@ilaomianyang/dsh-memory-graph 0.11.0 → 0.12.0

package/CHANGELOG.md

@@ -1,5 +1,24 @@
 # Changelog
 
+## 0.12.0 — 2026-11-03
+
+装完之后就不用再管它了：插件自己去问有没有新版本，积压的记忆自己一批批追平，视图跟着库走。
+
+**Added**
+
+- **自我更新**。设置页最上面一行「版本与更新」：当前版本、检查更新、立即更新。启动时查一次，之后每 `update.checkEveryHours` 小时（默认 12）最多一次——一次 `GET <registry>/@ilaomianyang%2fdsh-memory-graph/latest`，不带查询串、不带记忆库路径、不带任何对话内容；结果缓存进库里的 `.update-state.json`，离线时保留上一次的答案，不会把"查不到"显示成"已是最新"。
+- 安装走宿主自己的 `pluginManager.installBundle`（和 `dsh plugin` 同一条 pnpm 路径，尊重已配的 registry），装完照实说「需要重启 Harness 生效」还是「已生效」。**`link:` / `file:` 本地安装一律拒绝**，并说明会覆盖开发目录——这句是拒绝，不是报错。
+- `update.autoInstall`（默认关）：启动时发现新版本就自己装。它在任何对话开始之前动手，是唯一"换掉插件不会打断你正在说的话"的时刻；`link:` 安装同样拒绝。`memory_stats` 也报一行版本。
+- 新路由 `/api/memory-graph.update`：GET 读状态（`current/latest/behind/canInstall/installKind/busy/message`），POST 是唯一能装的方法，所以页面加载或预取永远不会误装。
+- **追平模式**（`auto.catchUp`，默认关）。原来的两条规则一次只消化一批，两千条积压要等几十天；打开追平后，只要还攒着**一整批**（≥ `maxEpisodesPerRun`）就按自己的节奏（`catchUpEveryMinutes`，默认 10 分钟）一批批蒸下去，一天最多 `auto.maxBatchesPerDay` 批（默认 6，0 = 不限），追平就自己停。每日批次计数存进 `.auto-state.json` 的 `batchDay`/`batchesToday`，重启不重置；批次在调用**之前**就计数，失败也算用掉一个名额——上限才是真的上限。静默时段压过追平。
+- `memory_backfill` 写完后顺手重画 `README.md` 和 `graph.html`（报告里 `refreshed`），导入两千条情节后打开视图看到的是当下的库，不是导入前的空图。
+- `test/update.mjs` 27 项 + `test/auto.mjs` 加 6 项追平检查。检查数 244 → 277。
+
+**Changed**
+
+- 设置页多了 7 个旋钮：`auto.catchUp` / `auto.catchUpEveryMinutes` / `auto.maxBatchesPerDay`，以及新的「版本与更新」分区（`update.enabled` / `checkEveryHours` / `autoInstall` / `registry`）。版本那一行不是 config 字段，是这一块自己 GET/POST 出来的 UI，刻意不占 `mg-card` / `mg-row` 这些类名。
+- `README.md` / `README.en.md`：「自动蒸馏」从两条规则写成三条，新增「它自己更新」一节，配置表 34 → 41 行，YAML 例子带上 `auto.catchUp` 与 `update`。
+
 ## 0.11.0 — 2026-11-03
 
 历史会话也能进库了，宿主侧的文案统一成中文。

package/README.en.md

@@ -96,6 +96,23 @@ After the restart it is **Settings → Memory Graph**, between Models and Plugin
 
 Uninstall with `plugin_manager action=remove_bundle target="@ilaomianyang/dsh-memory-graph"`. The vault is not touched — it lives in `~/dsh-memory`, independent of the plugin.
 
+## It updates itself
+
+The copy installed from npm gets a **Version & updates** row at the top of its Settings page: the version you are running, **Check**, and **Update**.
+
+![Version & updates in Settings: the current version, Check, Update, then four switches](docs/update-settings.png)
+
+That is the real page (on a development machine the install is a `link:`, so Update is greyed out and says why next to it): the version, whether it is current, which registry, when it last checked — and the four `update.*` switches below.
+
+- It checks once at startup, then at most every `update.checkEveryHours` hours (default 12). The request is one `GET` of `<registry>/@ilaomianyang%2fdsh-memory-graph/latest` — no query string, no vault path, no conversation content. The answer is cached in the vault's `.update-state.json`, so an offline machine keeps the last answer instead of showing "could not check" as "up to date".
+- Nothing is installed until you press Update. Installation goes through the host's own `pluginManager.installBundle` — the same path `dsh plugin` and the plugin manager use — so it respects your configured registry instead of inventing a second package manager.
+- **A `link:` install is refused**, and the refusal is a sentence rather than an error: "updating would overwrite your working copy". For a clone, updating is `git pull` and a reload, as above.
+- An install usually needs a Harness restart (same caching reason as `client.js` / `package.json`). The page says **restart Harness to apply** rather than pretending it is already running the new code.
+- `update.autoInstall` (off by default) installs a newer version at startup without being asked. It acts before any conversation starts, which is the one moment swapping the plugin cannot interrupt something you were saying. A `link:` install is still refused.
+- `memory_stats` reports the version too, so the model does not have to guess which copy it is talking to.
+
+Turn `update.enabled` off to stop it going online entirely: after that it sends nothing, and the page only shows the current version.
+
 ## What you get on disk
 
 ```
@@ -145,6 +162,13 @@ One file is one memory. Append-only. No binaries, no lockfiles. Run `git init` i
     retrieval:
       limit: 8
       hops: 1
+    auto:
+      enabled: true
+      catchUp: true             # drain a backlog, one batch at a time
+      maxBatchesPerDay: 6
+    update:
+      enabled: true             # check for a newer version at startup
+      autoInstall: false
 ```
 
 | Key | Default | Meaning | On the page |
@@ -183,37 +207,52 @@ One file is one memory. Append-only. No binaries, no lockfiles. Run `git init` i
 | `auto.minGapMinutes` | `30` | shared by both rules: minimum gap between two automatic runs | ✓ |
 | `auto.quietHours` | `''` | window in which nothing runs (e.g. `23:00-08:00`, may cross midnight) | ✓ |
 | `auto.project` | `''` | project automatic cards are filed under | ✓ |
+| `auto.catchUp` | `false` | catch-up: keep running batches while a full batch is still waiting | ✓ |
+| `auto.catchUpEveryMinutes` | `10` | minutes between catch-up batches (its own cadence) | ✓ |
+| `auto.maxBatchesPerDay` | `6` | most automatic batches per day, 0 lifts the cap; the one ceiling on automatic spending | ✓ |
+| `update.enabled` | `true` | ask npm for a newer version regularly (checks only, nothing about the vault leaves) | ✓ |
+| `update.checkEveryHours` | `12` | how long a check result stays good for | ✓ |
+| `update.autoInstall` | `false` | install a newer version at startup (a `link:` install is refused) | ✓ |
+| `update.registry` | `''` | which npm registry to ask; empty = the harness's, then npmjs | ✓ |
 | `retrieval.limit` | `8` | how many cards one search returns | ✓ |
 | `retrieval.hops` | `1` | entity expansion depth, `0` = keyword only | ✓ |
 | `retrieval.snippetChars` | `320` | how much of each hit the model gets to read | ✓ |
 | `retrieval.timelineLimit` | `20` | how far back `memory_timeline` looks by default | ✓ |
 
-Why only those 34 are editable there: the harness projects **only fields declared `volatile`** — the ones read each time they are used. Everything marked ✓ is read at call time, so one save changes the next call. `enabled` / `vaultDir` / `defaultProject` / `crossProject` are consumed during activation and would need a remount, so they are left out rather than shown as knobs that do nothing; the page's footer says exactly that.
+Why only those 41 are editable there: the harness projects **only fields declared `volatile`** — the ones read each time they are used. Everything marked ✓ is read at call time, so one save changes the next call. `enabled` / `vaultDir` / `defaultProject` / `crossProject` are consumed during activation and would need a remount, so they are left out rather than shown as knobs that do nothing; the page's footer says exactly that.
 
-### When it distils on its own: two rules
+### When it distils on its own: three rules
 
 By default it does **not**. With `auto.enabled` on, the plugin looks once a minute at whether a run is due — and that look is arithmetic only (how many episodes are waiting, how long since the last attempt, is this a quiet hour). **No model call, no cost.** Only a rule that matches starts a run, and it runs the same `distill()`.
 
-The two rules are independent: either one alone, or both.
+The three rules are independent: any one alone, or all of them.
 
 | Rule | Setting | Fires when |
 | --- | --- | --- |
 | Distil once enough pile up | `auto.minEpisodes` | undistilled episodes ≥ this number (say 10) |
 | Distil on a schedule | `auto.everyMinutes` | this many minutes have passed since the last automatic run (1440 = daily) |
+| Catch up | `auto.catchUp` | a **full batch** is still waiting (≥ `maxEpisodesPerRun`): run batch after batch, `auto.catchUpEveryMinutes` apart (default 10), at most `auto.maxBatchesPerDay` batches a day (default 6, 0 lifts the cap) |
+
+Catch-up is for a **backlog** — an imported history, a week away from the machine, two thousand undistilled episodes. The other two rules digest one batch per trigger and would take dozens of days; catch-up keeps going until less than a batch is left, then goes quiet.
+
+- It has **its own cadence** (`catchUpEveryMinutes`) and is not held back by `minGapMinutes`; the other two rules still are.
+- `maxBatchesPerDay` is the only knob that puts a **ceiling on automatic spending**: the default six batches is about 240 episodes a day. Raise it to drain faster, or set `quietHours` so it does not run at night.
+- An interrupted catch-up (restart, shutdown, another run holding the lock) loses nothing: the counter lives in `.auto-state.json`, the next interval continues, and nothing already distilled is distilled twice.
+- 2080 episodes at the defaults: six batches a day × 40 episodes ≈ nine days. Set the cap to 50 to finish in a day.
 
-Three gates both rules share:
+Three gates the rules share:
 
-- `auto.minGapMinutes` (default 30): never two automatic runs closer than this. It is checked before the rules — a vault that just crossed the threshold should not be asked again a minute later, and every ask costs money.
-- `auto.quietHours` (default empty): nothing runs inside this window; it may cross midnight (`23:00-08:00`). Unreadable text is ignored rather than silencing the feature forever.
+- `auto.minGapMinutes` (default 30): the two **non-catch-up** rules never fire closer than this. It is checked before them — a vault that just crossed the threshold should not be asked again a minute later, and every ask costs money.
+- `auto.quietHours` (default empty): nothing runs inside this window; it may cross midnight (`23:00-08:00`). Unreadable text is ignored rather than silencing the feature forever. Catch-up obeys it too.
 - `auto.project` (default empty): which project automatic cards are filed under; empty uses the newest episode's project.
 
 A few deliberate choices:
 
 - **The schedule clock starts when you switch it on**: it does not run on the spot — a daily schedule should not spend the moment you flip the switch. That clock is stored separately (`startedAt`) and is not treated as "the last run", so it does not push the threshold rule back by one minimum gap either.
-- **The gap is armed before the call**, so a run that produced nothing still counts and will not be re-asked the next minute. It is recorded in `.auto-state.json` in the vault (`lastAttemptAt` only), so **a restart does not reset it**.
+- **The gap is armed before the call**, so a run that produced nothing still counts and will not be re-asked the next minute. The catch-up counter works the same way: it is incremented before the call, so a failed batch still spends its slot and the daily cap stays a real cap. Both live in `.auto-state.json` in the vault (`lastAttemptAt`, `startedAt`, `batchDay`, `batchesToday`), so **a restart does not reset them**.
 - That file is not `.distill-state.json`: the latter is the **cursor** (which lines have been read), the former is the **attempt clock**. Merging them would make "read, but produced nothing" impossible to express.
 - Automatic runs, the button and `memory_distill` share **one lock**: while a run is in flight, the other callers are told so rather than queued, so two of them can never spend at once.
-- To stop it: `auto.enabled = false` (master switch), or set both rules to 0 (the switch stays on, no rule matches, nothing runs).
+- To stop it: `auto.enabled = false` (master switch), or set the rules to 0 (the switch stays on, nothing matches, nothing runs).
 
 ![Automatic distillation in Settings: the switch, the two rules, the minimum gap, quiet hours and the project](docs/auto-settings.png)
 
@@ -289,7 +328,7 @@ memory_backfill {dryRun: true}     → 133 sessions scanned / 2078 episodes woul
 memory_backfill {dryRun: false}    → 2078 episodes written; 85 already present, skipped
 ```
 
-What lands is **episodes**, not cards: distil once (`memory_distill`, the Settings button, or the daily auto run at 40 episodes a batch) and they enter the retrieval layer. Two thousand episodes is fifty model batches — that is a spend worth deciding on deliberately.
+What lands is **episodes**, not cards: distil once (`memory_distill`, the Settings button, or the daily auto run at 40 episodes a batch) and they enter the retrieval layer. Two thousand episodes is fifty model batches — that is a spend worth deciding on deliberately, and if you would rather have it drain itself, switch on `auto.catchUp`: it runs at most `auto.maxBatchesPerDay` batches a day and stops when it has caught up. A backfill also re-renders `README.md` and `graph.html` when it finishes, so those two files show the vault as it is now rather than as it was before the import.
 
 A session file is a **concatenated zstd container** (one frame per appended batch), so reading it means cutting it frame by frame — `src/history.js` carries a scanner equivalent to the one the harness itself uses.
 
@@ -404,7 +443,7 @@ Side benefit: an injected message no longer cuts a turn in half, so an episode's
 No network, no real vault:
 
 ```bash
-npm test        # 43 + 13 + 35 + 12 + 22 + 54 + 15 + 8 + 11 + 19 + 12 = 244 checks, plus the 25-question retrieval eval
+npm test        # 43 + 13 + 35 + 12 + 22 + 54 + 15 + 8 + 11 + 25 + 12 + 27 = 277 checks, plus the 25-question retrieval eval
 npm run eval    # retrieval eval only (offline, deterministic)
 ```
 
@@ -416,7 +455,8 @@ npm run eval    # retrieval eval only (offline, deterministic)
 - `test/tools.mjs` — runs `execute` + `render` for all nine tools
 - `test/page.mjs` — the route behind the sidebar panel: `/api/memory-graph.page` sits on the authenticated channel, an empty vault still answers with a page, it is **the same renderer as the `graph.html` on disk** (only the timestamp differs), `X-Frame-Options: SAMEORIGIN` so only the app can frame it, a broken vault prints the reason on the page (escaped), no vault at all is a 500 rather than a crash, and no vault path ever appears in the bytes
 - `test/distillapi.mjs` — the Distill now door: `/api/memory-graph.distill` on the authenticated `/api` channel, running the same `distill()` (live vault, the model selected in Settings, the live redaction rules), one press is one run, a concurrent second press is refused instead of queued, an extractor refusal is a 200 and not a 500, a throw is a 500 with the lock still released, views are re-rendered only when cards were written, a render failure only warns, the card list is capped at 40, and a route with no runner answers 500 instead of throwing at the browser
-- `test/auto.mjs` — the automatic distiller: each rule fires on its own, whichever matches first runs when both are on, the minimum gap is checked before the rules, a quiet window may cross midnight, a tick only decides (an off switch does not even read the vault), one tick is at most one run, the gap is armed before the call, a restart reads `.auto-state.json` instead of starting over, a run that wrote nothing still arms the gap, an in-flight run is not stacked on, failures are logged with their reason, the project reaches the run, an unreadable vault is a skip and not a crash, and the tick asks for enough episodes to see the threshold
+- `test/auto.mjs` — the automatic distiller: each rule fires on its own, whichever matches first runs when both are on, the minimum gap is checked before the rules, a quiet window may cross midnight, a tick only decides (an off switch does not even read the vault), one tick is at most one run, the gap is armed before the call, a restart reads `.auto-state.json` instead of starting over, a run that wrote nothing still arms the gap, an in-flight run is not stacked on, failures are logged with their reason, the project reaches the run, an unreadable vault is a skip and not a crash, and the tick asks for enough episodes to see the threshold; catch-up gets its own group: less than a full batch is not a backlog, it keeps its own ten-minute cadence instead of the minimum gap, it stops at the daily cap and starts again the next day
+- `test/update.mjs` — self-update: version comparison (a prerelease sorts below its release, a `v` prefix is ignored), a `link:`/`file:` install is recognised as a working copy and therefore refuses to update itself, the registry version and the status line in prose, one registry request inside the window, `force` does not re-ask within a minute, the cache expires on its own, an outage keeps the last answer instead of claiming "up to date", turning checks off sends nothing, a configured registry wins over the host's, an unusable plugin manager does not stop the check, installing asks for the exact version through `installBundle`, applied / restart-required / failed each read differently, a second install while one runs is refused rather than queued, `autoInstall` acts only when allowed and only on a registry install, and the route layer: GET is uncached, POST is the only verb that can install, no checker is a 503
 - `test/extract.mjs` — the extractor: `fetch` stubbed for the http route (prompt construction, key resolution order, tolerant JSON parsing, per-card project attribution, cursor advance, dry run writing nothing, secrets stopped before the write), then a fake `llm` service for the harness route (borrows the Settings model and thinking level, honours an explicit override, explains a missing selection, handles every terminal `finish`, reports cancellation, rejects an empty completion, and never touches `fetch`), and finally batching: the batch size derived from the output budget, a long backlog split across calls, a truncated batch halved and retried, earlier batches kept when a later one fails, and `dryRun` previewing only the first batch
 
 Two suites do touch the network (and cost money), so they are not in `npm test`:
@@ -443,7 +483,7 @@ The tag triggers `.github/workflows/publish.yml`, which does four things:
 
 1. checks that the tag and `package.json` agree — a mismatch fails instead of publishing a package whose version lies;
 2. checks whether that version is already on the registry — if so it skips, so re-running a tag is a no-op rather than an error;
-3. `npm publish --provenance --access public` (`prepublishOnly` runs the 244 checks again; a red build cannot reach the registry);
+3. `npm publish --provenance --access public` (`prepublishOnly` runs the 277 checks again; a red build cannot reach the registry);
 4. adds "Built and signed on GitHub Actions" to the npm page, proving the tarball came from this commit of this repository.
 
 Two ways to authenticate; the workflow accepts both (it uses `NPM_TOKEN` when the secret exists and falls back to OIDC when it does not):

package/README.md

@@ -90,6 +90,23 @@ plugin_manager action=install_bundle target="<clone 下来的绝对路径>"
 
 重启后打开 **设置 → 记忆图谱**（在「模型」和「内置插件」之间）。卸载：`plugin_manager action=remove_bundle target="@ilaomianyang/dsh-memory-graph"`。**记忆库不会被删**，它在 `~/dsh-memory`，独立于插件。
 
+## 它自己更新
+
+从 npm 装的那份，设置页最上面有「版本与更新」一行：当前版本号、**检查更新**、**立即更新**。
+
+![设置里的「版本与更新」：当前版本、检查更新、立即更新，下面是四个开关](docs/update-settings.png)
+
+上面这张是真机（开发机上是 `link:` 安装，所以「更新」是灰的，旁边就写着原因）：版本号、是否最新、用的哪个源、上次检查时间，下面四个开关是 `update.*`。
+
+- 插件启动时自己查一次，之后每 `update.checkEveryHours` 小时（默认 12）最多查一次。查的是 `<注册表>/@ilaomianyang%2fdsh-memory-graph/latest`——一次 GET，不带查询串、不带你的记忆库路径、不带任何对话内容。结果记在库里的 `.update-state.json`，离线时保留上一次的答案，不会把"查不到"显示成"已是最新"。
+- 点「立即更新」才装。安装走宿主自己的 `pluginManager.installBundle`，和 `dsh plugin` / 插件管理里那条路是同一条，尊重你配的 registry，不再造第二个包管理器。
+- **`link:` 本地安装会拒绝更新**，不是报错，是明说："这会覆盖你的开发目录"。从源码装的那份，更新就是 `git pull` + 重载插件，见上一节。
+- 装完通常要重启 Harness 才生效（和 `client.js` / `package.json` 同一类原因）。页面会直接写「需要重启 Harness 生效」，不会假装已经好了。
+- `update.autoInstall`（默认关）：启动时发现新版本就自己装，不用点。它在任何对话开始之前动手，是唯一"换掉插件不会打断你正在说的话"的时刻。`link:` 安装同样拒绝。
+- `memory_stats` 也会报一行版本，模型想知道自己手上是哪一版时不用猜。
+
+不想让它联网就把 `update.enabled` 关掉：那之后一次请求都不发，页面只显示当前版本。
+
 ## 你的记忆库长什么样
 
 ```
@@ -137,6 +154,13 @@ plugin_manager action=install_bundle target="<clone 下来的绝对路径>"
     retrieval:
       limit: 8
       hops: 1
+    auto:
+      enabled: true
+      catchUp: true             # 有积压就一批批追平（见下）
+      maxBatchesPerDay: 6
+    update:
+      enabled: true             # 启动时查一次有没有新版本
+      autoInstall: false
 ```
 
 | 键 | 默认 | 说明 | 在设置页 |
@@ -175,37 +199,52 @@ plugin_manager action=install_bundle target="<clone 下来的绝对路径>"
 | `auto.minGapMinutes` | `30` | 两条规则共用：两次自动蒸馏至少隔多久 | ✓ |
 | `auto.quietHours` | `''` | 不跑的时间段（如 `23:00-08:00`，可跨午夜） | ✓ |
 | `auto.project` | `''` | 自动蒸馏出的卡片记到哪个项目 | ✓ |
+| `auto.catchUp` | `false` | 追平模式：还攒着一整批就一批批蒸下去，直到追平 | ✓ |
+| `auto.catchUpEveryMinutes` | `10` | 追平模式每批之间隔多少分钟（自成一套节奏） | ✓ |
+| `auto.maxBatchesPerDay` | `6` | 一天最多自动蒸几批，0 = 不限；这是自动花钱的唯一上限 | ✓ |
+| `update.enabled` | `true` | 定期问 npm 有没有新版本（只查，不装，不带任何记忆库信息） | ✓ |
+| `update.checkEveryHours` | `12` | 查到的结果保持多久，期间不重复请求 | ✓ |
+| `update.autoInstall` | `false` | 启动时发现新版本就自己装（`link:` 安装会拒绝） | ✓ |
+| `update.registry` | `''` | 用哪个 npm 源查；空 = 宿主的源，再退到 npmjs | ✓ |
 | `retrieval.limit` | `8` | 一次检索最多返回几张卡片 | ✓ |
 | `retrieval.hops` | `1` | 联想扩散跳数，0 = 只做关键词 | ✓ |
 | `retrieval.snippetChars` | `320` | 每条结果给模型看多少字摘要 | ✓ |
 | `retrieval.timelineLimit` | `20` | `memory_timeline` 默认回看多少条 | ✓ |
 
-为什么只有这 34 个能在这页改？Harness 的设置机制只投影**声明为 volatile 的字段**，也就是"每次用到时才读"的那些。上面带 ✓ 的都是这种：改一次，下一次调用就换了。而 `enabled` / `vaultDir` / `defaultProject` / `crossProject` 是在激活时消费的，改了得重载插件，所以不放进页面装样子——页面底部会直接把这件事写出来，而不是给你几个点了没反应的开关。
+为什么只有这 41 个能在这页改？Harness 的设置机制只投影**声明为 volatile 的字段**，也就是"每次用到时才读"的那些。上面带 ✓ 的都是这种：改一次，下一次调用就换了。而 `enabled` / `vaultDir` / `defaultProject` / `crossProject` 是在激活时消费的，改了得重载插件，所以不放进页面装样子——页面底部会直接把这件事写出来，而不是给你几个点了没反应的开关。
 
-### 什么时候它会自己蒸：两条规则
+### 什么时候它会自己蒸：三条规则
 
 默认**不自动**。打开 `auto.enabled` 之后，插件每分钟看一次「现在该不该蒸」——这一眼只做算术（未蒸馏条数、距上次多久、是不是静默时段），**不调模型、不花钱**。命中规则才真的跑一次，跑的还是同一个 `distill()`。
 
-两条规则互相独立，可以只开一条，也可以都开：
+三条规则互相独立，可以只开一条，也可以都开：
 
 | 规则 | 配置 | 什么时候触发 |
 | --- | --- | --- |
 | 攒够就蒸 | `auto.minEpisodes` | 未蒸馏情节 ≥ 这个数（比如 10 条） |
 | 定时蒸 | `auto.everyMinutes` | 距上次自动蒸馏 ≥ 这么多分钟（1440 = 每天一次） |
+| 追平 | `auto.catchUp` | 还攒着**一整批**（≥ `maxEpisodesPerRun` 条）时就一批批蒸下去，每批之间隔 `auto.catchUpEveryMinutes`（默认 10），一天最多 `auto.maxBatchesPerDay` 批（默认 6，0 = 不限） |
+
+「追平」是给**积压**用的：导入了历史会话、或者一周没开机回来，未蒸馏的可能有两千条——前两条规则一次只消化一批，靠它们要等几十天。打开追平，它会自己一批批往前推，直到剩下不到一整批为止，然后安安静静等下一次。
+
+- 它**自己**有节奏（`catchUpEveryMinutes`），不受 `minGapMinutes` 限制；另外两条规则仍然受。
+- `maxBatchesPerDay` 是唯一给自动花钱**设上限**的旋钮：默认 6 批 ≈ 240 条情节/天。想快点追平就调大，不想它半夜跑就把 `quietHours` 设上。
+- 追平跑到一半被打断（重启、关机、锁被别人拿着）没关系：计数在 `.auto-state.json` 里，第二天（或下一个间隔）继续，已经蒸过的不会重蒸。
+- 2080 条积压 + 默认配置：6 批/天 × 40 条 ≈ 9 天追平；想一天追完就把上限调到 50。
 
-另外三个闸门，两条规则共用：
+另外三个闸门，规则之间共用：
 
-- `auto.minGapMinutes`（默认 30）：两次自动蒸馏至少隔这么久。它先于两条规则判断——刚攒够就被反复问是没意义的，而且每次问都要花钱。
-- `auto.quietHours`（默认空）：这段时间里不跑，可以跨午夜（`23:00-08:00`）。写不对就当没写，不会因此永远不跑。
+- `auto.minGapMinutes`（默认 30）：两条**非追平**规则共用：两次自动蒸馏至少隔这么久。它先于那两条规则判断——刚攒够就被反复问是没意义的，而且每次问都要花钱。
+- `auto.quietHours`（默认空）：这段时间里不跑，可以跨午夜（`23:00-08:00`）。写不对就当没写，不会因此永远不跑。追平也听它。
 - `auto.project`（默认空）：自动蒸出来的卡片记到哪个项目；留空用最近那条情节的项目。
 
 几个刻意的选择：
 
 - **定时蒸的时钟从打开那一刻算起**：刚打开不会立刻跑一次——「每天一次」不该在你按下开关的瞬间就花掉一次调用。时钟单独存（`startedAt`），不当作"上次运行"，所以它不会顺带把「攒够就蒸」推迟一个最短间隔。
-- **间隔在调用之前就记下**，所以一次跑空了（没有新卡片）也算数，不会因为"没产出"而在下一分钟再问一遍。记录在库里的 `.auto-state.json`（只有 `lastAttemptAt`），**重启不重置**。
+- **间隔在调用之前就记下**，所以一次跑空了（没有新卡片）也算数，不会因为"没产出"而在下一分钟再问一遍。追平的批次计数同理：调用之前就 +1，失败也算用掉一个名额——上限才是真的上限。记录在库里的 `.auto-state.json`（`lastAttemptAt` / `startedAt` / `batchDay` / `batchesToday`），**重启不重置**。
 - 它和 `.distill-state.json` 是两件事：后者是**游标**（哪些行已经读过），前者是**尝试时间**。混在一起会让"读过了但没产出"变得无法表达。
 - 自动、按钮、`memory_distill` 共用**一把锁**：正在跑的时候，第二条路拿到的是"已经有一次在跑"，不会排队、不会两份一起花。
-- 想关掉：`auto.enabled=false`（总开关），或者把两条规则都设 0（留着开关但没规则，等于不跑）。
+- 想关掉：`auto.enabled=false`（总开关），或者把规则都设 0（留着开关但没规则，等于不跑）。
 
 ![设置里的「自动蒸馏」：开关、攒够就蒸、定时蒸、最短间隔、静默时段、记到哪个项目](docs/auto-settings.png)
 
@@ -281,7 +320,7 @@ memory_backfill {dryRun: true}     → 133 个会话 / 将写入 2078 条情节
 memory_backfill {dryRun: false}    → 已写入 2078 条情节；已存在跳过 85 条
 ```
 
-回填出来的是**情节**，还没变成卡片：蒸馏一次（`memory_distill`、设置页按钮，或让自动蒸馏一天跑一批，默认每批 40 条）才会进检索层。2000 条情节要花的就是 50 批模型调用，这件事该由你决定什么时候花。
+回填出来的是**情节**，还没变成卡片：蒸馏一次（`memory_distill`、设置页按钮，或让自动蒸馏一天跑一批，默认每批 40 条）才会进检索层。2000 条情节要花的就是 50 批模型调用，这件事该由你决定什么时候花——想让它自己慢慢消化完，打开 `auto.catchUp`（追平模式）：它一天最多跑 `auto.maxBatchesPerDay` 批，追平就自己停。回填结束时会顺手把 `README.md` 和 `graph.html` 按新情节重画一次，所以打开这两个文件看到的是当下的库，不是导入前的样子。
 
 会话文件是**拼接的 zstd 容器**（每次追加写一帧），所以读它要先把帧切开——`src/history.js` 里有一个和 harness 自己用的算法同构的帧扫描器。
 
@@ -397,7 +436,7 @@ Harness 会把运行时快照、技能目录、时间戳、审批通知、压缩
 零依赖，不联网，不碰真实记忆库：
 
 ```bash
-npm test        # 43 + 13 + 35 + 12 + 22 + 54 + 15 + 8 + 11 + 19 + 12 = 244 项检查，另加 25 题检索评测
+npm test        # 43 + 13 + 35 + 12 + 22 + 54 + 15 + 8 + 11 + 25 + 12 + 27 = 277 项检查，另加 25 题检索评测
 npm run eval    # 只跑检索评测（离线，确定性的）
 ```
 
@@ -409,7 +448,8 @@ npm run eval    # 只跑检索评测（离线，确定性的）
 - `test/tools.mjs` — 逐个跑 9 个工具的 `execute` + `render`；另外验证模型要的那次蒸馏走的是共享 runner（而不是绕开它直接调 `distill`），否则它可能和一次自动蒸馏撞在一起
 - `test/page.mjs` — 侧栏面板背后那条路由：`/api/memory-graph.page` 在受认证通道上、空库也给一页而不是报错、**和落到磁盘的 `graph.html` 是同一份渲染**（只差时间戳）、`X-Frame-Options: SAMEORIGIN` 所以只有应用能嵌、库读坏了把原因印在页面上（转义后）、没有库是 500 不是崩溃、页面里不带任何库路径
 - `test/distillapi.mjs` — 「蒸馏一次」那条路：`/api/memory-graph.distill` 在受认证的 `/api` 通道上、跑的是同一个 `distill()`（活库 + 设置里的模型 + 打码规则）、一次按下一次运行、并发第二次被拒（不排队）、抽取器的拒绝是 200 而不是 500、抛错是 500 且锁照样解开、写了卡片才重建视图、渲染失败只 warn、卡片列表截到 40、没有 runner 时回 500 而不是把异常扔给浏览器
-- `test/auto.mjs` — 自动蒸馏：两条规则各自触发、都开时谁先满足谁先跑、最短间隔先于两条规则判断、静默时段跨午夜、tick 只做决定（关着时连库都不读）、一次 tick 最多一次运行、间隔在调用前就记下、重启后读回 `.auto-state.json` 所以不会重跑、跑空了也记间隔、正在跑时不叠加、失败照原因记日志、项目名透传、库读不了是跳过不是崩、阈值大时请求的条数够看到它
+- `test/auto.mjs` — 自动蒸馏：三条规则各自触发、都开时谁先满足谁先跑、最短间隔先于两条非追平规则判断、静默时段跨午夜且压过追平、tick 只做决定（关着时连库都不读）、一次 tick 最多一次运行、间隔在调用前就记下、重启后读回 `.auto-state.json` 所以不会重跑（追平的每日批次计数也一样读回来）、跑空了也记间隔、正在跑时不叠加、失败照原因记日志、项目名透传、库读不了是跳过不是崩、阈值大时请求的条数够看到它；追平单独一组：不足一整批不算积压、按自己的节奏（10 分钟）而不是 30 分钟的间隔、到每日上限就停、第二天重置
+- `test/update.mjs` — 自我更新：版本比较（预发布低于正式、`v` 前缀）、`link:`/`file:` 安装被识别成"工作副本"因而拒绝更新、注册表版本与状态行的中文说法、窗口内只问一次注册表、`force` 一分钟内不重复请求、缓存到期自己失效、离线时保留上一次答案而不是显示"已是最新"、关掉检查就一次请求都不发、配置的源优先于宿主的源、宿主没有 pluginManager 也不影响检查、安装只走 `installBundle` 且要的是确切版本、装成功/需重启/失败各自的说法、正在装的时候第二次拒绝而不是排队、`autoInstall` 只在允许且是注册表安装时动手、路由层 GET 不缓存 / POST 是唯一能装的方法 / 没有 checker 时 503
 - `test/extract.mjs` — 抽取器：`fetch` 打桩验证 http 路线（提示词构造、key 解析顺序、JSON 解析容错、每张卡片归到自己的项目、游标前进、dry run 不落盘、密钥在写入前被拦住），再用一个假的 `llm` 服务验证 harness 路线（借用设置里的模型与思考档、显式覆盖、没有选中模型时的报错、`finish` 各种终态、取消、空回复、以及全程不碰 `fetch`），最后验证分批：按输出上限折算每批条数、长后台分多次调用、被截断就对半拆、某一批失败时前面几批的卡片留着、`dryRun` 只预览第一批
 
 下面两个测试会真的联网/花钱，所以不在 `npm test` 里，要自己跑：
@@ -436,7 +476,7 @@ npm version minor -m "chore(release): v%s" && git push --follow-tags   # minor /
 
 1. 核对 tag 和 `package.json` 的版本号——不一致就失败，不会发出一个版本号对不上的包；
 2. 查这个版本是不是已经在注册表上——在就跳过，重推同一个 tag 是空操作而不是报错；
-3. `npm publish --provenance --access public`（`prepublishOnly` 会再跑一遍那 244 项，红的进不去）；
+3. `npm publish --provenance --access public`（`prepublishOnly` 会再跑一遍那 277 项，红的进不去）；
 4. npm 页面上多一条「Built and signed on GitHub Actions」，证明这个 tarball 是从这个仓库的这个提交构建的。
 
 凭据有两条路，工作流两条都认（有 `NPM_TOKEN` 就用它，没有就走 OIDC）：