Chatgpt Pro Consult

ChatGPT Pro Consult

This is a backward-compatible skill alias. Prefer chatgpt-delegate for new work because it chooses between Chat, Work, and official Codex based on the job and uses the strict experience/configuration APIs.

Use this alias when an existing workflow or user explicitly requests the visible Chat Pro intelligence setting for a focused consultation. Typical uses are critical review, synthesis, planning review, and a long-form second opinion through the user's ChatGPT subscription.

This skill is a thin workflow over the codex-chatgpt-control plugin. Use the snippets below for ordinary consults, but switch to the full codex-chatgpt-control skill as soon as bridge bootstrap, tab reuse, file attach, send, wait, read, or selector diagnosis becomes the main issue.

Guardrails

• This sends prompt and attachment content to ChatGPT web. Do not send secrets, credentials, private source material, financial details, legal evidence, medical details, or sensitive personal data unless the user clearly approved that disclosure.
• Use only visible ChatGPT web through the Codex/browser bridge. Do not replicate private ChatGPT network calls, read cookies, inspect localStorage/sessionStorage, or extract hidden auth headers.
• Make Pro selection explicit with experience: "chat" and configuration: { intelligence: "Pro" }. The strict configuration API handles older and newer visible pickers and verifies the postcondition. If the SDK cannot verify Pro, stop and report the blocker and visible candidate labels.
• Call experience.open({ experience: "chat" }) through the workflow even when Work is currently visible; do not assume the current ChatGPT pane is Chat.
• Prefer a fresh thread unless the user asked to continue a specific ChatGPT thread.
• Return Markdown by default. Use redacted reports by default; raw prompt/response content is opt-in only.
• Treat ChatGPT Pro output as another model's judgment, not verified truth. Verify current, legal, medical, financial, or high-stakes claims with primary sources.
• Keep each Codex tool call below the host call ceiling. Do not ask one JS call to wait for a long Pro response. Submit the prompt under Pro first, then poll/read in bounded chunks.

Staged Consult Boundary

Use the submit-then-poll pattern in this skill for expensive visible-browser work:

• Pro consultations.
• File-backed consults.
• Long Thinking or Deep Research-style answers.
• Recovery after a submit, wait, or read call times out.
• Any result whose completionState is generating or whose generationActive flag is true.

Do not force this staged protocol into ordinary short SDK calls. For normal chatgpt.ask(...), short askInThread(...), quick reads, and exact-string smokes, the smoother one-call wait: true / read: true flow is still the right default. Switch to staged recovery only when the answer is likely to exceed the Codex tool-call ceiling or when a partial/timeout result proves the prompt may already be submitted.

Runtime Loader

Resolve relative paths from this SKILL.md directory. The plugin runtime loader lives at:

../../runtime/import-chatgpt-control.mjs

From a bridge-enabled Codex Node runtime:

const loaderUrl = new URL(
  "../../runtime/import-chatgpt-control.mjs",
  "file:///absolute/path/to/plugins/codex-chatgpt-control/skills/chatgpt-pro-consult/SKILL.md"
);
const { importChatGPTControl } = await import(`${loaderUrl.href}?t=${Date.now()}`);
const { createChatGPT } = await importChatGPTControl();

const chatgpt = createChatGPT({
  agent: globalThis.agent,
  reporting: { enabled: true, includeContent: false }
});

Do not import from an older manually installed skill runtime; the plugin-bundled runtime is the intended source.

Bridge Bootstrap Handoff

This skill assumes the Codex Chrome bridge is already available in the JS runtime. If globalThis.agent is missing, do not stop after reporting hasAgent: false; initialize the Chrome plugin runtime using the bootstrap steps from the codex-chatgpt-control skill, then retry the consult.

After bootstrap, verify both values before using the snippets below:

JSON.stringify({
  hasAgent: !!globalThis.agent,
  hasBrowser: !!globalThis.browser
}, null, 2);

Only report browser_bridge_unavailable after the browser-control bootstrap itself fails or the bridge remains unavailable. If messages.ask, messages.wait, messages.status, messages.readLatest, messages.waitAndRead, mode selection, file attach, or existing-tab selection returns a partial/timeout result, preserve the raw structured result and continue under the full browser-control skill. The Pro wrapper is for consult intent; browser-control is the source of truth for low-level diagnosis.

Quick Consult

const POLL_CHUNK = {
  timeoutMs: 25_000,
  stableMs: 1_500,
  pollMs: 750,
  mode: "deep_research"
};

function isStillRunning(data) {
  return data?.completionState === "generating" || data?.generationActive === true;
}

const pro = chatgpt.agent({
  name: "chatgpt-pro-consultant",
  instructions: [
    "You are being consulted as ChatGPT Pro from a visible ChatGPT web session.",
    "Be critical, constructive, specific, and evidence-aware.",
    "Call out assumptions, risks, missing information, and concrete next steps.",
    "Return clear Markdown."
  ].join("\n"),
  defaults: {
    wait: false,
    read: false,
    report: { enabled: true, includeContent: false }
  }
});

async function readBoundedProAnswer() {
  const read = await chatgpt.messages.waitAndRead({
    ...POLL_CHUNK,
    role: "assistant",
    format: "markdown"
  });

  if (
    read.ok &&
    read.data?.complete !== false &&
    !isStillRunning(read.data)
  ) {
    return { status: "complete", read };
  }

  const progress = await chatgpt.messages.status({ maxPreviewChars: 500 });
  return {
    status: isStillRunning(progress.data) ? "pro_still_running" : "read_incomplete",
    read,
    progress
  };
}

const submitted = await chatgpt.runner.run(pro, {
  input: "Review this plan and recommend improvements:\n\n...",
  thread: { type: "new" },
  experience: "chat",
  configuration: { intelligence: "Pro" },
  report: { enabled: true, includeContent: false }
});

const configurationStep = submitted.steps?.find(step => step.id === "configuration");
if (!submitted.ok || configurationStep?.ok === false || configurationStep?.dataPreview?.verified !== true) {
  console.log(JSON.stringify({
    status: "blocked_before_submit_or_configuration_unverified",
    thread: submitted.state?.thread ?? submitted.data?.thread,
    configurationStep,
    interruptions: submitted.interruptions
  }, null, 2));
} else {
  const answer = await readBoundedProAnswer();

  if (answer.status !== "complete") {
    console.log(JSON.stringify({
      status: answer.status,
      message: "Prompt is already submitted. Do not resubmit; poll/read this same thread again.",
      thread: submitted.state?.thread ?? submitted.data?.thread,
      configurationStep,
      progress: answer.progress,
      read: answer.read
    }, null, 2));
  } else {
    console.log(answer.read.data?.responseText ?? "");
  }
}

Before describing the answer as a Pro-setting consultation, inspect configurationStep. It must succeed with dataPreview.verified === true. Do not infer an underlying model solely from the visible label. Do not pass response on the submit run; response formatting belongs on the separate bounded poll. If the poll times out, returns partial, or reports pro_still_running, do not submit the prompt again; reuse the same open thread and run another bounded messages.status / messages.waitAndRead call.