Skip to content

Releases: YUTA-fywoo/chatgpt-web-workers

v1.1.0 — 分阶段并行协作 / Staged parallel workers

Choose a tag to compare

@YUTA-fywoo YUTA-fywoo released this 15 Sep 18:49
cb33249

中文

chatgpt-web-workers v1.1.0:多 Chat 分担任务,分阶段独立推进。

  • **无需额外配置:**安装后使用 Codex 桌面端已有能力,不必另配 MCP 服务器或 API Key;需已登录 ChatGPT 且客户端与账号具备所需功能。
  • **内部读取更省 token:**直接读取回复,减少反复接收网页截图、界面结构与重复历史的主模型开销;配合缓存和重复过滤,保留首份完整成果。节省量因任务而异。
  • 较大独立任务优先4–6个普通极高 Chat,Astra 保留总体目标与最终验收。
  • 每轮一个主要成果,谁完成谁续轮;分阶段不缩水搜索量。
  • 内部工具读回复,所有发送和 Chat 操作走网页。
  • 每轮10分钟首读、每5分钟再读;15分钟首查网页、每10分钟再查。
  • 每轮90分钟仍无完整回复则保存进度、中断整个任务;不自动接管。
  • 来源限定 Chat 实际权限,原文缓存与首次完整接收保留,后续过滤重复内容。
  • 同步中英指南、插件描述和全部六个技能文件;通用机制与领域规则分开。

**升级提醒:**v1.0.0 的内部发消息路径已被取代,请一起更新六个技能文件。原文件上传和生成文件下载仍走网页。

下载下方 chatgpt-web-workers-v1.1.0.zip,解压安装其中的 chatgpt-web-workers 文件夹。ZIP 含六个技能文件与 MIT 许可证;也可从仓库安装 skills/chatgpt-web-workers。

中文说明 · 完整更新记录

本次检查覆盖发布文件一致性、格式、引用和安装包完整性;历史网页能力测试保留原日期。

English

No extra MCP/API-key setup; less host token overhead through internal reply reads.

Use existing Codex desktop capabilities with ChatGPT signed in and the required client/account features available. Internal reads reduce repeated screenshots, UI structure, and duplicate history in host context; caching and duplicate filtering preserve full first receipt. Savings vary by task.

More useful parallelism, bounded stages, and a durable final goal.

Prefer 4–6 ordinary Extra High Chats for substantial independent work, within the 1–6 total limit. Advance each owner as its complete stage is accepted, while preserving collection targets. Internal tools read replies; all sends and Chat controls use the webpage.

Each revision uses 10/5-minute reply scans, 15/10-minute webpage checks, and a 90-minute deadline with task-wide interruption when a complete reply cannot be verified. Keep collection within the assigned Chat's actual permissions, preserve full first receipt, and suppress unchanged later content.

Update all six skill files together when migrating from v1.0.0. Download the standalone ZIP below or install skills/chatgpt-web-workers from the repository.

v1.0.0 — Astra leads, Sol works / Astra 统筹,Sol 协作

Choose a tag to compare

@YUTA-fywoo YUTA-fywoo released this 12 Sep 08:53
80cb669

安装包 / Skill ZIP: chatgpt-web-workers-v1.0.0.zip

解压后直接安装里面的 chatgpt-web-workers 文件夹。包含完整技能文件和 MIT 许可证。

Extract and install the included chatgpt-web-workers folder. Includes all skill files and the MIT license.


中文

让 Astra 把额度花在关键判断上,让 Sol 承担实质工作。

chatgpt-web-workers v1.0.0:为 Astra 主智能体设计的 Codex 桌面技能,协调 1–6 个普通 Chat 的 Sol 极高子智能体。无需额外 MCP 服务器或 API Key。

  • 节约主任务开销: 把调研、分析、计算、长稿和代码方案交给 Sol,Astra 负责分工、跨部分判断和最终整合。
  • 减少页面操作: 浏览器创建并设置对话,日常文字用内置工具直接收发;文件上传下载走浏览器。
  • 保留完整成果: 完整报告首次接收后,后续过滤重复内容,让 Astra 使用细节完成综合判断。
  • 清晰分工与并行: 每块工作一个负责人,独立任务先派出再等待,同一子任务复用原对话。
  • 围绕交付验收: Sol 自检,Astra 按需检查具体问题,取得文件后核对内容与格式。
  • 便于安装与调整: 中英双语说明、使用示例,Plus 用户可按账号可用选项适配 Sol High,MIT 开源。

下载下方 Source code (zip),解压后取出 skills/chatgpt-web-workers 文件夹安装。也可让 Codex 的 $skill-installer 从该仓库安装此子目录。

中文完整说明 · 提出建议

欢迎提建议、反馈使用体验,一起让这个 Skill 更好用!

English

Let Astra spend its effort on key judgments. Let Sol do substantial work.

Version 1.0.0 coordinates 1–6 ordinary ChatGPT Sol Extra High workers from an Astra-led Codex desktop task, using existing browser and conversation tools. No extra MCP server or API key setup.

  • Delegate research, analysis, calculations, full drafts, and code proposals to reduce avoidable host work.
  • Use built-in tools for routine text and the browser for setup and files.
  • Receive complete reports once, then filter duplicate context and keep follow-ups focused.
  • Give each work unit one owner and run independent work concurrently.
  • Combine worker self-checks with focused host acceptance and usable final artifacts.
  • Includes bilingual guides, installation examples, Sol High adaptation guidance for Plus users, and the MIT license.

Download Source code (zip) below and install the inner skills/chatgpt-web-workers folder, or ask Codex's $skill-installer to install that subdirectory from this repository.

English guide · Suggestions and feedback

Ideas, improvements, and real-world use cases are welcome!