diff --git a/README.en.md b/README.en.md index 78f6fad9..7d269cc5 100644 --- a/README.en.md +++ b/README.en.md @@ -32,19 +32,30 @@ Current version **v2.4.0** (see [Changelog](CHANGELOG.md) and [Releases](https:/ ## Preview +> Lots of screenshots — all collapsed into the groups below (click to expand; GitHub Markdown has no tabs, so collapsible groups stand in for a unified showcase). + +
+Main UI — Dashboard / Providers / Settings / Logs + | Dashboard | Providers | |---|---| | ![Board](img/Board.png) | ![Providers](img/Providers.png) | | **Settings** | **Logs** | | ![Settings](img/Settings.png) | ![Logs](img/Logs.png) | -### Codex App in action +
+ +
+Codex App in action With any provider enabled, Codex App's model picker shows ` / `-style real model names. Tool loops / `previous_response_id` history replay / thinking-mode reasoning_content injection are all handled transparently by the local proxy: ![Codex App real conversation](img/codex-cli-real-chat.png) -### Codex Desktop background themes (optional) +
+ +
+Codex Desktop background themes (optional, 11 anime themes) Inject background image + frosted-glass panel CSS into Codex Desktop (the Electron client). Five built-in anime themes plus user upload. The Codex binary is not modified — runtime injection via Chromium DevTools Protocol. The toggle is a persistent preference marker: enabling it persists the setting and injects immediately (best-effort); if Codex wasn't launched via this tool (or its debug port is unavailable), a confirm dialog offers to restart Codex so the theme takes effect. Disabling it clears the saved preference and best-effort live-removes the injected theme via CDP (back to native UI), symmetric with enabling — likewise, if the debug port is unavailable the removal fails and the theme stays until the next Codex restart. @@ -56,12 +67,28 @@ Inject background image + frosted-glass panel CSS into Codex Desktop (the Electr A sixth theme (Carton) carries a floating mascot in the bottom-right that reacts to the cursor. **Custom backgrounds**: Theme page → "+ Add custom" → pick a JPG/PNG → 16:9 crop modal (drag + scroll to zoom) → apply. If the toggle is on at Codex launch, the selected theme auto-injects — no manual step needed. -### Usage panel inside Codex (optional) +
+ +
+Usage panel inside Codex (optional) Injects a standalone "Usage" section at the bottom of Codex Desktop's "Toggle pinned summary" popup: 5-hour / weekly / monthly plan quota bars (whitelisted providers: antigravity gemini series + GLM Coding Plan + Xiaomi MiMo Token Plan + **OpenCode Go** + **Kimi Code**), balance / usage numeric entries (DeepSeek / Kimi Moonshot / anyrouter), context usage, real-time token rate + cumulative total, and cache hit rate. The **Context** row expands into a Claude-style by-source breakdown dropdown (classified by what is sent upstream: tool calls & output / reasoning / developer instructions / conversation messages / tool definitions / system prompt), persisted per conversation for instant load. See "What it does" below. ![Usage panel inside Codex](img/codex-usage-breakdown.jpg) +
+ +
+Codex draft stash (optional) + +Stash/restore buttons next to the composer + a "Stash" list below the Usage panel: park pre-typed text + images to free up the box for a reply, then selectively restore to the composer or send directly (with multiple entries, restore opens a picker). + +| Stash panel (with images) | Quick-restore dropdown | +|---|---| +| ![Stash panel](img/codex-stash-panel.png) | ![Quick-restore dropdown](img/codex-stash-dropdown.png) | + +
+ ## What it does - Manage multiple providers; map OpenAI model names (`gpt-5.5` / `gpt-5.4` / `gpt-5.4-mini` / `gpt-5.3-codex` / `gpt-5.2`) to the provider's real model IDs @@ -94,6 +121,7 @@ Injects a standalone "Usage" section at the bottom of Codex Desktop's "Toggle pi - **Injected system prompts follow the UI language**: the `apply_patch` chat-path rules + autocompact summarization prompt that this project injects for non-OpenAI providers track the `语言 / Language` setting (Chinese users → Chinese prompts, avoiding mixed-language model thinking); V4A keywords (`*** Begin Patch` / `@@
` etc.) + Codex CLI error message originals stay in English (parser / matcher does not accept translations) - **Codex Desktop Theme (optional, off by default)**: Theme page ships 11 built-in anime themes (`carton` with a floating mascot, plus `changli` / `azurlane` / `nailin` / `zani` / `frost` / `nocturne` / `duet` / `rose` / `sonata` / `studio`), each individually colour-matched to its artwork (per-theme glass + accent). Injects design-token overrides (`--color-token-*` + the runtime `--color-*` layer) + a background image into Codex Desktop via CDP, covering chat / settings / collapsed-sidebar / popovers. Toggle is independent from Plugin Unlock; page reload re-applies automatically; disabling the toggle clears the saved preference and best-effort live-removes the injected theme (back to native UI), only staying until the next Codex restart if removal fails (debug port unavailable) - **Usage panel inside Codex Desktop (optional, off by default)** (MOC-204): Settings → "Show usage in Codex" injects a collapsible "Usage" section at the bottom of Codex's "Toggle pinned summary" popup (the panel that contains Environment / Sources sections), showing up to 4 rows: ① **5-hour quota / weekly quota / plan usage** — whitelisted providers only: **antigravity gemini series** reads from `cloudcode-pa.googleapis.com/v1internal:retrieveUserQuotaSummary` (dual-window 5h + weekly, remaining% = remainingFraction×100); **GLM Coding Plan** (`bigmodel.cn` / `z.ai` coding hosts) reads from `monitor/usage/quota/limit` (apiKey auth, no Bearer prefix), returning 5h + weekly TOKENS_LIMIT records, converted as remaining% = 100 − usage%; **Xiaomi MiMo Token Plan** (`platform.xiaomimimo.com`) shows a monthly-plan remaining% progress bar — the plan quota is only accessible via a MiMo web session (httpOnly cookie), so you must click "Sign in to Xiaomi account" in the provider edit page first: the app opens an embedded webview for login, captures the session cookie, and the daemon uses it to query `/api/v1/tokenPlan/usage`; **DeepSeek** (`api.deepseek.com`) shows a ¥X balance numeric entry, read from the official `/user/balance` endpoint using the same API key (Bearer); **Kimi (Moonshot PAYG, `api.moonshot.cn` / `.ai`)** shows balance numeric entries (available / cash / voucher, ¥/$ by host), read from the official `/v1/users/me/balance` using the same key (Bearer) — **the subscription-based `kimi-code` (`api.kimi.com/coding`) now shows plan quota via a console session**: quota is only available from the kimi.com console (not via the inference API key); authentication uses the `access_token` from localStorage, so you click "Sign in to Kimi" in the provider edit page — the app opens an embedded webview, injects JS to bridge the token out via a cookie, and the daemon uses it as a Bearer token to call the connect-RPC endpoint `kimi.gateway.membership.v2.MembershipService/GetSubscriptionStat` for subscriptionBalance / rate-limit; **OpenCode Go** (`opencode.ai/zen/go/v1`) shows **5-hour / weekly / monthly** plan remaining% across three tiers — quota is only available from the opencode.ai console (no API), so you click "Sign in to OpenCode" in the provider edit page: the app opens an embedded webview for login, captures the session cookie + workspace id, and the daemon queries the `/workspace//go` page SSR HTML to parse the three usagePercent values; **anyrouter** (`api.anyrouter.top`) shows a $X used-amount numeric entry, read from `/v1/dashboard/billing/usage` using the same key (Bearer; remaining balance is blocked by upstream anti-scraping so only the used amount is shown). Whitelist is determined by baseUrl host. Red warning ≤10% + reset time shown. Quota rows appear only when the active provider matches a whitelisted host; all others show no quota rows. ② **Context** — injected JS reads `contextUsage.usedTokens` + `contextWindow` directly from Codex's React fiber, available immediately for any existing conversation without a new turn; full window = contextWindow ÷ 0.95 (adds back the 5% reserve Codex hides); 1M models display "1M" not "1000k". ③ **Tokens (real-time rate · cumulative)** — rate estimated by a MutationObserver watching Codex's streaming text (2s sliding window, CJK-aware); cumulative total from Codex rollout. ④ **Cache hit rate** — from rollout cached_input/input. **③④ and the rate are all isolated per active conversation (MOC-230)**: injected JS reads the current `conversationId` from the React fiber and the daemon keys totals to that conversation's rollout (== filename uuid, not the most-recently-modified file), following conversation switches; shows "—" (never another conversation's data) when the id / its rollout can't be resolved. The "Usage" title is collapsible (chevron + localStorage-persisted). Injection uses periodic CDP pushes; re-attaches automatically after a Codex page reload or restart. Requires launching Codex through this app; restart Codex after toggling if already running. +- **Draft stash inside Codex Desktop (optional, off by default)**: Settings → "Enable draft stash in Codex" injects a stash/restore button pair next to the composer and a collapsible "Stash" section below the Usage panel. Solves the "pre-typed input vs. quick follow-up" conflict: park the current composer draft (push button) to free the box for an interjection, then selectively restore it to the composer (pop button — one stashed item restores directly, multiple opens a picker) or, from the Stash panel, restore / send / delete each entry. Restore uses swap semantics — if the composer is non-empty its current content is stashed first, so nothing is lost. More flexible than Codex's native steer (which can only auto-send queued messages in order). State is a global list in renderer localStorage (survives Codex reloads). Injection uses periodic CDP pushes and re-attaches after a Codex reload/restart; requires launching Codex through this app. - **System-proxy (VPN/ladder) connectivity detection** (MOC-114): the dashboard "Network Proxy" card shows live status — connected / disconnected / PAC auto-config / detecting. In relay real-account mode, the "Auto-unlock Codex Plugins" toggle gates on both conditions being met (valid account AND proxy reachable), preventing the silent-failure state where plugins spin and return 502s while the UI shows "logged in" because the proxy is down. Detection uses a short-timeout TCP connect to the proxy port only; chatgpt.com is never contacted. - **Built-in web fetch tool (web_fetch, MOC-144)**: Settings → "Built-in web fetch backend" — select `auto` (recommended; **defaults to `auto` since MOC-215, works out of the box**; **since MOC-256, a new install with no system Chrome / no downloaded shell defaults to `off`** to avoid a silent ~86 MB chrome-headless-shell download when web_fetch escalates to headless at runtime — new users with Chrome get web_fetch / web_search without manually enabling it, while users without Chrome confirm the download via the gate when they manually pick auto/headless; web_fetch uses curl/wreq and needs no Chrome, web_search is still gated on Chrome readiness and never silently downloads) / `curl` / `wreq` / `headless` (**independent of** the Codex sandbox network toggle). Transfer automatically registers a `web_fetch` MCP tool with Codex, which the model can call directly to fetch web pages — `curl` uses standard HTTP, `wreq` bypasses Cloudflare TLS challenges, `headless` drives a headless Chrome to retrieve JS-rendered DOM (selecting `headless` / `auto` in Settings first checks Chrome readiness: a system Chrome that passes the `--version` self-test, or an already-downloaded built-in chrome-headless-shell, enables it directly without re-downloading; otherwise it prompts to confirm downloading chrome-headless-shell, ~86 MB on demand. If a system proxy is configured but currently unreachable, it auto-downgrades to `wreq` with a notice). Beyond the three fetch backends, `web_fetch` also follows **HTML `meta refresh` / JS `location` redirects** (re-fetches the target URL, loop-protected to 3 hops) — curl/wreq/headless only follow HTTP 3xx and do not handle these client-side redirects; "placeholder" redirect pages (e.g. pages that bounce around Twitter/Substack blocks) are now automatically followed to the real destination (MOC-139). **`auto` tier (MOC-161)**: automatically escalates from curl → wreq → headless based on page-difficulty signals; remembers the last successful tier per origin so subsequent requests start there; downgrades to curl when no system proxy is reachable (wreq / headless rely on a proxy); first use of the headless tier still confirms the Chrome download. Switching tiers takes effect immediately (no restart needed); **toggling the feature on or off requires restarting Codex Desktop** for the network tools (web_fetch / web_search / read_url_local) to appear / disappear in Codex (since MOC-235 the MCP server stays registered to host `read_tool_artifact`; turning the network backend off just stops exposing those network tools rather than unloading the whole server). Fetched HTML is auto-converted to markdown before returning to the model (cleaner, fewer tokens; non-HTML responses pass through unchanged), and headless waits for networkIdle before capturing the rendered DOM (MOC-145). Headless fetches run with anti-detection stealth (strips `navigator.webdriver`, fakes `window.chrome`/plugins/WebGL, removes the `HeadlessChrome` UA token), passing passive-fingerprint / simple JS-challenge Cloudflare; interactive Turnstile/DataDome managed challenges still won't pass (MOC-152). On a CF JS-challenge page, headless now **waits in place for it to auto-clear** before reading (instead of returning the challenge page as content), and **persists the browser profile per origin** to reuse CF clearance cookies — a second fetch of the same site skips the repeat challenge and is faster (MOC-156). Before markdown conversion the page goes through **main-content extraction** (readability algorithm strips nav/header/footer/sidebar/ads, keeping only the article so large-page content is no longer crowded out by truncation; non-article pages fall back to the full page); **binary resources** (image / video / audio / PDF) and files over 16 MB are not downloaded and return a clear notice instead (no more garbage bytes / OOM) (MOC-152). `web_fetch` **returns the full extracted page text by default** (the current turn's tool output goes into the LLM context in full; the adapter layer automatically compresses older tool outputs to prevent context overflow; MOC-190) — no more pagination, no `offset` paging, no relevance-based `query` chunk selection, so precise content (code / schema / version numbers / figures) is never lost. If you fetched a URL earlier in the conversation and its content has since been folded/compressed in the context history, use **`read_url_local(url)`** to pull the full text from the in-process cache without re-fetching (cache TTL: 15 min). **More generally, when any tool's large output (shell / Feishu and other MCP / etc.) gets folded into a `[Tool output stored outside model context]` summary in history, the summary includes an `Artifact ID`, and the model can call `read_tool_artifact(artifact_id)` to retrieve that output's text** — read from the shared `tool_artifacts.db` (SQLite WAL, cross-process) that the proxy persists when compressing, so the model never re-runs a tool just to see history again; the retrieved content is visible only in the current turn and gets folded again next turn (no long-term context bloat); outputs over 90k chars are returned in pages (each below the proxy keep-full cap, with a trailer telling the model to page via `offset`) (MOC-235). These tools (`web_fetch` / `web_search` / `read_url_local` / `read_tool_artifact`) declare `readOnlyHint` (read-only), so Codex's auto-review guardian **skips approval** for them (`requires_mcp_tool_approval` short-circuits on the read-only hint) — network calls no longer incur a per-call risk-approval round-trip, removing that latency (MOC-172). - **Built-in web search tool (web_search, MOC-12)**: when the built-in web fetch backend is on (non-off) and the machine has Chrome ready, transfer registers a `web_search` tool with Codex — the model passes a query string and gets back a structured list of results (title + real URL + snippet), forming a **two-step search**: `web_search` to find sources, then `web_fetch` to read content, eliminating the need to guess URLs. **Why this matters**: Codex sends an OpenAI server-side `web_search` tool each turn, but third-party chat providers (MiniMax / DeepSeek / GLM / Kimi, etc.) don't support it — the adapter drops it, leaving the model to scrape search engines or guess URLs (real-world success rate ~17%). This tool queries **DuckDuckGo + Bing in parallel and merges the results, deduped by normalized URL** (no API key required, data-centre / VPN-exit IP friendly; the two indexes complement each other so single-call coverage is noticeably broader than a single source, MOC-215; previously Bing was only a fallback when DDG failed, MOC-186), and **always uses headless** internally — DDG / Bing block plain HTTP with anti-bot challenges regardless of TLS fingerprint, so a real browser is required; the parallel fetch keeps wall-time ≈ the slower single engine rather than the sum, and either engine being blocked / empty still leaves the other usable. `web_search` always uses headless internally, but its **exposure / invocation only requires Chrome to be ready** (system Chrome / Edge / Chromium, or an already-downloaded built-in chrome-headless-shell) — decoupled from the web_fetch tier: users with system Chrome can use search under any non-off tier (incl. curl / wreq) without triggering a download; if neither is present it stays hidden and a call returns a hint to pick the headless tier to complete the first-time download (MOC-190). Ad results are filtered out; blocked / no-results states return explicit error messages (never silently empty). **Pagination (MOC-215)**: `web_search` returns only the first page (~10-20 results, not fetching multiple pages at once to avoid excessive headless latency); when the model needs more / different sources it uses the separate **`web_search_more`** tool (same query, `page=2/3…`) to fetch the next batch (via Bing's `first=` deep pages), with a tail hint in the result steering the model to paginate rather than re-run the same query — numeric string arguments are parsed leniently (models often send `page` as the string `"2"`) so pagination never silently falls back to page 1. DDG HTML parsing borrows from `duckduckgo_search` (Python). diff --git a/README.md b/README.md index c21a20a1..c5b4d0e0 100644 --- a/README.md +++ b/README.md @@ -32,19 +32,30 @@ Codex App Transfer 是一个面向 **OpenAI Codex APP** 的轻量桌面配置 + ## 界面预览 +> 展示截图较多,统一收进下方折叠分组(默认折叠,点开查看;GitHub Markdown 不支持标签页,故用可折叠分组实现「整体展示区」)。 + +
+主界面 — 仪表盘 / 供应商 / 设置 / 日志 + | 仪表盘 | 供应商 | |---|---| | ![Board](img/Board.png) | ![Providers](img/Providers.png) | | **设置** | **日志** | | ![Settings](img/Settings.png) | ![Logs](img/Logs.png) | -### Codex APP 实际接入 +
+ +
+Codex APP 实际接入 启用任意供应商后,Codex APP 模型选择器会显示「 / 」形式的真实模型名,对话过程中工具循环 / `previous_response_id` 历史回放 / thinking 模式 reasoning_content 注入全部由本地代理透明处理: ![Codex APP 实际对话](img/codex-cli-real-chat.png) -### Codex Desktop 背景主题(可选) +
+ +
+Codex Desktop 背景主题(可选,11 套二次元主题) 为 Codex Desktop(Electron 客户端)注入背景图 + 半透明玻璃面板 CSS,内置 11 套二次元主题(每套按背景图独立配色)+ 自定义上传。不修改 Codex 的 binary,基于 Chromium DevTools Protocol 运行时注入。开关为持久化状态标记:开启时落盘保存并即时注入(best-effort),若当前 Codex 未经本工具启动 / 调试端口不可用,则弹确认提示是否重启 Codex 让主题生效;关闭时落盘清除偏好并 best-effort 即时 CDP 清除已注入主题(回原生 UI),与开启对称 —— 同样在调试端口不可用时清除失败,则保留至 Codex 下次重启移除。 @@ -56,12 +67,28 @@ Codex App Transfer 是一个面向 **OpenAI Codex APP** 的轻量桌面配置 + 第 6 套 Carton 自带右下角漂浮立绘(随鼠标动)。**自定义背景**:Theme 页 → "+ 添加自定义" → 选 JPG/PNG → 16:9 crop 弹窗自由选截取区域(拖拽 + 滚轮缩放)→ 应用。Codex 启动时如已开启 toggle 会自动注入已选主题,不需手动操作。 -### Codex 内用量显示(可选) +
+ +
+Codex 内用量显示(可选) 在 Codex Desktop 顶栏「Toggle pinned summary」弹窗底部注入独立「Usage」分区:5 小时 / 每周 / 月度套餐额度(白名单 provider:antigravity gemini 系 + GLM Coding Plan + 小米 MiMo Token Plan + **OpenCode Go** + **Kimi Code**)、余额 / 已用额度数值(DeepSeek / Kimi 月之暗面 / anyrouter)、上下文用量、实时 Tokens 速率与累计、缓存命中率。其中**上下文**行可展开 Claude 风格的 by-source 明细下拉(按发往上游的内容分类:工具调用与输出 / 推理 / 开发者指令 / 对话消息 / 工具定义 / 系统提示),数据按对话持久化、启动即用。详见下方「能做什么」。 ![Codex 内用量显示](img/codex-usage-breakdown.jpg) +
+ +
+Codex 草稿暂存 Stash(可选) + +输入框旁「暂存 / 恢复」按钮 + Usage 面板下方「Stash」列表:把预输入的文字 + 图片先暂存、腾出输入框去回复上一轮,随后选择性恢复到输入框或直接发送(多条时点恢复唤起下拉选)。 + +| Stash 面板(支持图片) | 快捷恢复下拉 | +|---|---| +| ![Stash 面板](img/codex-stash-panel.png) | ![快捷恢复下拉](img/codex-stash-dropdown.png) | + +
+ ## 能做什么 - 管理多套供应商,按 OpenAI 模型名(`gpt-5.5` / `gpt-5.4` / `gpt-5.4-mini` / `gpt-5.3-codex` / `gpt-5.2`)映射到供应商真实模型 ID @@ -94,6 +121,7 @@ Codex App Transfer 是一个面向 **OpenAI Codex APP** 的轻量桌面配置 + - **注入的 system prompts 跟随界面语言**:本项目对非 OpenAI provider 注入的 `apply_patch` chat-path 规则 + autocompact 总结提示词,跟设置里 `语言 / Language` 一致(中文用户 → 中文 prompt,避免模型中英混杂思考);V4A 关键字(`*** Begin Patch` / `@@
` 等)+ Codex CLI 错误消息原文保英文(parser / matcher 不接受翻译) - **Codex Desktop 主题(可选,默认关)**:Theme 页内置 11 套动漫主题(`carton` 含浮动看板娘,其余 `changli` / `azurlane` / `nailin` / `zani` / `frost` / `nocturne` / `duet` / `rose` / `sonata` / `studio`),每套按背景图独立调出暗玻璃 + 强调色。通过 CDP 向 Codex Desktop 注入设计令牌覆盖(`--color-token-*` + 运行时 `--color-*` 层)+ 背景图,覆盖聊天 / 设置页 / 折叠侧栏 / 弹层等各视图。开关跟 Plugin Unlock 独立,page reload 自动重应用;关闭开关落盘清除偏好 + best-effort 即时 CDP 清除已注入主题(回原生 UI),仅清除失败(调试端口不可用)才保留至 Codex 下次重启移除 - **Codex 内用量显示(可选,默认关)**(MOC-204):设置 → 「Codex 内显示用量信息」,在 Codex Desktop 顶栏「Toggle pinned summary」弹窗(含 Environment / Sources 等分区)底部注入独立「Usage」用量分区,最多 4 行:① **5 小时额度 / 每周额度 / 套餐用量**:仅白名单 provider 显示:**antigravity gemini 系**数据来自 `cloudcode-pa.googleapis.com/v1internal:retrieveUserQuotaSummary` 双窗口**剩余**额度(remainingFraction×100);**GLM Coding Plan**(`bigmodel.cn`/`z.ai` coding 系)数据来自 `monitor/usage/quota/limit` 端点(apiKey 直接鉴权,不带 Bearer),返回 5h / 每周 TOKENS_LIMIT,已用% → 剩余% = 100-已用;**小米 MiMo Token Plan**(`platform.xiaomimimo.com`)显示月度套餐剩余%进度条,需在 provider 编辑页点「登录小米账号」按钮——套餐用量只在小米控制台、走 httpOnly session cookie,app 内嵌 webview 登录后抓取 cookie 存本地,daemon 带该 cookie 查询 `/api/v1/tokenPlan/usage`;**DeepSeek**(`api.deepseek.com`)显示余额 ¥X 数值条目,调官方 `/user/balance` 接口、与推理同一把 API key(Bearer);**Kimi(月之暗面 / Moonshot PAYG,`api.moonshot.cn`/`.ai`)**显示余额数值条目(可用 / 现金 / 赠金,按 host 记 ¥/$),调官方 `/v1/users/me/balance`、与推理同一把 key(Bearer)——**订阅制 `kimi-code`(`api.kimi.com/coding`)现经控制台 session 显示套餐用量**:用量只在 kimi.com 控制台、鉴权用 localStorage 的 `access_token`(非 API key),需在 provider 编辑页点「登录 Kimi」内嵌 webview 登录后注入 JS 经 cookie 桥抓出 token,daemon 带它当 Bearer 调 connect-RPC `kimi.gateway.membership.v2.MembershipService/GetSubscriptionStat` 取 subscriptionBalance / rate-limit;**OpenCode Go**(`opencode.ai/zen/go/v1`)显示 **5 小时 / 每周 / 每月**三档套餐用量剩余%:用量只在 opencode.ai 控制台(无 API),需在 provider 编辑页点「登录 OpenCode」内嵌 webview 登录抓 session cookie + workspace id,daemon 带它查 `/workspace//go` 页 SSR HTML 解析三档 usagePercent;**anyrouter**(`api.anyrouter.top`)显示已用额度 $X 数值条目,调 `/v1/dashboard/billing/usage`、与推理同一把 key(Bearer;账户余额受上游反爬限制仅展示已用量)。白名单均按 baseUrl host 判定。≤10% 红色预警 + 重置时刻;仅活动 provider 命中白名单时显示额度行,其余不显。② **上下文**:注入脚本直接从 Codex React fiber 读 `contextUsage.usedTokens / contextWindow`,有历史对话即立即显示(不需新对话);满窗口 = contextWindow÷0.95(加回 Codex 隐藏的 5% reserve);1M 模型显「1M」而非「1000k」。③ **Tokens(实时速率·累计)**:速率由 MutationObserver 监测 Codex 流式文本增量估算(2s 滑窗,CJK 感知);累计量来自 Codex rollout 文件。④ **缓存命中率**:来自 rollout 的 cached_input/input。**③④ + 速率均按活动对话隔离(MOC-230)**:注入脚本从 React fiber 读当前 `conversationId`,daemon 按该 id 取对应 rollout(== 文件名 uuid,非「最近修改」的文件),切对话即跟随、不串号;读不到 id / 无对应 rollout 显「—」(绝不显示别的对话数据)。「Usage」标题可折叠(chevron + localStorage 持久)。注入走 CDP 周期推送,页面刷新 / 重启后自动重挂;需通过本应用启动 Codex,若 Codex 已在运行需重启生效 +- **Codex 内草稿暂存(可选,默认关)**:设置 →「Codex 内启用草稿暂存」,在输入框旁注入「暂存 / 恢复」按钮对,并在 Usage 面板下方注入可折叠「Stash」分区。解决「预输入内容 vs 临时插话」冲突:把当前输入框草稿先暂存(push 按钮)腾出输入框去补一句,随后选择性恢复到输入框(pop 按钮——1 条直接恢复,多条唤起下拉选)或在 Stash 面板里对每条 恢复 / 发送 / 删除。恢复采用 swap 语义——输入框非空时先把当前内容暂存,不丢草稿。比 Codex 原生 steer(预存后只能按队列顺序自动发送)更灵活。状态存于 renderer localStorage 的全局列表(跨 Codex 重载存活)。注入走 CDP 周期推送,页面刷新 / 重启后自动重挂;需通过本应用启动 Codex。 - **系统代理(梯子)连通性检测**(MOC-114):仪表盘「网络代理」卡实时显示系统代理是否活跃(已连接 / 未连接 / 自动配置 PAC / 检测中);relay 真实账号模式下「自动解锁 Codex Plugins」开关在账号有效且代理可达两条件同时满足时才激活,避免梯子没开时 plugins 静默全 502 却显示"已登录"的误导态。探测仅对代理端口做短超时 TCP 连通测试,不访问 chatgpt.com。 - **内置联网抓取工具(web_fetch,MOC-144)**:设置页 → 「内置联网抓取工具」选 `auto`(推荐,**MOC-215 起默认 `auto`、开箱即用**;**MOC-256 起:系统无 Chrome / 未下载内置 shell 的新装默认改为 `off`**,避免运行时 web_fetch 升 headless 时静默下载 ~86MB chrome-headless-shell —— 有 Chrome 的新装用户无需手动开启即可用 web_fetch / web_search,无 Chrome 用户手动选 auto/headless 时经门控确认下载;web_fetch 走 curl/wreq 不需 Chrome,web_search 仍受 Chrome 就绪 gate 保护、不静默下载) / `curl` / `wreq` / `headless`(**独立于** Codex 沙箱联网开关),transfer 自动往 Codex 注册 `web_fetch` MCP 工具,Codex 模型可直接调该工具抓取网页 —— `curl` 走标准 HTTP、`wreq` 绕 Cloudflare TLS 挑战、`headless` 驱动无头 Chrome 取 JS 渲染后 DOM(在设置页选 `headless` / `auto` 时先查 Chrome 就绪:系统 Chrome `--version` 自检通过、或已下载内置 chrome-headless-shell,即直接启用不重复下载;都没有才弹窗确认按需下载 chrome-headless-shell ~86 MB。若配了系统代理但当前连不上,会自动降级到 `wreq` 并提示)。三档之外,`web_fetch` 还能跟随 **HTML meta refresh / JS `location` 跳转**(重定向到目标 URL 重抓,防循环最多 3 跳)——curl/wreq/headless 只处理 HTTP 3xx,不跟这类客户端重定向;绕 Twitter/Substack 等封锁的"占位跳转页"会自动跟随到真实内容页(MOC-139)。**`auto` 档(MOC-161)**:按页面难度自动从 curl 升级到 wreq 再到 headless,对每个域名记住上次成功档位(下次从该档起步省试错);系统代理不可达时自动压制至 curl(wreq / headless 依赖代理);首次用 headless 档同样弹窗确认 Chrome 下载。切档即时生效(无需重启);**改"开/关"状态后需重启 Codex Desktop** 才会让联网工具(web_fetch / web_search / read_url_local)在 Codex 里出现 / 消失(MOC-235 起该 MCP server 始终注册以托管 `read_tool_artifact`,关闭联网档只是不再暴露这几个联网工具,不再卸载整个 server)。抓到的 HTML 会自动转成 markdown 返给模型(更省 token、更干净;非 HTML 响应原样透传),headless 用 networkIdle 等渲染落定再取(MOC-145)。headless 抓取启用反检测 stealth(抹 `navigator.webdriver`、伪造 `window.chrome`/插件/WebGL、UA 去 `HeadlessChrome` 标记),可过被动指纹 / 简单 JS 挑战类 Cloudflare;交互式 Turnstile/DataDome 托管挑战仍过不了(MOC-152)。headless 遇 CF JS 挑战页会**原地等其自动解出**再读(而非立即把挑战页当正文返回),并**按域名持久化浏览器 profile** 复用 CF 放行 cookie —— 同一站点二次抓取跳过重复挑战、更快(MOC-156)。抓到的页转 markdown 前先做**正文抽取**(readability 算法剥 nav/页眉/页脚/侧栏/广告,只留正文,大页正文不再被截断挤掉;非文章页自动回退整页);图片 / 视频 / 音频 / PDF 等**二进制资源**与超 16 MB 大文件不下载、直接返提示(不再吐乱码 / 防 OOM)(MOC-152)。`web_fetch` **默认直接返回抓取到的完整正文**(当前轮全文进 LLM 上下文、adapter 层自动把历史轮的 tool 输出压缩以防撑爆;MOC-190)—— 不再分页、不再按 `offset` 翻页、不再按 `query` 相关性选块,精确信息(代码 / schema / 版本 / 数字)不丢。若较早抓取的某 URL 正文在对话历史里被折叠 / 压缩、需要回看完整原文,用 **`read_url_local(url)`** 从进程内缓存取回,不必重新联网(缓存 15 min)。**更进一步,任意工具(shell / 飞书等 MCP / 其它)的大输出在历史里被压成 `[Tool output stored outside model context]` 摘要时,摘要会给出 `Artifact ID`,模型可调 `read_tool_artifact(artifact_id)` 取回该输出原文** —— 读 proxy 压缩时落盘的共享 `tool_artifacts.db`(SQLite WAL,跨进程读),不必为回看历史而重跑工具;取回内容仅当前轮可见、下一轮再被自动折叠不长期占上下文;超 90k 字符的大输出分页返回(每块低于 proxy keep-full 上限,末尾提示用 `offset` 逐块读完整)(MOC-235)。这些工具(`web_fetch` / `web_search` / `read_url_local` / `read_tool_artifact`)均声明 `readOnlyHint`(只读),Codex 的 auto-review guardian 据此**跳过审批**(`requires_mcp_tool_approval` 命中只读直接放行),联网调用不再逐次触发风险审批往返、消除审批延迟(MOC-172)。 - **内置 web_search 搜索工具(MOC-12)**:启用「内置联网抓取工具」(非 off)且本机 Chrome 就绪后,transfer 往 Codex 注册 `web_search` 工具 —— 模型给关键词即返回结构化结果列表(标题 + 真实 URL + 摘要),配合 `web_fetch` 组成**两段式联网**:先 `web_search` 找信息源、再 `web_fetch` 抓正文,免去模型瞎猜 URL。**为什么需要**:Codex 默认每轮发的 OpenAI server-side `web_search` 在第三方 chat provider(MiniMax / DeepSeek / GLM / Kimi 等)上游不被支持、被协议层 drop,模型只能退化到自己抓搜索引擎页 / 猜 URL(真机实测成功率仅 ~17%)。本工具走 **DuckDuckGo + Bing 双引擎并行检索、按 URL 归一化去重后交错合并**(免 key、对数据中心 / VPN 出口 IP 友好;两家索引互补、单次覆盖面较单源明显更全,MOC-215;此前 Bing 仅在 DDG 失败时兜底 MOC-186),且**内部固定 headless** 浏览器代搜 —— DDG / Bing 对纯 HTTP 请求反爬拦截(无论 TLS 指纹多真),必须真浏览器跑 JS;并行抓取故 wall-time ≈ 单家而非求和,任一引擎被拦 / 无结果时另一家仍可用。`web_search` 内部固定 headless,但其**暴露 / 调用只要求本机 Chrome 就绪**(系统装了 Chrome / Edge / Chromium,或已下载内置 chrome-headless-shell)—— 与 web_fetch 档位解耦:系统有 Chrome 的用户在任意非 off 档(含 curl / wreq)都能用 search 且不触发下载;两者皆无则不暴露、调用返回提示引导去 headless 档完成首次下载(MOC-190)。结果自动过滤广告;反爬拦截 / 无结果时返回明确提示(不静默吐空)。**翻页(MOC-215)**:`web_search` 只返第 1 页(约一二十条,不一次扩抓多页以免 headless 延迟过高);模型需要更多 / 不同来源时用独立工具 **`web_search_more`(同 query, page=2/3…)** 取下一批(走 Bing `first=` 深页),结果尾部附诱导提示引导模型主动翻页而非用同一 query 重复搜 —— 工具参数对数字字符串(模型常把 `page` 传成 `"2"`)做宽容解析,避免翻页静默退回第 1 页。DDG HTML 解析模式借鉴 `duckduckgo_search`(Python)上游。 diff --git a/frontend/src/i18n/en.ts b/frontend/src/i18n/en.ts index 3fcf3731..68860495 100644 --- a/frontend/src/i18n/en.ts +++ b/frontend/src/i18n/en.ts @@ -586,6 +586,8 @@ export default { "settings.autoUnlockCodexPluginsHint": "Unlocks Codex plugins. Requires launching Codex from this app.", "settings.codexQuotaEnabled": "Show usage in Codex", "settings.codexQuotaEnabledHint": "Shows a usage panel in Codex (context, token rate/total, quota). Only when Codex is launched from this app.", + "settings.codexStashEnabled": "Enable draft stash in Codex", + "settings.codexStashEnabledHint": "Stash composer drafts to restore or send later. Only when Codex is launched from this app.", "settings.pluginUnlockForce": "Force unlock", "settings.pluginUnlockForced": "Force-unlock triggered", "settings.pluginUnlockForceFailed": "Force unlock failed", diff --git a/frontend/src/i18n/zh.ts b/frontend/src/i18n/zh.ts index 5a1b40f1..29c6986b 100644 --- a/frontend/src/i18n/zh.ts +++ b/frontend/src/i18n/zh.ts @@ -668,6 +668,9 @@ export default { "settings.codexQuotaEnabled": "Codex 内显示用量信息", "settings.codexQuotaEnabledHint": "仅通过本应用启动 Codex 时生效，在 Codex 內显示用量面板:上下文占用、Tokens 速率与累计以及额度信息。", + "settings.codexStashEnabled": "Codex 内启用草稿暂存", + "settings.codexStashEnabledHint": + "仅通过本应用启动 Codex 时生效，可暂存输入框草稿，随后恢复到输入框或直接发送。", "settings.pluginUnlockForce": "强制开启", "settings.pluginUnlockForced": "已触发强制解锁", "settings.pluginUnlockForceFailed": "强制开启失败", diff --git a/frontend/src/pages/SettingsPage.vue b/frontend/src/pages/SettingsPage.vue index 17e42794..6dcea0ab 100644 --- a/frontend/src/pages/SettingsPage.vue +++ b/frontend/src/pages/SettingsPage.vue @@ -242,6 +242,7 @@ const autoApplyOnStart = toggle('autoApplyOnStart', true) const restoreCodexOnExit = toggle('restoreCodexOnExit', true) const autoWakeCodexPet = toggle('autoWakeCodexPet', true) const codexQuotaEnabled = toggle('codexQuotaEnabled', false) +const codexStashEnabled = toggle('codexStashEnabled', false) const codexNetworkAccess = toggle('codexNetworkAccess', false) const exposeAllProviderModels = toggle('exposeAllProviderModels', false) const showGrayProviders = toggle('showGrayProviders', false) @@ -617,6 +618,9 @@ const UPDATE_REPO_URL = 'https://github.com/Cmochance/codex-app-transfer' + + +
{{ t('theme.title') }}
diff --git a/img/codex-stash-dropdown.png b/img/codex-stash-dropdown.png new file mode 100644 index 00000000..d7737642 Binary files /dev/null and b/img/codex-stash-dropdown.png differ diff --git a/img/codex-stash-panel.png b/img/codex-stash-panel.png new file mode 100644 index 00000000..783ab28f Binary files /dev/null and b/img/codex-stash-panel.png differ diff --git a/src-tauri/src/admin/services/desktop/process.rs b/src-tauri/src/admin/services/desktop/process.rs index 8cdf5cac..2adec517 100644 --- a/src-tauri/src/admin/services/desktop/process.rs +++ b/src-tauri/src/admin/services/desktop/process.rs @@ -432,7 +432,14 @@ fn should_attach_debug_port() -> Vec { .and_then(|s| s.get("codexQuotaEnabled")) .and_then(|v| v.as_bool()) .unwrap_or(false); - if !plugin_unlock_needs_port && !theme_enabled && !quota_enabled { + // 草稿暂存(Stash)注入 daemon 同样走 CDP,只开 stash 不开其它时也要端口。 + let stash_enabled = cfg + .as_ref() + .and_then(|c| c.get("settings")) + .and_then(|s| s.get("codexStashEnabled")) + .and_then(|v| v.as_bool()) + .unwrap_or(false); + if !plugin_unlock_needs_port && !theme_enabled && !quota_enabled && !stash_enabled { return vec![]; } diff --git a/src-tauri/src/codex_quota_injector.rs b/src-tauri/src/codex_quota_injector.rs index 1ab2854a..26b66409 100644 --- a/src-tauri/src/codex_quota_injector.rs +++ b/src-tauri/src/codex_quota_injector.rs @@ -56,7 +56,7 @@ const INSTALL_SCRIPT: &str = r##" // [MOC-230] 版本化幂等 guard:同版本跳过(常态);版本变(应用升级后 INSTALL_SCRIPT 改了) // → 先拆旧 observer + DOM 节点再重装,使新逻辑无需重启 Codex 即覆盖旧注入。旧版只有 // __catQuotaInstalled、无 __catQuotaVersion(undefined ≠ 当前版本)→ 同样触发重装。 - var VERSION = 3; // [MOC-231] 加 breakdown caret/下拉 + convId 守卫 → bump,升级后免重启 Codex 即覆盖旧注入 + var VERSION = 4; // 加「Stash 面板在场时让位坐到其前」的占位协调 → bump,升级后免重启 Codex 即覆盖旧注入 if (window.__catQuotaInstalled) { if (window.__catQuotaVersion === VERSION) return; try { if (window.__catQuotaObserver) window.__catQuotaObserver.disconnect(); } catch (e) {} @@ -582,9 +582,15 @@ const INSTALL_SCRIPT: &str = r##" node.id = 'cat-quota-entry'; scroller.appendChild(node); fresh = true; + } + // 末尾归位:React 后续追加 section 会把条目挤到中间(appendChild/insertBefore 对已存在 + // 节点是移动)。默认恒在 scroller 末尾;但若 Stash 面板(cat-stash-entry,codex_stash_injector) + // 在场,本额度条目须坐到它**之前**(「Usage 在 Stash 上面」),否则两个注入器每 tick 争抢 + // lastElementChild。稳定不动时不写,避免自触发 observer churn。 + var __stash = document.getElementById('cat-stash-entry'); + if (__stash && __stash.parentElement === scroller) { + if (node.nextElementSibling !== __stash) scroller.insertBefore(node, __stash); } else if (node !== scroller.lastElementChild) { - // React 后续往 scroller 追加 section 会把条目挤到中间;appendChild - // 对已存在节点是移动,保持条目恒在弹窗末尾 scroller.appendChild(node); } // 内容只在数据变化(或节点新建)时重建,避免 observer 高频 re-render churn @@ -1850,7 +1856,7 @@ mod tests { assert!(INSTALL_SCRIPT.contains("ctxUsed")); // 缓存上下文绝对值(used/window) // [MOC-231] breakdown 的 convId 守卫(切对话不串)+ 版本化 guard 已 bump(升级免重启覆盖) assert!(INSTALL_SCRIPT.contains("cqbdmismatch")); - assert!(INSTALL_SCRIPT.contains("var VERSION = 3")); + assert!(INSTALL_SCRIPT.contains("var VERSION = 4")); // bump:Stash 面板在场时让位占位协调 } #[test] diff --git a/src-tauri/src/codex_stash_injector.rs b/src-tauri/src/codex_stash_injector.rs new file mode 100644 index 00000000..c5d333d1 --- /dev/null +++ b/src-tauri/src/codex_stash_injector.rs @@ -0,0 +1,696 @@ +//! Codex 草稿暂存(Stash)注入器 —— 经 CDP 向 Codex 渲染进程注入「输入框草稿暂存区」。 +//! +//! 解决的痛点:用户常在输入框预输入下一轮内容,但上一轮助手回复需要先补充追问时, +//! 只能删掉预输入内容去回复,回头再重打。Stash 提供一个「先存起来、随后选择性恢复/发送」 +//! 的草稿暂存区,比 Codex 原生 steer(预存后只能按队列顺序自动发送)更灵活。 +//! +//! 三个交互面(全部 renderer 本地,纯 DOM + localStorage,无 Rust↔JS 数据流): +//! - **暂存按钮**(composer 工具栏):存当前输入框内容 → 清空输入框;输入框空时禁用。 +//! - **快捷恢复按钮**(composer 工具栏):0 条隐藏 / 1 条点击直接恢复 / >1 条点击唤起下拉选。 +//! - **Stash 面板**(pinned-summary 弹窗里、Usage 面板下面):列全部暂存项,每行 恢复/发送/删除。 +//! +//! 架构与 [`crate::codex_quota_injector`] 同构:daemon 每 tick 经 CDP `Runtime.evaluate` 跑 +//! 幂等 `INSTALL_SCRIPT`(版本化 guard,升级即覆盖旧注入,免重启 Codex)。所有状态在 renderer +//! 侧 localStorage(`catStash`),跨 Codex reload 存活(与 quota 的 usageCache 同机制)。 +//! +//! 存储 schema:`localStorage['catStash'] = [{ id, text, ts, images }]`(全局,不按会话隔离; +//! `images` 为 `[{ src(dataURL), filename }]`,可序列化)。 +//! +//! 恢复语义(swap,不丢草稿):输入框空 → 直接填入目标项;输入框非空 → 先把当前内容暂存, +//! 再填入目标项。恢复/发送均「消费」该条(git stash pop 模型)。 +//! +//! 面板挂载点与 quota 面板同一 scroller。为保证「Stash 在 Usage 下面」的稳定顺序 +//! `[...React sections, quota, stash]`,[`crate::codex_quota_injector`] 的占位逻辑做了配合: +//! 当 `#cat-stash-entry` 在场时,quota 面板把自己 `insertBefore` 到 stash 之前(否则两者 +//! 每 tick 争抢 lastElementChild)。 + +use futures::{SinkExt, StreamExt}; +use serde_json::json; +use tokio::time::Duration; +use tokio_tungstenite::{connect_async, tungstenite::Message as WsMessage}; + +use crate::codex_theme_injector::{drain_until_response, locate_main_window_ws, make_msg}; + +const INSTALL_SCRIPT: &str = r##" +(function() { + // 版本化幂等 guard(对齐 quota injector):同版本仅补一次 ensure 后返回;版本变(应用升级 + // 后本脚本改了)→ 先拆旧 observer + 注入节点再重装,新逻辑免重启 Codex 即覆盖旧注入。 + var VERSION = 10; // review r7:不支持附件(文件/非 data: 图)统一拦 push/restore/send + 配额溢出整体中止 + if (window.__catStashInstalled) { + if (window.__catStashVersion === VERSION) { try { window.__catStashEnsure && window.__catStashEnsure(); } catch (e) {} return; } + try { if (window.__catStashObserver) window.__catStashObserver.disconnect(); } catch (e) {} + ['cat-stash-entry', 'cat-stash-bar', 'cat-stash-menu', 'cat-stash-style'].forEach(function(id){ var n = document.getElementById(id); if (n) n.remove(); }); + } + window.__catStashVersion = VERSION; + window.__catStashSig = null; // 面板 body 重建签名(见 ensurePanel),重装时清空强制首刷 + + // ── 存储(全局 localStorage,跨 reload 存活)── + var KEY = 'catStash'; + // 解析缓存:load() 会 JSON.parse 整个 blob(含图片 dataURL,可达数 MB)。热路径(ensureBar / + // ensurePanel 每 tick + 每次打字经 observer→rAF 调 ensure)若每次都重解析就会在敲键时反复 + // parse 数 MB → 卡顿。用一个会话内单调写计数 __rev 做缓存:仅在 save() 后失效一次,其余调用 + // 直接返回上次解析结果(纯内存,连 localStorage 都不读)。单渲染进程、所有写都走 save(),故 + // 缓存与磁盘一致;reload 后 __rev 归 0、首个 load() 重新解析一次。 + // **不可变契约**:load() 返回的数组视为只读,mutate 它的 caller(stashCurrent)须先 .slice()。 + var __rev = 0, __loadRev = -1, __loadVal = []; + // load 时逐项规整:id/text 强制 string、ts 兜底,屏蔽 schema 漂移或外部写入的非字符串值 + // (否则 renderPanel 的 it.text.replace 会 throw 进而整个面板/按钮静默消失)。 + function load() { + if (__loadRev === __rev) return __loadVal; // 缓存命中:无写发生,直接复用上次解析结果 + var out = []; + try { + var s = localStorage.getItem(KEY); var a = s ? JSON.parse(s) : []; + if (Array.isArray(a)) { + out = a.filter(function (x) { return x && x.id != null; }).map(function (x) { + return { + id: String(x.id), + text: x.text == null ? '' : String(x.text), + ts: x.ts || 0, + // 图片附件:{src:dataURL, filename}(可序列化;只留带 data: 前缀的)。文件附件注入侧无法 + // 干净恢复(setComposerStateField 不可达 / 合成 drop 无 fs path),不纳入,见 followup CAT-260。 + images: Array.isArray(x.images) ? x.images + .filter(function (im) { return im && typeof im.src === 'string' && im.src.indexOf('data:') === 0; }) + .map(function (im) { return { src: im.src, filename: String(im.filename || 'image.png') }; }) : [] + }; + }); + } + } catch (e) { out = []; } + __loadVal = out; __loadRev = __rev; + return out; + } + // save 返回是否成功:图片 dataURL 较大,localStorage 配额(~5-10MB)可能溢出,caller 据此降级 + // (绝不在保存失败时清空 composer 丢内容)。写成功才 bump __rev 让缓存失效(失败则缓存仍有效)。 + function save(a) { try { localStorage.setItem(KEY, JSON.stringify(a)); __rev++; return true; } catch (e) { return false; } } + // id 用「ms + 会话内单调序号 + 随机」:序号保证同毫秒/同 random 也不碰撞(碰撞会让 + // filter(x=>x.id!==id) 误删两条)。 + var __seq = 0; + function uid() { return Date.now().toString(36) + '-' + (++__seq).toString(36) + Math.random().toString(36).slice(2, 6); } + function findItem(arr, id) { for (var i = 0; i < arr.length; i++) { if (arr[i].id === id) return arr[i]; } return null; } + + // ── composer 操作(.ProseMirror,选择子均 CDP 实证)── + function composerEl() { return document.querySelector('.ProseMirror'); } + // 取草稿全文(保留换行):ProseMirror 多行草稿是单 textblock 内的 hard break(
)/多 block, + // textContent 把文本节点直接拼接、丢掉所有换行(`a
b` → "ab")。用 innerText 还原渲染所见 + // 的换行(块边界 /
→ \n),不可用时回退 textContent。**捕获/恢复/校验三处都走它**保持同一 + // 文本空间;空值判断用更廉价的 composerTextRaw(见 ensureBar 热路径)。 + function composerText() { var e = composerEl(); if (!e) return ''; var v = e.innerText; return (v != null ? v : (e.textContent || '')); } + // 仅判空(textContent 即可,无需 innerText 触发 reflow):热路径用,避免每 tick reflow。 + function composerTextRaw() { var e = composerEl(); return e ? (e.textContent || '') : ''; } + // 暂存用的规范文本:顶层块(`

` 等)即「行」,各块 textContent 为该行文本(本 composer 无 + // hard break,Shift+Enter = 新块);按**单** `\n` join。空块 = 空行,故**保留用户的有意空行**。 + // 这恰是 paste 期望的「每换行一个 `\n`」规范 → restore paste 重建同构块,往返幂等(CDP 实证: + // `a\n\nb` 空行往返不丢)。**不可用 innerText**(段间 `\n\n`、空段落 `\n` 数非线性,且 paste 回去 + // 会翻倍膨胀)、**不可用 textContent**(丢全部换行 → 多行拼成一行)。 + // 单块取文本:块内**非 trailing** `
` → `\n`(防御富文本粘贴/未来版本在块内产生 hard break; + // CDP 实证本 composer 把 `
`/HTML 断行规整成独立段落,正常不触发此分支);否则直接 textContent + // (空块 = 空行,textContent 为 ''→ blank line 保留;用 innerText 会让空块变 '\n' 破坏空行)。 + function blockText(b) { + if (b.querySelector && b.querySelector('br:not(.ProseMirror-trailingBreak)')) { + var c = b.cloneNode(true), brs = c.querySelectorAll('br:not(.ProseMirror-trailingBreak)'); + for (var j = 0; j < brs.length; j++) { brs[j].parentNode.replaceChild(document.createTextNode('\n'), brs[j]); } + return c.textContent || ''; + } + return b.textContent || ''; + } + function captureText() { + var e = composerEl(); if (!e) return ''; + var ch = e.children; + if (!ch || !ch.length) return (e.innerText != null ? e.innerText : (e.textContent || '')); + var lines = []; + for (var i = 0; i < ch.length; i++) lines.push(blockText(ch[i])); + return lines.join('\n'); + } + // 写入校验比较:两侧都用块级规范文本(captureText)比对,**保留全部空行(含前导/尾随)**:只归一 + // `\r` + 去各行尾的水平空白(空格/Tab,用户多半无意、PM 可能规整),**绝不折叠或裁剪任何 `\n`**。 + // 折叠/裁剪换行会让 `a\n\nb`→`a\nb`、或丢前导空行也判通过 → 消费/发送被破坏草稿(见 review r5/r6)。 + function sameText(a, b) { + function f(s) { return (s == null ? '' : String(s)).replace(/\r/g, '').replace(/[ \t]+$/gm, ''); } + return f(a) === f(b); + } + function fiberOf(el) { if (!el) return null; for (var k in el) { if (k.indexOf('__reactFiber$') === 0) return el[k]; } return null; } + // 把文本合成 text/plain paste 进 composer(ProseMirror 的 paste 路径把 \n→hard break、\n\n→段落, + // 是多行最可靠的注入方式,与图片合成 paste 同机制)。 + function pasteText(pm, text) { + try { + var dt = new DataTransfer(); dt.setData('text/plain', text); + var ev = new ClipboardEvent('paste', { bubbles: true, cancelable: true }); + Object.defineProperty(ev, 'clipboardData', { value: dt }); + pm.dispatchEvent(ev); + } catch (x) {} + } + // 设输入框内容:focus → 全选删除(CDP 实证可靠清空)→ 写入(空则只清)。**返回是否真落值** + // (caller 据此决定是否消费 stash,绝不在没落值时丢草稿):目标空 → 须真空;目标非空 → 块级规范 + // 文本须与目标一致(sameText,**保留空行**;不折叠换行,否则丢空行也判通过 → 消费被破坏草稿)。 + // **写入机制实证(CDP)**:本 composer 的 Shift+Enter = 新建 `

` 段落(非 `
` hard break), + // 多行草稿即多个 `

`。text/plain 合成 paste 恰好按 `\n` 切段、重建同构 `

`,**稳定、与原草稿 + // 同构、空行不丢**,故作首选。execCommand insertLineBreak 会插入裸 `\n` 被 PM 规整成空格(丢换行), + // **不可用**;回退仅用 insertText / InputEvent。校验用块级 captureText(与捕获/存储同一空间)。 + function setComposer(text) { + var e = composerEl(); if (!e) return false; + try { e.focus(); } catch (x) {} + try { document.execCommand('selectAll', false, null); } catch (x) {} + try { document.execCommand('delete', false, null); } catch (x) {} + if (text) { + pasteText(e, text); // 首选:多行安全(重建

段落,与 Shift+Enter 同构,空行保留) + if (!sameText(captureText(), text)) { // 回退 1:execCommand insertText(paste 被拦时;单行可靠) + try { document.execCommand('selectAll', false, null); document.execCommand('delete', false, null); } catch (x) {} + try { document.execCommand('insertText', false, text); } catch (x) {} + } + if (!sameText(captureText(), text)) { // 回退 2:单次 InputEvent + try { + e.dispatchEvent(new InputEvent('beforeinput', { inputType: 'insertText', data: text, bubbles: true, cancelable: true })); + e.dispatchEvent(new InputEvent('input', { inputType: 'insertText', data: text, bubbles: true })); + } catch (x) {} + } + } + var now = captureText(); + return text ? (now.trim() !== '' && sameText(now, text)) : (now.trim() === ''); + } + // 提交:优先点 compose 按钮,但**运行态(按钮为 Stop)不点**——点它会停掉当前轮次; + // 改走 fiber handleSubmit()(空闲时发送,运行时入队为 steer,符合「运行中也能发送」语义)。 + // disabled 同样跳过点按钮、回退 fiber。 + function submitComposer() { + var btn = document.querySelector('button[class*="size-token-button-compose"]'); + var running = !!(btn && /stop/i.test(btn.getAttribute('aria-label') || '')); + if (btn && !btn.disabled && !running) { try { btn.click(); return true; } catch (e) {} } + var pm = composerEl(); if (!pm) return false; + var el = pm, f = null; while (el && !(f = fiberOf(el))) el = el.parentElement; + var d = 0; while (f && d < 24) { var p = f.memoizedProps; if (p && typeof p.handleSubmit === 'function') { try { p.handleSubmit(); return true; } catch (e) { return false; } } f = f.return; d++; } + return false; + } + + // ── 图片附件(CDP 实证:附件行上方 fiber props 持有 imageAttachments + onRemoveImage)── + // 从附件行 / composer 往上爬 fiber,拿持有 imageAttachments 数组的 props。无附件时附件行不渲染 + // → 返回 null(视作无图)。 + function composerAtt() { + var anchor = document.querySelector('[data-composer-attachments-row]') || composerEl(); + if (!anchor) return null; + var el = anchor, f = null; while (el && !(f = fiberOf(el))) el = el.parentElement; + var d = 0; + while (f && d < 40) { + var p = f.memoizedProps; + if (p && typeof p === 'object' && Array.isArray(p.imageAttachments)) { + return { images: p.imageAttachments, onRemoveImage: typeof p.onRemoveImage === 'function' ? p.onRemoveImage : null }; + } + f = f.return; d++; + } + return null; + } + // 读当前图片为可序列化 {src(dataURL), filename}(只留 dataURL 的,能恢复的才存) + function readImages() { + var a = composerAtt(); if (!a || !a.images) return []; + return a.images + .filter(function (im) { return im && typeof im.src === 'string' && im.src.indexOf('data:') === 0; }) + .map(function (im) { return { src: im.src, filename: im.filename || 'image.png' }; }); + } + // 清掉 composer 里**可序列化的(data:)**图片(那些已被 readImages 存进 stash 的);**不动**非 data: + // 图片(如 blob 上传)—— readImages 存不了它们,这里也不删,避免「没存上又删掉 = 丢图」(见 review)。 + // 也**不动文件**(文件不在 stash 范围,见 CAT-260)。 + function clearImages() { + var a = composerAtt(); if (!a || !a.onRemoveImage || !a.images) return; + a.images.forEach(function (im) { + if (im && typeof im.src === 'string' && im.src.indexOf('data:') === 0) { + try { a.onRemoveImage(im.id); } catch (e) {} + } + }); + } + // 把保存的图片合成 paste 回 composer(atob 解 base64→File→ClipboardEvent;fetch(dataURL) 被 CSP 挡) + function addImages(imgs) { + if (!imgs || !imgs.length) return; + var pm = composerEl(); if (!pm) return; + try { pm.focus(); } catch (e) {} + imgs.forEach(function (im) { + try { + var arr = im.src.split(','), mime = (arr[0].match(/:(.*?);/) || [])[1] || 'image/png'; + var bstr = atob(arr[1] || ''), n = bstr.length, u8 = new Uint8Array(n); + while (n--) u8[n] = bstr.charCodeAt(n); + var file = new File([u8], im.filename || 'image.png', { type: mime }); + var dt = new DataTransfer(); dt.items.add(file); + var ev = new ClipboardEvent('paste', { bubbles: true, cancelable: true }); + Object.defineProperty(ev, 'clipboardData', { value: dt }); + pm.dispatchEvent(ev); + } catch (e) {} + }); + } + // 合成 paste 是异步落值(FileReader+persist),发送前轮询图片就位(最多 ~2.5s)。 + // **回调带 ok**:就位 → cb(true);超时 → cb(false)。caller 必须区分 —— 超时不能照发(会发出 + // 不带图片的消息),更不能在确认前消费 stash。 + function waitImages(target, cb) { + var tries = 0; + var iv = setInterval(function () { + tries++; + if (readImages().length >= target) { clearInterval(iv); cb(true); } + else if (tries > 25) { clearInterval(iv); cb(false); } + }, 100); + } + // composer 是否还挂着**任何**附件(图片或文件)。附件行 `[data-composer-attachments-row]` 是 + // **条件渲染**的:有任意附件才在 DOM,空态整行移除(CDP 实证)。故它在场 = 还有附件。比只数 + // imageAttachments 更全 —— **也能挡住文件**:文件不在 stash 范围(curEntry 不存、clearImages + // 不清,见 CAT-260),但 send 不该把残留文件一起发出(见 review),所以清空判定必须把它算进去。 + function hasAttachments() { return !!document.querySelector('[data-composer-attachments-row]'); } + // 全部附件数(图片 + 文件):每个附件 = 附件行 flex 容器(`row.firstElementChild`)下的一个直接 + // 子节点(CDP 实证:1 图→1 子、2 图→2 子)。**不能**用 `querySelectorAll('.composer-attachment- + // surface')` —— 该类有嵌套,每个附件命中 2 个(实证 1 图→4 命中错误)。 + function attTotal() { + var row = document.querySelector('[data-composer-attachments-row]'); + var inner = row && row.firstElementChild; + return inner ? inner.children.length : 0; + } + // composer 是否挂着**任何不可暂存的附件**。可序列化的只有 data: 图(readImages)。 + // imageAttachments 总数 > data: 图数 → 有非 data: 图(blob / 上传中,readImages 存不了); + // 总附件数 > 图片附件数 → 有文件。任一为真 = 有不支持附件,push/restore/send 须中止保草稿 + // (否则只存一部分、把不支持的附件孤儿化、被下一条原生发送带走,见 review r6/r7)。 + function unsupportedAtt() { + var a = composerAtt(); + var imgN = (a && a.images) ? a.images.length : 0; + return (imgN > readImages().length) || (attTotal() > imgN); + } + // 图片恢复/发送的完整时序(确保只带对的图、且写失败/未就位时不丢暂存),纯文本也走它(等旧附件清空): + // 1) 先等当前附件(图片**和**文件)全部清空(clearImages 异步;否则旧图仍在 React 态、或残留 + // 文件,submit 会把它们一起发出,见 review)——附件行消失才 2);文件清不掉 → 超时安全中止。 + // 2) paste 目标图(纯文本时 imgs=[] → 无操作);3) 等目标图就位(imgs=[] 时 waitImages(0) 立即 true)。 + // 成功 cb(true);清空或就位超时 cb(false)(caller 据此不消费/不发)。 + function settleImages(imgs, cb) { + var n = 0; + var iv = setInterval(function () { + n++; + if (!hasAttachments()) { clearInterval(iv); addImages(imgs); waitImages(imgs.length, cb); } + else if (n > 25) { clearInterval(iv); cb(false); } // 旧附件清空超时(如残留文件)= 失败,不消费/不发 + }, 100); + } + // 最小 toast(暂存失败/降级时告知;无侵入,2.5s 自removed) + function notify(msg) { + try { + var t = el('div', null, msg); t.id = 'cat-stash-toast'; + t.style.cssText = 'position:fixed;left:50%;bottom:80px;transform:translateX(-50%);z-index:99999;padding:8px 14px;border-radius:8px;background:rgba(20,24,36,.92);color:#fff;font-size:13px;box-shadow:0 6px 24px rgba(0,0,0,.4);pointer-events:none'; + var old = document.getElementById('cat-stash-toast'); if (old) old.remove(); + document.body.appendChild(t); + setTimeout(function () { try { t.remove(); } catch (e) {} }, 2500); + } catch (e) {} + } + + // ── 核心动作 ── + // restore/sendItem 共享 composer 且异步(settleImages):全局串行锁,防同条重复点 / 不同条并发 / + // 处理期间 push 互相 clobber composer + 重复提交(见 review)。每条路径(含早退)必复位。 + var __busy = false; + // 当前 composer 内容(文本 + 图片)打包成一条 swap 用的 entry;全空返 null + function curEntry() { + var t = captureText(), imgs = readImages(); + if ((!t || !t.trim()) && !imgs.length) return null; + return { id: uid(), text: t || '', ts: Date.now(), images: imgs }; + } + // 当前内容 swap 进 stash 但**暂不消费目标项**:先存一份「原列表 + cur」(cur 不丢),目标项 + // 仍在列表里;真正落值/发送成功后再单独移除目标项。返回该 swap 是否 save 成功(配额够)。 + function swapInCurrent(arr) { + var cur = curEntry(); + if (!cur) return true; // 输入框空,无需 swap + var next = arr.slice(); next.push(cur); + return save(next); + } + function consumeItem(id) { save(load().filter(function (x) { return x.id !== id; })); } + // 恢复:把目标(文本+图片)放进 composer;输入框非空则先把当前内容 swap 进 stash(不丢)。 + // **先 swap-save(配额够)再动 composer**——保存失败就安全中止。**setComposer 真落值才消费目标项**; + // 有图时**进一步等图片真就位**才消费(写/paste 失败 → 目标项保留在 stash,绝不出现「消费了却没 + // 完整落到输入框」)。 + function restore(id) { + if (__busy) return; // 处理中,忽略重复点(防 double restore/send,见 review) + var arr = load(); var item = findItem(arr, id); if (!item) return; + if (unsupportedAtt()) { notify('输入框含不支持的附件(文件/非 data: 图),无法恢复;请先移除'); return; } // swap 前拦,保草稿不拆分 + __busy = true; + if (!swapInCurrent(arr)) { __busy = false; notify('暂存空间不足,恢复已取消'); return; } // composer 还没动,安全 + clearImages(); // 清掉当前 data: 图片(已 swap 进 stash);不动非 data: / 文件 + if (!setComposer(item.text)) { __busy = false; notify('写入输入框失败,恢复已取消'); closeMenu(); ensure(); return; } // 目标项仍在 stash + // 纯文本也走 settleImages([]):先等旧附件清空再消费(避免恢复后还残留上一稿的图/文件);有图则等 paste 就位。 + settleImages(item.images || [], function (ok) { + if (ok) consumeItem(id); else notify('图片未就位,已保留暂存'); + __busy = false; closeMenu(); ensure(); + }); + } + // 发送:同 restore 把目标放进 composer。**仅在文本落值 + 图片就位后才消费并提交**: + // 写失败 → 不消费、不提交;图片超时未就位 → 不消费、不提交(避免发出不带图片的消息),目标项 + // 留在 stash 可重试。失败均无损(当前内容已 swap 进 stash,目标项仍在)。 + function sendItem(id) { + if (__busy) return; // 处理中,忽略重复点(防重复发送,见 review) + var arr = load(); var item = findItem(arr, id); if (!item) return; + if (unsupportedAtt()) { notify('输入框含不支持的附件(文件/非 data: 图),无法发送;请先移除'); return; } // swap 前拦,保草稿不拆分 + __busy = true; + if (!swapInCurrent(arr)) { __busy = false; notify('暂存空间不足,发送已取消'); return; } + clearImages(); + if (!setComposer(item.text)) { __busy = false; notify('写入输入框失败,发送已取消'); ensure(); return; } // 目标项仍在 stash + // 纯文本也走 settleImages([]):**先等旧附件清空再提交**,否则上一稿的图/文件(异步未清完)会被一起 + // 发出(见 review);有图则等目标图 paste 就位。**提交成功才消费**(submitComposer 返回 false——按钮/ + // fiber 路径变化或暂不可提交——则保留暂存,不静默吞掉,见 review);失败均保留可重试。 + settleImages(item.images || [], function (ok) { + if (!ok) notify('图片未就位,已保留暂存(未发送)'); + else if (submitComposer()) consumeItem(id); + else notify('发送失败,已保留暂存'); + __busy = false; ensure(); + }); + } + function del(id) { save(load().filter(function (x) { return x.id !== id; })); ensure(); } + // 暂存当前(文本+图片)→ 清空输入框。**先 save 成功才清**,避免清了没存上丢内容。 + // load() 返回只读缓存,故先 .slice() 再 push(不污染缓存)。 + function stashCurrent() { + if (__busy) return; // restore/send 异步进行中,别动 composer(见 review) + var t = captureText(), imgs = readImages(); + if ((!t || !t.trim()) && !imgs.length) return; + // 含不支持的附件(文件 / 非 data: 图):中止 push,保草稿不动 —— 它们无法暂存/恢复(CAT-260), + // 只存一部分会把不支持的附件孤儿化、被下一条原生发送带走(见 review)。提示用户先移除。 + if (unsupportedAtt()) { notify('草稿含不支持的附件(文件/非 data: 图),暂不支持暂存;请先移除'); return; } + var entry = { id: uid(), text: t || '', ts: Date.now(), images: imgs }; + var arr = load().slice(); arr.push(entry); + if (!save(arr)) { + // 配额溢出(图片 dataURL 太大):**整体中止、不动 composer**,保草稿完整(文字+图片)不拆分。 + // 不再降级为「只存文字、图片留输入框」—— 那会把草稿拆成 stash 文字 + 残留图,残留图还会被下一条 + // 原生发送带走(见 review)。 + notify('暂存失败:存储空间不足(图片过大?草稿保持不变)'); return; + } + if (t && t.trim()) setComposer(''); // 存成功才清:文字 + 图片 + clearImages(); + // 刚暂存若含图:等附件真正清空再解锁(clearImages 异步;否则刚暂存的图还挂着,用户紧接着发送的 + // 插话会把它一起带出去,见 review)。__busy 期间我方 push/restore/send 不可再入。 + if (imgs.length) { + __busy = true; + var n = 0; + var iv = setInterval(function () { n++; if (!hasAttachments() || n > 25) { clearInterval(iv); __busy = false; ensure(); } }, 100); + } + ensure(); + } + + // ── 样式(scoped,跟随 Codex 主题 token + 注入主题 accent)── + function ensureStyle() { + if (document.getElementById('cat-stash-style')) return; + var st = document.createElement('style'); st.id = 'cat-stash-style'; + st.textContent = + // 面板(对齐 quota 面板视觉) + '#cat-stash-entry{display:block;padding:0 0 6px;user-select:none}' + + '#cat-stash-entry .cshdr{display:flex;align-items:center;height:28px;padding:0 10px 2px 16px;background:var(--color-token-dropdown-background,rgba(20,24,36,.78))}' + + '#cat-stash-entry .csbtn{display:inline-flex;align-items:center;gap:6px;cursor:pointer;border-radius:6px;padding:2px 4px 2px 0}' + + '#cat-stash-entry .cstt{font-size:14px;font-weight:430;color:var(--color-token-text-tertiary,rgba(238,241,247,.56))}' + + '#cat-stash-entry .cscount{font-size:12px;color:var(--color-token-text-tertiary,rgba(238,241,247,.45));font-variant-numeric:tabular-nums}' + + '#cat-stash-entry .cschev{width:14px;height:14px;opacity:0;transition:opacity .12s ease,transform .15s ease}' + + '#cat-stash-entry .csbtn:hover .cschev{opacity:1}' + + '#cat-stash-entry.cscol .cschev{transform:rotate(-90deg)}' + + '#cat-stash-entry.cscol .csbody{display:none}' + + '#cat-stash-entry .csbody{padding-top:3px;display:flex;flex-direction:column;gap:3px}' + + '#cat-stash-entry .csempty{padding:5px 16px;font-size:12.5px;color:var(--color-token-text-tertiary,rgba(238,241,247,.45))}' + + '#cat-stash-entry .csrow{display:flex;align-items:center;gap:8px;padding:5px 16px}' + + '#cat-stash-entry .cstext{flex:1 1 auto;min-width:0;font-size:13px;color:var(--color-token-text-primary,#ededed);white-space:nowrap;overflow:hidden;text-overflow:ellipsis}' + + '#cat-stash-entry .csact{flex:0 0 auto;display:inline-flex;align-items:center;justify-content:center;width:22px;height:22px;border-radius:5px;cursor:pointer;color:var(--color-token-text-secondary,#8c8782);opacity:.75}' + + '#cat-stash-entry .csact:hover{opacity:1;background:rgba(128,128,128,.16);color:var(--cl-accent,#6c83c4)}' + + '#cat-stash-entry .csact svg{width:14px;height:14px}' + + // 行内图片计数 chip + '#cat-stash-entry .csimg{flex:0 0 auto;display:inline-flex;align-items:center;gap:3px;font-size:11px;color:var(--color-token-text-tertiary,rgba(238,241,247,.5));font-variant-numeric:tabular-nums}' + + '#cat-stash-entry .csimg svg{width:13px;height:13px}' + + // composer 工具栏按钮组 + '#cat-stash-bar{display:inline-flex;align-items:center;gap:4px;margin-right:4px}' + + '#cat-stash-bar button{position:relative;display:inline-flex;align-items:center;justify-content:center;width:30px;height:30px;border-radius:8px;border:1px solid var(--color-token-border,rgba(128,128,128,.28));background:transparent;color:var(--color-token-text-secondary,#8c8782);cursor:pointer;padding:0}' + + '#cat-stash-bar button:hover:not(:disabled){color:var(--cl-accent,#6c83c4);border-color:color-mix(in srgb,var(--cl-accent,#6c83c4) 40%,transparent)}' + + '#cat-stash-bar button:disabled{opacity:.35;cursor:default}' + + '#cat-stash-bar button svg{width:16px;height:16px}' + + '#cat-stash-bar .csbadge{position:absolute;top:-4px;right:-4px;min-width:14px;height:14px;padding:0 3px;border-radius:7px;background:var(--cl-accent,#6c83c4);color:#fff;font-size:9px;line-height:14px;text-align:center;font-variant-numeric:tabular-nums}' + + // 下拉选单 + '#cat-stash-menu{position:fixed;z-index:99999;min-width:220px;max-width:360px;max-height:300px;overflow-y:auto;padding:4px;border-radius:10px;background:var(--color-token-dropdown-background,#23262e);border:1px solid var(--color-token-border,rgba(128,128,128,.3));box-shadow:0 8px 28px rgba(0,0,0,.4)}' + + '#cat-stash-menu .csmi{display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:6px;cursor:pointer;font-size:13px;color:var(--color-token-text-primary,#ededed)}' + + '#cat-stash-menu .csmi:hover{background:rgba(128,128,128,.16)}' + + '#cat-stash-menu .csmi .t{flex:1 1 auto;min-width:0;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}'; + (document.head || document.documentElement).appendChild(st); + } + + function el(tag, cls, txt) { var e = document.createElement(tag); if (cls) e.className = cls; if (txt != null) e.textContent = txt; return e; } // textContent sink,杜绝 XSS + var ICON = { + push: '', + pop: '', + restore: '', + send: '', + del: '', + image: '' + }; + function iconBtn(cls, ic, title, onClick) { + var b = el('span', cls); b.innerHTML = ICON[ic]; b.title = title; + b.addEventListener('click', function (ev) { ev.stopPropagation(); onClick(); }); + return b; + } + + // ── 折叠态(localStorage 记忆)── + function isCollapsed() { try { return localStorage.getItem('catStashCollapsed') === '1'; } catch (e) { return false; } } + function setCollapsed(v) { try { localStorage.setItem('catStashCollapsed', v ? '1' : '0'); } catch (e) {} } + + // ── 面板挂载点(同 quota 的 pinned-summary scroller)── + // 注意:此选择子与 codex_quota_injector 的 findScroller 完全一致,是 Codex DOM 耦合点(Codex + // 升级改 section header 标记即失效)。**改这里务必同步改 codex_quota_injector 的同名函数**,否则 + // 升级后会出现「Usage 面板正常、Stash 面板 + 按钮静默消失」的半坏态。 + function findScroller() { + var btns = document.querySelectorAll('section header button[class~="group/section-toggle"]'); + for (var i = 0; i < btns.length; i++) { var sec = btns[i].closest('section'); if (sec && sec.parentElement) return sec.parentElement; } + return null; + } + var CHEV = ''; + function renderPanel(node, arr) { + node.textContent = ''; + node.classList.toggle('cscol', isCollapsed()); + var h = el('div', 'cshdr'); + var btn = el('span', 'csbtn'); + btn.appendChild(el('span', 'cstt', 'Stash')); + btn.appendChild(el('span', 'cscount', arr.length ? String(arr.length) : '')); + var cw = document.createElement('span'); cw.innerHTML = CHEV; if (cw.firstChild) btn.appendChild(cw.firstChild); + btn.addEventListener('click', function () { var col = !node.classList.contains('cscol'); node.classList.toggle('cscol', col); setCollapsed(col); }); + h.appendChild(btn); node.appendChild(h); + var body = el('div', 'csbody'); + if (!arr.length) { + body.appendChild(el('div', 'csempty', '暂无暂存草稿')); + } else { + arr.forEach(function (it) { + var row = el('div', 'csrow'); + var nimg = (it.images || []).length; + var label = (it.text || '').replace(/\s+/g, ' ').trim(); + if (!label && nimg) label = '(图片)'; + var t = el('div', 'cstext', label); t.title = it.text || ''; + row.appendChild(t); + if (nimg) { + var chip = el('span', 'csimg'); chip.innerHTML = ICON.image; + chip.appendChild(el('span', null, String(nimg))); chip.title = nimg + ' 张图片'; + row.appendChild(chip); + } + row.appendChild(iconBtn('csact', 'restore', '恢复到输入框', function () { restore(it.id); })); + row.appendChild(iconBtn('csact', 'send', '直接发送', function () { sendItem(it.id); })); + row.appendChild(iconBtn('csact', 'del', '删除', function () { del(it.id); })); + body.appendChild(row); + }); + } + node.appendChild(body); + } + function ensurePanel() { + var scroller = findScroller(); + var node = document.getElementById('cat-stash-entry'); + if (!scroller) { if (node) node.remove(); window.__catStashSig = null; return; } + var fresh = !node; + if (!node) { node = document.createElement('section'); node.id = 'cat-stash-entry'; } + // 顺序:始终置于 scroller 末尾(在 quota 之后)。quota injector 配合把自己 insertBefore 到 stash 前。 + if (node.parentElement !== scroller || node !== scroller.lastElementChild) { scroller.appendChild(node); } + // 仅在数据变化(或新建)时重建 body:renderPanel 做 textContent='' + 重建子节点(childList + // 变更),本模块 observer 监听 body childList → 若每 tick 无条件重建就是 ~60fps 自循环(空闲也跑)。 + // 用签名 guard 跳过无变化重建(对齐 quota 的 __catQuotaSig)。折叠态由 header 点击直接 toggle class, + // 不经重建,故不入签名。 + // 廉价签名:条目不可变(只增删/消费),id 列表 + 各自图片数即可表征变化;**不** stringify 整个 + // load()(图片 dataURL 可达数 MB,每 tick stringify 会很热)。load() 已带解析缓存(见上), + // 此处与 renderPanel 共用同一份(不重复解析)。 + var arr = load(); + var sig = arr.map(function (e) { return e.id + '#' + ((e.images || []).length); }).join(','); + if (fresh || sig !== window.__catStashSig) { renderPanel(node, arr); window.__catStashSig = sig; } + } + + // ── 下拉选单(pop 多条时)── + function closeMenu() { var m = document.getElementById('cat-stash-menu'); if (m) m.remove(); document.removeEventListener('mousedown', onDocDown, true); } + function onDocDown(ev) { var m = document.getElementById('cat-stash-menu'); if (m && !m.contains(ev.target)) closeMenu(); } + function openMenu(anchor) { + closeMenu(); + var arr = load(); if (!arr.length) return; + var m = el('div'); m.id = 'cat-stash-menu'; + arr.forEach(function (it) { + var mi = el('div', 'csmi'); + var sp = document.createElement('span'); sp.innerHTML = ICON.restore; sp.style.opacity = '.7'; sp.style.flex = '0 0 auto'; + var svg = sp.firstChild; if (svg) { svg.setAttribute('width', '14'); svg.setAttribute('height', '14'); mi.appendChild(svg); } + var lbl = (it.text || '').replace(/\s+/g, ' ').trim(); + var ni = (it.images || []).length; + if (!lbl && ni) lbl = '(图片)'; + if (ni) lbl += ' ·图' + ni; + mi.appendChild(el('span', 't', lbl)); + mi.title = it.text || ''; + mi.addEventListener('click', function (ev) { ev.stopPropagation(); restore(it.id); }); + m.appendChild(mi); + }); + document.body.appendChild(m); + var r = anchor.getBoundingClientRect(); + // 弹在按钮上方(composer 在底部),右对齐按钮 + var top = Math.max(8, r.top - m.offsetHeight - 6); + var left = Math.max(8, Math.min(r.right - m.offsetWidth, window.innerWidth - m.offsetWidth - 8)); + m.style.top = top + 'px'; m.style.left = left + 'px'; + setTimeout(function () { document.addEventListener('mousedown', onDocDown, true); }, 0); + } + + // ── composer 工具栏按钮组(push + pop)── + function ensureBar() { + var send = document.querySelector('button[class*="size-token-button-compose"]'); + if (!send || !send.parentElement) { var b = document.getElementById('cat-stash-bar'); if (b) b.remove(); return; } + var host = send.parentElement; // div.flex.shrink-0(Dictate + Send 同排,CDP 实证) + var bar = document.getElementById('cat-stash-bar'); + if (!bar) { + bar = el('span'); bar.id = 'cat-stash-bar'; + var push = el('button'); push.innerHTML = ICON.push; bar.appendChild(push); + var pop = el('button'); pop.innerHTML = ICON.pop; + var badge = el('span', 'csbadge'); pop.appendChild(badge); + bar.appendChild(pop); + push.addEventListener('click', function (ev) { ev.preventDefault(); ev.stopPropagation(); stashCurrent(); }); + pop.addEventListener('click', function (ev) { + ev.preventDefault(); ev.stopPropagation(); + var arr = load(); if (!arr.length) return; + if (arr.length === 1) restore(arr[0].id); else openMenu(pop); + }); + bar.__push = push; bar.__pop = pop; bar.__badge = badge; + } + if (bar.parentElement !== host || host.firstChild !== bar) { host.insertBefore(bar, host.firstChild); } + // 状态刷新(仅变化时写,避免自触发 observer churn)。load() 带解析缓存,热路径不重复 parse。 + var arr = load(); + // 有文字或有图片(均走廉价检查:文字用 textContent 判空而非 innerText 触发 reflow;图片查 + // 附件行里的 img 缩略图,避免每 tick 爬 fiber)即可暂存 + var hasImg = !!document.querySelector('[data-composer-attachments-row] img'); + var pushDisabled = !(composerTextRaw().trim()) && !hasImg; + if (bar.__push.disabled !== pushDisabled) bar.__push.disabled = pushDisabled; + bar.__push.title = '暂存当前输入(文字 + 图片)'; + var popShow = arr.length > 0 ? '' : 'none'; + if (bar.__pop.style.display !== popShow) bar.__pop.style.display = popShow; + bar.__pop.title = arr.length > 1 ? ('恢复暂存(' + arr.length + ' 条)') : '恢复暂存'; + var cnt = arr.length > 1 ? String(arr.length) : ''; + if (bar.__badge.textContent !== cnt) bar.__badge.textContent = cnt; + bar.__badge.style.display = cnt ? '' : 'none'; + } + + function ensure() { try { ensureStyle(); ensurePanel(); ensureBar(); } catch (e) {} } + window.__catStashEnsure = ensure; + + // rAF 合并的 observer:DOM 变更(含 composer 文本变化 / React re-render 驱逐注入节点)即 re-ensure。 + var scheduled = false; + var mo = new MutationObserver(function () { if (scheduled) return; scheduled = true; requestAnimationFrame(function () { scheduled = false; ensure(); }); }); + mo.observe(document.body, { childList: true, subtree: true, characterData: true }); + window.__catStashObserver = mo; + window.__catStashInstalled = true; + ensure(); +})(); +"##; + +/// 卸载脚本:断 observer、删注入 DOM(面板/工具栏按钮/下拉/style)、清全局态。幂等。 +const REMOVE_SCRIPT: &str = r#" +(function() { + try { if (window.__catStashObserver) window.__catStashObserver.disconnect(); } catch (e) {} + ['cat-stash-entry', 'cat-stash-bar', 'cat-stash-menu', 'cat-stash-style'].forEach(function(id){ var n = document.getElementById(id); if (n) n.remove(); }); + try { + delete window.__catStashInstalled; delete window.__catStashVersion; + delete window.__catStashObserver; delete window.__catStashEnsure; delete window.__catStashSig; + } catch (e) {} +})(); +"#; + +/// 读 registry config 的 `settings.codexStashEnabled`(默认关)。对齐 quota 的 `quota_enabled`。 +fn stash_enabled() -> bool { + crate::admin::registry_io::load() + .ok() + .as_ref() + .and_then(|c| c.get("settings")) + .and_then(|s| s.get("codexStashEnabled")) + .and_then(serde_json::Value::as_bool) + .unwrap_or(false) +} + +/// CDP push 失败分类(对齐 quota injector 的 `PushError`): +/// - `Connect`:Codex 没跑 / 端口未就绪 = 注入还连不上,**常态**,不告警。 +/// - `Evaluate`:已连上但发送 / 注入 JS 抛错(选择子随 Codex 升级失效等)= 真异常,warn-once。 +enum StashPushError { + Connect(String), + Evaluate(String), +} + +/// 一次 CDP `Runtime.evaluate`。复用 [`crate::codex_theme_injector`] 的共享 CDP helper。 +async fn evaluate_once(script: &str) -> Result<(), StashPushError> { + let ws_url = locate_main_window_ws() + .await + .map_err(|e| StashPushError::Connect(e.to_string()))?; + let (ws_stream, _) = connect_async(&ws_url) + .await + .map_err(|e| StashPushError::Connect(e.to_string()))?; + let (mut write, mut read) = ws_stream.split(); + let (msg, _) = make_msg( + 1, + "Runtime.evaluate", + json!({ "expression": script, "returnByValue": true }), + ); + write + .send(WsMessage::Text(msg.into())) + .await + .map_err(|e| StashPushError::Evaluate(e.to_string()))?; + drain_until_response(&mut read, 1) + .await + .map_err(StashPushError::Evaluate)?; + let _ = write.close().await; + Ok(()) +} + +/// Stash 注入 daemon:`main.rs` spawn 一次,独立于 quota 开关。每 tick 推幂等 install(纯保活, +/// 覆盖 Codex reload/导航);关闭时推一次 remove 清残留(成功才复位 needs_remove)。 +/// CDP 不可达(Codex 没跑 / 端口未就绪)→ 静默跳过本 tick(debug 级,常态)。 +pub async fn run_stash_daemon() { + const TICK: Duration = Duration::from_secs(5); + let mut needs_remove = true; + let mut evaluate_warned = false; + loop { + tokio::time::sleep(TICK).await; + if stash_enabled() { + needs_remove = true; + match evaluate_once(INSTALL_SCRIPT).await { + Ok(()) => evaluate_warned = false, + // Connect 失败 = Codex 没跑 / 端口未就绪,常态,只 debug(默认不输出)不刷屏。 + Err(StashPushError::Connect(e)) => { + tracing::debug!(error = %e, "[Stash] CDP 未就绪(Codex 没跑 / 端口未就绪),常态"); + } + // Evaluate 失败 = Codex 在跑但注入异常(选择子失效等),warn-once,便于定位 + // 「Codex 没启动」vs「注入坏了」(refute 这条会丢掉 quota 早就有的诊断信号)。 + Err(StashPushError::Evaluate(e)) => { + if !evaluate_warned { + tracing::warn!(error = %e, "[Stash] 注入 evaluate 失败 — Codex 在跑但注入异常(选择子可能随 Codex 升级失效)"); + evaluate_warned = true; + } + } + } + } else if needs_remove { + match evaluate_once(REMOVE_SCRIPT).await { + Ok(()) => needs_remove = false, + Err(_) => {} // CDP 瞬时不可达 → 保持 needs_remove,下 tick 重试 + } + } + } +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn install_script_has_idempotent_guard() { + assert!(INSTALL_SCRIPT.contains("window.__catStashInstalled")); + assert!(INSTALL_SCRIPT.contains("var VERSION =")); + } + + #[test] + fn remove_script_clears_nodes_and_globals() { + assert!(REMOVE_SCRIPT.contains("cat-stash-entry")); + assert!(REMOVE_SCRIPT.contains("cat-stash-bar")); + assert!(REMOVE_SCRIPT.contains("delete window.__catStashInstalled")); + } + + #[test] + fn raw_scripts_have_no_terminator_collision() { + // r##"..."## / r#"..."# 定界符不能在脚本体里出现 + assert!(!INSTALL_SCRIPT.contains("\"##")); + assert!(!REMOVE_SCRIPT.contains("\"#")); + } +} diff --git a/src-tauri/src/main.rs b/src-tauri/src/main.rs index ea479510..282c70af 100644 --- a/src-tauri/src/main.rs +++ b/src-tauri/src/main.rs @@ -8,6 +8,7 @@ mod anyrouter_quota; mod codex_plugin_unlocker; mod codex_quota_injector; mod codex_real_account; +mod codex_stash_injector; mod codex_theme_injector; mod deepseek_quota; mod glm_quota; @@ -221,6 +222,11 @@ fn main() { // 开关关 / CDP 不可达时 tick 内静默跳过,常驻无负担。 tauri::async_runtime::spawn(codex_quota_injector::run_quota_daemon()); + // 草稿暂存(Stash)注入 daemon:每 tick 读 settings.codexStashEnabled,经 CDP 推 + // 幂等注入(composer 工具栏 push/pop 按钮 + Usage 下方 Stash 面板)。独立于额度开关; + // 开关关 / CDP 不可达时 tick 内静默跳过。 + tauri::async_runtime::spawn(codex_stash_injector::run_stash_daemon()); + // [MOC-231] GC 旧的上下文明细缓存(context-breakdown/.json,每对话一个; // >14 天的陈旧对话删除,下次有请求会重建)。fire-and-forget,不阻塞 startup。 tauri::async_runtime::spawn(async {