dsh-tacit 0.2.0 → 0.2.2

package/README.md

@@ -6,14 +6,14 @@ A plugin for <a href="https://www.npmjs.com/package/@deepseek-ai/dsh">DeepSeek H
 <p align="center">
   <a href="https://www.npmjs.com/package/dsh-tacit"><img alt="npm" src="https://img.shields.io/npm/v/dsh-tacit"></a>
   <a href="https://github.com/hackernotfound/dsh-tacit/actions/workflows/test.yml"><img alt="test" src="https://github.com/hackernotfound/dsh-tacit/actions/workflows/test.yml/badge.svg"></a>
-  <a href="LICENSE"><img alt="MIT" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <a href="README.zh.md">中文</a>
+  <a href="https://github.com/hackernotfound/dsh-tacit/blob/main/LICENSE"><img alt="MIT" src="https://img.shields.io/badge/license-MIT-blue"></a>
+  <a href="https://github.com/hackernotfound/dsh-tacit/blob/main/docs/README.zh.md">中文</a>
 </p>
 
 You keep writing prompts the way you already do. Tacit watches how each turn
 *actually* went — retries, tool errors, and above all the message you send
 when the agent got it wrong — and turns your habits into a few directives the
-agent follows in every new session. No clicks, about $0.001 per lesson.
+agent follows in every new conversation. No clicks, $0.001–0.003 per lesson.
 
 ```
 you:    "make the login page better"
@@ -21,63 +21,79 @@ agent:  …stumbles…
 you:    "no, I meant the Next.js app under apps/web"
 
 Tacit:  learns → "The user often omits which app they mean — check apps/web first."
-        injects it into every later session's system prompt.
+        injects it into every later conversation's system prompt.
         You never have to say it again.
 ```
 
+<p align="center">
+  <img alt="Install Tacit, open its conversation tab, review learning controls, and edit the directive it gives the agent" src="https://raw.githubusercontent.com/hackernotfound/dsh-tacit/main/docs/assets/tacit-demo.gif" width="960">
+</p>
+
 ## Install — 30 seconds
 
 ```bash
-dsh plugin --profile web add dsh-tacit
+npx @deepseek-ai/dsh plugin --profile web add dsh-tacit
 ```
 
-Restart `dsh web`, refresh the page. Optional kick-start: **Settings → Tacit →
-*Learn from my last 20 turns*** (≈ $0.02, once).
+If `dsh` is already installed globally, `dsh plugin --profile web add dsh-tacit`
+is the equivalent shorthand.
+
+Start (or restart) with `npx @deepseek-ai/dsh web`, then refresh the page.
+Optional kick-start: **Settings → Tacit →
+*Learn from my last 20 turns*** (≈ $0.02–0.05, once).
 
 Requires DeepSeek Harness `>= 0.1.1-rc.1`, Node `>= 22`, and a DeepSeek API key
 already configured in the harness (Tacit never reads it).
 
 ## What you get
 
-- **Zero-click learning** — messy turns and your own corrections are analyzed
-  in the background, with the previous turn as context. Capped per day
-  (30 by default). *≈ $0.001 per analysis.*
-- **Directives that earn their place** — learned directives are injected as a
-  small system-prompt section you can read, edit, toggle or delete. A new one
-  starts as a *candidate* and is retired automatically if your messy-turn
-  rate gets worse. *free*
-- **✨ Improve** — a composer button that rewrites your current draft using
-  what Tacit has learned, with a before/after preview and 👍/👎. *≈ $0.001 per click*
-- **Measured, not guessed** — Settings shows your real trend: messy-turn rate
-  and tokens per turn, first 20 turns vs. latest 20. *free*
+| | What | Cost |
+| --- | --- | --- |
+| **Zero-click learning** | messy turns and your own corrections are analyzed in the background, with the previous turn as context; automatic analyses are capped per day (30 by default) | $0.001–0.003 each |
+| **Directives that earn their place** | learned directives are injected as a short system-prompt section you can read, edit, toggle or delete; a new one starts as a *candidate* and is retired if your messy-turn rate gets worse | free |
+| **✨ Improve** | a composer button that rewrites your current draft using what Tacit has learned, with a before/after preview and 👍/👎 | $0.001–0.002 per click |
+| **Measured, not guessed** | Settings shows your real trend — messy-turn rate and tokens per turn, first 20 turns vs. latest 20 | free |
+
+<p align="center">
+  <img alt="Tacit settings with automatic learning controls and an editable active directive" src="https://raw.githubusercontent.com/hackernotfound/dsh-tacit/main/docs/assets/tacit-settings.png" width="960">
+</p>
+<p align="center"><sub>Settings → Tacit, captured from a clean local profile with a synthetic example directive.</sub></p>
 
 ## How it works
 
-1. A pure fold over the session's events produces a bounded digest per turn.
-2. When a turn ends messy, or your next message reads as a correction, one
-   small `deepseek-v4-flash` call (low reasoning effort, structured output)
-   analyzes the prompt.
-3. Every few analyses, one more call distills the findings into 2–4 directives.
-4. The directives are rendered as a ≤300-token section of the system prompt —
-   the exact text is visible in Settings, and nothing else about your prompts
-   leaves the machine.
+1. Tacit keeps a small, bounded digest of every turn: the prompt, what tools ran,
+   what went wrong, how it ended.
+2. When a turn ends messy, or your next message reads as a correction, one small
+   `deepseek-v4-flash` call analyzes that prompt and records what was missing.
+3. Every few analyses, one more call distills the findings into 1–4 one-sentence
+   directives for the agent. Each new directive is on trial for 10 turns.
+4. The directives become a ~300-token section of the system prompt in every new
+   conversation. The exact text is visible in Settings; only clipped digests of
+   your turns ever leave the machine.
+
+The full walkthrough with every number and a diagram:
+[How it works](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/how-it-works.md).
 
 ## Privacy & cost, in one paragraph
 
 Tacit never sees your API key (every call goes through the harness's own model
-service), only calls the allowlisted official model, keeps reports and
-directives in `~/.dsh/storages/tacit/`, refuses cross-site requests to its own
-routes, and never deletes anything but its own reports. The dollar figures are
-estimates at list price; a cost plugin such as `dsh-cost-meter` shows the real
-number. Full details, including the honest list of limitations:
-[docs/privacy-and-cost.md](docs/privacy-and-cost.md).
-
-## More
-
-[Configuration](docs/configuration.md) ·
-[Privacy, cost & limitations](docs/privacy-and-cost.md) ·
-[Architecture](docs/architecture.md) ·
-[Contributing](CONTRIBUTING.md) ·
-[Changelog](CHANGELOG.md)
+service), only calls the allowlisted official models over your session's own
+provider route, keeps reports and directives in `~/.dsh/storages/tacit/`, refuses
+cross-site requests to its own routes, and never deletes anything but its own
+reports. Dollar figures are estimates at list price; a cost plugin such as
+`dsh-cost-meter` shows the real number. The full data-flow and cost tables, and
+the honest list of limitations:
+[Privacy, cost & limitations](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/privacy-and-cost.md).
+
+## Documentation
+
+| | |
+| --- | --- |
+| [Getting started](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/getting-started.md) | install, check it's on, bootstrap, where things are in the UI, troubleshooting |
+| [How it works](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/how-it-works.md) | the pipeline step by step, with a diagram and a glossary |
+| [Privacy, cost & limitations](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/privacy-and-cost.md) | what stays local, what is sent, what each call costs, what it can't do yet |
+| [Configuration](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/configuration.md) | every setting, defaults and ranges |
+| [Architecture](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/architecture.md) | for contributors: modules, hooks, routes, storage |
+| [Contributing](https://github.com/hackernotfound/dsh-tacit/blob/main/CONTRIBUTING.md) · [Changelog](https://github.com/hackernotfound/dsh-tacit/blob/main/CHANGELOG.md) | |
 
 MIT © hackernotfound

package/client/client.js

@@ -12,7 +12,7 @@
  *    slot, beside Chat/Trajectory) listing every turn's digest with an
  *    Analyze button and the analysis report;
  *  - a small ✨ Improve button in the composer tool row
- *    (`conversation.input.left`) once Tacit has learned enough patterns;
+ *    (`conversation.input.left`) while it is enabled in Settings;
  *  - a before/after preview popup in `conversation.input.overlay` whose
  *    Apply action replaces the composer draft via `inputActions.setDraft`.
  *
@@ -59,7 +59,7 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
       'trigger.bootstrap': '引导',
       'bootstrap.btn': '从我最近 20 轮中学习',
       'bootstrap.running': '学习中… {done}/{total}',
-      'bootstrap.hint': '≈ $0.02，一次性。已顺利完成的提示词会被跳过。',
+      'bootstrap.hint': '约 $0.02–0.05，一次性。已顺利完成的提示词会被跳过。',
       'steer.enrich': '发送前补充学到的上下文（实验性，每次发送一次小调用）',
       'steer.title': '智能体被告知的关于你的信息',
       'steer.desc': '这些指令会注入每个新会话的系统提示词，让智能体替你补上你通常没说的内容。',
@@ -166,7 +166,7 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
       'trigger.bootstrap': 'bootstrap',
       'bootstrap.btn': 'Learn from my last 20 turns',
       'bootstrap.running': 'Learning… {done}/{total}',
-      'bootstrap.hint': '≈ $0.02, one time. Skips prompts that already went fine on their own.',
+      'bootstrap.hint': '≈ $0.02–0.05, one time. Skips prompts that already went fine on their own.',
       'steer.enrich': 'Add learned context before each send (experimental — one small call per send)',
       'steer.title': 'What the agent is told about you',
       'steer.desc': 'These directives ride every new session\'s system prompt so the agent fills in what you usually leave unsaid.',
@@ -1220,8 +1220,9 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
     function PreviewOverlay(kit) {
       const { t } = kit
       return function PreviewOverlayView(props) {
-        const hooks = props.hooks !== null && typeof props.hooks === 'object' ? props.hooks : null
-        const store = hooks !== null ? hooks.tacitStore : null
+        const store = props.tacitStore !== null && typeof props.tacitStore === 'object'
+          ? props.tacitStore
+          : null
         useStoreVersion(store)
         if (store === null || !store.preview.open) return null
 
@@ -1258,7 +1259,7 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
       }
     }
 
-    // ── Feedback strip (under the composer, live stage only) ────────────────
+    // ── Feedback strip (under the composer, Improve enabled only) ───────────
 
     function FeedbackStrip(kit) {
       const { t } = kit
@@ -1646,11 +1647,11 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
         id: 'tacit-preview',
         order: 10,
         locale: NS,
-        inject: (sessionId) => ({ hooks: { tacitStore: storeFor(sessionId) } }),
+        inject: (sessionId) => ({ tacitStore: storeFor(sessionId) }),
       }, (props) => h(PreviewOverlayView, props)))
 
       // The post-apply 👍/👎 strip lives in the band UNDER the composer card
-      // (composer.dock), only for the live stage (threshold reached).
+      // (composer.dock), only while the Improve button is enabled.
       const FeedbackStripView = FeedbackStrip(kit)
       ctx.slots.inject('conversation.composer.dock', () => ctx.slots.register({
         name: 'conversation.composer.dock',
@@ -1661,8 +1662,7 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
       }, (props) => h(FeedbackStripView, props)))
 
       // Own Settings page under "Cost": a nav section in the Settings panel,
-      // on its own. Reuses the coach panel (progress, model, threshold, live,
-      // clear reports, coached prompts). Order 32 places it right after
+      // on its own. Reuses the coach panel. Order 32 places it right after
       // cost-meter's "Cost" section (order 30/31).
       const SettingsSectionView = SettingsSection(kit)
       ctx.slots.inject('settings.section', () => ctx.slots.register({
@@ -1682,6 +1682,7 @@ if (typeof window === 'undefined' || window.__ModuleLoader__ === undefined || ty
       __test: {
         css,
         applyImproved,
+        closePreview,
         closeFeedback,
         fadeFeedback,
         rootStore,

package/docs/README.md

@@ -0,0 +1,21 @@
+# Tacit documentation
+
+Tacit is a plugin for [DeepSeek Harness](https://www.npmjs.com/package/@deepseek-ai/dsh)
+(`dsh`). It watches how your conversations with the agent *actually* go, learns what
+you habitually leave unsaid, and tells the agent for you — through a short, editable
+section of the system prompt.
+
+Pick the question you have:
+
+| I want to… | Read |
+| --- | --- |
+| Install it and see it working in five minutes | [Getting started](getting-started.md) |
+| Understand what it does, step by step, and what the words mean | [How it works](how-it-works.md) |
+| Know what leaves my machine, what it costs, and what it can't do yet | [Privacy, cost & limitations](privacy-and-cost.md) |
+| Change a setting | [Configuration](configuration.md) |
+| Read or change the code | [Architecture](architecture.md) · [Contributing](../CONTRIBUTING.md) |
+| See what changed between versions | [Changelog](../CHANGELOG.md) |
+
+Reading order for a newcomer: Getting started → How it works → Privacy & cost.
+
+中文：[README.zh.md](README.zh.md)（仅 README 有中文版）。

package/docs/README.zh.md

@@ -0,0 +1,82 @@
+<h1 align="center">Tacit</h1>
+
+<p align="center"><b>学习你在提示词里没说出口的东西——并替你告诉智能体。</b><br>
+一个 <a href="https://www.npmjs.com/package/@deepseek-ai/dsh">DeepSeek Harness</a> 插件。</p>
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/dsh-tacit"><img alt="npm" src="https://img.shields.io/npm/v/dsh-tacit"></a>
+  <a href="https://github.com/hackernotfound/dsh-tacit/actions/workflows/test.yml"><img alt="test" src="https://github.com/hackernotfound/dsh-tacit/actions/workflows/test.yml/badge.svg"></a>
+  <a href="https://github.com/hackernotfound/dsh-tacit/blob/main/LICENSE"><img alt="MIT" src="https://img.shields.io/badge/license-MIT-blue"></a>
+  <a href="https://github.com/hackernotfound/dsh-tacit/blob/main/README.md">English</a>
+</p>
+
+> 本文由英文版翻译而来，尚未经过母语者校对；以 [README.md](https://github.com/hackernotfound/dsh-tacit/blob/main/README.md) 为准，欢迎提交修正。
+
+你照常写提示词。Tacit 在后台观察每一轮*实际*的走向——重试、工具报错，尤其是
+智能体做错之后你发的那条消息——把你的习惯变成几条指令，让智能体在每个新对话里
+自动遵守。无需点击，每次学习 $0.001–0.003。
+
+```
+你:     "把登录页做好一点"
+智能体: …走偏…
+你:     "不对，我是说 apps/web 下的 Next.js 应用"
+
+Tacit:  学到 → "用户常常不说明是哪个应用——先查 apps/web。"
+        注入之后每个对话的系统提示。
+        你再也不用重复说了。
+```
+
+## 安装——30 秒
+
+```bash
+npx @deepseek-ai/dsh plugin --profile web add dsh-tacit
+```
+
+用 `npx @deepseek-ai/dsh web` 启动（或重启），然后刷新页面。可选的快速起步：**设置 → Tacit → *从我最近 20 轮中学习***
+（约 $0.02–0.05，一次性）。
+
+需要 DeepSeek Harness `>= 0.1.1-rc.1`、Node `>= 22`，以及已在 harness 里配置好的
+DeepSeek API Key（Tacit 从不读取它）。
+
+## 你会得到什么
+
+| | 是什么 | 费用 |
+| --- | --- | --- |
+| **零点击学习** | 不顺的轮次和你自己的纠正会在后台被分析，并带上上一轮作为上下文；自动分析每天有上限（默认 30 次） | 每次 $0.001–0.003 |
+| **指令要靠表现留下** | 学到的指令作为一小段系统提示注入，你可以查看、编辑、开关或删除；新指令先是*候选*，如果你的不顺轮次比例变差就会退役 | 免费 |
+| **✨ 改进** | 输入框里的按钮，用 Tacit 学到的东西重写当前草稿，带前后对比预览和 👍/👎 | 每次点击 $0.001–0.002 |
+| **测量而非猜测** | 设置页显示你的真实趋势：不顺轮次比例和每轮 token 数，最早 20 轮 vs 最近 20 轮 | 免费 |
+
+## 它是怎么工作的
+
+1. Tacit 为每一轮保留一份很小的有界摘要：提示词、跑了哪些工具、哪里出错、怎么结束。
+2. 当一轮结束得不顺、或你的下一条消息像是纠正，就用一次很小的 `deepseek-v4-flash`
+   调用分析那条提示词，记下缺了什么。
+3. 每隔几次分析，再用一次调用把发现蒸馏成 1–4 条一句话的指令。每条新指令试用 10 轮。
+4. 指令成为每个新对话系统提示里约 300 token 的一段——原文在设置里可见；离开本机的
+   只有你各轮的裁剪摘要。
+
+完整流程（含每个数字和一张图，英文）：
+[How it works](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/how-it-works.md)。
+
+## 隐私与成本，一段话
+
+Tacit 从不接触你的 API Key（所有调用都经过 harness 自己的模型服务），只通过你会话
+自己的 provider 路由调用白名单里的官方模型，报告和指令都存放在
+`~/.dsh/storages/tacit/`，拒绝对自身路由的跨站请求，除了自己的报告之外从不删除任何
+东西。美元数字是按公开价格的估算；`dsh-cost-meter` 这样的成本插件会显示真实花费。
+完整的数据流与成本表、以及坦白的限制清单（英文）：
+[Privacy, cost & limitations](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/privacy-and-cost.md)。
+
+## 文档（均为英文）
+
+| | |
+| --- | --- |
+| [Getting started](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/getting-started.md) | 安装、确认已启用、快速起步、界面各处在哪、排障 |
+| [How it works](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/how-it-works.md) | 逐步流程，带图和术语表 |
+| [Privacy, cost & limitations](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/privacy-and-cost.md) | 什么留在本地、什么会发出去、每次调用的花费、目前做不到的事 |
+| [Configuration](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/configuration.md) | 所有设置项、默认值和取值范围 |
+| [Architecture](https://github.com/hackernotfound/dsh-tacit/blob/main/docs/architecture.md) | 面向贡献者：模块、钩子、路由、存储 |
+| [Contributing](https://github.com/hackernotfound/dsh-tacit/blob/main/CONTRIBUTING.md) · [Changelog](https://github.com/hackernotfound/dsh-tacit/blob/main/CHANGELOG.md) | |
+
+MIT © hackernotfound

package/lib/analyze.js

@@ -314,8 +314,8 @@ export const IMPROVE_REPAIR_SYSTEM_PROMPT = [
 ].join('\n')
 
 /**
- * Style-rule distillation: the ONLY new paid model call of the v2 loop,
- * fired rarely (every 3+ unreviewed 👎 reasons) with a hard 300-token cap.
+ * Style-rule distillation: fired rarely (every 3+ unreviewed 👎 reasons),
+ * capped at DISTILL_MAX_TOKENS output tokens.
  */
 export const DISTILL_SYSTEM_PROMPT = [
   'You distill user feedback into durable prompt-writing style rules for a',

package/lib/service.js

@@ -8,8 +8,9 @@
  * an `ok` flag and a stable error `code` the client localizes. Expected
  * failures never throw; unexpected ones are mapped by the route layer.
  *
- * Only user-triggered model calls exist here (analyzeTurn / improveDraft):
- * there is no background work, no polling, no telemetry.
+ * Model calls are either user-triggered (analyzeTurn / improveDraft /
+ * bootstrap) or automatic analyses of messy and corrected turns, the latter
+ * capped by `autoDailyBudget`. There is no polling and no telemetry.
  */
 
 import {
@@ -285,7 +286,7 @@ export function createCoachService(ctx, store, effectiveConfig) {
    * Returns the profile with fresh rules on success, the input profile on a
    * soft no-op/failure.
    */
-  const maybeDistill = async (profile, provider) => {
+  const maybeDistill = async (profile, provider, sessionId) => {
     if (profile.pendingDistill < 3 || distillInFlight) return profile
     distillInFlight = true
     try {
@@ -299,6 +300,7 @@ export function createCoachService(ctx, store, effectiveConfig) {
         maxTokens: DISTILL_MAX_TOKENS,
         timeoutMs: DISTILL_TIMEOUT_MS,
         tool: DISTILL_TOOL,
+        sessionId,
       })
       const rules = normalizeDistillRules(text)
       if (rules.length === 0) return profile
@@ -843,7 +845,7 @@ export function createCoachService(ctx, store, effectiveConfig) {
       // Distillation also fires on user-triggered improve calls (soft, in-flight
       // deduped, never awaited: an improve call is never blocked by it).
       if (profile.pendingDistill >= 3) {
-        maybeDistill(profile, providerForSession(sessionId)).catch(() => {})
+        maybeDistill(profile, providerForSession(sessionId), sessionId).catch(() => {})
       }
       // Only trusted (or still-inexperienced) patterns reach the prompt;
       // style rules + the last 3 verbatim down-reasons ride along for free
@@ -948,7 +950,7 @@ export function createCoachService(ctx, store, effectiveConfig) {
       // user action. Awaited so the returned profile already carries the
       // fresh style rules; a failure is soft-silent and retries later.
       if (profile.pendingDistill >= 3) {
-        profile = await maybeDistill(profile, providerForSession(record.sessionId))
+        profile = await maybeDistill(profile, providerForSession(record.sessionId), record.sessionId)
       }
       return { ok: true, profile, code: '', detail: '' }
     },

package/package.json

@@ -1,10 +1,11 @@
 {
   "name": "dsh-tacit",
-  "version": "0.2.0",
+  "version": "0.2.2",
   "description": "Tacit learns what you leave unsaid in your prompts — from messy turns and your own corrections, with zero clicks — and tells the agent how to compensate, on every turn, via a system-prompt section you can read and edit.",
   "author": "hackernotfound",
   "license": "MIT",
   "type": "module",
+  "packageManager": "pnpm@11.23.0",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/hackernotfound/dsh-tacit.git"
@@ -31,11 +32,14 @@
     "client",
     "cordis.patch.yml",
     "README.md",
-    "README.zh.md",
-    "LICENSE"
+    "LICENSE",
+    "docs/README.zh.md"
   ],
   "scripts": {
     "test": "node --test",
+    "check": "pnpm test && pnpm check:docs && pnpm check:package",
+    "check:docs": "node scripts/check-doc-links.mjs",
+    "check:package": "node scripts/check-package.mjs",
     "smoke": "node scripts/smoke.mjs"
   },
   "publishConfig": {
@@ -87,8 +91,8 @@
     "zod": "^4.4.3"
   },
   "devDependencies": {
-    "react": "^18.3.1",
-    "react-dom": "^18.3.1"
+    "react": "^19.2.8",
+    "react-dom": "^19.2.8"
   },
   "keywords": [
     "deepseek",

package/README.zh.md

@@ -1,77 +0,0 @@
-<h1 align="center">Tacit</h1>
-
-<p align="center"><b>学习你在提示词里没说出口的东西——并替你告诉智能体。</b><br>
-一个 <a href="https://www.npmjs.com/package/@deepseek-ai/dsh">DeepSeek Harness</a> 插件。</p>
-
-<p align="center">
-  <a href="https://www.npmjs.com/package/dsh-tacit"><img alt="npm" src="https://img.shields.io/npm/v/dsh-tacit"></a>
-  <a href="https://github.com/hackernotfound/dsh-tacit/actions/workflows/test.yml"><img alt="test" src="https://github.com/hackernotfound/dsh-tacit/actions/workflows/test.yml/badge.svg"></a>
-  <a href="LICENSE"><img alt="MIT" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <a href="README.md">English</a>
-</p>
-
-> 本文由英文版翻译而来，尚未经过母语者校对；以 [README.md](README.md) 为准，欢迎提交修正。
-
-你照常写提示词。Tacit 在后台观察每一轮*实际*的走向——重试、工具报错，尤其是
-智能体做错之后你发的那条消息——把你的习惯变成几条指令，让智能体在每个新会话里
-自动遵守。无需点击，每次学习约 $0.001。
-
-```
-你:     "把登录页做好一点"
-智能体: …走偏…
-你:     "不对，我是说 apps/web 下的 Next.js 应用"
-
-Tacit:  学到 → "用户常常不说明是哪个应用——先查 apps/web。"
-        注入之后每个会话的系统提示。
-        你再也不用重复说了。
-```
-
-## 安装——30 秒
-
-```bash
-dsh plugin --profile web add dsh-tacit
-```
-
-重启 `dsh web`，刷新页面。可选的快速起步：**设置 → Tacit → *从我最近 20 轮中学习***
-（约 $0.02，一次性）。
-
-需要 DeepSeek Harness `>= 0.1.1-rc.1`、Node `>= 22`，以及已在 harness 里配置好的
-DeepSeek API Key（Tacit 从不读取它）。
-
-## 你会得到什么
-
-- **零点击学习**——不顺的轮次和你自己的纠正会在后台被分析，并带上上一轮作为上下文。
-  每天有上限（默认 30 次）。*每次分析约 $0.001*
-- **指令要靠表现留下**——学到的指令作为一小段系统提示注入，你可以查看、编辑、开关或
-  删除。新指令先是*候选*，如果你的不顺轮次比例变差就会自动退役。*免费*
-- **✨ 改进**——输入框里的按钮，用 Tacit 学到的东西重写当前草稿，带前后对比预览和
-  👍/👎。*每次点击约 $0.001*
-- **测量而非猜测**——设置页显示你的真实趋势：不顺轮次比例和每轮 token 数，最早 20 轮
-  vs 最近 20 轮。*免费*
-
-## 它是怎么工作的
-
-1. 对会话事件做纯折叠，得到每一轮的有界摘要。
-2. 当一轮结束得不顺、或你的下一条消息像是纠正，就用一次很小的 `deepseek-v4-flash`
-   调用（低推理强度，结构化输出）分析那条提示词。
-3. 每隔几次分析，再用一次调用把发现蒸馏成 2–4 条指令。
-4. 指令被渲染成系统提示里 ≤300 token 的一段——原文在设置里可见，除此之外你的提示词
-   不会离开本机。
-
-## 隐私与成本，一段话
-
-Tacit 从不接触你的 API Key（所有调用都经过 harness 自己的模型服务），只调用白名单里的
-官方模型，报告和指令都存放在 `~/.dsh/storages/tacit/`，拒绝对自身路由的跨站请求，
-除了自己的报告之外从不删除任何东西。美元数字是按公开价格的估算；`dsh-cost-meter`
-这样的成本插件会显示真实花费。完整细节（包括坦白的限制清单，英文）：
-[docs/privacy-and-cost.md](docs/privacy-and-cost.md)。
-
-## 更多
-
-[配置](docs/configuration.md) ·
-[隐私、成本与限制](docs/privacy-and-cost.md) ·
-[架构](docs/architecture.md) ·
-[参与贡献](CONTRIBUTING.md) ·
-[更新日志](CHANGELOG.md)（均为英文）
-
-MIT © hackernotfound