| name | cgpro |
| description | Use the local cgpro CLI to consult ChatGPT 5.5 Pro (extended thinking + live web search) AND to mirror the current Claude Code project into a ChatGPT Project so conversations land in the right folder with persistent project memory. Reach for this skill when the user wants a second opinion, deep reasoning, fresh web info, multi-turn chats, or wants to make a directory's GPT-5.5 conversations stay organised inside a named ChatGPT Project (auto-routed, with a memory.md that's prepended to every new chat). |
cgpro: ChatGPT 5.5 Pro from the terminal
cgpro is a CLI that drives the user's authenticated ChatGPT session
against chatgpt.com to expose GPT-5.5 Pro with extended thinking
and live web search — capabilities that are not exposed by the public
OpenAI API or by Codex CLI as of April 2026.
Use this skill when the user wants any of:
- A second opinion on a hard architectural / debugging question
- Extended reasoning on a problem that benefits from deliberation
- Live web search results with citations (current events, fresh docs)
- A multi-turn ChatGPT thread driven from the terminal
How to call it
cgpro is a single binary on the user's PATH. All calls are plain
shell — use the Bash tool (or your platform equivalent) to run them.
Single question
cgpro ask "your question"
Streams the answer to stdout. Live web search is always on by
default — policy choice for this CLI, freshness + sources matter
more than determinism. Use --json for NDJSON event stream when you
want to parse programmatically.
Long-running turns (>10 min)
GPT-5.5 Pro extended-thinking can take 5-30 min for non-trivial
questions and over an hour for hard reasoning. cgpro defaults
--timeout to 7200 (2 h). When invoking from a sandboxed shell
(Claude Code Bash, IDE task runner, etc.) that has a per-call
ceiling, use run_in_background and poll the output stream — the
turn keeps running browser-side regardless of who's listening.
Multi-turn
cgpro auto-resumes the most recent shell-session conversation for
30 minutes, so back-to-back cgpro ask calls thread automatically.
For an explicit multi-turn REPL, use cgpro chat. For a clean slate,
pass --new-session.
cgpro ask "explain CRDTs in 3 bullets"
cgpro ask "now sketch a Rust trait for a state-based one"
cgpro ask --new-session "totally different question"
Resume a saved conversation
cgpro thread list
cgpro thread list --remote --refresh
cgpro ask --resume <name|uuid> "follow-up"
cgpro ask --save <name> "..."
Project routing (keep your Recents clean)
cgpro mirrors the current cwd into a ChatGPT Project so conversations
land there and out of the global Recents list. Identity is the git
remote (when present) or the cwd path.
cgpro project list
cgpro project link "MyProj"
cgpro project show
cgpro project digest
cgpro ask "anything"
cgpro ask --no-project "..."
Project memory lives at ~/.cgpro/projects/<gizmoId>/memory.md and is
appended (not overwritten) by digest.
Speed: daemon mode
Cold-starting Chromium adds ~5-10s per call. Start the daemon once and
all subsequent asks use the warm browser:
cgpro daemon start
cgpro ask "..."
cgpro daemon status
cgpro daemon stop
cgpro ask auto-detects the daemon. No client-side change needed.
When NOT to use cgpro
- Quick lookups Claude already knows. Don't waste a Pro turn on
trivial questions.
- Tasks where the user wants Claude's own answer. If they ask
you to do something, do it. Don't reflexively bounce to GPT-5.5.
- When
cgpro status shows the user is not signed in. Tell the
user to run cgpro login (interactive) or cgpro adopt (imports
the desktop app's session). Don't try to authenticate yourself.
Health check
If cgpro returns Not signed in, Cloudflare challenge, or
Selector broken, surface the error to the user — those need
human intervention. For selector breakage, cgpro doctor audits
the live DOM and points at the broken selector.
Conformance
cgpro uses the user's own ChatGPT subscription via the standard
sign-in flow — same credentials, same session, same rate limits as
the desktop app. Personal CLI use only.