diff --git a/README.md b/README.md
index 8677925..4c9a7a1 100644
--- a/README.md
+++ b/README.md
@@ -1,231 +1,3 @@
 [English](./README.md) · [简体中文](./README.zh-CN.md)
 
-# auto-github-contributor
-
-![auto-github-contributor](./assets/cover_en.png)
-
-A Claude Code skill + slash command that takes any GitHub repo and drives a full OSS contribution pipeline end-to-end: discover a quick-win candidate (labeled issues + repo scan), run a TDD dev-loop until lint/typecheck/tests all pass, then open a PR via `gh` and print the URL.
-
-One slash command, full pipeline, any repo.
-
----
-
-## Overview
-
-When you invoke `/auto-contribute` inside Claude Code, the skill runs an 11-step interactive playbook:
-
-1. **Prereqs** — verify `gh`, `git`, `jq` are installed and `gh auth status` is green.
-2. **Target repo** — accept `owner/name` or a GitHub URL (or ask interactively).
-3. **Discover candidates in parallel**:
-   - Labeled issues (`good first issue`, `help wanted`, `documentation`, …)
-   - Repo-scan quick wins: typos, missing tests, i18n gaps, actionable TODOs
-4. **Ranked picklist** with estimated time + LLM cost per item.
-5. **Wait for explicit user confirmation** before touching anything.
-6. **Isolate** a workdir + feature branch under `$AGC_WORK_ROOT`.
-7. **Write SPEC.md + TODO.md** from templates.
-8. **TDD dev-loop** per todo (red → green → refactor), capped at 20 iterations.
-9. **Final verification**: clean install + lint + typecheck + tests + build.
-10. **Commit, push, open PR** via `gh pr create`. Prefers `TARGET_FORK` over upstream.
-11. **Print the PR URL** on its own line.
-
-Safety rails: never force-push, never commit to `main`/`master`/`develop`, never `rm -rf` outside `$AGC_WORK_ROOT`, never post to external systems unasked.
-
-## Requirements
-
-- `gh` (GitHub CLI) — authenticated (`gh auth login`, `repo` scope)
-- `git`
-- `jq`
-- Claude Code
-
-Optional but nice to have:
-
-- `rg` (ripgrep) — the scanner falls back to `grep` when absent
-- `pnpm` / `yarn` / `npm` / `bun` — autodetected from the target repo's lockfile
-- `python3` — used by `create-pr.sh` to render the PR body template
-
-## Installation
-
-This repo ships the slash command (`commands/auto-contribute.md`) and the skill (`skills/auto-github-contributor/`) as plain files — no plugin manifest, no marketplace registration. Drop them into Claude Code's user or project directory and they're live.
-
-**A. User-level (recommended — available in every project)**
-
-```bash
-git clone https://github.com/nexu-io/auto-github-contributor.git /tmp/auto-github-contributor
-mkdir -p ~/.claude/commands ~/.claude/skills
-cp    /tmp/auto-github-contributor/commands/auto-contribute.md  ~/.claude/commands/
-cp -R /tmp/auto-github-contributor/skills/auto-github-contributor ~/.claude/skills/
-```
-
-Restart Claude Code. `/auto-contribute` appears as a slash command, and Claude can also invoke the `auto-github-contributor` skill by name.
-
-**B. Project-level (scoped to one repo)**
-
-From the target repo's root:
-
-```bash
-git clone https://github.com/nexu-io/auto-github-contributor.git /tmp/auto-github-contributor
-mkdir -p .claude/commands .claude/skills
-cp    /tmp/auto-github-contributor/commands/auto-contribute.md  .claude/commands/
-cp -R /tmp/auto-github-contributor/skills/auto-github-contributor .claude/skills/
-```
-
-The command and skill are then available only inside that project.
-
-**C. Skill-only (no slash command)**
-
-If you don't want the `/auto-contribute` command — just the skill — drop `skills/auto-github-contributor/` into `~/.claude/skills/` (or `.claude/skills/`) alone. Invoke it by asking Claude to "auto-contribute to `<repo>`" or similar; the description in `SKILL.md` handles trigger words.
-
-## Usage
-
-```text
-/auto-contribute                               # skill will ask for the repo
-/auto-contribute cli/cli                       # owner/name form
-/auto-contribute https://github.com/cli/cli    # URL form
-```
-
-Override defaults via env vars before launching Claude Code (or export them in your shell profile):
-
-| Variable | Default | Purpose |
-|---|---|---|
-| `TARGET_REPO` | _(prompted)_ | `owner/name` of the upstream |
-| `TARGET_FORK` | _(empty → push to origin)_ | Where to push the feature branch |
-| `AGC_BASE_BRANCH` | `main` | Base branch for the PR |
-| `AGC_WORK_ROOT` | `$HOME/auto-gh-contrib-work` | Where workdirs get cloned |
-| `AGC_LABELS` | `good first issue,help wanted,documentation,good-first-issue` | Issue labels to search |
-| `AGC_ISSUE_LIMIT` | `30` | Max issues per label |
-| `AGC_INSTALL_CMD` / `AGC_LINT_CMD` / `AGC_TYPECHECK_CMD` / `AGC_TEST_CMD` / `AGC_BUILD_CMD` | pnpm-flavored | Fallback dev-loop commands (overridden by lockfile detection) |
-| `AGC_DEV_URL` | `http://localhost:5173` | URL for browser-verify stub |
-
-## How to test
-
-The skill has three testable layers. Start with Layer 1 — it's cheap, deterministic, and covers most of the real surface area.
-
-### Layer 1 — Script smoke tests (read-only, safe)
-
-All scripts live in `skills/auto-github-contributor/scripts/` and emit `KEY=VAL` lines or JSON. You can drive each one directly.
-
-```bash
-SKILL_DIR=skills/auto-github-contributor/scripts
-
-# 1. Prerequisites
-bash "$SKILL_DIR/check-prereqs.sh"
-# expect: ✓ gh / ✓ git / ✓ jq / ✓ gh authed as <you>; exit 0
-
-# 2. Issue discovery against a real repo
-TARGET_REPO=cli/cli AGC_ISSUE_LIMIT=3 bash "$SKILL_DIR/fetch-issues.sh" \
-  | jq '.[] | {number, title, score, labels}'
-# expect: ranked array of open issues matching the configured labels
-
-# 3. Quick-win scanner against any clone
-SCRATCH=/tmp/agc-smoke && rm -rf "$SCRATCH" \
-  && git clone --depth 1 https://github.com/expressjs/express.git "$SCRATCH"
-bash "$SKILL_DIR/scan-quick-wins.sh" --workdir "$SCRATCH" --max 20 \
-  | jq 'group_by(.kind) | map({kind: .[0].kind, count: length})'
-# expect: an array with some of {typo, missing-test, i18n, todo}
-
-# 4. Parse-check the scripts that mutate state
-bash -n "$SKILL_DIR/setup-workspace.sh"
-bash -n "$SKILL_DIR/dev-loop-check.sh"
-bash -n "$SKILL_DIR/create-pr.sh"
-bash -n "$SKILL_DIR/browser-verify.sh"
-```
-
-### Layer 2 — Slash command + skill loading
-
-In a Claude Code session at the repo root, type `/auto-contribute` and verify:
-
-- the slash command autocompletes (proves `commands/auto-contribute.md` is registered),
-- Claude immediately invokes the `auto-github-contributor` skill,
-- it runs `check-prereqs.sh` first,
-- it asks for the target repo via `AskUserQuestion`,
-- it fetches issues and scans quick-wins in parallel,
-- it renders a picklist with time/cost estimates and **stops for confirmation**.
-
-You can cancel at the picklist — no workdir is created beyond a `_scratch` shallow clone.
-
-### Layer 3 — End-to-end on a sandbox repo
-
-Create a throwaway repo you own (e.g. `your-handle/agc-sandbox`) with a known quick-win seeded — for instance, write `# Teh setup guide` (note `Teh`) in `README.md`. Then:
-
-```text
-/auto-contribute your-handle/agc-sandbox
-```
-
-Pick the typo from the list, approve, and let the skill drive through to a real PR. Close the PR once you've verified the flow.
-
-Tip: set `TARGET_FORK=your-handle/agc-sandbox` so it pushes to the same repo (branch) instead of prompting about origin.
-
-## How it works
-
-```
-commands/auto-contribute.md         # /auto-contribute slash command (thin wrapper)
-skills/auto-github-contributor/
-  SKILL.md                          # orchestration playbook Claude reads
-  scripts/
-    config.sh                       # shared env + helpers
-    check-prereqs.sh                # gh/git/jq + auth check
-    fetch-issues.sh                 # label-based issue discovery + ranking
-    scan-quick-wins.sh              # typo / missing-test / i18n / TODO scanner
-    setup-workspace.sh              # clone + feature branch
-    dev-loop-check.sh               # red/green/final phases, lockfile-aware
-    browser-verify.sh               # visual-verification stub
-    create-pr.sh                    # commit, push, open PR, render body
-  templates/
-    SPEC.template.md                # problem / acceptance / approach / risk
-    TODO.template.md                # atomic todos mirrored to TaskCreate
-    PR-BODY.template.md             # rendered into .auto-pr/PR-BODY.md
-```
-
-The skill file (`SKILL.md`) is the contract Claude follows. The scripts do the real work; Claude only orchestrates, reads their output, and makes decisions at the branching points.
-
-## How to contribute
-
-Contributions welcome. The most valuable additions right now:
-
-1. **New quick-win kinds** — extend `scan-quick-wins.sh`. Good targets:
-   - broken markdown links
-   - unused exports (ts-prune-style)
-   - missing JSDoc on public APIs
-   - outdated dependency pins in a package.json section
-2. **Browser-verify backend** — `browser-verify.sh` is a deliberate stub. Wire up a Playwright / Puppeteer / chrome-devtools MCP backend so UI changes get screenshotted before the PR.
-3. **Language/framework dev-loops** — `dev-loop-check.sh` currently detects JS package managers. Add Python (`uv`/`poetry`/`pip`), Go, Rust, etc.
-4. **Cost estimator** — the picklist cost numbers are hand-waved rules of thumb. A smarter estimator (token-weighted file size × action kind) would help users pick more confidently.
-
-### Guidelines
-
-- **Scripts emit machine-readable output.** Stdout is `KEY=VAL` lines or JSON. Human logs go to stderr via `agc::log` / `agc::err`. Don't mix them.
-- **No silent stubs.** If a feature isn't wired up yet, print `stub: <reason>` so the PR body can surface the caveat.
-- **Safety rails stay on.** Never force-push, never commit to `main`/`master`/`develop`, never `rm -rf` outside `$AGC_WORK_ROOT`. If you add a new mutation path, add the guard too.
-- **Config via env vars.** New tunables go into `config.sh` with a sensible default and an entry in the README table above.
-- **Shell style.** `set -euo pipefail` (relax only where explicitly justified, as in `scan-quick-wins.sh`). Prefer `$(...)` over backticks. Quote expansions.
-
-### Local dev loop
-
-```bash
-# parse-check all scripts
-for f in skills/auto-github-contributor/scripts/*.sh; do bash -n "$f" || exit 1; done
-
-# smoke the scanner against a local clone
-bash skills/auto-github-contributor/scripts/scan-quick-wins.sh --workdir /path/to/some/repo | jq .
-
-# run shellcheck if you have it installed
-shellcheck skills/auto-github-contributor/scripts/*.sh
-```
-
-### Filing issues
-
-- Include the Claude Code version, OS, and `gh --version`.
-- Paste the stderr log from the failing script (the `[auto-gh]` lines are the useful ones).
-- If the issue is in discovery, include `TARGET_REPO` — different repos expose very different label sets.
-
-### PR checklist
-
-- [ ] `bash -n` passes for every touched script
-- [ ] Script output still follows the stdout-is-data / stderr-is-log split
-- [ ] New env vars documented in `README.md` and defaulted in `config.sh`
-- [ ] Safety rails not weakened
-- [ ] If you touched `SKILL.md`, re-read it end-to-end — Claude follows it literally
-
-## License
-
-MIT — see [LICENSE](LICENSE).
+# auto-github-shitter
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 8433a43..50d451c 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -1,231 +1,3 @@
 [English](./README.md) · [简体中文](./README.zh-CN.md)
 
-# auto-github-contributor
-
-![auto-github-contributor](./assets/cover_zh.png)
-
-一套 Claude Code skill + slash 命令：对任意 GitHub 仓库端到端跑通一次开源贡献流程 —— 自动发现 quick-win 候选（带标签的 issue + 仓库扫描），按 TDD 走 red/green/refactor 直到 lint/typecheck/tests 全绿，然后通过 `gh` 开 PR 并输出链接。
-
-一条 slash 命令，覆盖完整流水线，适用于任何仓库。
-
----
-
-## 简介
-
-在 Claude Code 中调用 `/auto-contribute` 后，skill 会按 11 步交互式 playbook 执行：
-
-1. **环境检查** —— 确认 `gh` / `git` / `jq` 已安装，`gh auth status` 正常。
-2. **锁定目标仓库** —— 接受 `owner/name` 或 GitHub URL（缺参时交互式询问）。
-3. **并行发现候选**：
-   - 带标签的 issue（`good first issue`、`help wanted`、`documentation` 等）
-   - 仓库扫描的 quick-win：拼写错误、缺测试、i18n 缺 key、可执行的 TODO
-4. **排序挑选列表**，每一项都带时间和 LLM 成本估算。
-5. **等待用户明确确认** 后才动手改代码。
-6. **隔离工作区**：在 `$AGC_WORK_ROOT` 下建独立目录和功能分支。
-7. **写 SPEC.md + TODO.md**，用模板生成。
-8. **TDD 开发循环**（red → green → refactor），每个 todo 最多 20 轮。
-9. **最终验证**：clean install + lint + typecheck + tests + build。
-10. **提交、推送、开 PR**（`gh pr create`），优先推到 `TARGET_FORK`，否则推到 upstream。
-11. **单独一行输出 PR URL**，方便用户点击。
-
-安全约束：绝不 force-push；绝不提交到 `main`/`master`/`develop`；绝不在 `$AGC_WORK_ROOT` 外 `rm -rf`；未经授权绝不向外部系统（Slack、邮件等）发消息。
-
-## 环境要求
-
-- `gh`（GitHub CLI）—— 已登录（`gh auth login`，需要 `repo` scope）
-- `git`
-- `jq`
-- Claude Code
-
-可选但推荐：
-
-- `rg`（ripgrep）—— 没有时扫描脚本会回退到 `grep`
-- `pnpm` / `yarn` / `npm` / `bun` —— 根据目标仓库的 lockfile 自动识别
-- `python3` —— `create-pr.sh` 渲染 PR 正文模板时用到
-
-## 安装
-
-本仓库只包含 slash 命令（`commands/auto-contribute.md`）和 skill（`skills/auto-github-contributor/`），都是普通文件 —— 不需要 plugin manifest，也不需要注册 marketplace。拷到 Claude Code 的 user 或 project 目录下即可生效。
-
-**A. 用户级安装（推荐，所有项目通用）**
-
-```bash
-git clone https://github.com/nexu-io/auto-github-contributor.git /tmp/auto-github-contributor
-mkdir -p ~/.claude/commands ~/.claude/skills
-cp    /tmp/auto-github-contributor/commands/auto-contribute.md  ~/.claude/commands/
-cp -R /tmp/auto-github-contributor/skills/auto-github-contributor ~/.claude/skills/
-```
-
-重启 Claude Code，`/auto-contribute` 就会出现在 slash 命令列表里，`auto-github-contributor` skill 也能被 Claude 按名字调用。
-
-**B. 项目级安装（只在单个仓库生效）**
-
-在目标项目根目录下：
-
-```bash
-git clone https://github.com/nexu-io/auto-github-contributor.git /tmp/auto-github-contributor
-mkdir -p .claude/commands .claude/skills
-cp    /tmp/auto-github-contributor/commands/auto-contribute.md  .claude/commands/
-cp -R /tmp/auto-github-contributor/skills/auto-github-contributor .claude/skills/
-```
-
-这样命令和 skill 只在这个项目里可用。
-
-**C. 只要 skill、不要 slash 命令**
-
-只把 `skills/auto-github-contributor/` 拷到 `~/.claude/skills/`（或 `.claude/skills/`）即可。之后用自然语言触发（例如 "auto-contribute to `<repo>`"）——`SKILL.md` 里的 description 字段会匹配关键词。
-
-## 使用
-
-```text
-/auto-contribute                               # 不带参数，会交互式问仓库
-/auto-contribute cli/cli                       # owner/name 形式
-/auto-contribute https://github.com/cli/cli    # URL 形式
-```
-
-在启动 Claude Code 前通过环境变量覆盖默认值（也可以放进 shell profile）：
-
-| 变量 | 默认值 | 用途 |
-|---|---|---|
-| `TARGET_REPO` | _（运行时询问）_ | upstream 的 `owner/name` |
-| `TARGET_FORK` | _（空即推到 origin）_ | 功能分支推往哪个 fork |
-| `AGC_BASE_BRANCH` | `main` | PR 的 base 分支 |
-| `AGC_WORK_ROOT` | `$HOME/auto-gh-contrib-work` | 克隆工作区根目录 |
-| `AGC_LABELS` | `good first issue,help wanted,documentation,good-first-issue` | 搜索 issue 用的标签 |
-| `AGC_ISSUE_LIMIT` | `30` | 每个标签最多拉多少条 |
-| `AGC_INSTALL_CMD` / `AGC_LINT_CMD` / `AGC_TYPECHECK_CMD` / `AGC_TEST_CMD` / `AGC_BUILD_CMD` | pnpm 风格 | dev-loop 兜底命令（lockfile 识别会覆盖它们） |
-| `AGC_DEV_URL` | `http://localhost:5173` | 浏览器可视化校验的 URL |
-
-## 如何测试
-
-整个 skill 可以分三层来测。建议从 Layer 1 起 —— 成本最低、确定性最强，覆盖真正容易出问题的代码路径。
-
-### Layer 1 —— 脚本冒烟测试（只读、安全）
-
-所有脚本都在 `skills/auto-github-contributor/scripts/` 下，输出都是 `KEY=VAL` 或 JSON，可独立调用：
-
-```bash
-SKILL_DIR=skills/auto-github-contributor/scripts
-
-# 1. 环境检查
-bash "$SKILL_DIR/check-prereqs.sh"
-# 预期: ✓ gh / ✓ git / ✓ jq / ✓ gh authed as <你的用户名>；exit 0
-
-# 2. 用真实仓库跑 issue 发现
-TARGET_REPO=cli/cli AGC_ISSUE_LIMIT=3 bash "$SKILL_DIR/fetch-issues.sh" \
-  | jq '.[] | {number, title, score, labels}'
-# 预期: 按 score 排序的 open issue 数组
-
-# 3. 对任意克隆跑 quick-win 扫描
-SCRATCH=/tmp/agc-smoke && rm -rf "$SCRATCH" \
-  && git clone --depth 1 https://github.com/expressjs/express.git "$SCRATCH"
-bash "$SKILL_DIR/scan-quick-wins.sh" --workdir "$SCRATCH" --max 20 \
-  | jq 'group_by(.kind) | map({kind: .[0].kind, count: length})'
-# 预期: 数组里包含 {typo, missing-test, i18n, todo} 中的若干种
-
-# 4. 对会改状态的脚本做语法检查
-bash -n "$SKILL_DIR/setup-workspace.sh"
-bash -n "$SKILL_DIR/dev-loop-check.sh"
-bash -n "$SKILL_DIR/create-pr.sh"
-bash -n "$SKILL_DIR/browser-verify.sh"
-```
-
-### Layer 2 —— slash 命令 + skill 装载
-
-在本仓库根目录启动 Claude Code，输入 `/auto-contribute`，检查：
-
-- slash 命令能自动补全（说明 `commands/auto-contribute.md` 被识别）
-- Claude 立即通过 Skill 工具加载 `auto-github-contributor`
-- 先跑 `check-prereqs.sh`
-- 通过 `AskUserQuestion` 交互式问目标仓库
-- 并行拉 issue 与扫描 quick-win
-- 展示带时间/成本估算的挑选列表，**然后停下来等确认**
-
-在挑选列表处直接取消即可，不会产生 `_scratch` 之外的任何副作用。
-
-### Layer 3 —— 真实端到端
-
-建一个你自己的实验仓库（例如 `your-handle/agc-sandbox`），刻意埋一个 quick-win —— 比如在 `README.md` 里写 `# Teh setup guide`（注意 `Teh` 是错的）。然后：
-
-```text
-/auto-contribute your-handle/agc-sandbox
-```
-
-从挑选列表里选中这个 typo、同意继续，让 skill 走完完整流水线开出一个真 PR。验证完关掉 PR 就行。
-
-小技巧：同时设 `TARGET_FORK=your-handle/agc-sandbox`，这样会直接推到同仓库的分支，不会提示推 origin 的确认。
-
-## 工作原理
-
-```
-commands/auto-contribute.md         # /auto-contribute slash 命令（薄壳）
-skills/auto-github-contributor/
-  SKILL.md                          # Claude 要读的 orchestration playbook
-  scripts/
-    config.sh                       # 共享环境变量 + 工具函数
-    check-prereqs.sh                # gh/git/jq + 登录态检查
-    fetch-issues.sh                 # 按标签发现 + 排序 issue
-    scan-quick-wins.sh              # typo / 缺测试 / i18n / TODO 扫描
-    setup-workspace.sh              # 克隆 + 创建功能分支
-    dev-loop-check.sh               # red/green/final 阶段，识别 lockfile
-    browser-verify.sh               # 浏览器可视化验证（stub）
-    create-pr.sh                    # commit/push/开 PR/渲染正文
-  templates/
-    SPEC.template.md                # 问题/验收标准/方案/风险
-    TODO.template.md                # 会镜像到 TaskCreate 的原子 todo
-    PR-BODY.template.md             # 渲染到 .auto-pr/PR-BODY.md
-```
-
-`SKILL.md` 是 Claude 要遵守的契约；真正干活的是 scripts/。Claude 只做编排 —— 读脚本输出、在分支点做决定。
-
-## 如何贡献
-
-欢迎 PR。目前最有价值的方向：
-
-1. **新增 quick-win 类型** —— 扩展 `scan-quick-wins.sh`，候选：
-   - markdown 死链检查
-   - 未使用的 export（ts-prune 风格）
-   - 缺 JSDoc 的 public API
-   - package.json 里过时的依赖版本 pin
-2. **浏览器验证后端** —— `browser-verify.sh` 目前是有意保留的 stub。接上 Playwright / Puppeteer / chrome-devtools MCP 后端，让 UI 改动在开 PR 前先截一张图。
-3. **多语言/框架 dev-loop** —— `dev-loop-check.sh` 当前只识别 JS 包管理器。欢迎加 Python（`uv` / `poetry` / `pip`）、Go、Rust 等。
-4. **成本估算器** —— 挑选列表里的成本数字目前只是经验拍的。更靠谱的算法（按 token 加权的文件大小 × 动作类型）能帮用户更有信心地选。
-
-### 贡献约定
-
-- **脚本输出是机器可读的。** stdout 只写 `KEY=VAL` 或 JSON；人类日志走 stderr，通过 `agc::log` / `agc::err`。别混。
-- **不允许静默 stub。** 新功能没接完就打印 `stub: <原因>`，让 PR 正文能把 caveat 带出来。
-- **安全约束保持开启。** 绝不 force-push、绝不提交到 `main`/`master`/`develop`、绝不在 `$AGC_WORK_ROOT` 外 `rm -rf`。新加会改状态的代码路径必须顺手加 guard。
-- **配置通过环境变量。** 新的可调项统一进 `config.sh`，给个合理默认值，并在本文件的环境变量表里补上一行。
-- **Shell 风格。** 默认 `set -euo pipefail`（只在有明确理由时放宽，如 `scan-quick-wins.sh`）。用 `$(...)` 代替反引号，注意对扩展加引号。
-
-### 本地开发循环
-
-```bash
-# 对全部脚本做语法检查
-for f in skills/auto-github-contributor/scripts/*.sh; do bash -n "$f" || exit 1; done
-
-# 对本地任意克隆跑扫描器
-bash skills/auto-github-contributor/scripts/scan-quick-wins.sh --workdir /path/to/some/repo | jq .
-
-# 装了 shellcheck 的话
-shellcheck skills/auto-github-contributor/scripts/*.sh
-```
-
-### 提 issue
-
-- 附上 Claude Code 版本、操作系统、`gh --version`。
-- 粘出失败脚本的 stderr 日志（特别是 `[auto-gh]` 开头的那些行，信息量最大）。
-- 如果是发现阶段的问题，把 `TARGET_REPO` 写进来 —— 不同仓库的标签集差异很大。
-
-### PR checklist
-
-- [ ] 所有改动过的脚本 `bash -n` 通过
-- [ ] 输出仍然遵守「stdout 是数据 / stderr 是日志」的分工
-- [ ] 新环境变量在 `README.md` 里有文档、在 `config.sh` 里有默认值
-- [ ] 安全约束没有被削弱
-- [ ] 改过 `SKILL.md` 的话，从头到尾读一遍 —— Claude 是严格按它执行的
-
-## 许可证
-
-MIT —— 详见 [LICENSE](LICENSE)。
+# 自动拉屎贡献者
diff --git a/assets/cover_en.png b/assets/cover_en.png
deleted file mode 100644
index ce92852..0000000
Binary files a/assets/cover_en.png and /dev/null differ
diff --git a/assets/cover_zh.png b/assets/cover_zh.png
deleted file mode 100644
index aee4cc0..0000000
Binary files a/assets/cover_zh.png and /dev/null differ
diff --git a/commands/auto-contribute.md b/commands/auto-contribute.md
deleted file mode 100644
index 7c558d9..0000000
--- a/commands/auto-contribute.md
+++ /dev/null
@@ -1,23 +0,0 @@
----
-description: Find a quick-win contribution in a GitHub repo and open a PR end-to-end.
-argument-hint: "[repo-url-or-owner/name]"
----
-
-You are entering the **auto-github-contributor** flow.
-
-User input (may be empty): `$ARGUMENTS`
-
-## What to do right now
-
-1. **Load the skill** by invoking the `auto-github-contributor` skill via the Skill tool. The skill owns the full execution playbook — do not reimplement it inline.
-
-2. **Pass the user input forward**:
-   - If `$ARGUMENTS` looks like a repo (`https://github.com/<owner>/<name>` or `<owner>/<name>`), set it as `TARGET_REPO` and skip the "ask for repo" step in the skill.
-   - Otherwise, the skill will interactively ask for the repo URL via `AskUserQuestion`.
-
-3. **Honor the interactive contract**:
-   - Always run the prerequisite check first (`gh` installed + authed). If it fails, surface the install/auth hint and stop — do not try workarounds.
-   - Always present the discovered candidates (labeled issues + repo-scan quick wins) with estimated time/cost, and **wait for explicit user confirmation** before starting the dev-loop.
-   - At the end, print the PR URL on its own line so the user can click through.
-
-Begin by invoking the skill now.
diff --git a/skills/auto-github-contributor/SKILL.md b/skills/auto-github-contributor/SKILL.md
deleted file mode 100644
index 4ecabf7..0000000
--- a/skills/auto-github-contributor/SKILL.md
+++ /dev/null
@@ -1,229 +0,0 @@
----
-name: auto-github-contributor
-description: Interactively pick a quick-win contribution in any GitHub repo — either a labeled issue ("good first issue", "help wanted", docs) or a repo-scan candidate (typo, missing tests, i18n gap, actionable TODO). Runs a TDD dev-loop, opens a PR via gh, and returns the PR URL. Use when the user wants to auto-contribute to an open-source project end-to-end. Trigger words: auto-contribute, open a PR for me, find a good first issue, contribute to <repo>.
-allowed-tools:
-  - Bash
-  - Read
-  - Write
-  - Edit
-  - AskUserQuestion
-  - TaskCreate
-  - TaskUpdate
-  - WebFetch
----
-
-# auto-github-contributor — interactive OSS PR agent
-
-One command, full pipeline, any repo:
-
-1. **Check prerequisites** (`gh`, `git`, `jq`, gh auth) — guide the user to fix anything missing.
-2. **Ask for the target repo** (`owner/name` or GitHub URL). Default fork to the authed `gh` user.
-3. **Discover candidates** from two sources in parallel:
-   - Labeled issues (`good first issue`, `help wanted`, `documentation`, etc.)
-   - Repo scan for quick wins (typos, missing tests, i18n gaps, actionable TODOs)
-4. **Present a ranked picklist** with estimated time + LLM cost per item.
-5. **Wait for explicit user confirmation** before touching anything.
-6. **Run the TDD dev-loop** (red → green → refactor) until lint / typecheck / tests all pass.
-7. **Push + open PR** via `gh`. **Print the PR URL on its own line** so the user can click.
-
-Scripts live next to this file under `scripts/`. Each script emits machine-readable `KEY=VAL` lines on stdout and logs to stderr. Source the shared helpers from every script:
-
-```bash
-source "$(dirname "$0")/config.sh"
-```
-
-Override config via env vars on invocation:
-
-```bash
-TARGET_REPO=owner/name TARGET_FORK=me/name bash <script>
-```
-
-## Execution steps — follow in order
-
-### Step 1 — Prerequisite check (always first)
-
-```bash
-bash "$SKILL_DIR/scripts/check-prereqs.sh"
-```
-
-Reads `GH_USER=...` from stdout — capture it as the default fork owner.
-
-- Exit code **0**: continue.
-- Exit code **2**: surface the printed install / auth hint **verbatim** to the user and stop. Do **not** attempt token workarounds, `sudo` installs, or other shortcuts. The user runs the fix and restarts the flow.
-
-### Step 2 — Resolve the target repo
-
-Priority order:
-
-1. If the slash command passed `TARGET_REPO` via `$ARGUMENTS`, normalize and use it.
-2. Otherwise, ask the user via **AskUserQuestion** with one free-text prompt titled "Target repo" and hint `owner/name or https://github.com/owner/name`. If the user gives a URL, strip the protocol + trailing `.git`.
-3. Ask for fork owner (optional) — default to `$GH_USER`. If the user confirms no fork, `TARGET_FORK` stays empty and we push to origin (create-pr.sh will warn 3s before doing so).
-
-Validate: `gh repo view "$TARGET_REPO"` must succeed. If the repo doesn't exist or the user lacks access, stop and surface the gh error.
-
-Export `TARGET_REPO` and `TARGET_FORK` for the rest of the session.
-
-### Step 3 — Discover candidates (parallel)
-
-**3a. Labeled issues**
-
-```bash
-bash "$SKILL_DIR/scripts/fetch-issues.sh" > /tmp/agc-issues.json
-```
-
-This queries each label in `$AGC_LABELS` (default: `good first issue, help wanted, documentation, good-first-issue`), merges, de-dupes, and ranks by a score that weights label specificity + freshness + body clarity. Top of the list = easiest high-quality pick.
-
-**3b. Repo quick-win scan**
-
-We need a shallow clone first so scan-quick-wins.sh has files to read. Use a scratch clone rather than the per-issue workdir (no branch yet):
-
-```bash
-SCRATCH="$AGC_WORK_ROOT/$(echo $TARGET_REPO | tr / -)/_scratch"
-if [[ ! -d "$SCRATCH/.git" ]]; then
-  git clone --depth 1 "https://github.com/$TARGET_REPO.git" "$SCRATCH"
-fi
-bash "$SKILL_DIR/scripts/scan-quick-wins.sh" --workdir "$SCRATCH" > /tmp/agc-quickwins.json
-```
-
-Returns an array of `{kind, title, summary, file, line, estimated_minutes, slug}`. Kinds: `typo`, `missing-test`, `i18n`, `todo`.
-
-### Step 4 — Present the picklist
-
-Render a single merged table. For each entry compute:
-
-- **Estimated time** — issues: flat 60min (we'll refine after reading the body); quick-wins: `estimated_minutes` from the scanner.
-- **Estimated cost** — use this rule of thumb, cheap and transparent:
-  - typo, tiny doc fix: **~$0.30**
-  - i18n, small test add: **~$1.50**
-  - missing-test for a non-trivial module: **~$3**
-  - TODO resolution or labeled issue: **~$5–8** depending on scope
-  - UI-affecting issue that needs visual verification: **+$2**
-  The point isn't precision — it's letting the user pick a ceiling.
-
-Format as markdown so the user can scan quickly:
-
-```
-Discovered candidates in owner/name:
-
-Quick wins (repo scan)
-  1. [typo]         Fix "teh" → "the" in docs/intro.md:42           ~5 min  ~$0.30
-  2. [missing-test] Add tests for src/utils/parse.ts                 ~30 min ~$1.50
-  3. [i18n]         Fill 12 missing keys in locales/zh-CN.json       ~45 min ~$2.00
-  ...
-
-Issues (labeled)
-  a. #124  Clarify CONTRIBUTING steps for monorepo setup             ~60 min ~$3.00
-     labels: documentation · good first issue
-  b. #131  Add --json flag to `cli list` command                     ~90 min ~$6.00
-     labels: good first issue · help wanted
-  ...
-```
-
-Then call **AskUserQuestion** titled "Pick one" with options: top quick-wins (numbered) + top issues (lettered) + "Other: paste issue # or slug" fallback + "Cancel".
-
-If the user picks "Cancel", stop cleanly. No work, no clones beyond `_scratch`.
-
-### Step 5 — Isolate the workdir
-
-```bash
-# For issue pick:
-bash "$SKILL_DIR/scripts/setup-workspace.sh <ISSUE_NUMBER>"
-
-# For quick-win pick:
-bash "$SKILL_DIR/scripts/setup-workspace.sh <SLUG> --title "<short title>"
-```
-
-Capture `WORKDIR=...`, `BRANCH=...`, `ISSUE_TITLE=...` from stdout.
-
-### Step 6 — Write the spec
-
-Fill `templates/SPEC.template.md` → `$WORKDIR/.auto-pr/SPEC.md`:
-
-- **Problem**: restate in your own words.
-- **Acceptance criteria**: testable bullets.
-- **Approach**: concrete change plan.
-- **Files likely touched**: from a quick `rg`/`find` pass against the target repo.
-- **Risk / blast radius**.
-- **Test plan**: unit + integration + visual.
-
-For quick-win picks, the spec can be minimal (1–2 acceptance criteria). Don't over-invest for a typo fix.
-
-### Step 7 — Generate TODO breakdown + mirror to TaskCreate
-
-Fill `templates/TODO.template.md` → `$WORKDIR/.auto-pr/TODO.md`, then create one TaskCreate entry per atomic todo. Prefer `failing test → minimal implementation → refactor` triples. Mark each TaskUpdate `in_progress` before starting, `completed` when dev-loop green.
-
-### Step 8 — TDD dev-loop (per todo)
-
-1. **Red**: write/update the failing test.
-2. ```bash
-   bash "$SKILL_DIR/scripts/dev-loop-check.sh" --phase red --workdir "$WORKDIR"
-   ```
-   Must exit 0 (tests fail as expected).
-3. **Green**: minimal implementation.
-4. ```bash
-   bash "$SKILL_DIR/scripts/dev-loop-check.sh" --phase green --workdir "$WORKDIR"
-   ```
-   Runs install + lint + typecheck + test. Fail → diagnose → fix → re-run. **Do not mark the todo complete until this exits 0.**
-5. **Refactor** if warranted — re-run step 4.
-6. **Visual verification** (UI-affecting only):
-   ```bash
-   bash "$SKILL_DIR/scripts/browser-verify.sh" --url "$AGC_DEV_URL" --out "$WORKDIR/.auto-pr/screenshots/<todo-slug>.png"
-   ```
-   If the script is still a stub, continue — note `visual verify: stub — manual check pending` in the PR body.
-
-Iteration cap: **20 loops per todo**. On exceed, write the blocker to `$WORKDIR/.auto-pr/BLOCKERS.md` and move on; the PR will be marked draft and call out the blocker.
-
-### Step 9 — Final verification
-
-```bash
-bash "$SKILL_DIR/scripts/dev-loop-check.sh" --phase final --workdir "$WORKDIR"
-```
-
-Clean install + lint + typecheck + full tests + build. Must exit 0.
-
-### Step 10 — Commit, push, open PR
-
-```bash
-# Issue-driven:
-bash "$SKILL_DIR/scripts/create-pr.sh <ISSUE_NUMBER>"
-
-# Quick-win:
-bash "$SKILL_DIR/scripts/create-pr.sh <SLUG> --title "<short title>"
-```
-
-`create-pr.sh` commits, pushes to `$TARGET_FORK` (falls back to origin with a 3s abort window), renders the PR body from `SPEC.md` + `TODO.md` + screenshots + blockers, and opens the PR via `gh pr create`.
-
-**Print the PR URL on its own line** to the user, e.g.:
-
-```
-PR opened: https://github.com/owner/name/pull/456
-```
-
-### Step 11 — Wrap up
-
-Print a one-line recap: what was picked, PR URL, any stub hooks still pending. No extra summary files unless the user asks.
-
-## Flow variations
-
-- **User passes repo URL via slash command**: skip the "Ask for repo" prompt in Step 2.
-- **User picks "Other" in Step 4**: fall back to a free-text prompt; accept either `#<issue>` or a slug. If `#<issue>`, route through the issue branch; otherwise treat as a quick-win slug and ask for a one-line title.
-- **Discovery returns zero candidates**: tell the user the repo looks "clean" for the configured labels and scan heuristics. Offer to broaden labels (`AGC_LABELS`) or accept a manual issue number.
-
-## Stub policy
-
-Every script accepts `--dry-run` where it makes sense. Stub sections are marked:
-
-```
-# TODO(auto-gh): <what's missing>
-```
-
-Never silently skip verification — always emit a visible `stub: <reason>` line so the PR body includes the caveat.
-
-## Safety rails
-
-- **Never force-push.** `create-pr.sh` uses plain `git push -u <remote> <branch>`.
-- **Never commit to upstream base branch.** Hard-coded guard: if the current branch is `main`/`master`/`develop`, abort.
-- **Never `rm -rf`** outside `$AGC_WORK_ROOT`. Workdir paths are validated before any cleanup.
-- **Never write to shared systems** (Slack, email, external services) unless the user explicitly asks.
-- If `gh auth status` fails mid-flow, surface the error and stop — no token workarounds.
-- Ask before doing anything destructive the user didn't explicitly request (force-push, branch deletion, rebase onto a shifted base).
diff --git a/skills/auto-github-contributor/scripts/browser-verify.sh b/skills/auto-github-contributor/scripts/browser-verify.sh
deleted file mode 100755
index 0967223..0000000
--- a/skills/auto-github-contributor/scripts/browser-verify.sh
+++ /dev/null
@@ -1,87 +0,0 @@
-#!/usr/bin/env bash
-# Browser-based visual verification.
-#
-# This script drives a headless browser to capture screenshots / DOM dumps so
-# the agent can eyeball the change before opening the PR. The actual automation
-# body is intentionally a STUB — wire it up to whichever tool you prefer:
-#
-#   - Playwright    (pnpm dlx playwright screenshot ...)
-#   - Puppeteer     (node ./scripts/puppet.js ...)
-#   - chrome-devtools MCP   (mcp__chrome_devtools__take_screenshot)
-#   - Browser Use MCP
-#
-# Contract:
-#   Input  : --url <url> --out <png-path> [--selector <css>] [--wait <ms>]
-#   Output : writes PNG to --out, prints "SCREENSHOT=<abs-path>" on stdout
-#   Stubs  : when the automation body is not yet implemented, the script
-#            still creates the output directory, writes a placeholder note,
-#            and exits 0 with "stub: <reason>" so the agent flow continues.
-
-set -euo pipefail
-source "$(dirname "$0")/config.sh"
-
-URL=""
-OUT=""
-SELECTOR=""
-WAIT_MS="1500"
-
-while (($#)); do
-  case "$1" in
-    --url) URL="$2"; shift 2 ;;
-    --out) OUT="$2"; shift 2 ;;
-    --selector) SELECTOR="$2"; shift 2 ;;
-    --wait) WAIT_MS="$2"; shift 2 ;;
-    *) agc::die "unknown flag: $1" ;;
-  esac
-done
-
-[[ -n "$URL" ]] || agc::die "--url required"
-[[ -n "$OUT" ]] || agc::die "--out required"
-
-mkdir -p "$(dirname "$OUT")"
-
-agc::log "browser-verify url=${URL} out=${OUT} viewport=${AGC_BROWSER_VIEWPORT} headless=${AGC_BROWSER_HEADLESS}"
-
-# --------------------------------------------------------------------------
-# Implementation hook — pick ONE backend and uncomment.
-# Left as stubs on purpose; user will plug in preferred tooling.
-# --------------------------------------------------------------------------
-
-# === Backend A: Playwright CLI =============================================
-# agc::require npx
-# VIEWPORT_W="${AGC_BROWSER_VIEWPORT%x*}"
-# VIEWPORT_H="${AGC_BROWSER_VIEWPORT#*x}"
-# npx --yes playwright@latest screenshot \
-#   --viewport-size="${VIEWPORT_W},${VIEWPORT_H}" \
-#   --wait-for-timeout "$WAIT_MS" \
-#   ${SELECTOR:+--selector "$SELECTOR"} \
-#   "$URL" "$OUT"
-
-# === Backend B: Puppeteer via node =========================================
-# node "$(dirname "$0")/puppet.mjs" --url "$URL" --out "$OUT" \
-#   --selector "$SELECTOR" --wait "$WAIT_MS" \
-#   --viewport "$AGC_BROWSER_VIEWPORT" --headless "$AGC_BROWSER_HEADLESS"
-
-# === Backend C: chrome-devtools MCP ========================================
-# The agent should call the MCP tool directly rather than shelling out.
-# If routed through this script, print the hint so the agent knows to switch:
-# agc::log "use mcp__chrome_devtools__take_screenshot instead of this shell"
-
-# --------------------------------------------------------------------------
-# Fallback: stub mode
-# --------------------------------------------------------------------------
-if [[ ! -s "$OUT" ]]; then
-  cat > "${OUT%.png}.stub.txt" <<EOF
-# TODO(auto-gh): browser-verify backend not wired up yet.
-# Requested: $URL
-# Would write screenshot to: $OUT
-# Selector: ${SELECTOR:-<full page>}
-# Wait: ${WAIT_MS}ms
-# Viewport: ${AGC_BROWSER_VIEWPORT}
-EOF
-  agc::log "stub: browser backend not configured (see ${OUT%.png}.stub.txt)"
-  printf 'SCREENSHOT=%s\n' "${OUT%.png}.stub.txt"
-  exit 0
-fi
-
-printf 'SCREENSHOT=%s\n' "$OUT"
diff --git a/skills/auto-github-contributor/scripts/check-prereqs.sh b/skills/auto-github-contributor/scripts/check-prereqs.sh
deleted file mode 100755
index 4f2b284..0000000
--- a/skills/auto-github-contributor/scripts/check-prereqs.sh
+++ /dev/null
@@ -1,74 +0,0 @@
-#!/usr/bin/env bash
-# Verify required CLI tools + gh authentication before the agent starts work.
-# Exits 0 when everything is ready.
-# Exits 2 + prints a structured hint when something is missing — the skill
-# surfaces the hint to the user verbatim and stops.
-#
-# Usage: check-prereqs.sh
-
-set -uo pipefail
-# Deliberately not -e: we want to continue checking all tools even when one fails.
-
-# shellcheck disable=SC1091
-source "$(dirname "$0")/config.sh"
-
-STATUS=0
-MISSING=()
-HINTS=()
-
-check_bin() {
-  local bin="$1" install_hint="$2"
-  if command -v "$bin" >/dev/null 2>&1; then
-    printf '  ✓ %s\n' "$bin" >&2
-  else
-    printf '  ✗ %s (not installed)\n' "$bin" >&2
-    MISSING+=("$bin")
-    HINTS+=("$install_hint")
-    STATUS=2
-  fi
-}
-
-printf '[auto-gh] checking prerequisites...\n' >&2
-
-OS="$(uname -s)"
-case "$OS" in
-  Darwin) GH_HINT="brew install gh" ;;
-  Linux)  GH_HINT="see https://github.com/cli/cli#installation (e.g. 'sudo apt install gh' or 'brew install gh')" ;;
-  *)      GH_HINT="see https://github.com/cli/cli#installation" ;;
-esac
-
-check_bin gh   "$GH_HINT"
-check_bin git  "install git for your OS"
-check_bin jq   "$( [[ $OS == Darwin ]] && echo 'brew install jq' || echo 'sudo apt install jq  (or brew install jq)' )"
-
-if ((${#MISSING[@]} > 0)); then
-  printf '\n[auto-gh][error] missing required tools: %s\n' "${MISSING[*]}" >&2
-  printf '\nInstall hints:\n' >&2
-  for i in "${!MISSING[@]}"; do
-    printf '  - %s: %s\n' "${MISSING[$i]}" "${HINTS[$i]}" >&2
-  done
-  exit 2
-fi
-
-# gh auth
-if ! gh auth status >/dev/null 2>&1; then
-  cat >&2 <<'EOF'
-
-[auto-gh][error] gh is installed but not authenticated.
-
-Run this in a terminal, then retry:
-
-  gh auth login
-
-Pick: GitHub.com → HTTPS → authenticate via browser (recommended) or token.
-You need at least `repo` scope to open pull requests.
-EOF
-  exit 2
-fi
-
-GH_USER="$(gh api user --jq .login 2>/dev/null || echo '?')"
-printf '  ✓ gh authed as %s\n' "$GH_USER" >&2
-
-# Surface the authed user so the skill can default TARGET_FORK.
-printf 'GH_USER=%s\n' "$GH_USER"
-printf 'READY=1\n'
diff --git a/skills/auto-github-contributor/scripts/config.sh b/skills/auto-github-contributor/scripts/config.sh
deleted file mode 100755
index eaf85d8..0000000
--- a/skills/auto-github-contributor/scripts/config.sh
+++ /dev/null
@@ -1,88 +0,0 @@
-#!/usr/bin/env bash
-# Shared config for the auto-github-contributor skill.
-# Override any value by exporting env vars before invoking a script.
-#
-# REQUIRED at runtime (no default — the skill prompts the user):
-#   TARGET_REPO   "<owner>/<name>"  e.g. refly-ai/refly
-#
-# OPTIONAL:
-#   TARGET_FORK   "<owner>/<name>"  push branches here (else upstream)
-#   AGC_BASE_BRANCH                 default: main
-#   AGC_WORK_ROOT                   default: $HOME/auto-gh-contrib-work
-#   AGC_LABELS                      comma-separated labels for issue search
-#                                   default: "good first issue,help wanted,documentation,good-first-issue"
-#   AGC_ISSUE_LIMIT                 default: 30
-
-set -euo pipefail
-
-: "${TARGET_REPO:=}"
-: "${TARGET_FORK:=}"
-
-: "${AGC_BASE_BRANCH:=main}"
-: "${AGC_WORK_ROOT:="$HOME/auto-gh-contrib-work"}"
-
-: "${AGC_LABELS:=good first issue,help wanted,documentation,good-first-issue}"
-: "${AGC_ISSUE_LIMIT:=30}"
-
-# Dev-loop commands. Auto-detected per-repo when possible; these are fallbacks.
-: "${AGC_INSTALL_CMD:=pnpm install --frozen-lockfile}"
-: "${AGC_LINT_CMD:=pnpm lint}"
-: "${AGC_TYPECHECK_CMD:=pnpm typecheck}"
-: "${AGC_TEST_CMD:=pnpm test}"
-: "${AGC_BUILD_CMD:=pnpm build}"
-: "${AGC_DEV_URL:=http://localhost:5173}"
-
-# Browser verification defaults.
-: "${AGC_SCREENSHOT_DIR:=.auto-pr/screenshots}"
-: "${AGC_BROWSER_HEADLESS:=1}"
-: "${AGC_BROWSER_VIEWPORT:=1440x900}"
-
-export TARGET_REPO TARGET_FORK
-export AGC_BASE_BRANCH AGC_WORK_ROOT AGC_LABELS AGC_ISSUE_LIMIT
-export AGC_INSTALL_CMD AGC_LINT_CMD AGC_TYPECHECK_CMD AGC_TEST_CMD AGC_BUILD_CMD AGC_DEV_URL
-export AGC_SCREENSHOT_DIR AGC_BROWSER_HEADLESS AGC_BROWSER_VIEWPORT
-
-agc::log() { printf '[auto-gh] %s\n' "$*" >&2; }
-agc::err() { printf '[auto-gh][error] %s\n' "$*" >&2; }
-agc::die() { agc::err "$*"; exit 1; }
-
-agc::require() {
-  local bin="$1"
-  command -v "$bin" >/dev/null 2>&1 || agc::die "missing dependency: $bin"
-}
-
-agc::require_repo() {
-  [[ -n "$TARGET_REPO" ]] || agc::die "TARGET_REPO is empty. Export TARGET_REPO=<owner>/<name> before invoking."
-  case "$TARGET_REPO" in
-    */*) ;;
-    *) agc::die "TARGET_REPO must be in <owner>/<name> form (got: $TARGET_REPO)" ;;
-  esac
-}
-
-agc::repo_slug() {
-  printf '%s' "$TARGET_REPO" | tr '/' '-'
-}
-
-agc::workdir_for() {
-  # $1 = issue number OR a slug for non-issue work
-  printf '%s/%s/%s\n' "$AGC_WORK_ROOT" "$(agc::repo_slug)" "$1"
-}
-
-agc::slugify() {
-  local s="${1:-}"
-  s="$(printf '%s' "$s" | tr '[:upper:]' '[:lower:]')"
-  s="$(printf '%s' "$s" | sed -E 's/[^a-z0-9]+/-/g; s/^-+//; s/-+$//')"
-  printf '%s' "${s:0:40}"
-}
-
-# Normalize "https://github.com/owner/repo[.git]" or "owner/repo" → "owner/repo".
-agc::normalize_repo() {
-  local input="$1"
-  input="${input%/}"
-  input="${input%.git}"
-  case "$input" in
-    https://github.com/*) input="${input#https://github.com/}" ;;
-    git@github.com:*)     input="${input#git@github.com:}" ;;
-  esac
-  printf '%s' "$input"
-}
diff --git a/skills/auto-github-contributor/scripts/create-pr.sh b/skills/auto-github-contributor/scripts/create-pr.sh
deleted file mode 100755
index 0c81173..0000000
--- a/skills/auto-github-contributor/scripts/create-pr.sh
+++ /dev/null
@@ -1,154 +0,0 @@
-#!/usr/bin/env bash
-# Commit pending changes, push the branch, and open a PR via gh.
-# Usage: create-pr.sh <issue-number-or-slug> [--workdir <dir>] [--draft] [--title "..."]
-# Env:   TARGET_REPO required, TARGET_FORK optional.
-#
-# Reads:
-#   <workdir>/.auto-pr/SPEC.md
-#   <workdir>/.auto-pr/TODO.md
-#   <workdir>/.auto-pr/screenshots/*.png  (optional)
-#   <workdir>/.auto-pr/BLOCKERS.md        (optional)
-#   <workdir>/.auto-pr/issue.json         (optional — only if started from an issue)
-# Writes:
-#   <workdir>/.auto-pr/PR-BODY.md  (rendered)
-# Prints PR URL on success.
-
-set -euo pipefail
-source "$(dirname "$0")/config.sh"
-
-KEY="${1:?issue number or slug required}"
-shift || true
-
-WORKDIR=""
-DRAFT=""
-TITLE_OVERRIDE=""
-while (($#)); do
-  case "$1" in
-    --workdir) WORKDIR="$2"; shift 2 ;;
-    --draft) DRAFT="--draft"; shift ;;
-    --title) TITLE_OVERRIDE="$2"; shift 2 ;;
-    *) agc::die "unknown flag: $1" ;;
-  esac
-done
-
-agc::require_repo
-agc::require gh
-agc::require git
-
-WORKDIR="${WORKDIR:-$(agc::workdir_for "$KEY")}"
-[[ -d "$WORKDIR/.git" ]] || agc::die "not a git workdir: $WORKDIR"
-
-cd "$WORKDIR"
-BRANCH="$(git rev-parse --abbrev-ref HEAD)"
-
-# Safety: never commit/push directly to base branch.
-case "$BRANCH" in
-  main|master|develop) agc::die "refusing to push base branch '$BRANCH'" ;;
-esac
-
-ISSUE_JSON_PATH="$WORKDIR/.auto-pr/issue.json"
-ISSUE_NUMBER=""
-ISSUE_TITLE=""
-ISSUE_URL=""
-if [[ -f "$ISSUE_JSON_PATH" ]]; then
-  ISSUE_NUMBER="$(jq -r '.number' "$ISSUE_JSON_PATH")"
-  ISSUE_TITLE="$(jq -r '.title' "$ISSUE_JSON_PATH")"
-  ISSUE_URL="$(jq -r '.url' "$ISSUE_JSON_PATH")"
-fi
-
-PR_TITLE=""
-COMMIT_PREFIX=""
-if [[ -n "$TITLE_OVERRIDE" ]]; then
-  PR_TITLE="$TITLE_OVERRIDE"
-  COMMIT_PREFIX="chore"
-elif [[ -n "$ISSUE_NUMBER" ]]; then
-  PR_TITLE="fix(#${ISSUE_NUMBER}): ${ISSUE_TITLE}"
-  COMMIT_PREFIX="fix"
-else
-  PR_TITLE="chore: ${KEY}"
-  COMMIT_PREFIX="chore"
-fi
-
-# 1) Stage + commit if there are changes.
-if ! git diff --quiet || ! git diff --cached --quiet; then
-  git add -A
-  if [[ -n "$ISSUE_NUMBER" ]]; then
-    COMMIT_MSG="${COMMIT_PREFIX}(#${ISSUE_NUMBER}): ${ISSUE_TITLE}
-
-Fixes #${ISSUE_NUMBER}
-
-Ref: ${ISSUE_URL}
-"
-  else
-    COMMIT_MSG="${PR_TITLE}
-"
-  fi
-  git commit -m "$COMMIT_MSG"
-  agc::log "created commit"
-else
-  agc::log "no uncommitted changes — assuming work was already committed"
-fi
-
-# 2) Decide push remote. Prefer fork when configured.
-PUSH_REMOTE="origin"
-if [[ -n "${TARGET_FORK}" ]] && git remote | grep -q '^fork$'; then
-  PUSH_REMOTE="fork"
-else
-  agc::log "no fork configured (TARGET_FORK empty) — pushing to origin (${TARGET_REPO}). Ctrl+C within 3s to abort."
-  sleep 3 || true
-fi
-agc::log "pushing to ${PUSH_REMOTE}/${BRANCH}"
-git push -u "$PUSH_REMOTE" "$BRANCH"
-
-# 3) Render PR body.
-TEMPLATE_DIR="$(cd "$(dirname "$0")/../templates" && pwd)"
-PR_BODY="$WORKDIR/.auto-pr/PR-BODY.md"
-
-SPEC_SNIPPET="$(cat "$WORKDIR/.auto-pr/SPEC.md" 2>/dev/null || echo '_spec missing_')"
-TODO_SNIPPET="$(cat "$WORKDIR/.auto-pr/TODO.md" 2>/dev/null || echo '_todo missing_')"
-SHOT_LIST=""
-for f in "$WORKDIR"/.auto-pr/screenshots/*.png; do
-  [[ -e "$f" ]] || continue
-  SHOT_LIST+=$'\n'"- \`$(basename "$f")\`"
-done
-[[ -z "$SHOT_LIST" ]] && SHOT_LIST=$'\n'"_No screenshots captured (stub or non-UI change)._"
-
-BLOCKERS_SNIPPET=""
-if [[ -f "$WORKDIR/.auto-pr/BLOCKERS.md" ]]; then
-  BLOCKERS_SNIPPET="$(cat "$WORKDIR/.auto-pr/BLOCKERS.md")"
-fi
-
-export ISSUE_NUMBER ISSUE_TITLE ISSUE_URL SPEC_SNIPPET TODO_SNIPPET SHOT_LIST BLOCKERS_SNIPPET
-
-python3 - "$TEMPLATE_DIR/PR-BODY.template.md" "$PR_BODY" <<'PY'
-import os, pathlib, sys
-src, dst = sys.argv[1], sys.argv[2]
-body = pathlib.Path(src).read_text()
-body = (body
-  .replace("{{ISSUE_NUMBER}}", os.environ.get("ISSUE_NUMBER", ""))
-  .replace("{{ISSUE_TITLE}}", os.environ.get("ISSUE_TITLE", ""))
-  .replace("{{ISSUE_URL}}", os.environ.get("ISSUE_URL", ""))
-  .replace("{{SPEC}}", os.environ.get("SPEC_SNIPPET", ""))
-  .replace("{{TODO}}", os.environ.get("TODO_SNIPPET", ""))
-  .replace("{{SCREENSHOTS}}", os.environ.get("SHOT_LIST", ""))
-  .replace("{{BLOCKERS}}", os.environ.get("BLOCKERS_SNIPPET", "")))
-pathlib.Path(dst).write_text(body)
-PY
-
-# 4) Create PR.
-HEAD_REF="$BRANCH"
-if [[ "$PUSH_REMOTE" == "fork" && -n "${TARGET_FORK}" ]]; then
-  HEAD_REF="${TARGET_FORK%%/*}:${BRANCH}"
-fi
-
-agc::log "opening PR against ${TARGET_REPO}:${AGC_BASE_BRANCH} from ${HEAD_REF}"
-PR_URL="$(gh pr create \
-  -R "$TARGET_REPO" \
-  --base "$AGC_BASE_BRANCH" \
-  --head "$HEAD_REF" \
-  --title "$PR_TITLE" \
-  --body-file "$PR_BODY" \
-  ${DRAFT} 2>&1 | tee /dev/stderr | tail -n 1)"
-
-agc::log "PR created: $PR_URL"
-printf 'PR_URL=%s\n' "$PR_URL"
diff --git a/skills/auto-github-contributor/scripts/dev-loop-check.sh b/skills/auto-github-contributor/scripts/dev-loop-check.sh
deleted file mode 100755
index 00ed63a..0000000
--- a/skills/auto-github-contributor/scripts/dev-loop-check.sh
+++ /dev/null
@@ -1,113 +0,0 @@
-#!/usr/bin/env bash
-# Phase-aware dev-loop checker. Invoked from the repo workdir.
-# Phases:
-#   --phase red    → expect the newest test(s) to FAIL (proves coverage gap)
-#   --phase green  → install + lint + typecheck + test (must all pass)
-#   --phase final  → clean install + lint + typecheck + full tests + build
-#
-# Usage:
-#   dev-loop-check.sh --phase green [--workdir <dir>]
-
-set -euo pipefail
-source "$(dirname "$0")/config.sh"
-
-PHASE=""
-WORKDIR="$(pwd)"
-
-while (($#)); do
-  case "$1" in
-    --phase) PHASE="$2"; shift 2 ;;
-    --workdir) WORKDIR="$2"; shift 2 ;;
-    *) agc::die "unknown flag: $1" ;;
-  esac
-done
-
-[[ -n "$PHASE" ]] || agc::die "--phase required (red|green|final)"
-[[ -d "$WORKDIR/.git" ]] || agc::die "not a git workdir: $WORKDIR"
-
-cd "$WORKDIR"
-
-run() {
-  local label="$1"; shift
-  agc::log "▶ ${label}: $*"
-  if "$@"; then
-    agc::log "✓ ${label}"
-    return 0
-  else
-    local rc=$?
-    agc::err "✗ ${label} (exit ${rc})"
-    return $rc
-  fi
-}
-
-# Pick the right package manager command based on lockfile.
-detect_install_cmd() {
-  if [[ -f pnpm-lock.yaml ]]; then echo "pnpm install --frozen-lockfile";
-  elif [[ -f yarn.lock ]]; then echo "yarn install --frozen-lockfile";
-  elif [[ -f package-lock.json ]]; then echo "npm ci";
-  elif [[ -f bun.lockb ]]; then echo "bun install --frozen-lockfile";
-  else echo "${AGC_INSTALL_CMD}";
-  fi
-}
-
-# Pick the JS package manager runner based on lockfile.
-detect_runner() {
-  if [[ -f pnpm-lock.yaml ]]; then echo "pnpm run";
-  elif [[ -f yarn.lock ]]; then echo "yarn";
-  elif [[ -f bun.lockb ]]; then echo "bun run";
-  else echo "npm run";
-  fi
-}
-
-detect_script() {
-  # Check if package.json defines a given script; echo cmd if present, else fallback.
-  local name="$1" fallback="$2"
-  if [[ -f package.json ]] && jq -e --arg n "$name" '.scripts[$n] // empty' package.json >/dev/null 2>&1; then
-    printf '%s %s' "$(detect_runner)" "$name"
-  else
-    echo "$fallback"
-  fi
-}
-
-INSTALL_CMD="$(detect_install_cmd)"
-LINT_CMD="$(detect_script lint "$AGC_LINT_CMD")"
-TYPECHECK_CMD="$(detect_script typecheck "$AGC_TYPECHECK_CMD")"
-TEST_CMD="$(detect_script test "$AGC_TEST_CMD")"
-BUILD_CMD="$(detect_script build "$AGC_BUILD_CMD")"
-
-case "$PHASE" in
-  red)
-    # Red: run only the test suite. Expect non-zero (failing new test) — we
-    # invert the exit code so a "correctly failing" red phase reports success.
-    agc::log "red phase: expecting test failure"
-    if eval "$TEST_CMD"; then
-      agc::err "tests passed in red phase — you have not written a failing test yet"
-      exit 1
-    else
-      agc::log "tests failed as expected (red phase OK)"
-      exit 0
-    fi
-    ;;
-
-  green)
-    run "install" bash -lc "$INSTALL_CMD"
-    run "lint" bash -lc "$LINT_CMD"
-    run "typecheck" bash -lc "$TYPECHECK_CMD"
-    run "test" bash -lc "$TEST_CMD"
-    ;;
-
-  final)
-    run "install (clean)" bash -lc "$INSTALL_CMD"
-    run "lint" bash -lc "$LINT_CMD"
-    run "typecheck" bash -lc "$TYPECHECK_CMD"
-    run "test" bash -lc "$TEST_CMD"
-    run "build" bash -lc "$BUILD_CMD"
-    # TODO(auto-gh): repo-specific `verify` / e2e hook. Wire a shell command here
-    # once the target repo ships a stable e2e target (e.g. `pnpm test:e2e`).
-    agc::log "final verification complete"
-    ;;
-
-  *)
-    agc::die "invalid phase: $PHASE (expected red|green|final)"
-    ;;
-esac
diff --git a/skills/auto-github-contributor/scripts/fetch-issues.sh b/skills/auto-github-contributor/scripts/fetch-issues.sh
deleted file mode 100755
index bcb0b8f..0000000
--- a/skills/auto-github-contributor/scripts/fetch-issues.sh
+++ /dev/null
@@ -1,89 +0,0 @@
-#!/usr/bin/env bash
-# Fetch open beginner-friendly issues from the target repo.
-# Queries each label in $AGC_LABELS (comma-separated) separately, merges, de-dupes,
-# ranks by a heuristic score (label match strength + freshness + body length),
-# and emits a JSON array on stdout.
-#
-# Env: TARGET_REPO required. AGC_LABELS + AGC_ISSUE_LIMIT optional.
-
-set -euo pipefail
-source "$(dirname "$0")/config.sh"
-
-agc::require gh
-agc::require jq
-agc::require_repo
-
-if ! gh auth status >/dev/null 2>&1; then
-  agc::err "gh is not authenticated. Run: gh auth login"
-  exit 2
-fi
-
-IFS=',' read -r -a LABELS <<<"$AGC_LABELS"
-
-agc::log "searching ${TARGET_REPO} for labels: ${AGC_LABELS}"
-
-TMP="$(mktemp)"
-trap 'rm -f "$TMP"' EXIT
-
-printf '[]' > "$TMP"
-
-for raw_label in "${LABELS[@]}"; do
-  label="${raw_label#"${raw_label%%[![:space:]]*}"}"
-  label="${label%"${label##*[![:space:]]}"}"
-  [[ -z "$label" ]] && continue
-
-  agc::log "  → label='${label}'"
-
-  # Per-label fetch. Some repos won't have a given label — ignore those errors.
-  PARTIAL="$(gh issue list \
-    -R "$TARGET_REPO" \
-    --label "$label" \
-    --state open \
-    --limit "$AGC_ISSUE_LIMIT" \
-    --search "sort:updated-desc no:assignee" \
-    --json number,title,url,labels,updatedAt,body 2>/dev/null || echo '[]')"
-
-  jq --argjson add "$PARTIAL" '. + $add' "$TMP" > "$TMP.next" && mv "$TMP.next" "$TMP"
-done
-
-# Quick-label weights (more specific beginner-friendly labels score higher).
-jq --arg limit "$AGC_ISSUE_LIMIT" '
-  def label_score(ls):
-    (ls | map(ascii_downcase) | map(
-      if . == "good first issue" or . == "good-first-issue" then 3
-      elif . == "help wanted" or . == "help-wanted" then 2
-      elif . == "documentation" or . == "docs" then 2
-      elif . == "typo" then 3
-      elif . == "i18n" or . == "l10n" or . == "translation" then 2
-      elif . == "testing" or . == "tests" then 2
-      else 1 end
-    ) | max // 0);
-
-  def freshness(updatedAt):
-    ((now - (updatedAt | fromdateiso8601)) / 86400) as $days
-    | if $days < 7 then 3
-      elif $days < 30 then 2
-      elif $days < 90 then 1
-      else 0 end;
-
-  def body_clarity(body):
-    ((body // "") | length) as $len
-    | if $len > 200 and $len < 2000 then 2
-      elif $len >= 2000 then 1
-      elif $len > 50 then 1
-      else 0 end;
-
-  # De-dupe by number; keep first occurrence.
-  unique_by(.number)
-  | map({
-      number,
-      title,
-      url,
-      updatedAt,
-      labels: [.labels[].name],
-      excerpt: ((.body // "") | gsub("\r"; "") | split("\n") | map(select(length > 0)) | .[0:3] | join(" ") | .[0:240]),
-      score: (label_score([.labels[].name]) + freshness(.updatedAt) + body_clarity(.body))
-    })
-  | sort_by(-.score, .updatedAt)
-  | .[0:($limit | tonumber)]
-' "$TMP"
diff --git a/skills/auto-github-contributor/scripts/scan-quick-wins.sh b/skills/auto-github-contributor/scripts/scan-quick-wins.sh
deleted file mode 100755
index 895673e..0000000
--- a/skills/auto-github-contributor/scripts/scan-quick-wins.sh
+++ /dev/null
@@ -1,250 +0,0 @@
-#!/usr/bin/env bash
-# Scan a checked-out repo for quick-win contribution candidates.
-# Emits a JSON array on stdout. Each entry:
-#   { kind, title, summary, file, line, estimated_minutes, slug }
-#
-# Kinds covered:
-#   typo        — common English misspellings in docs / comments
-#   missing-test — source files with no matching test file
-#   i18n        — locale JSON files missing keys present in the reference locale
-#   todo        — TODO/FIXME/XXX comments (actionable picks only)
-#
-# Usage: scan-quick-wins.sh [--workdir <path>] [--max <n>]
-# Notes:
-#   - Read-only. Never modifies the repo.
-#   - Uses ripgrep when available, falls back to grep.
-
-# shellcheck disable=SC1091
-source "$(dirname "$0")/config.sh"
-# Deliberately loosen after sourcing config.sh (which sets -e). This is a
-# best-effort scanner: individual checks returning empty (e.g. `grep -Ev` with
-# no matches) must not abort the run.
-set +e
-set -uo pipefail
-
-WORKDIR="$(pwd)"
-MAX=40
-while (($#)); do
-  case "$1" in
-    --workdir) WORKDIR="$2"; shift 2 ;;
-    --max) MAX="$2"; shift 2 ;;
-    *) agc::die "unknown flag: $1" ;;
-  esac
-done
-
-[[ -d "$WORKDIR/.git" ]] || agc::die "not a git workdir: $WORKDIR"
-cd "$WORKDIR"
-
-if command -v rg >/dev/null 2>&1; then
-  BACKEND="rg"
-  SEARCH() {
-    rg -n --no-heading --hidden \
-      --glob '!node_modules' --glob '!dist' --glob '!build' --glob '!.git' --glob '!.next' \
-      "$@"
-  }
-else
-  BACKEND="grep"
-  SEARCH() {
-    grep -rEn --binary-files=without-match \
-      --exclude-dir=node_modules --exclude-dir=dist --exclude-dir=build --exclude-dir=.git \
-      "$@"
-  }
-fi
-
-# Backend-specific file-type filter for the typo scan.
-# rg uses -t <name>; grep uses --include='*.ext'.
-if [[ "$BACKEND" == "rg" ]]; then
-  # rg covers .jsx/.tsx via 'js'/'ts'; rust/ruby are the long names (not rs/rb).
-  TYPO_FILE_FILTER=(-t md -t txt -t rst -t js -t ts -t py -t go -t rust -t java -t ruby)
-else
-  TYPO_FILE_FILTER=(
-    --include='*.md' --include='*.mdx' --include='*.txt' --include='*.rst'
-    --include='*.js' --include='*.jsx' --include='*.mjs' --include='*.cjs'
-    --include='*.ts' --include='*.tsx' --include='*.cts' --include='*.mts'
-    --include='*.py' --include='*.go' --include='*.rs' --include='*.java' --include='*.rb'
-  )
-fi
-
-TMP="$(mktemp -d)"
-trap 'rm -rf "$TMP"' EXIT
-
-# --- 1) Typo scan --------------------------------------------------------
-# Curated list: common → correct. Case-insensitive regex; we only report the
-# first match per file to keep noise down.
-TYPO_MAP=(
-  "teh|the"
-  "recieve|receive"
-  "seperate|separate"
-  "occured|occurred"
-  "occurence|occurrence"
-  "definately|definitely"
-  "wierd|weird"
-  "accomodate|accommodate"
-  "untill|until"
-  "refered|referred"
-  "acheive|achieve"
-  "arguement|argument"
-  "begining|beginning"
-  "calender|calendar"
-  "commited|committed"
-  "enviroment|environment"
-  "existance|existence"
-  "familliar|familiar"
-  "foward|forward"
-  "independant|independent"
-  "occassion|occasion"
-  "persistant|persistent"
-  "priviledge|privilege"
-  "publically|publicly"
-  "reccomend|recommend"
-  "succesful|successful"
-  "sucessful|successful"
-  "tommorow|tomorrow"
-  "writting|writing"
-  "lenght|length"
-  "heigth|height"
-  "widht|width"
-  "wether|whether"
-  "alot|a lot"
-  "intial|initial"
-  "paramter|parameter"
-  "paramaters|parameters"
-  "parmeter|parameter"
-  "responsability|responsibility"
-  "transfered|transferred"
-  "proccess|process"
-)
-
-echo "[]" > "$TMP/typo.json"
-for entry in "${TYPO_MAP[@]}"; do
-  bad="${entry%%|*}"
-  good="${entry#*|}"
-  # Word-boundary regex, case-insensitive. Scan text-ish files.
-  matches="$(SEARCH -i -w "${TYPO_FILE_FILTER[@]}" "\b${bad}\b" 2>/dev/null | head -n 2 || true)"
-  [[ -z "$matches" ]] && continue
-  while IFS=: read -r file line _rest; do
-    [[ -z "$file" ]] && continue
-    jq --arg bad "$bad" --arg good "$good" --arg file "$file" --arg line "$line" \
-      '. + [{
-        kind: "typo",
-        title: "Typo: \($bad) → \($good) in \($file)",
-        summary: "Fix spelling \"\($bad)\" → \"\($good)\" at \($file):\($line).",
-        file: $file,
-        line: ($line | tonumber? // 0),
-        estimated_minutes: 5,
-        slug: "typo-\($bad)-\($file | gsub("[^A-Za-z0-9]+"; "-"))"
-      }]' "$TMP/typo.json" > "$TMP/typo.next" && mv "$TMP/typo.next" "$TMP/typo.json"
-  done <<<"$matches"
-done
-
-# --- 2) Missing-test scan (JS/TS) ----------------------------------------
-# For each src/**/*.{ts,tsx,js,jsx} that looks like a module (not a test and not an index re-export),
-# check whether a sibling test file exists (same basename + .test.* / .spec.*).
-echo "[]" > "$TMP/missing.json"
-if [[ -d src ]] || [[ -d packages ]] || [[ -d lib ]]; then
-  CANDIDATES="$(
-    {
-      find src packages lib -type f \( -name '*.ts' -o -name '*.tsx' -o -name '*.js' -o -name '*.jsx' \) 2>/dev/null || true
-    } | grep -Ev '(\.test\.|\.spec\.|__tests__|__mocks__|\.d\.ts$|node_modules|/dist/|/build/)' | head -n 200
-  )"
-  count=0
-  while IFS= read -r src_file; do
-    [[ -z "$src_file" ]] && continue
-    ((count > 6)) && break
-    dir="$(dirname "$src_file")"
-    base="$(basename "$src_file")"
-    stem="${base%.*}"
-    ext="${base##*.}"
-    # Skip trivial re-exports (file is just `export * from ...`).
-    if [[ "$(wc -l < "$src_file" 2>/dev/null || echo 0)" -lt 10 ]]; then
-      continue
-    fi
-    # Only flag files that export something.
-    grep -qE '^(export |module\.exports|exports\.)' "$src_file" 2>/dev/null || continue
-    # Look for a matching test file.
-    if ! find "$dir" -maxdepth 2 -type f \( -name "${stem}.test.${ext}" -o -name "${stem}.spec.${ext}" \) 2>/dev/null | grep -q . ; then
-      if ! find "$dir" -maxdepth 2 -type d -name '__tests__' -exec find {} -name "${stem}.*" \; 2>/dev/null | grep -q . ; then
-        jq --arg file "$src_file" --arg stem "$stem" \
-          '. + [{
-            kind: "missing-test",
-            title: "Add unit tests for \($file)",
-            summary: "Exported module \($file) has no co-located test file. Add a minimal test covering its public API.",
-            file: $file,
-            line: 1,
-            estimated_minutes: 30,
-            slug: "add-tests-\($stem | gsub("[^A-Za-z0-9]+"; "-"))"
-          }]' "$TMP/missing.json" > "$TMP/missing.next" && mv "$TMP/missing.next" "$TMP/missing.json"
-        ((count++)) || true
-      fi
-    fi
-  done <<<"$CANDIDATES"
-fi
-
-# --- 3) i18n gap scan ----------------------------------------------------
-# Heuristic: find folders named "locales", "lang", "i18n" or "translations"
-# containing multiple JSON files; flag any file that has fewer top-level keys
-# than the largest sibling (likely missing translations).
-echo "[]" > "$TMP/i18n.json"
-LOCALE_DIRS="$(find . -type d \( -name 'locales' -o -name 'locale' -o -name 'lang' -o -name 'i18n' -o -name 'translations' \) 2>/dev/null | grep -Ev 'node_modules|/dist/|/build/' | head -n 10)"
-while IFS= read -r dir; do
-  [[ -z "$dir" ]] && continue
-  JSONS="$(find "$dir" -maxdepth 2 -type f -name '*.json' 2>/dev/null | head -n 20)"
-  [[ -z "$JSONS" ]] && continue
-  # Find the file with the most keys (reference locale).
-  ref=""; ref_count=0
-  while IFS= read -r f; do
-    [[ -z "$f" ]] && continue
-    n="$(jq 'paths(scalars) | length' "$f" 2>/dev/null | wc -l | tr -d ' ')"
-    if (( n > ref_count )); then
-      ref="$f"; ref_count="$n"
-    fi
-  done <<<"$JSONS"
-  [[ -z "$ref" ]] && continue
-  while IFS= read -r f; do
-    [[ -z "$f" ]] && continue
-    [[ "$f" == "$ref" ]] && continue
-    n="$(jq 'paths(scalars) | length' "$f" 2>/dev/null | wc -l | tr -d ' ')"
-    if (( n > 0 )) && (( n < ref_count * 9 / 10 )); then
-      missing=$((ref_count - n))
-      jq --arg file "$f" --arg ref "$ref" --arg missing "$missing" \
-        '. + [{
-          kind: "i18n",
-          title: "Fill missing i18n keys in \($file)",
-          summary: "Locale \($file) appears to be missing ~\($missing) keys present in \($ref).",
-          file: $file,
-          line: 1,
-          estimated_minutes: 45,
-          slug: "i18n-\($file | gsub("[^A-Za-z0-9]+"; "-"))"
-        }]' "$TMP/i18n.json" > "$TMP/i18n.next" && mv "$TMP/i18n.next" "$TMP/i18n.json"
-    fi
-  done <<<"$JSONS"
-done <<<"$LOCALE_DIRS"
-
-# --- 4) TODO/FIXME scan --------------------------------------------------
-echo "[]" > "$TMP/todo.json"
-# Match only TODO/FIXME that look actionable (have a hint of what to do).
-# Pattern: TODO|FIXME|XXX followed by a non-alphanum then 15+ chars on the line.
-TODOS="$(SEARCH -e '(TODO|FIXME|XXX)[^A-Za-z0-9].{15,200}' 2>/dev/null | head -n 15 || true)"
-while IFS=: read -r file line rest; do
-  [[ -z "$file" ]] && continue
-  # Clean up the excerpt.
-  excerpt="$(printf '%s' "$rest" | sed -E 's/^[[:space:]]+//' | cut -c1-160)"
-  jq --arg file "$file" --arg line "$line" --arg excerpt "$excerpt" \
-    '. + [{
-      kind: "todo",
-      title: "Resolve TODO in \($file):\($line)",
-      summary: $excerpt,
-      file: $file,
-      line: ($line | tonumber? // 0),
-      estimated_minutes: 60,
-      slug: "todo-\($file | gsub("[^A-Za-z0-9]+"; "-"))-\($line)"
-    }]' "$TMP/todo.json" > "$TMP/todo.next" && mv "$TMP/todo.next" "$TMP/todo.json"
-done <<<"$TODOS"
-
-# --- Merge + cap ---------------------------------------------------------
-jq --arg max "$MAX" -s '
-  (.[0] + .[1] + .[2] + .[3])
-  | unique_by(.slug)
-  | sort_by(.estimated_minutes, .kind)
-  | .[0:($max | tonumber)]
-' "$TMP/typo.json" "$TMP/missing.json" "$TMP/i18n.json" "$TMP/todo.json"
diff --git a/skills/auto-github-contributor/scripts/setup-workspace.sh b/skills/auto-github-contributor/scripts/setup-workspace.sh
deleted file mode 100755
index 68bca7f..0000000
--- a/skills/auto-github-contributor/scripts/setup-workspace.sh
+++ /dev/null
@@ -1,96 +0,0 @@
-#!/usr/bin/env bash
-# Clone (or reuse) the target repo in an isolated workdir and create a feature branch.
-# Usage: setup-workspace.sh <issue-number-or-slug> [--title "<branch title>"]
-# Env:   TARGET_REPO required, TARGET_FORK optional.
-# Prints (machine-readable):
-#   WORKDIR=<abs path>
-#   BRANCH=<branch name>
-#   ISSUE_TITLE=<string>   (only when an issue number was given)
-
-set -euo pipefail
-source "$(dirname "$0")/config.sh"
-
-KEY="${1:?issue number or slug required}"
-shift || true
-
-TITLE_OVERRIDE=""
-while (($#)); do
-  case "$1" in
-    --title) TITLE_OVERRIDE="$2"; shift 2 ;;
-    *) agc::die "unknown flag: $1" ;;
-  esac
-done
-
-agc::require_repo
-agc::require gh
-agc::require git
-agc::require jq
-
-WORKDIR="$(agc::workdir_for "$KEY")"
-mkdir -p "$AGC_WORK_ROOT"
-
-# Safety: never operate outside $AGC_WORK_ROOT.
-case "$WORKDIR" in
-  "$AGC_WORK_ROOT"/*) ;;
-  *) agc::die "refusing to operate on path outside AGC_WORK_ROOT: $WORKDIR" ;;
-esac
-
-ISSUE_JSON=""
-ISSUE_TITLE=""
-BRANCH_PREFIX="fix"
-if [[ "$KEY" =~ ^[0-9]+$ ]]; then
-  agc::log "fetching issue metadata for #$KEY"
-  ISSUE_JSON="$(gh issue view "$KEY" -R "$TARGET_REPO" --json number,title,body,labels,url)"
-  ISSUE_TITLE="$(printf '%s' "$ISSUE_JSON" | jq -r '.title')"
-  SLUG="$(agc::slugify "$ISSUE_TITLE")"
-  BRANCH="${BRANCH_PREFIX}/issue-${KEY}-${SLUG}"
-else
-  # Quick-win mode: KEY is already a slug (e.g. "typo-readme").
-  ISSUE_TITLE="${TITLE_OVERRIDE:-$KEY}"
-  SLUG="$(agc::slugify "$KEY")"
-  BRANCH_PREFIX="chore"
-  BRANCH="${BRANCH_PREFIX}/${SLUG}"
-fi
-
-CLONE_URL="https://github.com/${TARGET_REPO}.git"
-
-if [[ -d "$WORKDIR/.git" ]]; then
-  agc::log "reusing existing workdir: $WORKDIR"
-  git -C "$WORKDIR" fetch origin --prune
-else
-  agc::log "cloning $CLONE_URL → $WORKDIR"
-  git clone --depth 50 "$CLONE_URL" "$WORKDIR"
-fi
-
-# Make sure base branch is up to date.
-git -C "$WORKDIR" checkout "$AGC_BASE_BRANCH"
-git -C "$WORKDIR" pull --ff-only origin "$AGC_BASE_BRANCH"
-
-# Configure fork remote if provided.
-if [[ -n "${TARGET_FORK}" ]]; then
-  if git -C "$WORKDIR" remote | grep -q '^fork$'; then
-    git -C "$WORKDIR" remote set-url fork "https://github.com/${TARGET_FORK}.git"
-  else
-    git -C "$WORKDIR" remote add fork "https://github.com/${TARGET_FORK}.git"
-  fi
-fi
-
-# Create or reset branch off latest base.
-if git -C "$WORKDIR" show-ref --verify --quiet "refs/heads/$BRANCH"; then
-  agc::log "branch $BRANCH exists — switching to it"
-  git -C "$WORKDIR" checkout "$BRANCH"
-else
-  git -C "$WORKDIR" checkout -b "$BRANCH" "$AGC_BASE_BRANCH"
-fi
-
-mkdir -p "$WORKDIR/.auto-pr/screenshots"
-
-if [[ -n "$ISSUE_JSON" ]]; then
-  printf '%s\n' "$ISSUE_JSON" > "$WORKDIR/.auto-pr/issue.json"
-fi
-printf '%s\n' "$TARGET_REPO" > "$WORKDIR/.auto-pr/target-repo.txt"
-
-agc::log "workspace ready"
-printf 'WORKDIR=%s\n' "$WORKDIR"
-printf 'BRANCH=%s\n' "$BRANCH"
-printf 'ISSUE_TITLE=%s\n' "$ISSUE_TITLE"
diff --git a/skills/auto-github-contributor/templates/PR-BODY.template.md b/skills/auto-github-contributor/templates/PR-BODY.template.md
deleted file mode 100644
index ff7f822..0000000
--- a/skills/auto-github-contributor/templates/PR-BODY.template.md
+++ /dev/null
@@ -1,32 +0,0 @@
-## Summary
-
-Fixes #{{ISSUE_NUMBER}} — _{{ISSUE_TITLE}}_
-Source issue: {{ISSUE_URL}}
-
-## Spec
-
-{{SPEC}}
-
-## Work breakdown (TDD)
-
-{{TODO}}
-
-## Visual verification
-
-{{SCREENSHOTS}}
-
-## Blockers / open questions
-
-{{BLOCKERS}}
-
-## Test plan
-
-- [x] `pnpm lint`
-- [x] `pnpm typecheck`
-- [x] `pnpm test`
-- [x] `pnpm build`
-- [ ] Manual smoke on the affected surface (see screenshots)
-
----
-
-_Generated by the `auto-github-contributor` skill. The agent followed a red-green-refactor loop and captured visual artefacts under `.auto-pr/screenshots/`._
diff --git a/skills/auto-github-contributor/templates/SPEC.template.md b/skills/auto-github-contributor/templates/SPEC.template.md
deleted file mode 100644
index 0e2ed28..0000000
--- a/skills/auto-github-contributor/templates/SPEC.template.md
+++ /dev/null
@@ -1,40 +0,0 @@
-# Spec — Issue #{{ISSUE_NUMBER}}: {{ISSUE_TITLE}}
-
-> Source: {{ISSUE_URL}}
-
-## Problem
-
-<!-- Restate the issue in your own words. 2–4 sentences. -->
-
-## Acceptance criteria
-
-<!-- Each bullet must be testable. Ideally maps 1:1 with a test case. -->
-
-- [ ] …
-- [ ] …
-
-## Approach
-
-<!-- What you'll change and why this approach over alternatives. -->
-
-## Files likely touched
-
-<!-- Inferred from ripgrep on symbols named in the issue. Update as you go. -->
-
-- `path/to/file.ts` — …
-
-## Risk / blast radius
-
-- **Scope**: small / medium / large
-- **Breaking**: yes / no
-- **Migration needed**: yes / no
-
-## Test plan
-
-- **Unit**: …
-- **Integration**: …
-- **Visual**: page(s) / selector(s) to screenshot via `browser-verify.sh`
-
-## Out of scope
-
-<!-- Explicitly note things you're NOT fixing so the PR stays reviewable. -->
diff --git a/skills/auto-github-contributor/templates/TODO.template.md b/skills/auto-github-contributor/templates/TODO.template.md
deleted file mode 100644
index fc88a11..0000000
--- a/skills/auto-github-contributor/templates/TODO.template.md
+++ /dev/null
@@ -1,31 +0,0 @@
-# TODO — Issue #{{ISSUE_NUMBER}}
-
-> Each item pairs a **failing test** with the **minimal implementation** that makes it pass (TDD).
-
-## Setup
-
-- [ ] Read full issue body + linked discussions
-- [ ] Run `dev-loop-check.sh --phase green` once on clean `main` to confirm baseline is green
-- [ ] Identify the module(s) that need changes (record in SPEC.md)
-
-## TDD loop
-
-- [ ] **Red 1** — Write failing test for `<behavior>`: …
-  - [ ] **Green 1** — Minimal implementation passes the test
-  - [ ] **Refactor 1** — Tidy without breaking green
-
-- [ ] **Red 2** — …
-  - [ ] **Green 2** — …
-  - [ ] **Refactor 2** — …
-
-## Verification
-
-- [ ] `dev-loop-check.sh --phase green` → 0
-- [ ] `browser-verify.sh` screenshots captured (or stub noted) for each visual acceptance item
-- [ ] Self-review diff: no debug prints, no commented-out code, no unrelated changes
-
-## Wrap-up
-
-- [ ] `dev-loop-check.sh --phase final` → 0
-- [ ] `create-pr.sh` to push + open PR
-- [ ] Paste PR URL into recap memo