content, merged_content, reason, evidence,
|
||
rule_version, score_min, score_max, status, idempotency_key, run_id, created_at
|
||
FROM memory_tidy_candidates WHERE id = ?`
|
||
|
||
const selectTidyRunSQL = `SELECT id, started_at, finished_at, scanned, candidates,
|
||
applied, skipped, errors, vetoed, veto_reasons, rule_version
|
||
FROM memory_tidy_runs WHERE id = ?`
|
||
|
||
// selectLastTidyRunSQL 取最近一轮整理的开始时刻（切片 TIDY-1 的调度判据）：
|
||
// 「上次 run 的时刻」是**持久事实**，不是进程内记忆 —— 后端两天重启十几次
|
||
// 正是旧实现（只看内存 lastRun）永远走不完 24h 周期的根因。
|
||
const selectLastTidyRunSQL = `SELECT started_at FROM memory_tidy_runs
|
||
ORDER BY started_at DESC, rowid DESC LIMIT 1`
|
||
|
||
// incrementTidyRunAppliedSQL 在批量应用之后把该轮的 applied 计数累加上去
|
||
// （design/23 §10 的 memory_tidy_runs.applied）。运行时那一轮写的是 0（当轮
|
||
// 绝不自动应用，§7），这里补的是「事后来应用了几条」。
|
||
const incrementTidyRunAppliedSQL = `UPDATE memory_tidy_runs SET applied = applied + ? WHERE id = ?`
|
||
|
||
const listTidyRunsSQL = `SELECT id, started_at, finished_at, scanned, candidates,
|
||
applied, skipped, errors, vetoed, veto_reasons, rule_version
|
||
FROM memory_tidy_runs ORDER BY started_at DESC, rowid DESC LIMIT ?`
|
||
|
||
// listTidyPartitions 返回所有含在册条目的分区。DISTINCT 让每个分区的
|
||
// 扫描上限（design §9 每分区 5000 条）各自独立。
|
||
func (s *Store) listTidyPartitions(ctx context.Context) ([]tidyPartition, error) {
|
||
rows, err := s.db.QueryContext(ctx, listTidyPartitionsSQL, string(tidyStatusActive))
|
||
if err != nil {
|
||
return nil, fmt.Errorf("memory: list tidy partitions: %w", err)
|
||
}
|
||
defer func() { _ = rows.Close() }()
|
||
|
||
parts := make([]tidyPartition, 0)
|
||
for rows.Next() {
|
||
var p tidyPartition
|
||
var scope string
|
||
if err := rows.Scan(&p.personaID, &scope, &p.projectID); err != nil {
|
||
return nil, fmt.Errorf("memory: scan tidy partition: %w", err)
|
||
}
|
||
p.scope = Scope(scope)
|
||
parts = append(parts, p)
|
||
}
|
||
if err := rows.Err(); err != nil {
|
||
return nil, fmt.Errorf("memory: iterate tidy partitions: %w", err)
|
||
}
|
||
return parts, nil
|
||
}
|
||
|
||
// listTidyPartitionItems 读取一个分区内的在册条目（最新在前，最多 limit 条）。
|
||
func (s *Store) listTidyPartitionItems(ctx context.Context, p tidyPartition, limit int) ([]MemoryItem, error) {
|
||
rows, err := s.db.QueryContext(ctx, listTidyPartitionItemsSQL,
|
||
p.personaID, string(p.scope), p.projectID, string(tidyStatusActive), limit)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("memory: list tidy partition items: %w", err)
|
||
}
|
||
return scanItems(rows, "tidy partition item")
|
||
}
|
||
|
||
// countTidyPartitionItems 数一个分区里在册条目的**真总数**。只在扫描上限命中时
|
||
// 调用一次：日志里的「总数」必须是真数，而不是被 LIMIT 截出来的 len(items)
|
||
// （那是**下限**，把「至少 5000」写成「正好 5000」就是一句谎）。
|
||
func (s *Store) countTidyPartitionItems(ctx context.Context, p tidyPartition) (int, error) {
|
||
var total int
|
||
if err := s.db.QueryRowContext(ctx, countTidyPartitionItemsSQL,
|
||
p.personaID, string(p.scope), p.projectID, string(tidyStatusActive)).Scan(&total); err != nil {
|
||
return 0, fmt.Errorf("memory: count tidy partition items: %w", err)
|
||
}
|
||
return total, nil
|
||
}
|
||
|
||
// insertTidyCandidate 落库一条候选，返回它是否是**新建**的。幂等键
|
||
// （design/23 §9）唯一：同一批数据重复运行命中已存在行时返回 (false, nil)，
|
||
// 绝不重复产候选。
|
||
func (s *Store) insertTidyCandidate(ctx context.Context, c TidyCandidate) (bool, error) {
|
||
c.IdempotencyKey = tidyIdempotencyKey(c)
|
||
sourceJSON, err := json.Marshal(c.SourceIDs)
|
||
if err != nil {
|
||
return false, fmt.Errorf("memory: encode tidy source ids: %w", err)
|
||
}
|
||
evidenceJSON, err := json.Marshal(c.Evidence)
|
||
if err != nil {
|
||
return false, fmt.Errorf("memory: encode tidy evidence: %w", err)
|
||
}
|
||
if c.CreatedAt.IsZero() {
|
||
c.CreatedAt = time.Now().UTC()
|
||
}
|
||
res, err := s.db.ExecContext(ctx, insertTidyCandidateSQL,
|
||
c.ID, c.PersonaID, string(c.Scope), c.ProjectID, string(c.Kind), string(c.Action),
|
||
string(sourceJSON), c.TargetID, c.ProposedContent, c.MergedContent, c.Reason,
|
||
string(evidenceJSON), c.RuleVersion, c.ScoreMin, c.ScoreMax, c.Status,
|
||
c.IdempotencyKey, c.RunID, formatTime(c.CreatedAt))
|
||
if err != nil {
|
||
return false, fmt.Errorf("memory: insert tidy candidate: %w", err)
|
||
}
|
||
n, err := res.RowsAffected()
|
||
if err != nil {
|
||
return false, fmt.Errorf("memory: tidy candidate rows affected: %w", err)
|
||
}
|
||
return n > 0, nil
|
||
}
|
||
|
||
// persistTidyRun 写一条运行留痕（design/23 §10 memory_tidy_runs）。
|
||
func (s *Store) persistTidyRun(ctx context.Context, run TidyRun) error {
|
||
if run.ID == "" {
|
||
run.ID = mustNewTidyID()
|
||
}
|
||
reasons, err := json.Marshal(run.VetoReasons)
|
||
if err != nil {
|
||
return fmt.Errorf("memory: encode tidy veto reasons: %w", err)
|
||
}
|
||
if _, err := s.db.ExecContext(ctx, insertTidyRunSQL,
|
||
run.ID, formatTime(run.StartedAt), formatTime(run.FinishedAt),
|
||
run.Scanned, run.Candidates, run.Applied, run.Skipped, run.Errors,
|
||
run.Vetoed, string(reasons), run.RuleVersion); err != nil {
|
||
return fmt.Errorf("memory: insert tidy run: %w", err)
|
||
}
|
||
return nil
|
||
}
|
||
|
||
// ListTidyCandidates 返回指定状态的整理候选（最新在前，最多 limit 条，
|
||
// 非正 limit 取 defaultTidyCandidateLimit）。它是读侧入口：候选正文含隐私，
|
||
// 只应由 owner-only 的读侧使用（design/23 §11）。
|
||
func (s *Store) ListTidyCandidates(ctx context.Context, status string, limit int) ([]TidyCandidate, error) {
|
||
return s.listTidyCandidates(ctx, status, "", limit)
|
||
}
|
||
|
||
// ListTidyCandidatesByRun 返回某一轮整理产出的候选（status 默认 pending）。
|
||
// runID 为空串表示**不按轮次筛** —— 即「全部待定应用」那一档（切片 TIDY-1）。
|
||
func (s *Store) ListTidyCandidatesByRun(ctx context.Context, runID, status string, limit int) ([]TidyCandidate, error) {
|
||
return s.listTidyCandidates(ctx, status, runID, limit)
|
||
}
|
||
|
||
// listTidyCandidates 是上面两个读侧入口的共用实现：runID 为空串时走「不筛轮次」
|
||
// 的 SQL 分支（`? = ” OR run_id = ?`），因此只有一个查询形状要维护。
|
||
func (s *Store) listTidyCandidates(ctx context.Context, status, runID string, limit int) ([]TidyCandidate, error) {
|
||
if strings.TrimSpace(status) == "" {
|
||
status = tidyCandidatePending
|
||
}
|
||
if limit <= 0 {
|
||
limit = defaultTidyCandidateLimit
|
||
}
|
||
rows, err := s.db.QueryContext(ctx, listTidyCandidatesByRunSQL, status, runID, runID, limit)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("memory: list tidy candidates: %w", err)
|
||
}
|
||
defer func() { _ = rows.Close() }()
|
||
|
||
out := make([]TidyCandidate, 0)
|
||
for rows.Next() {
|
||
c, err := scanTidyCandidate(rows)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("memory: scan tidy candidate: %w", err)
|
||
}
|
||
out = append(out, c)
|
||
}
|
||
if err := rows.Err(); err != nil {
|
||
return nil, fmt.Errorf("memory: iterate tidy candidates: %w", err)
|
||
}
|
||
return out, nil
|
||
}
|
||
|
||
// GetTidyCandidate 按 id 读一条候选（含已决定的行）。不存在时返回 ErrNotFound。
|
||
func (s *Store) GetTidyCandidate(ctx context.Context, id string) (TidyCandidate, error) {
|
||
if strings.TrimSpace(id) == "" {
|
||
return TidyCandidate{}, ErrNotFound
|
||
}
|
||
row := s.db.QueryRowContext(ctx, selectTidyCandidateSQL, id)
|
||
c, err := scanTidyCandidate(row)
|
||
if errors.Is(err, sql.ErrNoRows) {
|
||
return TidyCandidate{}, ErrNotFound
|
||
}
|
||
if err != nil {
|
||
return TidyCandidate{}, fmt.Errorf("memory: get tidy candidate: %w", err)
|
||
}
|
||
return c, nil
|
||
}
|
||
|
||
// countTidyCandidates 数某状态下（可选按轮次）的候选总数。批量应用用它并把
|
||
// 「真总数 vs 本轮应用上限」如实回报给调用方（超限绝不静默）。
|
||
func (s *Store) countTidyCandidates(ctx context.Context, status, runID string) (int, error) {
|
||
if strings.TrimSpace(status) == "" {
|
||
status = tidyCandidatePending
|
||
}
|
||
var total int
|
||
if err := s.db.QueryRowContext(ctx, countTidyCandidatesByRunSQL, status, runID, runID).
|
||
Scan(&total); err != nil {
|
||
return 0, fmt.Errorf("memory: count tidy candidates: %w", err)
|
||
}
|
||
return total, nil
|
||
}
|
||
|
||
// GetTidyRun 按 id 读一条运行留痕。不存在时返回 ErrTidyRunRecordNotFound ——
|
||
// 面板拿一个拼错的 runId 来批量应用时，必须是「查不到这一轮」而不是「这一轮
|
||
// 没有待定候选」。
|
||
func (s *Store) GetTidyRun(ctx context.Context, id string) (TidyRun, error) {
|
||
if strings.TrimSpace(id) == "" {
|
||
return TidyRun{}, ErrTidyRunRecordNotFound
|
||
}
|
||
row := s.db.QueryRowContext(ctx, selectTidyRunSQL, id)
|
||
r, err := scanTidyRun(row)
|
||
if errors.Is(err, sql.ErrNoRows) {
|
||
return TidyRun{}, ErrTidyRunRecordNotFound
|
||
}
|
||
if err != nil {
|
||
return TidyRun{}, fmt.Errorf("memory: get tidy run: %w", err)
|
||
}
|
||
return r, nil
|
||
}
|
||
|
||
// LastTidyRunStartedAt 返回最近一轮整理的开始时刻；一次都没跑过时返回零值
|
||
// 时间与 nil（**不是**错误：那是「还没有历史」这一正常状态，调度据此判断
|
||
// 是否需要补跑）。
|
||
func (s *Store) LastTidyRunStartedAt(ctx context.Context) (time.Time, error) {
|
||
var startedAt string
|
||
err := s.db.QueryRowContext(ctx, selectLastTidyRunSQL).Scan(&startedAt)
|
||
if errors.Is(err, sql.ErrNoRows) {
|
||
return time.Time{}, nil
|
||
}
|
||
if err != nil {
|
||
return time.Time{}, fmt.Errorf("memory: read last tidy run: %w", err)
|
||
}
|
||
t, err := parseTime(startedAt)
|
||
if err != nil {
|
||
return time.Time{}, err
|
||
}
|
||
return t, nil
|
||
}
|
||
|
||
// ListTidyRuns 返回整理运行留痕（最新在前，最多 limit 条）。非正 limit 取
|
||
// 一个保守的默认条数。
|
||
func (s *Store) ListTidyRuns(ctx context.Context, limit int) ([]TidyRun, error) {
|
||
if limit <= 0 {
|
||
limit = defaultTidyCandidateLimit
|
||
}
|
||
rows, err := s.db.QueryContext(ctx, listTidyRunsSQL, limit)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("memory: list tidy runs: %w", err)
|
||
}
|
||
defer func() { _ = rows.Close() }()
|
||
|
||
out := make([]TidyRun, 0)
|
||
for rows.Next() {
|
||
r, err := scanTidyRun(rows)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("memory: scan tidy run: %w", err)
|
||
}
|
||
out = append(out, r)
|
||
}
|
||
if err := rows.Err(); err != nil {
|
||
return nil, fmt.Errorf("memory: iterate tidy runs: %w", err)
|
||
}
|
||
return out, nil
|
||
}
|
||
|
||
// scanTidyRun 读取一行运行留痕：veto_reasons 以 JSON 存储，时间戳为定宽 UTC 文本。
|
||
func scanTidyRun(sc rowScanner) (TidyRun, error) {
|
||
var (
|
||
r TidyRun
|
||
startedAt string
|
||
finishedAt string
|
||
reasonsJSON string
|
||
)
|
||
if err := sc.Scan(&r.ID, &startedAt, &finishedAt, &r.Scanned, &r.Candidates,
|
||
&r.Applied, &r.Skipped, &r.Errors, &r.Vetoed, &reasonsJSON, &r.RuleVersion); err != nil {
|
||
return TidyRun{}, err
|
||
}
|
||
var err error
|
||
if r.StartedAt, err = parseTime(startedAt); err != nil {
|
||
return TidyRun{}, err
|
||
}
|
||
if r.FinishedAt, err = parseTime(finishedAt); err != nil {
|
||
return TidyRun{}, err
|
||
}
|
||
if err := json.Unmarshal([]byte(reasonsJSON), &r.VetoReasons); err != nil {
|
||
return TidyRun{}, fmt.Errorf("memory: decode tidy veto reasons: %w", err)
|
||
}
|
||
return r, nil
|
||
}
|
||
|
||
// scanTidyCandidate 读取一行候选：source_ids 与 evidence 以 JSON 存储。
|
||
func scanTidyCandidate(sc rowScanner) (TidyCandidate, error) {
|
||
var (
|
||
c TidyCandidate
|
||
scope, kind string
|
||
action string
|
||
sourceJSON string
|
||
evidenceRaw string
|
||
createdAt string
|
||
)
|
||
if err := sc.Scan(&c.ID, &c.PersonaID, &scope, &c.ProjectID, &kind, &action,
|
||
&sourceJSON, &c.TargetID, &c.ProposedContent, &c.MergedContent, &c.Reason,
|
||
&evidenceRaw, &c.RuleVersion, &c.ScoreMin, &c.ScoreMax, &c.Status,
|
||
&c.IdempotencyKey, &c.RunID, &createdAt); err != nil {
|
||
return TidyCandidate{}, err
|
||
}
|
||
c.Scope = Scope(scope)
|
||
c.Kind = Kind(kind)
|
||
c.Action = TidyAction(action)
|
||
if err := json.Unmarshal([]byte(sourceJSON), &c.SourceIDs); err != nil {
|
||
return TidyCandidate{}, fmt.Errorf("memory: decode tidy source ids: %w", err)
|
||
}
|
||
if c.SourceIDs == nil {
|
||
c.SourceIDs = []string{}
|
||
}
|
||
if err := json.Unmarshal([]byte(evidenceRaw), &c.Evidence); err != nil {
|
||
return TidyCandidate{}, fmt.Errorf("memory: decode tidy evidence: %w", err)
|
||
}
|
||
t, err := parseTime(createdAt)
|
||
if err != nil {
|
||
return TidyCandidate{}, err
|
||
}
|
||
c.CreatedAt = t
|
||
return c, nil
|
||
}
|