chatgpt-pro-line
Use the packaged ChatGPT Pro communication line for blocking AI-to-AI reasoning through the persistent browser session.
pauljunsukhan · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add pauljunsukhan/codex-chatgpt-pro-plugin --skill chatgpt-pro-line技能说明
name: chatgpt-pro-line description: Use the packaged ChatGPT Pro communication line for blocking AI-to-AI reasoning through the persistent browser session.
ChatGPT Pro Line
Intended Use
Use ChatGPT Pro for higher-level intelligence tasks where a second strong model can materially improve the work:
- system architecture and technical strategy
- product and feature design decisions
- specification drafting and critique
- research synthesis, including online research when appropriate
- risk analysis, tradeoff review, and planning across files/repos/agents
- debugging strategy for confusing failures after Codex has gathered evidence
Do not burn the line on trivial syntax, mechanical edits, or proof-only smoke tests except when validating the transport itself. Transport tests prove the line works; they are not the product.
Ask for high-leverage judgment, not merely the smallest next patch. It is fine to request the most intelligent or bold implementation direction given the evidence; Codex can still execute that direction in small, testable slices. Invite ChatGPT Pro to disagree, ask clarifying questions, propose a different architecture, or recommend larger prerequisite work when that is the professional answer. The line is for senior planning and system judgment, not for rubber-stamping the current implementation.
Canonical Command
When this skill is installed from the Codex plugin, resolve the package root as
two directories above this SKILL.md file and prefer the packaged CLI at:
<plugin-root>/bin/chatgpt-pro
Use chatgpt-pro directly only when that command is already on PATH. Inside
this source repo, ./bin/chatgpt-pro ... and npm run chatgpt:call -- ...
are development entrypoints for the same behavior.
First-run package flow:
chatgpt-pro init
chatgpt-pro doctor --warm
# complete visible login if prompted
chatgpt-pro doctor --live
chatgpt-pro status --alias=main
doctor --warm opens or reuses the dedicated visible ChatGPT browser profile
without sending a prompt. doctor --live verifies login, composer, model, and
intelligence state without sending a prompt, and refreshes the model cache.
chatgpt-pro call \
--alias=main \
--prompt-file=prompt.md \
--upload-file=.devspace/context/current/repo-context.md \
--response-mode=blocking
Equivalent minimal form:
chatgpt-pro call --alias=main --prompt="..."
Before a call, agents may inspect repo-local state without touching the browser:
chatgpt-pro status --alias=main
status is a registry/local command. It reports the current project id, room
binding, latest receipt/transcript paths, browser/profile lock owner if busy,
project-state lock owner if busy, and model-cache state. It must remain safe to
run while another agent holds the live browser lock.
By default, chatgpt-pro call selects the live Pro intelligence level when
the website exposes it. Use --level=... / --intelligence=... or
CHATGPT_LEVEL to choose another live label. Use --no-default-pro only for
transport debugging where changing the selector would obscure the failure.
Required Live Thread Output
When using this line in a Codex conversation, repeat the exchange verbatim in the thread using these exact headings:
## Message Sent To ChatGPT Pro
...
## Message Received From ChatGPT Pro
...
The run artifact transcript.md must use the same headings. The canonical
renderer is src/transcript.mjs.
chatgpt-pro call prints this exact thread-echo block by default after the
receipt paths. Copy the printed block into the Codex thread. Do not summarize,
paraphrase, trim, or rewrite the Message Sent To ChatGPT Pro or
Message Received From ChatGPT Pro sections. Additional commentary may come
before or after the exact block, but not inside it.
When files are uploaded, the Message Sent To ChatGPT Pro block is the exact
text typed into the composer. The uploaded file bodies are not pasted into the
Codex thread; their paths, bytes, hashes, and upload status belong in the
receipt.
chatgpt-pro read prints the exact received message. Machine callers may set
CHATGPT_THREAD_ECHO=0, but interactive Codex use should leave echoing on and
receipts must record threadEcho.mode, threadEcho.transcriptSha256,
threadEcho.sentSha256, and threadEcho.receivedSha256.
Verify transcript integrity when needed:
chatgpt-pro transcript verify --receipt=.devspace/runs/<run-id>/receipt.json
Response Mode
Default and only implemented mode:
blocking: send the prompt, wait for the anchored assistant response, write artifacts, then continue.
Reserved for later:
event: send and return a run id immediately, then read/poll later.
If a caller requests anything except blocking, the command should fail rather
than pretending event mode exists.
Context
Default posture: keep the composer text small and attach bulky context
deliberately. chatgpt-pro call defaults to --repo-context=auto: it detects
prompts that clearly ask for repo architecture, whole-codebase context, security
review, or similar cross-file reasoning. Generated repo context is never uploaded
or inlined unless the call also includes --confirm-repo-context-upload or
CHATGPT_CONFIRM_REPO_CONTEXT_UPLOAD=1. It skips repo upload for trivial
transport checks and when the caller already supplied --context-dir,
--context-file, or --upload-file.
Use --repo-context=upload --confirm-repo-context-upload when broad repo context
is definitely needed. Use --repo-context=inline --confirm-repo-context-upload
only for tiny/debug contexts. Use --no-repo-context or --repo-context=off for
narrow questions. The bundle lives under
.devspace/context-bundles/<id>/ and includes:
repo-context.md: one markdown file containing repo structure, LOC, file contents, and a source mapmanifest.json: machine-readable file and directory records, including line ranges