thatgfsj-code 3.0.3 → 3.0.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thatgfsj-code",
-  "version": "3.0.3",
+  "version": "3.0.6",
   "description": "Thatgfsj Code - AI Coding Assistant with Reasonix-style prompt caching",
   "main": "dist/cmd/index.js",
   "type": "module",
@@ -8,13 +8,29 @@
     "thatgfsj": "./dist/cmd/index.js",
     "gfcode": "./dist/cmd/index.js"
   },
+  "files": [
+    "dist"
+  ],
+  "engines": {
+    "node": ">=20.19"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Thatgfsj/thatgfsj-code.git"
+  },
+  "bugs": {
+    "url": "https://github.com/Thatgfsj/thatgfsj-code/issues"
+  },
+  "homepage": "https://github.com/Thatgfsj/thatgfsj-code#readme",
   "scripts": {
     "build": "tsc",
     "start": "node dist/cmd/index.js",
     "dev": "tsc && node dist/cmd/index.js",
     "link": "npm link",
     "test": "vitest run",
-    "test:watch": "vitest"
+    "test:watch": "vitest",
+    "smoke": "npm run build && node tests/smoke-tool.mjs && node tests/smoke-tool-stream.mjs && node tests/smoke-pollution.mjs && node tests/smoke-thinking.mjs",
+    "prepublishOnly": "npm run build && npm test"
   },
   "keywords": [
     "ai",
@@ -30,13 +46,9 @@
     "ink": "^7.1.0",
     "ink-select-input": "^6.2.0",
     "ink-text-input": "^6.0.0",
-    "inquirer": "^13.3.2",
     "marked": "^15.0.12",
     "marked-terminal": "^7.3.0",
-    "ora": "^7.0.1",
-    "playwright": "^1.61.1",
-    "react": "^19.2.7",
-    "readline": "^1.3.0"
+    "react": "^19.2.7"
   },
   "devDependencies": {
     "@types/marked-terminal": "^6.1.1",

package/CHANGELOG.md

@@ -1,192 +0,0 @@
-# Changelog
-
-All notable changes to **Thatgfsj Code** are documented here. The format follows
-[Keep a Changelog](https://keepachangelog.com/) and the project adheres to
-[Semantic Versioning](https://semver.org/).
-
-## [3.0.0 / 产品 0.5.0] - 2026-08-08  - Reasonix 风格 Prompt Caching 架构
-
-> **重大版本变更**：流式协议 `AsyncGenerator<string>` → `AsyncGenerator<StreamChunk>`。
-> 自定义依赖 `LLMService.chatStream` / `App.streamResponse` 的下游代码需要适配。
->
-> 参考：DeepSeek-Reasonix (esengine/DeepSeek-Reasonix) 的"不可变前缀 + 追加日志 + 易变草稿"三支柱。
-
-### Added
-- **M1 — 流式协议重构**：干掉 `@@TOOL@@` 哨兵字符串；`StreamChunk` 升级为结构化 union（text / tool_calls / thinking / usage）。
-- **M2 — Anthropic prompt cache**：`system` 改为 block 数组，最后一块挂 `cache_control: { type: 'ephemeral' }`；工具列表最后一个也挂 marker；新增 `anthropic-beta: prompt-caching-2024-07-31` header；修「第二条 system 消息被丢弃」bug。
-- **M3 — 缓存统计与 UI**：`CacheStatsStore` 持久化到 `~/.thatgfsj/cache-stats.json`；TUI Header 显示 `⚡ 命中率 87%`；新增 `/cache` 命令显示 24h 命中率曲线。
-- **M4 — 高级特性**：`VolatileScratch` 隔离 thinking；Tool-call Repair（append `[TOOL_REPAIR]` 而非修改既有消息）；`smartModel.shouldDowngrade` 智能路由；`Config.cache` 字段 + InitWizard 两步缓存配置。
-- **OpenAI provider 缓存友好**：`stableStringify` 保证 JSON 字节级稳定；流末尾 yield `{ type: 'usage' }` chunk。
-- **28 个 vitest 单元测试**（`tests/cache/`）。
-
-### Changed
-- `SessionManager.autoCompact` 不再 mutate messages，改为通过 `onSuggestNewSession` 回调通知用户调 `/new`。Reasonix 原则：超长开新会话，保护上游缓存前缀。
-- `system prompt` 片段顺序调整：NWT history 和 Date 移到末尾（易变区），前 6 段是 Immutable Prefix。
-- `LLMService.fromConfig` 接收 `Config.cache` 透传给 provider。
-
-## [2.2.7 / 产品 0.4.4] - 2026-07-06  - 烟囱测试驱动的 anti-pollution 过滤收紧
-
-> 双版本号方案:
->
-> | 位置 | 版本 | 含义 |
-> |---|---|---|
-> | `package.json` `"version"` | `2.2.7` | npm 包版本,正常递增 (+0.0.1) |
-> | `src/cmd/index.tsx` `.version(...)` | `0.4.4` | **产品版本** (`gfcode --version` 看到的) |
-
-### Fixed
-
-**Pollution 过滤误杀正常对话** — v2.2.4 引入的 anti-pollution 过滤
-(`SessionManager.looksPolluted`) 有两条过宽的模式:
-
-- `/\u5df2\u4e2d\u65ad[^\n]{0,40}/` — 把所有包含"已中断"字样的中文
-  都当污染,包括 `我看到这个进程已中断了` 这种合法对话
-- `/response (was )?truncated/i` — 把所有提到"response truncated"的句子
-  都当污染,包括用户问 `Why was the response truncated last time?`
-
-烟囱测试 `tests/smoke-pollution.mjs` 直接捕获了这两个误杀。
-
-### 修复策略 — 两层模式分级
-
-```
-STRONG (always drop):
-  [已中断]              必有方括号,模型原样输出
-  [interrupted]         English 等价
-
-WEAK (drop only when ALL gates pass):
-  ^...\[已中断\]       行首标记 (带可选 markdown 装饰)
-  ^...\[interrupted\]   行首标记
-  \bresponse (was )?(truncated|cut off|interrupted)\b
-  \boutput (was )?(truncated|cut off|interrupted)\b
-
-  + 消息长度 < 200 字符 (截断的响应通常较短)
-  + 不以 ? 结尾 (疑问句是用户讨论过往行为)
-  + 不包含 "last time" / "earlier" / "before" / "previously" / "yesterday"
-    (时间副词表明用户在回顾过去,不是污染)
-```
-
-### Smoke 测试新增
-
-- `tests/smoke-thinking.mjs` — 28 个用例覆盖 `splitThinking` / `compressThinking`
-  所有边缘情况 (空输入、多 delimiter、Chinese/English 混合、tag 跨行)
-- `tests/smoke-tool.mjs` — 18 个用例覆盖 ToolCall 渲染路径 (空结果、
-  pending、截断计数、错误、超长单行)
-- `tests/smoke-pollution.mjs` — 17 个用例覆盖 anti-pollution 过滤的
-  drop / keep 双向场景
-
-### Test summary (本 patch 后)
-```
-thinking:   28 / 28 pass
-tool:       18 / 18 pass
-pollution:  17 / 17 pass
-total:      63 / 63 pass
-```
-
----
-
-## [2.2.6 / 产品 0.4.3] - 2026-07-06  - 修复工具结果不显示
-
-> 双版本号方案:
->
-> | 位置 | 版本 | 含义 |
-> |---|---|---|
-> | `package.json` `"version"` | `2.2.6` | npm 包版本,正常递增 (+0.0.1) |
-> | `src/cmd/index.tsx` `.version(...)` | `0.4.3` | **产品版本** (`gfcode --version` 看到的) |
-
-### Fixed
-
-**Bug: 工具调用结果不显示** — 在 1.0.4 源码栈下用户报告"⚙ shell date" 这条
-命令发起后,**结果完全看不到**,只看到 AI 后续的文字回复 ("好嘞，那我给你
-演示一组工具——并行调用"). 模型明明看到了工具结果 (因为它的后续回复引用了
-结果),但 UI 上不显示结果文本。
-
-通过代码 review + 单测复现,定位到三个潜在原因,全部修复:
-
-1. **`src/tui/components/ToolCall.tsx`** — `tool.result ? ... : null` 用 truthy
-   判断,**空字符串结果被当 falsy 跳过渲染**。改成 `tool.result !== undefined ? ... : null`。
-   同样修复: `wrap="truncate"` 在 Ink 里会把超长行**静默切掉不显示**,
-   改成 `wrap="wrap"` 让长行自动换行。还把 `result.truncated` 的指示从
-   模糊的 `"..."` 改成 `(+N more lines)` 显示实际被截掉的行数。
-2. **加 ⏳ pending 指示器** — 工具调用发起到结果返回之间的中间状态,
-   之前没有视觉反馈,看起来像是"卡住了"。现在显示 `⏳ running...`
-3. **`src/tui/hooks/useChat.ts`** — 流结束 push 到 messages 时,assemble
-   一个 `[tool: name → result]` 摘要附加到 assistant 的 content 里。
-   这是 belt-and-suspenders 兜底: 即使 Ink 的 `<ToolCall/>` 渲染路径
-   因为任何边缘 case 失败 (例如太长的输出 wrap 问题),用户依然能在
-   AI 的文字回复里看到工具结果。
-
-### 仍然不修的 (但要意识到)
-- 用户的 paste 里还有另一条命令 `echo "=== 当前时间 ===" && date && echo
-  "=== 系统信息 ===" && system` — 这里 `system` 不是合法命令,
-  会返回错误。这不是代码 bug,是用户/模型自己写的命令有问题。本次 patch
-  不修改这个。
-
-### 不在本次范围
-- Anthropic / Gemini provider 的 `delta.tool_calls` 流式解析 (仍是 TODO)
-- 工具输入参数命名前缀 (`command` vs `cmd` 不一致) — 下次 patch
-
----
-
-## [2.2.5 / 产品 0.4.2] - 2026-07-06  - 压缩 <think> 思考块 (类似 opencode)
-
-> 双版本号方案:
->
-> | 位置 | 版本 | 含义 |
-> |---|---|---|
-> | `package.json` `"version"` | `2.2.5` | npm 包版本,正常递增 (+0.0.1) |
-> | `src/cmd/index.tsx` `.version(...)` | `0.4.2` | **产品版本** (`gfcode --version` 看到的) |
-
-### Fixed
-
-**Bug: 思考过程太长,刷屏** — 在 1.0.4 源码栈下,即使是"非推理"模型
-(Qwen / DeepSeek-V3 chat / Kimi 等) 也会在回复里发整段 `<think>...</think>`
-块,经常是几十行内部独白。这块内容对用户来说没有可读价值,只
-是噪音 — 用户在 v2.2.4 的会话日志里能看到这问题特别严重。
-
-参考 opencode 的做法 (`packages/opencode/src/session/prompt.ts:244`
-里的 `<think>[\s\S]*?<\/think>` 正则),引入压缩模式:
-
-- 新文件 `src/utils/thinking.ts`,导出三个 helper:
-  - `splitThinking(content)` — 把 `<think>...` / `<reasoning>...</reasoning>` /
-    `[THINK]...[/THINK]` 三种块格式剥离开,返回 `{thinking, conclusion,
-    thinkingLines, thinkingHint}`。
-  - `summarizeThinking(split)` — 输出 `💭 thought for N lines: <first hint>`
-    单行摘要。
-  - `compressThinking(content, showThinking)` — `showThinking=false` 时把
-    思考块压成一行 + 结论;`showThinking=true` 时原样返回。
-- 三个调用点全部接入压缩:
-  - `src/cmd/index.tsx`: 流式阶段**不**实时打印 AI 文字 (避免半个 <think>
-    块被切到屏幕上),流结束后用 `splitThinking + summarizeThinking` 输出
-    一行摘要 + 结论。**持久化时只保存结论**,让 history 不被思考块撑大。
-  - `src/app/index.ts:runPrompt`: 持久化前 `compressThinking(fullResponse,
-    showThinking)`。
-  - `src/tui/hooks/useChat.ts`: 同上,持久化 + 显示内容都走压缩。
-
-### Added
-
-- **CLI flag `--show-thinking`**: 在命令行下开启完整思考块显示 (调试用)
-- **`App.showThinking` 字段**: 默认 `false`,可通过 REPL 命令切换
-- **REPL 命令 `/thinking on|off`** (中文别名 `/思考`): 在 Ink TUI 模式下
-  切换思考块显示
-
-### 设计选择
-
-为什么不沿用 opencode 的"协议级分离 reasoning"?
-- opencode 把 reasoning 当成单独的 stream part 处理 (`MessageV2.ReasoningPart`),
-  那需要模型支持 reasoning API (Anthropic extended thinking / OpenAI o1-style)。
-- 我们用的多数 provider (Qwen / DeepSeek / Kimi) 把 reasoning 直接当作
-  content 的一部分输出,所以必须在**显示层**做正则剥离。
-
-为什么不实时 (mid-stream) 压缩?
-- 流式 chunk 是任意切分的,`<think>` 标签可能跨 chunk。实时检测需要
-  state machine,延迟和边界处理都很脆弱。
-- 流结束后压缩更可靠,而且不影响用户对结论的实时阅读(用户其实只在乎结论)。
-
-### 不在本次范围
-- `/thinking` 命令目前是全局开关,不支持 per-message 切换
-- 没有提供"展开"按钮 — 用户要看完整内容只能 `--show-thinking` 跑一遍
-- 持久化时去掉了 thinking 内容,这意味着**重放历史时也看不到完整思考**;
-  这是有意为之 (context window 优先),如需调试可以 `app.showThinking = true` 后
-  重新发起对话
-- Anthropic / Gemini SSE tool_call 流式解析 仍是 TODO
-
----

package/DEVELOPMENT.md

@@ -1,286 +0,0 @@
-# Development Guide — Thatgfsj Code
-
-This document is for contributors / maintainers. End users should read
-[`README.md`](./README.md) instead.
-
----
-
-## 1. Overview
-
-**Thatgfsj Code** is a Claude Code–style interactive AI coding assistant that
-runs entirely in your terminal. It is written in TypeScript, published as an
-ESM npm package, distributed under MIT, and targets **Node.js ≥ 18**.
-
-The tool exposes:
-
-- A non-interactive single-shot CLI (`gfcode "..."`).
-- An interactive REPL (`gfcode` with no args).
-- Subcommands: `init`, `explain`, `debug`, `chat`, `template`.
-
-The release pipeline is: **bug fix → docs → smoke test → git commit + tag →
-push → GitHub release → `npm publish`**. Everything in `src/` compiles to
-`dist/` via `tsc`; the `package.json` `files` whitelist decides what ships
-in the npm tarball.
-
----
-
-## 2. Architecture
-
-```
-┌──────────────────────────────────────────────────────────────┐
-│                         src/index.ts                          │
-│   CLI entry · commander program · subcommand dispatch         │
-│   (init / explain / debug / chat / template / default REPL)  │
-└─────────────────────────┬────────────────────┬───────────────┘
-                          │                    │
-              ┌───────────▼──────────┐ ┌───────▼─────────────┐
-              │   AI Engine + Tools  │ │      REPL Loop       │
-              │ src/core/ai-engine   │ │ src/repl/{loop,...}  │
-              │ src/tools/{file,…}   │ │ @inquirer/input      │
-              └───────────┬──────────┘ └───────┬─────────────┘
-                          │                    │
-              ┌───────────▼────────────────────▼───────────────┐
-              │        Core / Utils / Agent / MCP              │
-              │ src/core/{config,session,types,subagent,…}     │
-              │ src/utils/{diff-preview,memory,project-…}     │
-              │ src/agent/{core,intent,streaming}             │
-              │ src/mcp/client.ts                              │
-              └────────────────────────────────────────────────┘
-```
-
-### Module responsibilities
-
-| Layer     | Files                                        | Job                                                                |
-|-----------|----------------------------------------------|--------------------------------------------------------------------|
-| Entry     | `src/index.ts`                               | commander program, subcommand handlers, REPL bootstrap             |
-| REPL      | `src/repl/{input,output,loop,welcome}.ts`    | interactive prompt, ANSI output, SIGINT, welcome/init wizard       |
-| Core      | `src/core/{ai-engine,config,session,…}.ts`   | AI provider abstraction, config loading, session state            |
-| Tools     | `src/tools/{file,shell,git,search}.ts`       | file I/O, shell exec, git ops, code search                         |
-| Agent     | `src/agent/{core,intent,streaming}.ts`       | intent classification, agent loop, streamed terminal output        |
-| Utils     | `src/utils/{diff-preview,memory,…}.ts`       | side-effect-free helpers                                           |
-| MCP       | `src/mcp/client.ts`                          | Model Context Protocol stdio client                                |
-
-### Data flow on a single prompt
-
-```
-User ──▶ commander ──▶ executeTask(prompt)
-                       │
-                       ├─▶ ConfigManager.load()         (process.env + ~/.thatgfsj/config.json)
-                       │
-                       ├─▶ AIEngine.chatStream(messages) ──▶ provider base URL /v1/chat/completions
-                       │                                  or Anthropic /messages
-                       │                                  or Gemini :generateContent
-                       │
-                       └─▶ process.stdout.write(chunks) ──▶ terminal (scrollable history)
-```
-
----
-
-## 3. Build & install
-
-### Prerequisites
-
-- Node.js ≥ 18 (project uses native `fetch`, `AbortController`, top-level `await`,
-  `node --test`).
-- npm ≥ 9 (so `npm publish` understands the `files` whitelist).
-- Git (for the install scripts and during release).
-
-### Local development
-
-```bash
-git clone https://github.com/Thatgfsj/thatgfsj-code.git
-cd thatgfsj-code
-npm install
-npm run build           # produces ./dist
-npm test                # node --test tests/
-npm start               # node dist/index.js (interactive REPL)
-npm run dev             # tsc && node dist/index.js
-
-# Optional: link globally so `gfcode` is in your PATH
-npm link
-```
-
-`npm install` after cloning will trigger `prepublishOnly` if you have it
-configured (we do). Always run `npm run build` after editing source so the
-`dist/` matches `src/`.
-
-### Platform notes
-
-- **Windows**: `src/index.ts` invokes `chcp 65001` synchronously before any
-  output to force UTF-8 in the legacy Windows console. Done via
-  `createRequire(import.meta.url)` so it works under `"type": "module"`.
-- **Numeric keypad arrows**: the REPL uses `@inquirer/input` instead of Node's
-  built-in `readline.question()` because the latter doesn't reliably translate
-  numeric-keypad ANSI escape sequences on Windows terminals (Bug #1 history).
-- **Streaming output**: `process.stdout.write` directly, never `rl.question`
-  inside the loop — the latter would reset the terminal and lose scrollback.
-
----
-
-## 4. Adding a provider
-
-1. Add the env-var map to `src/core/config.ts` (`envKeys`).
-2. Add a `*_MODELS` array to `src/repl/welcome.ts` and wire it through
-   `getModelsForProvider()` and `interactiveSetup()` (numbered choice).
-3. Update `WelcomeScreen.hasApiKey()` in `src/repl/welcome.ts` to include the
-   new env-var name.
-4. If the provider uses a non-OpenAI request format, extend the dispatcher
-   in `src/core/ai-engine.ts::streamRequest`.
-5. Update `docs/API_KEY_GUIDE.md` and `README.md`'s provider list.
-
-No other files should care about the specific provider.
-
----
-
-## 5. Adding a built-in CLI command
-
-`src/index.ts` is the single `commander` program. Register a new command
-**before** `program.parse(process.argv)`:
-
-```ts
-program
-  .command('foo')
-  .description('...')
-  .argument('<bar>', '...')
-  .action(async (bar) => { /* ... */ });
-```
-
-For REPL-internal slash-style commands (visible only inside the interactive
-loop), add a new `case` to `src/repl/loop.ts::handleCommand`.
-
-### Built-in REPL commands (current)
-
-| Command          | Behavior                                                                         |
-| ---------------- | -------------------------------------------------------------------------------- |
-| `help`           | Show built-in command list                                                       |
-| `exit` / `quit`  | Leave the REPL                                                                   |
-| `clear`          | Clear the screen                                                                 |
-| `context`        | Show the current project context                                                 |
-| `history`        | Show the command history for this session                                         |
-| `tools`          | List registered tools                                                            |
-| `models`         | Read-only listing of the current provider's models                               |
-| `providers`      | Read-only listing of all providers                                               |
-| `/model`         | **Interactive picker — actually switches the active model** for the current provider. Persists to `~/.thatgfsj/config.json` and calls `AIEngine.updateConfig()`. |
-| `/provider`      | **Interactive picker — switches provider + chains into `/model`** for the new one. Warns if the corresponding env-var / saved API key is missing. |
-| `Ctrl+C`         | Aborts the in-flight stream once, exits after two empty-prompt cancels.          |
-
-Both `/model` and `/provider` accept either a numeric index or the exact
-provider/model id; pressing Enter with no input keeps the current value. After
-a switch, `REPLLoop` injects a one-line `[system: ... switched to ...]`
-message into the session so the LLM can see the change in the next turn.
-
----
-
-## 6. Adding a tool
-
-Tools implement the `Tool` interface declared in `src/core/types.ts`:
-
-```ts
-export interface Tool {
-  name: string;
-  description: string;
-  parameters: ToolParameter[];
-  execute(params: any, ctx: ToolContext): Promise<ToolResult>;
-}
-```
-
-Steps:
-1. Create `src/tools/<name>.ts` exporting a class implementing `Tool`.
-2. Register it in `src/tools/index.ts::getBuiltInTools()` so both the REPL
-   `REPLLoop.init` and `executeTask` pick it up automatically.
-
-> **Known limitation (0.2.2):** `src/core/ai-engine.ts::extractToolCalls`
-> currently returns `undefined`. Tools are registered but never actually
-> invoked from a streamed response. Wiring this up requires parsing each
-> provider's native `tool_calls` delta and is tracked under "Unreleased" in
-> `CHANGELOG.md`. Until then, `ToolRegistry` is a forward-compatible seam.
-
----
-
-## 7. MCP integration
-
-`src/mcp/client.ts` defines a small stdio MCP client. It spawns a child
-process and exchanges JSON-RPC via newline-delimited messages. The current
-shape is intentionally minimal — see the file for the protocol mapping.
-
----
-
-## 8. Cross-platform & UX details
-
-- **Encoding**: Windows-only `chcp 65001` in `src/index.ts:9-15`. Skipped
-  silently if it fails.
-- **Color**: `chalk` v5 is ESM-only; the project is `"type": "module"` so
-  this is fine.
-- **Spinner**: `ora` v7 with `dots` spinner; cyan.
-- **History**: local to a single REPL session (not persisted across runs).
-
----
-
-## 9. Tests
-
-```bash
-npm test                # runs node --test tests/  (full TAP output)
-npm run test:silent     # dot reporter, for noisy CI logs
-```
-
-Tests live in `tests/` and use Node's built-in `node:test` runner — no
-vitest / jest dev-dep. They cover:
-
-- Smoke tests of the compiled `dist/index.js` binary (`--version`, `--help`,
-  subcommand empty-arg guards, `bash -n install.sh`).
-- Boundary tests of pure modules (`SessionManager`, `DiffPreview`,
-  `REPLOutput`) — empty input, unicode, CRLF, very long input.
-- No network. No mocked API keys. No shell-out side-effects other than the
-  install script's `bash -n` parse-check.
-
-CI integration: if you set this up later, run `npm test && npm run build`
-in the job; the `prepublishOnly` script already chains them.
-
----
-
-## 10. Release process
-
-1. Edit `package.json` `version` (semver).
-2. Add a new section at the top of `CHANGELOG.md`.
-3. `npm test && npm run build` locally — make sure both are green.
-4. `git add -A && git commit -m "release: vX.Y.Z"`
-5. `git tag vX.Y.Z`
-6. `git push origin main --follow-tags`
-7. `gh release create vX.Y.Z --title "vX.Y.Z" --notes-file CHANGELOG.md`
-   (the section for the new version is the top entry; trim if needed).
-8. `npm publish --access public` (a `prepublishOnly` step in `package.json`
-   re-runs tests + build, so you can't accidentally publish a stale build).
-
-### Semver policy
-
-- **patch** (`Z`): bug fixes, docs, refactors, no public API change.
-- **minor** (`Y`): new public command, new provider, new tool.
-- **major** (`X`): breaking change to the CLI surface or the JS API surface
-  (currently no JS API is stable).
-
----
-
-## 11. Troubleshooting
-
-### `ReferenceError: require is not defined`
-
-You're editing TypeScript and accidentally called `require(...)` outside of a
-`createRequire(import.meta.url)` shim. The project is ESM. Use `import x from 'y'`
-at the top of the file.
-
-### `npm publish` ships no `dist/`
-
-You forgot to run `npm run build` (or the `prepublishOnly` script was
-removed). The `files` whitelist in `package.json` requires the `dist/`
-directory to exist locally when publishing.
-
-### `gfcode --version` shows the wrong number
-
-Edit it in `package.json` only — `src/index.ts` and `src/repl/output.ts` both
-read `VERSION` from `package.json` via `import pkg from '../package.json' …`.
-
-### REPL keypresses / Ctrl+C misbehave on Windows
-
-Make sure you're on Windows Terminal or any TTY that emits ANSI escape
-sequences natively. The legacy `cmd.exe` console has incomplete ANSI
-support — upgrade to Windows Terminal or use PowerShell.

package/ROADMAP.md

@@ -1,91 +0,0 @@
-# Thatgfsj Code - Development Roadmap
-
----
-
-## ✅ v0.9.0 - 稳定版 (当前版本)
-
-### 核心架构
-- [x] Ink (React) TUI 框架
-- [x] LLM Provider 抽象层 (OpenAI/Anthropic/Gemini)
-- [x] 15 个 Provider + 自定义中转站
-- [x] 模块化目录结构 (cmd/app/config/llm/session/tools/tui/skills/hooks/prompts)
-
-### 工具系统
-- [x] file (读/写/列表/删除)
-- [x] shell (执行命令)
-- [x] git (操作)
-- [x] search (搜索)
-- [x] nwt (项目记忆，30天自动归档)
-
-### Skills 系统
-- [x] 16 个内置 Skills (规划/调试/TDD/架构/浏览器等)
-- [x] /skills 命令管理
-- [x] 自动激活
-
-### UI/UX
-- [x] Ink TUI 流式输出
-- [x] Markdown 渲染 (marked-terminal)
-- [x] 工具调用显示 (opencode 风格)
-- [x] 状态栏
-- [x] 命令系统 (/model, /new, /compact, /skills, /mcp, /help)
-- [x] /model 交互式选择框
-
-### 智能功能
-- [x] 自动读取 CLAUDE.md/SKILLS.md (通用路径)
-- [x] NWT 项目历史自动注入系统提示
-- [x] 上下文自动压缩 (超限触发)
-- [x] API 错误检测 (401/403/429)
-- [x] 启动时检测 API Key
-
-### 安装
-- [x] npm install -g thatgfsj-code
-- [x] Windows PowerShell 一键安装
-- [x] Linux/macOS Bash 一键安装
-- [x] 自动下载 Node.js (如果没有)
-
-### 测试通过
-- [x] 简单任务: 读文件 + 回答
-- [x] 中等任务: 修改文件 + 验证
-- [x] 困难任务: 多步骤 + NWT 记录
-- [x] Git + Search + NWT 联合测试
-- [x] 多文件创建和修改
-
----
-
-## 📋 未来计划
-
-### v1.0.0
-- [ ] SQLite 会话持久化
-- [ ] Token 使用量显示
-- [ ] 命令 Tab 补全
-- [ ] 多会话管理
-- [ ] LSP 集成
-- [ ] 图片支持
-
----
-
-## 历史版本
-
-### ✅ v0.8.0 - 命令系统
-- /model, /new, /compact, /skills, /mcp 命令
-
-### ✅ v0.7.0 - Skills + NWT
-- 16 个内置 Skills, NWT 项目记忆
-
-### ✅ v0.6.0 - Ink TUI
-- React 组件化 UI, Markdown 渲染
-
-### ✅ v0.5.0 - Provider 重写
-- 15 个 Provider, 自定义中转站
-
-### ✅ v0.4.0 - UI 优化
-- 流式输出, 工具调用显示
-
-### ✅ v0.3.0 - 架构重构
-- 模块化目录, Provider 抽象
-
-### ✅ v0.2.0 - 工具 + LLM
-- Git/搜索工具, 多 Provider
-
-### ✅ v0.1.0 - REPL 基础
-- 彩色 UI, 命令提示符