gpt-relay

name	gpt-relay
description	Use when the user asks to use GPT 5.5 Pro, 5.5 pro, GPT 5.5, Extended Pro, Pro Extended, pro extend think, or similar wording to send a prompt or files to ChatGPT and bring the answer back into Codex.

GPT Relay

This skill relays a Codex task to ChatGPT through the user's existing Chrome session. By default it keeps the user's current visible ChatGPT Intelligence selection. Only change the model, mode, or effort when the user explicitly requests a visible combination such as 5.5 Pro extended, 5.5 Thinking light, or 5.4 Instant.

What It Does

Use the helper script at ../../scripts/chatgpt_relay.mjs to:

1. Open a new ChatGPT tab.
2. Ensure the composer is available.
3. Select the requested ChatGPT Intelligence model/mode/effort only when explicitly requested; otherwise keep the current ChatGPT selection.
4. Attach optional local images/files through ChatGPT's visible upload controls first.
5. For image attachments, wait until the attachment UI is stable for 5 seconds and no upload/processing status is visible before sending.
6. Send the user's prompt.
7. Wait until ChatGPT finishes answering; do not answer the user's task from Codex while ChatGPT is still running.
8. Return the complete assistant delivery to Codex: full text with Markdown formatting preserved when possible, generated images, Deep Research reports, and any saved artifact paths.
9. Keep the ChatGPT tab open by default for follow-up use.
10. Store session metadata, URL, summary, keywords, status, and artifacts.
11. Save a pending session immediately after ChatGPT creates the /c/... conversation URL, before waiting for the final answer.
12. For Create image tasks, treat generated image artifacts as a valid completion signal even when the assistant text is empty.
13. Save generated images locally and return imageMarkdown lines that can be rendered in the Codex conversation.
14. For Deep Research tasks, treat Research completed in ... · ... citations · ... searches as the completion signal, then open the report viewer, export Markdown, validate it, and return the Markdown artifact.

Required Setup

This skill depends on the Chrome plugin and its browser-client runtime.

The helper auto-bootstraps the bundled Chrome plugin when globalThis.browser is missing. If auto-bootstrap fails, report the helper error code and ask the user to ensure the Chrome plugin is installed and the Codex Chrome extension is connected.

Do not inspect cookies, local storage, passwords, or browser session stores.

The preferred session store is ~/.codex/gpt-relay/sessions.json. If the runtime cannot write there, the helper falls back to nodeRepl.tmpDir and returns session.stateWarning; this is degraded persistence, not a relay failure. Later session lookup also checks that fallback store when no explicit statePath is supplied.

Trigger Handling

When the user mentions phrases such as:

• 5.5 pro
• GPT 5.5
• GPT 5.5 Pro
• Extended Pro
• Pro Extended
• pro extend think
• 5.5 thinking light
• 5.4 instant
• 5.5 pro standard
• 5.5 pro extended

interpret that as a request to use this relay, unless the user is only asking a conceptual question.

When the user asks to continue, reopen, list, search, resume, poll, or check an earlier GPT 5.5 Pro conversation, use the exported session helpers instead of starting a new unrelated chat.

Safety Boundaries

• The user's prompt is authorized to send when they explicitly ask to relay it to GPT 5.5 Pro.
• For files, only attach local absolute file paths that the user provided or that are clearly attached to the current request.
• Images and documents are uploaded through ChatGPT's visible Add photos & files / file input controls by default.
• Do not pre-compress images unless the user explicitly asks for compression.
• Clipboard image paste is only a fallback for small images after upload paths fail.
• For large images, do not fall back to native clipboard paste automatically; report the upload permission blocker so the user can fix the real upload path.
• If the relay returns ATTACHMENT_UPLOAD_PERMISSION_REQUIRED, report the two permission gates: Codex Chrome Uploads and Chrome extension Allow access to file URLs.
• If the relay returns ATTACHMENT_UPLOAD_NOT_CONFIRMED or IMAGE_ATTACHMENT_NOT_CONFIRMED, report that ChatGPT did not visibly accept the attachment and do not answer the attachment task from Codex.
• Text-like files are uploaded by default. To embed a text file into the prompt instead, pass attachments: [{ path: "...", inline: true }].
• If a file path is ambiguous, ask for the path.
• If ChatGPT shows login, CAPTCHA, payment, permission, or account prompts, stop and report the visible state.
• If the relay fails, do not answer with another model, web search, or local reasoning unless the user explicitly asks for a fallback.
• If ChatGPT is still answering, keep waiting or poll the same session; never substitute a local Codex answer for the requested GPT 5.5 Pro answer.
• Do not solve CAPTCHA or accept browser permission prompts automatically.
• Do not claim the backend model name from hidden state. Report only the visible ChatGPT Intelligence selection that the helper requested or observed.
• Not every model exposes every reasoning mode. If ChatGPT does not show a requested combination, report INTELLIGENCE_COMBINATION_UNAVAILABLE instead of silently choosing a different model or effort.
• If the user does not explicitly ask to change the model, mode, or effort, keep the current ChatGPT selection. Do not force 5.5 Pro Extended.
• Pro mode is the paid ChatGPT Pro Intelligence mode. It only exposes Standard and Extended effort. Do not request Light or Heavy for Pro mode.
• If the user requests Pro but the visible ChatGPT account does not expose Pro, report that Pro is unavailable and include the available models/modes/efforts from the helper error instead of falling back to another model.

Node REPL Usage

const { runExtendedProRelay } = await import("/absolute/path/to/plugin/scripts/chatgpt_relay.mjs");
const result = await runExtendedProRelay({
  prompt: "User prompt here",
  filePaths: [],
  keepTab: true,
  waitChunkMs: 90000,
  timeoutMs: 6 * 60 * 60 * 1000,
  returnPending: false
});
nodeRepl.write(result.finalDeliveryText ?? result.finalResponseText);

CRITICAL FINAL OUTPUT RULE: If any relay helper returns status: "complete", mustReturnFinalDelivery: true or mustReturnVerbatim: true, and a non-empty finalDeliveryText, the Codex final answer MUST be exactly result.finalDeliveryText. Do not add a summary before it. Do not shorten it. Do not rewrite it. Do not omit lines. Do not wrap it in JSON or Markdown commentary. Do not replace it with key points. The finalDeliveryText field already contains the complete ChatGPT assistant text, generated image Markdown, saved non-image artifact paths, and the conversation link on the final line. For ordinary assistant answers, it should preserve ChatGPT-rendered Markdown structure such as headings, lists, tables, links, inline formatting, and original code blocks. Do not wrap the whole answer in a code block. finalResponseText is kept only as a text-only compatibility/debug field; do not use it as the final answer when finalDeliveryText is available.

Before sending the final response, perform this check:

• Did this turn use runExtendedProRelay, pollRelaySession, or continueExtendedProRelay?
• Is (result.mustReturnFinalDelivery === true || result.mustReturnVerbatim === true) and result.finalDeliveryText available?
• If yes, final answer = result.finalDeliveryText verbatim.
• If result.imageMarkdown, result.reportMarkdown, or result.artifacts exist, confirm they are represented in result.finalDeliveryText.
• If you are about to summarize, stop and paste result.finalDeliveryText instead.

When calling this helper through node_repl, set the tool call timeout high enough for the expected ChatGPT runtime. Do not omit the tool timeout for long Pro jobs. Use at least 30 minutes when the user asks for patient waiting.

Select a visible ChatGPT Intelligence combination explicitly:

const result = await runExtendedProRelay({
  prompt: "User prompt here",
  model: "5.5",
  mode: "thinking",
  effort: "light",
  keepTab: true,
  timeoutMs: 30 * 60 * 1000
});
nodeRepl.write(result.finalDeliveryText ?? result.finalResponseText);

The helper can also infer common phrases from the prompt, such as GPT 5.5 Thinking light, 5.4 Instant, or Pro extended. The plugin name GPT Relay alone does not count as a request to change models. If a requested combination is not visible in ChatGPT, poll or report the error with the available options; do not resubmit the task with a different model.

If the outer tool call still times out, do not answer the user's task from Codex. Use listRelaySessions or pollRelaySession to recover the stored pending conversation and continue waiting.

Use an absolute import path resolved from this skill file:

<plugin-root>/scripts/chatgpt_relay.mjs

Continue a previous conversation:

const { continueExtendedProRelay } = await import("/absolute/path/to/plugin/scripts/chatgpt_relay.mjs");
const result = await continueExtendedProRelay({
  query: "高達模型",
  prompt: "請接著整理模型比例與入門建議。"
});
nodeRepl.write(result.finalDeliveryText ?? result.finalResponseText);

List or search stored sessions:

const { listRelaySessions } = await import("/absolute/path/to/plugin/scripts/chatgpt_relay.mjs");
const sessions = await listRelaySessions({ query: "高達", limit: 10 });
nodeRepl.write(JSON.stringify(sessions, null, 2));

Check a long-running session after an outer tool timeout or when the user explicitly asked to background the work:

const { pollRelaySession } = await import("/absolute/path/to/plugin/scripts/chatgpt_relay.mjs");
const result = await pollRelaySession({
  query: "高達",
  timeoutMs: 30 * 60 * 1000
});
nodeRepl.write(result.finalDeliveryText ?? result.finalResponseText);

If status is still pending, repeat pollRelaySession later. Do not resubmit the original prompt unless the user explicitly requests a new GPT run.

Start a long-running session and return a resumable pending result after one short wait:

const { startExtendedProRelay } = await import("/absolute/path/to/plugin/scripts/chatgpt_relay.mjs");
const result = await startExtendedProRelay({
  prompt: "Long prompt here",
  keepTab: true
});
nodeRepl.write(JSON.stringify(result, null, 2));

Use a GPT App mention:

await runExtendedProRelay({
  appName: "Canva",
  prompt: "Create a clean poster concept for this campaign."
});

Use ChatGPT tools:

await runExtendedProRelay({
  feature: "deep-research",
  prompt: "Research the topic and return a structured report."
});

await runExtendedProRelay({
  feature: "create-image",
  prompt: "Create an image of a futuristic Gundam hangar."
});

await runExtendedProRelay({
  projectName: "Image Gen",
  prompt: "Continue inside this project."
});

When feature: "deep-research" completes, return finalDeliveryText as the final Codex answer. It contains the full report Markdown and the conversation URL. The helper also returns reportMarkdown, deepResearch, and an artifact like:

{
  "kind": "deep-research-report",
  "localPath": "/absolute/path/deep-research-report.md",
  "mimeType": "text/markdown"
}

If a Deep Research job was started earlier, call pollRelaySession with the session id, title, or keywords. The helper first tries to claim the original Chrome tab before opening the stored conversation URL.

When feature: "create-image" completes, return finalDeliveryText as the final Codex answer. It contains the generated image Markdown line(s), optional assistant text, and the conversation URL. The helper also returns artifacts with localPath and imageMarkdown for inspection:

![generated image 1](/absolute/path/to/image-01.png)

If status is pending but the user's browser visibly shows a generated image, call pollRelaySession again. The optimized helper can complete image-only responses once the generated image artifact is stable.

Attach an original image or document:

await runExtendedProRelay({
  prompt: "Analyze this image.",
  attachments: [{ path: "/absolute/path/image.png" }],
  keepTab: true,
  waitChunkMs: 90000,
  timeoutMs: 6 * 60 * 60 * 1000,
  returnPending: false
});

Response Shape

The helper returns:

{
  "ok": true,
  "status": "complete",
  "mode": "5.5 Thinking Light",
  "intelligence": {
    "model": "5.5",
    "mode": "thinking",
    "effort": "light"
  },
  "assistantText": "...",
  "reportMarkdown": "...",
  "finalResponseText": "...\\n\\nConversation URL: https://chatgpt.com/c/...",
  "finalDeliveryText": "...\\n\\n![generated image 1](/absolute/path/image-01.png)\\n\\nArtifacts:\\n- report: /absolute/path/report.md\\n\\nConversation URL: https://chatgpt.com/c/...",
  "mustReturnFinalDelivery": true,
  "finalDeliveryField": "finalDeliveryText",
  "deepResearch": {},
  "conversationUrl": "https://chatgpt.com/c/...",
  "title": "...",
  "session": {},
  "artifacts": [],
  "imageMarkdown": [],
  "messages": []
}

Only when returnPending: true is explicitly requested, the helper may return:

{
  "ok": true,
  "status": "pending",
  "assistantText": "",
  "conversationUrl": "https://chatgpt.com/c/...",
  "session": {}
}

When status is pending, keep the tab open and later call pollRelaySession. Do not write an answer to the user's original prompt from Codex.

If it cannot continue, it throws an error with a stable code such as:

• CHROME_BROWSER_MISSING
• CHROME_BROWSER_CLIENT_MISSING
• CHATGPT_COMPOSER_MISSING
• INTELLIGENCE_COMBINATION_UNAVAILABLE
• ATTACHMENT_UPLOAD_FAILED
• ATTACHMENT_UPLOAD_PERMISSION_REQUIRED
• ATTACHMENT_UPLOAD_NOT_CONFIRMED
• IMAGE_CLIPBOARD_TOO_LARGE when an explicit image size limit was configured for fallback paste
• IMAGE_CLIPBOARD_PASTE_FAILED
• IMAGE_ATTACHMENT_NOT_CONFIRMED
• FEATURE_CONTROL_UNAVAILABLE
• PROJECT_UNAVAILABLE
• SESSION_NOT_FOUND
• CHATGPT_RESPONSE_TIMEOUT
• DEEP_RESEARCH_COMPLETED_CARD_NOT_OPENABLE
• DEEP_RESEARCH_VIEWER_OPEN_FAILED
• DEEP_RESEARCH_EXPORT_MENU_MISSING
• DEEP_RESEARCH_MARKDOWN_EXPORT_MISSING
• DEEP_RESEARCH_MARKDOWN_DOWNLOAD_TIMEOUT
• DEEP_RESEARCH_REPORT_VALIDATION_FAILED

Report these errors in natural language and include the practical next step.