ai-browser-bridge 0.7.0 → 0.7.1

package/README.es.md

@@ -179,7 +179,7 @@ Flow requiere un plan **Google AI Pro/Ultra**. Como los renders de Veo tardan mi
 
 ## Soporte de Claude Design
 
-El bridge controla **[Claude Design](https://claude.ai/design)** desde tu Chrome del bridge con sesión iniciada. Las lecturas y los cambios de proyectos/archivos usan las propias peticiones de la app desde la pestaña ([ADR 0021](docs/adr/current/0021-claude-design-app-rpc.md)); los turnos, plantillas y modelo/esfuerzo pasan por la interfaz real, una pestaña por proyecto.
+El bridge controla **[Claude Design](https://claude.ai/design)** desde tu Chrome del bridge con sesión iniciada. Las lecturas y los cambios de proyectos/archivos usan las propias peticiones de la app desde la pestaña; los turnos, plantillas y modelo/esfuerzo pasan por la interfaz real, una pestaña por proyecto.
 
 ```bash
 bridge chrome start --provider design

package/README.he.md

@@ -211,7 +211,7 @@ bridge flow project-delete --yes         # מחיקה לצמיתות של הפר
 
 ## תמיכה ב-Claude Design
 
-ה-bridge מפעיל את **[Claude Design](https://claude.ai/design)** מתוך ה-Chrome של ה-bridge שבו אתם מחוברים. קריאות ושינויים בפרויקטים ובקבצים עוברים דרך הבקשות של האפליקציה עצמה מתוך הלשונית ([ADR 0021](docs/adr/current/0021-claude-design-app-rpc.md)); סבבי שיחה, תבניות, מודל ורמת מאמץ עוברים דרך הממשק האמיתי, לשונית אחת לכל פרויקט.
+ה-bridge מפעיל את **[Claude Design](https://claude.ai/design)** מתוך ה-Chrome של ה-bridge שבו אתם מחוברים. קריאות ושינויים בפרויקטים ובקבצים עוברים דרך הבקשות של האפליקציה עצמה מתוך הלשונית; סבבי שיחה, תבניות, מודל ורמת מאמץ עוברים דרך הממשק האמיתי, לשונית אחת לכל פרויקט.
 
 ```bash
 bridge chrome start --provider design

package/README.md

@@ -374,7 +374,7 @@ Flow requires a **Google AI Pro/Ultra** plan. Because Veo renders take minutes,
 
 ## Claude Design support
 
-The bridge drives **[Claude Design](https://claude.ai/design)** — Claude's project workspace for slides, prototypes, and documents — from your signed-in bridge Chrome. Reads and project/file changes use the app's own requests from inside the tab (one call each, [ADR 0021](docs/adr/current/0021-claude-design-app-rpc.md)); turns, templates, and model/effort go through the real UI, one tab per project.
+The bridge drives **[Claude Design](https://claude.ai/design)** — Claude's project workspace for slides, prototypes, and documents — from your signed-in bridge Chrome. Reads and project/file changes use the app's own requests from inside the tab (one call each); turns, templates, and model/effort go through the real UI, one tab per project.
 
 ```bash
 bridge chrome start --provider design            # sign in at claude.ai

package/README.zh.md

@@ -179,7 +179,7 @@ Flow 需要 **Google AI Pro/Ultra** 套餐。由于 Veo 渲染需要数分钟，
 
 ## Claude Design 支持
 
-bridge 可以在已登录的 bridge Chrome 中驱动 **[Claude Design](https://claude.ai/design)**。读取以及项目/文件的修改通过标签页内应用自身的请求完成（[ADR 0021](docs/adr/current/0021-claude-design-app-rpc.md)）；对话轮次、模板、模型和推理强度则通过真实界面操作，每个项目一个标签页。
+bridge 可以在已登录的 bridge Chrome 中驱动 **[Claude Design](https://claude.ai/design)**。读取以及项目/文件的修改通过标签页内应用自身的请求完成；对话轮次、模板、模型和推理强度则通过真实界面操作，每个项目一个标签页。
 
 ```bash
 bridge chrome start --provider design

package/SKILL.md

@@ -0,0 +1,235 @@
+---
+name: ai-browser-bridge
+description: Drive ChatGPT, Gemini, Claude, DeepSeek, Grok, Perplexity, Duck.ai, Arena, Google Flow, or Claude Design (claude.ai/design) through a signed-in Chrome session, either from the bridge CLI or its outbound MCP tools. Use when Claude Code, Codex, or another agent should ask a browser-hosted model, upload local files or screenshots, search, resume, or batch-organize browser conversations, fan out across providers, create or edit Claude Design projects (slides, prototypes, documents), or use bridge-managed ChatGPT and Flow capabilities.
+---
+
+# ai-browser-bridge
+
+Drive ChatGPT, Gemini, Claude, DeepSeek, Grok, Perplexity, Duck.ai, Arena, Google Flow, or Claude Design in a real browser from any agent — one provider or fanned out. Exposes sandboxed local repo tools to ChatGPT, Claude, and Grok over MCP, and serves outbound MCP `ask`, `search_conversations`, ChatGPT, Flow, and Claude Design tools so agents can call browser surfaces natively.
+
+## Prerequisites
+
+- macOS
+- Node.js ≥ 22
+- Google Chrome
+- `cloudflared` (optional, for ChatGPT, Claude, and Grok MCP tools)
+- Signed-in providers: run `bridge chrome start --provider <name>` and sign in if needed
+
+## Install & setup
+
+```bash
+npm install -g ai-browser-bridge   # installs `bridge` and ships this SKILL.md
+```
+
+The package folder is also the skill folder. Link it once into each agent's
+global skills directory; every later `npm install -g ai-browser-bridge` then
+updates the skill too:
+
+```bash
+PKG="$(npm root -g)/ai-browser-bridge"
+mkdir -p ~/.claude/skills ~/.agents/skills
+ln -sfn "$PKG" ~/.claude/skills/ai-browser-bridge   # Claude Code
+ln -sfn "$PKG" ~/.agents/skills/ai-browser-bridge   # Codex
+```
+
+## How to use as a tool
+
+### MCP stdio (Claude Code, Codex, Kiro, any MCP client)
+
+```bash
+bridge serve
+```
+
+Exposes tools over stdio:
+- `ask({ prompt, providers?, timeoutSeconds? })`
+- `search_conversations({ query, providers?, limit? })`
+- `design_*` for Claude Design: `design_state`, `design_list_projects`, `design_catalog`,
+  `design_choose_model`, `design_create_project`, `design_send`, `design_read_conversation`,
+  `design_new_conversation`, `design_rename_conversation`, `design_list_files`,
+  `design_put_files`, `design_remove_files`, `design_download`, `design_share`,
+  `design_update_project`, `design_duplicate_project`, `design_delete_project`,
+  `design_open_project`. `design_delete_project` and `design_remove_files` need
+  `confirm: true`; `design_put_files` replaces an existing path without asking.
+- `flow_*` for Google Flow: `flow_generate`, `flow_extend_clip`, `flow_reuse_clip`,
+  `flow_list_clips`, `flow_list_projects`, `flow_list_ingredients`, `flow_download_clips`,
+  `flow_rename_clip`, `flow_rename_project`, `flow_delete_clip`, `flow_delete_project`,
+  `flow_remove_ingredient`, `flow_clear_ingredients`
+- `chatgpt_render_state` for the live ChatGPT render (streaming, image progress, limits)
+
+### CLI (Codex, scripts, any shell-based agent)
+
+```bash
+# One provider
+bridge ask "summarize this repo" --provider chatgpt --json
+
+# Fan out across multiple
+bridge ask "compare approaches" --provider claude,deepseek,grok --json
+```
+
+`--json` emits machine-readable output. Never hangs in a pipe.
+
+## Claude Design
+
+Claude Design (claude.ai/design) is a project workspace for slides, prototypes,
+and documents. The bridge drives it in the signed-in bridge Chrome, one tab per
+project. It is not a fan-out provider, so `ask` does not reach it; use the
+`design_*` tools or `bridge design` instead.
+
+1. `bridge chrome start --provider design` and sign in at claude.ai if needed.
+2. `design_state` first: where Claude Design is, whether a turn is running, and
+   which actions are available.
+3. `design_catalog` for templates, models with effort levels, and design systems.
+4. New work: `design_create_project` with a prompt (and optional template, model,
+   effort, attachments). It waits for Claude's first reply.
+5. Follow-ups: `design_send` in the project's Conversation, then
+   `design_read_conversation` to read replies. For long turns pass `wait: false`
+   and poll `design_read_conversation`.
+6. Files: `design_list_files`, `design_put_files` (repo files only),
+   `design_download` into `.bridge/downloads/design`.
+
+Same surface from a shell:
+
+```bash
+bridge design state --json
+bridge design create --template Slides --prompt "our Q3 launch" --json
+bridge design send --project <id> --message "make the title bolder" --json
+bridge design read --project <id> --json
+```
+
+Left to the human in the UI: Publish as artifact, Send to Claude Code,
+PNG/video/PDF/PowerPoint export, comments, and version restore.
+
+## Per-agent setup
+
+### Claude Code
+
+```bash
+claude mcp add --transport stdio --scope user ai-browser-bridge -- bridge serve
+```
+
+### Codex
+
+```bash
+codex mcp add ai-browser-bridge -- bridge serve
+```
+
+This writes `[mcp_servers.ai-browser-bridge]` to `~/.codex/config.toml`. Codex can
+also call the CLI directly, for example `bridge ask "your question" --provider chatgpt --json`.
+
+### Kiro
+
+Add to your MCP config:
+```jsonc
+{
+  "mcpServers": {
+    "ai-browser-bridge": { "command": "bridge", "args": ["serve"] }
+  }
+}
+```
+
+### Cursor
+
+Add to `.cursor/mcp.json`:
+```json
+{
+  "mcpServers": {
+    "ai-browser-bridge": { "command": "bridge", "args": ["serve"] }
+  }
+}
+```
+
+## Available commands
+
+| Command | Purpose |
+|---------|---------|
+| `bridge` (bare) | Interactive TUI |
+| `bridge ask <prompt>` | One-shot send + reply |
+| `bridge chrome start --provider <name>` | Start existing Chrome profile with debug port |
+| `bridge status` / `bridge chrome status` | Show Chrome debug-port status |
+| `bridge cache list\|prune` | Inspect/prune safe generated Chrome cache |
+| `bridge serve` | Outbound MCP ask/search tools (stdio) |
+| `bridge download` | Download conversation attachments |
+| `bridge sessions` | List stored sessions |
+| `bridge stop` | Kill warm Chrome |
+| `bridge project list\|create\|rename\|delete` | Manage ChatGPT Projects |
+| `bridge chat list\|search\|move\|organize\|archive` | List, search, batch-organize & archive conversations |
+| `bridge chat organize status\|pause\|resume` | Inspect or control the latest persisted organization queue |
+| `bridge task list\|create` | Schedule ChatGPT Tasks |
+| `bridge chatgpt` | Inspect the live ChatGPT render |
+| `bridge flow` | Generate and manage Google Flow clips and ingredients |
+| `bridge design` | Claude Design projects, templates, model/effort, files, Conversations, share |
+
+## Keep conversations organized (ChatGPT)
+
+Before starting a **new** ChatGPT conversation, check whether it belongs in an
+existing Project instead of adding one more loose chat:
+
+1. `bridge project list` — the Projects that already exist.
+2. `bridge chat search "<topic>"` — a related past chat (and where it lives).
+3. If a Project fits, ask/resume there, then file the chat:
+   `bridge chat move "<idOrTitle>" --project "<Project>"`.
+4. Only leave a chat loose when nothing fits. The first time a **second**
+   related chat appears, `bridge project create "<Project>"` and move both in.
+
+For agents driving the bridge:
+
+- One Project per topic / repo / deliverable — reuse before you create.
+- `bridge chat list --orphans` shows only loose (project-less) chats; a growing list
+  there is the signal to file them into Projects.
+- Archive dead or scratch chats with `bridge chat archive "<idOrTitle>"`
+  (reversible — hides from the sidebar) to keep it lean. Batch with `--id`.
+- Don't spawn throwaway or test conversations in the signed-in account. Use an
+  isolate profile (`bridge ask --fan-out` with `isolate`) for scratch runs, and
+  clean up anything you create.
+
+For a full-history cleanup, scan only loose chats with
+`bridge chat list --orphans --json`, group clear recurring topics, and leave
+ambiguous or uncommon chats loose. Reuse existing Projects first; create a new
+Project only for a stable topic with at least two chats. Put every accepted move
+into one JSON plan of `{ "conversation": "<idOrTitle>", "project": "<name>" }`
+items, dry-run it, then start the persisted queue:
+
+```bash
+bridge chat organize --plan @organization-plan.json --dry-run
+bridge chat organize --plan @organization-plan.json \
+  --interval 60 --cooldown 600 --max-attempts 4 --json
+
+# Later lifecycle operations do not need the plan again
+bridge chat organize status --json
+bridge chat organize pause
+bridge chat organize resume --json
+```
+
+The queue persists under `.bridge/chat-organization-queues/`. Running the exact
+same plan again resumes pending work without replaying completed moves. Rate
+limits and closed Chrome sessions are deferred without consuming a move attempt;
+the queue relaunches its browser session automatically. Adaptive pacing starts at
+60 seconds, speeds up 20% after every three successful moves to a 15-second floor,
+and slows 50% after each rate limit to a 180-second ceiling, in addition to the
+600-second cooldown. Pace is persisted across pause/resume; use `--no-adaptive`
+only when a fixed/random interval is specifically required.
+
+At completion, the queue automatically runs `--orphans` discovery to the proven
+stable bottom of ChatGPT history and reports both intentional remaining orphans
+and any planned Conversation still loose. The scanner throws rather than silently
+accepting a partial inventory. Planned Conversations that remain loose are retried
+automatically up to `--max-attempts`. Start the queue, confirm it acquires the lock and
+processes or defers one item, then let it continue unattended—continuous agent
+watching is unnecessary. Use `bridge chat organize pause` instead of killing the
+process, `resume` to continue the latest queue without reconstructing its plan,
+and `--restart` only when the whole plan should intentionally be replayed.
+
+## Constraints
+
+- macOS only (hardcoded Chrome path, pbcopy/lsof)
+- Each provider needs Chrome started with `bridge chrome start --provider <name>` and a signed-in browser session
+- File operations are sandboxed to the target repo (no escape)
+- No raw shell — only validated MCP tools
+- Browser selectors may break when provider UIs update
+
+## Fan-out behavior
+
+- `--provider a,b,c` runs all in parallel
+- Partial-failure tolerant: exits non-zero only when ALL fail
+- `--strict`: exit non-zero if ANY fails
+- Replies keyed by provider in JSON output

package/dist/bridge.js

@@ -75,7 +75,12 @@ var PROVIDER_CONFIG = parseConfig(ProviderConfigTableSchema, {
     defaultUrl: "https://chatgpt.com",
     defaultModel: "ChatGPT",
     selectors: {
-      composer: '#prompt-textarea, [contenteditable="true"]',
+      composer: [
+        '[data-testid="prompt-textarea"]',
+        "#prompt-textarea",
+        'form[data-chatgpt-composer] [data-composer-markdown][contenteditable="true"][role="textbox"]',
+        'form[data-chatgpt-composer] .ProseMirror[contenteditable="true"]'
+      ].join(", "),
       assistant: '[data-message-author-role="assistant"]'
     }
   },
@@ -2857,11 +2862,13 @@ var injectPrompt = async (page, text) => {
   await runInjectPromptAttempts({ page, text });
 };
 var readComposerText = async (ctx) => {
-  const text = await ctx.page.evaluate(() => {
-    const prompt = document.querySelector("#prompt-textarea");
-    if (prompt === null || prompt.innerText === void 0) return "";
-    return prompt.innerText.trim();
-  });
+  const text = await ctx.page.evaluate((selector) => {
+    const prompt = document.querySelector(selector);
+    if (prompt === null) return "";
+    const value = prompt.value;
+    if (typeof value === "string") return value.trim();
+    return (prompt.innerText ?? prompt.textContent ?? "").trim();
+  }, SELECTORS.promptInput);
   if (text === null || text === void 0) return "";
   return text;
 };
@@ -18722,7 +18729,7 @@ var subcommandOpts = (command) => {
 // src/features/terminal/registerCli.ts
 var PROVIDER_OPTION = `Browser provider: ${PROVIDER_IDS.join(", ")} (default: ${DEFAULT_PROVIDER})`;
 var registerCliCommands = (program) => {
-  program.name("bridge").description("Terminal CLI that bridges ChatGPT or Gemini with local tools via MCP").version("0.7.0").option("-r, --repo <path>", "Path to the target repository (default: cwd)").option("-p, --port <number>", "MCP server port (default: 8765)").option("--provider <name>", PROVIDER_OPTION).option("--no-browser", "Skip Chrome browser connection").action(async (_options, command) => {
+  program.name("bridge").description("Terminal CLI that bridges ChatGPT or Gemini with local tools via MCP").version("0.7.1").option("-r, --repo <path>", "Path to the target repository (default: cwd)").option("-p, --port <number>", "MCP server port (default: 8765)").option("--provider <name>", PROVIDER_OPTION).option("--no-browser", "Skip Chrome browser connection").action(async (_options, command) => {
     if (process.stdin.isTTY !== true || process.stdout.isTTY !== true) {
       command.outputHelp({ error: true });
       process.exitCode = 1;