chatgpt-web-workers README

项目的 README,来自 github.com/YUTA-fywoo/chatgpt-web-workers。合并为一个 markdown 文件:README.md。

ChatGPT Web Workers

Let Astra coordinate the goal. Let each Chat finish one useful stage.

简体中文 · Download v1.1.0 · Changelog · MIT license

A Codex desktop skill that delegates work to 1–6 ordinary ChatGPT chats at Extra High / 极高, proactively using 4–6 for substantial independent work. Astra owns the overall goal, stage acceptance, cross-part decisions, and final delivery. Workers are called Sol as a project prompting convention.

No extra MCP server or API key configuration after installation. Use the Codex desktop app's existing in-app browser and internal conversation reader, with ChatGPT signed in and the required client/account capabilities available.

Internal reply reads reduce host-model token overhead. Retrieve conversation content directly instead of repeatedly loading webpage screenshots, UI structure, and duplicate history just to read replies. Compact status probes, cached originals, and duplicate filtering keep context focused on results and evidence. Full first receipt, webpage sends, necessary health checks, and file transfers remain intact. Savings vary by task; no fixed percentage is claimed.

What's new in v1.1.0

Improvement Behavior
Active distribution Plan independent scopes before dispatch; prefer 4–6 for substantial work. Total distinct chats, including replacements, remains 1–6.
One primary result per prompt End each stage at an assessable result. Do not combine discovery, exhaustive verification, ranking, and final reporting in one huge prompt. Closely coupled operations and self-checks may stay together.
Independent continuation Accept each completed reply at a scheduled check and advance that owner while other chats continue. No full-wave barrier; at most one unanswered assignment per chat.
Preserve collection volume Splitting stages does not shrink assigned targets into a small pilot or introduce arbitrary search caps. Later stages adapt to real gaps and the final goal.
Internal reads, webpage sends Use internal tools for replies. Initial prompts, follow-ups, corrections, retries, and Chat controls all use the in-app browser. Internal sending is no longer an allowed fallback.
Explicit timing Reply scans start at minute 10, then every 5 minutes. Webpage checks start at minute 15, then every 10. Each revision has a 90-minute deadline.
Ownership and actual access New external evidence collection belongs to the assigned Chat. Host-only accounts, plugins, APIs, or local paths do not establish Chat access.
Durable task state Persist the objective, revisions, dependencies, accepted evidence, gaps, and next actions. Receipt, acceptance, and final task completion are distinct.
Complete results, less repetition Cache raw replies, use compact probes, deliver each new complete reply once, then read relevant cached sections or actual changes. Full first receipt remains required.
Focused instruction loading Transport for worker operation; file guidance for transfer; model rationale for maintenance. Domain-specific standards belong to dependent skills.

How it works

flowchart TD
    A[User goal] --> B[Astra: scopes and stage targets]
    B --> C[In-app browser: create Chat and send prompts]
    C --> D[Workers generate concurrently]
    D --> E[Internal reader: match complete replies]
    E --> F[Astra: accept, integrate, track gaps]
    F -->|Next useful stage for this owner| C
    F -->|Final requirements satisfied| G[Checked final deliverable]
    E -->|No verified complete reply at final deadline| H[Preserve progress and stop entire task]

Browser creation and sends are sequential; worker generation is concurrent. Timers are anchored to each revision's actual submission. The host can classify, deduplicate, calculate, resolve conflicts, assemble artifacts, and run required local checks using supplied evidence. New external collection or source-page checks return to Chat unless explicitly authorized for the host by the user. Slow replies and blocked sources do not authorize silent takeover.

Waiting and recovery

The 90-minute rule supersedes the earlier one-hour setting discussed during development. It is per revision, not a total limit for a multi-stage task. A completed reply reporting a limitation is a returned result. The full procedure is in transport.md.

Install or update

$skill-installer Install skills/chatgpt-web-workers from
https://github.com/YUTA-fywoo/chatgpt-web-workers

Alternatively, download chatgpt-web-workers-v1.1.0.zip from Releases, extract it, and install the contained chatgpt-web-workers folder. Source code (zip) instead contains the repository layout; its installable folder is skills/chatgpt-web-workers.

The tested Windows desktop installation uses $CODEX_HOME/skills/chatgpt-web-workers, normally ~/.codex/skills/chatgpt-web-workers. Use the location your client discovers. Preserve local customizations when upgrading and avoid duplicate copies with the same skill name. The repository also provides .codex-plugin/plugin.json for plugin distribution.

Migration from v1.0.0: update all six skill files together. Remove local overrides that still permit internal sends, assume a one-hour timeout, cap substantial work at three chats, or wait for all chats before continuing an owner. Keep browser file routes.

Use

Use $chatgpt-web-workers to research [topic]. Use 4–6 chats if there are
enough independent scopes. Give each chat one assessable stage per prompt,
preserve the full collection targets, and advance completed owners independently.
Deliver one integrated report with evidence and limitations.
Use $chatgpt-web-workers to analyze these files. Give each owner only its
required inputs, preserve the final goal across stages, and deliver a checked artifact.
Use $chatgpt-web-workers for code proposals and focused critique of these
components. Keep scopes separate; review, execute necessary local checks,
and integrate the accepted results.

Supports research, extraction, writing, analysis, calculations, code proposals, and critique. No fixed research pipeline or step count is required. Use 1–3 chats for small or tightly coupled work, or an actual resource constraint.

Prompting and quality

Each packet has a task/revision marker, a brief overall-goal anchor, one stage objective, relevant inputs and constraints, and a return contract. Follow-ups carry relevant changes rather than whole transcripts or unchanged tables. Shorter prompts must remain sufficient for the stage.

The design applies OpenAI's Astra skill guidance to concise descriptions, conditional reference loading, and outcome boundaries. Worker prompts use direct goals and clear input boundaries following general reasoning guidance. See model-guidance.md for Sol-specific sources. The 4–6 preference and exact timers are project policies, not official OpenAI concurrency recommendations.

Verify the visible ordinary Chat Extra High / 极高 setting. A Pro subscription badge is not Pro reasoning; prompts cannot activate API-only settings. If your account offers another effort, request an explicit local adaptation rather than silently substituting it.

The current policy preserves full first receipt, rather than replacing all reports with summaries. Later probes omit unchanged content; global comparisons can still require the full merged result. Original uploads and newly generated downloads use the browser, with actual bytes and content checked: files.md.

The aim is to reduce duplicate host work while preserving evidence. No savings percentage or universal no-regression claim has been measured. Historical validation retains its original dates; this release is not a new live test of every capability.

Feedback and sanitized examples are welcome in Issues. MIT licensed.

chatgpt-web-workers (YUTA-fywoo/chatgpt-web-workers)

正在搜索数据库…

这是什么?这是 Codex 桌面版的一个技能,让名为 Astra 的主控角色可以把工作委派给 1-6 个普通的 ChatGPT 浏览器对话(设置为 Extra High 推理强度),由它协调各阶段并接收结果,而不是自己包揽所有工作。
它是独立应用,还是在 Codex 内运行?它不是独立应用,而是作为一个技能文件夹安装到 Codex 桌面版的 skills 目录中,并通过 Codex 现有的应用内浏览器和内部对话读取器运行。
使用它需要设置 API 密钥或 MCP 服务器吗?不需要。它无需额外配置 MCP 服务器或 API 密钥,只需在 Codex 自带的应用内浏览器中登录 ChatGPT 账号即可。
如果 ChatGPT 对话卡住或停止回复,会发生什么?每次修订有 90 分钟的时限,从第 10 分钟开始检查回复,之后每 5 分钟检查一次;如果期间没有收到经过确认的完整回复,整个任务将停止,只有在用户明确下达指令后才会恢复。
Astra 是什么,Sol 又是什么?Astra 是 Codex 中的主控角色,负责整体目标、接收各阶段成果并完成最终交付;'Sol' 则是本项目对其所派发任务的工作型 ChatGPT 对话的命名方式。
它如何避免因重复读取而浪费 token?它使用 Codex 内部的对话读取器直接提取 ChatGPT 的回复,而不是反复读取网页截图和重复的历史记录;同时缓存原始回复,并使用简洁的状态探测,使上下文集中在新的结果上。
它通常会协调多少个 ChatGPT 对话?总共 1-6 个;处理大量相互独立的工作时优先使用 4-6 个,处理小型或紧密耦合的任务时保留 1-3 个。
ChatGPT Pro 订阅能保证这里有更好的结果吗?不能。该技能只依赖普通对话中可见的 Extra High / 极高 强度设置;Pro 订阅标识并不等同于 Pro 推理,提示词也无法启用仅限 API 的设置。
大小
0.1 MB
星标
329
时长
2个月
提交 · 90 天