ask-pro

ask-pro is a Codex plugin and CLI that lets coding agents ask ChatGPT Pro for a focused second opinion through a human-controlled browser session.

Use it for architecture calls, production-risk reviews, migrations, debugging strategy, and implementation planning. The calling agent still owns the work: ask-pro collects the consult, stores the answer, and never applies generated code automatically.

Install

Add this repository as a Codex plugin marketplace:

codex plugin marketplace add https://github.com/Pimpmuckl/ask-pro --ref main

Then open /plugins, install ask-pro, and start a new Codex thread. Refresh later with:

codex plugin marketplace upgrade ask-pro

If the marketplace does not appear after adding or upgrading it, restart Codex. The repository marketplace exposes the root plugin through a Git-backed plugin entry, so no manual copy into ~/.codex/plugins/cache is needed.

For local development from a checkout:

pnpm install
pnpm run build
pnpm run plugin:refresh

pnpm run plugin:refresh updates the local Codex plugin cache from the source checkout. By default it refreshes the installed Git marketplace cache under ~/.codex/plugins/cache/ask-pro/ask-pro/<version>, so no separate local marketplace install is needed for development. Do not edit ~/.codex/plugins/cache/... by hand.

Git marketplace installs cache the source checkout. If ask-pro is not on PATH, agents should use the cached runner under ~/.codex/plugins/cache/<marketplace>/ask-pro/<version>/scripts/run-cached-cli.mjs. The first runner call may install dependencies and build dist in that cache.

Requirements

• Node.js 24+
• pnpm 10+
• Chrome
• A ChatGPT account with Pro access

Authentication is manual. ask-pro never asks for, types, reads, or logs passwords, MFA codes, recovery codes, session cookies, or raw auth tokens.

The default persistent browser profile is:

~/.agents/skills/ask-pro/browser-profile

Each new profile may need a human login once. On Windows, fresh runs for an auth-ready managed profile start minimized; login, resume/recovery, stale-auth, and debug paths stay visible or are restored for human action.

Quick Use

Ask for an inline markdown consult:

ask-pro --no-temporary --prompt-file question.md --files src --files tests

Use --extended only for deep, long-running architecture or production-risk questions:

ask-pro --extended --no-temporary --prompt-file architecture.md --files docs --files src

Request generated files only when you really need an implementation package:

ask-pro --artifacts --prompt-file implementation-plan.md --files src

Harvest the answer:

ask-pro --harvest <session-id>

Agent Guidance

• Prefer --prompt-file for multiline prompts.
• If ask-pro is not on PATH, use the cached plugin runner instead of a mutable source checkout.
• Treat ChatGPT Pro as a cold oracle: it does not know the repo, user, prior decisions, or Codex thread context unless you include that in the prompt or attached files.
• Provide more background than feels strictly necessary: product goal, current state, constraints, files attached, options considered, and the exact output you need.
• Prefer --no-temporary for repo advisories, review rounds, large bundles, or anything where recovery matters.
• Keep bundles focused: relevant source, focused tests, docs that define the contract, recent changes, and validation status.
• Add .ask-pro/ to consuming repos' .gitignore.
• Treat INCOMPLETE_ANSWER / preamble_without_artifacts as not done; resume or rerun with a tighter prompt and bundle.
• Never execute generated zip contents automatically.

Useful advisory prompt starter:

Return final markdown only. Do not answer with a preamble. Do not produce an implementation package. Rank findings by severity. Treat attached bundle as authoritative. Call out uncertainty.

CLI

ask-pro [options] [question...]

Common options:

• --files <pattern>: add a file, directory, or glob to CONTEXT.zip.
• --prompt-file <path>: read the question from a UTF-8 file; use - for stdin.
• --artifacts / --response-zip: ask for ask-pro-response.zip.
• --resume [session-id]: resume a prepared, waiting, or auth-gated session.
• --status [session-id]: print compact session state.
• --harvest [session-id]: print ANSWER.md for answer-bearing sessions.
• --extended: request Extended Pro thinking.
• --temporary: require ChatGPT Temporary Chat.
• --no-temporary: use normal ChatGPT for better recovery.
• --verbose: print browser automation diagnostics.

Session data lives under .ask-pro/sessions/<session-id>/.

Development

Fast checks:

pnpm run build
pnpm run lint
pnpm run test:ask-pro
pnpm run format:check
pnpm pack --dry-run

Manual browser smokes are opt-in because they open a real ChatGPT session. See docs/manual-tests.md.