- name
- chatgpt-skill
- description
- Automate ChatGPT Web conversations through a local browser session with persistent login state, one-shot prompts, reusable multi-turn sessions, and conversation metadata storage. Use when OpenClaw/Codex needs ChatGPT-style normal dialogue through the web UI instead of the OpenAI API, including login setup, continued conversations, session reuse, or browser-based troubleshooting.
ChatGPT Web Skill
Use this skill to drive the ChatGPT Web UI through a local browser profile.
Core Purpose
- Reuse ChatGPT login state stored inside the skill directory
- Start a fresh conversation or reopen an existing one
- Ask one-off questions through a normal browser page
- Keep a long-lived browser session for multi-turn dialogue
Always Use python3 scripts/run.py
Always run commands through the wrapper:
python3 scripts/run.py auth_manager.py status
python3 scripts/run.py auth_manager.py validate
python3 scripts/run.py auth_manager.py setup
python3 scripts/run.py ask_chatgpt.py --question "你好"
python3 scripts/run.py session_manager.py createThe wrapper ensures the local .venv exists, installs dependencies, and runs the target script with the correct interpreter.
Authentication Flow
- Check local state:
python3 scripts/run.py auth_manager.py status- Validate the real browser state:
python3 scripts/run.py auth_manager.py validate- If needed, complete login manually in a visible browser:
python3 scripts/run.py auth_manager.py setup- To clear local auth/profile data:
python3 scripts/run.py auth_manager.py logoutBasic Chat Flow
One-shot question:
python3 scripts/run.py ask_chatgpt.py --question "你好"Debug in visible browser:
python3 scripts/run.py ask_chatgpt.py --question "总结这段文本" --show-browserContinue a known conversation id:
python3 scripts/run.py ask_chatgpt.py --conversation-id <conversation_id> --question "继续"Ask with an explicit model workflow and save a proof screenshot:
python3 scripts/run.py ask_chatgpt.py \
--new-chat \
--model "GPT 5.4 Thinking" \
--extended-thinking \
--proof-screenshot \
--question "请你推荐最近一个月，RLVR领域的论文"Persistent Session Flow
Create a reusable session:
python3 scripts/run.py session_manager.py create
python3 scripts/run.py session_manager.py create --conversation-id <conversation_id>Ask follow-up questions in the same tab:
python3 scripts/run.py session_manager.py ask --session-id <session_id> --question "继续问"You can also switch model, force a fresh chat, enable Extended thinking, and save a proof screenshot in the same call:
python3 scripts/run.py session_manager.py ask --session-id <session_id> --new-chat --model "GPT 5.4 Thinking" --extended-thinking --proof-screenshot --question "请你推荐最近一个月，RLVR领域的论文"Inspect or maintain sessions:
python3 scripts/run.py session_manager.py list
python3 scripts/run.py session_manager.py info --session-id <session_id>
python3 scripts/run.py session_manager.py reset --session-id <session_id>
python3 scripts/run.py session_manager.py close --session-id <session_id>
python3 scripts/run.py session_manager.py gcData Storage
All data stays inside the skill directory:
data/auth_info.json— auth metadata and last validation statusdata/browser_state/— cookies, storage state, persistent browser profiledata/conversations.json— discovered/opened conversation metadatadata/sessions.json— persistent session metadatadata/session_runtime/— daemon socket, pid, and runtime artifactsdata/screenshots/— debug screenshots captured on failures
Error Model
The scripts return explicit machine-readable errors for cases such as:
- not logged in
- redirected to login
- account chooser encountered
- CAPTCHA / 2FA / human verification required
- page load failed
- prompt input missing
- send path unavailable
- reply timeout
- empty reply
- page structure changed
- session not found
- invalid conversation id
- session daemon unavailable
- browser profile in use
- network or proxy failure
Best Practices
- Validate auth before deeper automation
- Use
--show-browserwhen the UI changes - Prefer persistent sessions for multi-turn workflows
- Keep proxy env vars explicit on Linux when network behavior is unstable
- Inspect
data/screenshots/and JSON error payloads before patching selectors
Limitations
- ChatGPT Web selectors can change without notice
- A blank new chat may still not expose a stable server conversation id until the first message is sent; the skill also probes frontend state, storage, and resource history, and may fall back to a client thread id such as
WEB:...when ChatGPT keeps the URL at/ - Some login or risk-review flows require manual intervention
- Conversation deletion is intentionally left unimplemented until the web flow is stable enough