chatgpt-delegate

name	chatgpt-delegate
description	Use when Codex should delegate a visible, user-directed task to ChatGPT Chat or Work, including configuration discovery, long-running Work progress, steering, response capture, files, and artifacts.

ChatGPT Delegate

Use this skill for outcome-oriented delegation to the visible ChatGPT product. Choose Chat for a conversational answer or review. Choose Work for a longer task with progress, steering, files, or deliverable artifacts. Use the official Codex SDK or CLI—not this skill—for repository editing, terminal execution, sandboxing, and code deployment.

This is a workflow over the codex-chatgpt-control plugin. It operates visible UI only and must not call private ChatGPT endpoints, inspect credentials or browser storage, bypass login or confirmation, or submit sensitive material without the user's approval.

Runtime Loader

Resolve relative paths from this SKILL.md. Load the plugin-bundled runtime:

const loaderUrl = new URL(
  "../../runtime/import-chatgpt-control.mjs",
  "file:///absolute/path/to/plugins/codex-chatgpt-control/skills/chatgpt-delegate/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 }
});

If globalThis.agent is absent, use the bridge-bootstrap instructions in the broad codex-chatgpt-control skill. An ordinary-shell browser_bridge_unavailable result is expected and should be reported, not worked around.

Choose The Experience

• chat: reviews, questions, synthesis, brainstorming, search-oriented answers, or continuing an existing conversation.
• work: longer research or production tasks that benefit from progress checks, steering, files, or artifacts.
• official Codex: local code changes, commands, tests, branches, deployments, or repository-native agents.

Do not choose a surface solely from a model name. Availability varies by account, workspace, locale, and rollout. Inspect the visible capability graph and apply only controls that are actually present.

When the user requests Chat or Work, call experience.open for that surface before inspecting configuration or submitting. Do not assume the currently visible pane is the requested one. The SDK handles the current Chat/Work radio selector and returning from an active Work task to the home selector, while retaining compatibility fallbacks for older UI shapes.

const detected = await chatgpt.experience.detect();
const configuration = await chatgpt.configuration.inspect({
  experience: detected.data?.experience === "unknown"
    ? undefined
    : detected.data?.experience
});

Chat Delegation

Use the runner for a visible Chat review:

const reviewer = chatgpt.agent({
  name: "chatgpt-reviewer",
  instructions: [
    "Be critical, constructive, specific, and evidence-aware.",
    "Call out assumptions, risks, and concrete next steps.",
    "Return clear Markdown."
  ].join("\n"),
  defaults: {
    experience: "chat",
    wait: false,
    read: false,
    report: { enabled: true, includeContent: false }
  }
});

const submitted = await chatgpt.runner.run(reviewer, {
  input: "Review this plan:\n\n...",
  thread: { type: "new" },
  experience: "chat",
  configuration: { intelligence: "Pro" }
});

The configuration is a visible preference, not a guaranteed underlying model identifier. Inspect the configuration sequence step and require verified: true before claiming that the requested setting was active. If the account does not expose that option, report the structured blocker and visible candidates.

For long answers, submit once and poll the same thread:

const progress = await chatgpt.messages.status({ maxPreviewChars: 500 });
const read = await chatgpt.messages.waitAndRead({
  timeoutMs: 25_000,
  stableMs: 1_500,
  pollMs: 750,
  role: "assistant",
  format: "markdown"
});

partial, completionState: "generating", or generationActive: true means the prompt may already be running. Never resubmit it blindly.

Work Delegation

Start a fresh Work task with explicit, strictly verified configuration:

const started = await chatgpt.work.start({
  prompt: "Research the options and produce a decision-ready implementation brief.",
  newTask: true,
  configuration: {
    model: "GPT-5.6 Sol",
    effort: "High",
    speed: "Standard"
  },
  wait: false,
  read: false
});

newTask defaults to true. If an existing task is loaded and no unique new-task control can be verified, the SDK blocks rather than appending the prompt accidentally. Pass newTask: false only when intentionally continuing the currently visible Work task.

Poll without resubmission:

const status = await chatgpt.work.status({ includeArtifacts: true });
const waited = await chatgpt.work.wait({
  timeoutMs: 25_000,
  stableMs: 1_500,
  pollMs: 750,
  responseContent: "metadata"
});

Steer the same task:

await chatgpt.work.steer({
  prompt: "Prioritize the migration risks and add a two-week execution sequence.",
  wait: false,
  read: false
});

Read the latest response and enumerate artifacts:

const latest = await chatgpt.work.readLatest({ format: "markdown" });
const artifacts = await chatgpt.work.artifacts.listLatest({});

Use chatgpt.work.artifacts.wait(...) and downloadLatest(...) for visible downloadable deliverables. Attach only user-approved absolute host-local paths.

When the deliverable has a known filename, use the file workflow with an exact case-insensitive regular expression:

await chatgpt.files.downloadLatest({
  destDir: "/absolute/output/dir",
  filenamePattern: "^report\\.csv$"
});

This supports direct file links and filename-labelled artifact previews. A filename mismatch is a blocker, not permission to accept another visible image or older file.

Compatibility

The legacy mode runner input, chatgpt.modes.set/get, and chatgpt-pro-consult skill still work for existing callers. New work should use experience plus configuration, because Chat and Work expose different nested axes and strict postcondition verification.

Blockers And Reporting

Stop and report:

• browser_bridge_unavailable
• login_required
• captcha
• rate_limit
• permission
• needs_confirmation
• selector_drift

Preserve the SDK's real ok, status, warnings, blocker, error, steps, and task/thread reference. Do not wrap a failed result as success. Reports remain redacted by default.

Treat delegated output as another model's judgment, not verified truth. Verify current or high-stakes claims with primary sources.