GPT KeepGoing

Auto-continue for the ChatGPT Windows desktop app.

English | 简体中文

---

The problem

You leave a long task running in the ChatGPT desktop app. It stops because you hit the usage limit, and tells you to try again at 00:45. Nobody is awake at 00:45. In the morning the task is exactly where it stopped.

The app will not fix this by itself, because:

1. It does not refresh its quota while the window is in the background. We measured 24 samples over 9 minutes across a reset point: the banner never changed.
2. Even when it refreshes, someone still has to click Resume goal or type "continue".

What this does

GPT KeepGoing is a small Windows tray app. It watches one conversation and continues it for you.

• Finds the client automatically. No configuration, no window title matching.
• Checks every minute. Reads the state from the client's accessibility tree.
• Continues the right way — goal first.
  • If the conversation has a goal, it clicks Resume goal, and confirms the follow-up dialog if one appears.
  • If there is no goal, it types a continuation message and sends it.
• Continues after a finished task, not only after a quota reset.
• Speaks your language. 7 UI languages, and the sent message is localized too.
• Shows your quota. 5-hour and weekly windows, with reset times.
• Stays out of the way. Tray icon, optional autostart, a 20-line log, and notifications only when you are actually needed.

It never touches your account, never calls an API, and never modifies the client.

Languages

The interface and the continuation message both follow the selected language. Language names are always written in their own script.

Language	Interface	Message sent
简体中文	自动续跑助手	继续
English	Auto Resume	Continue
日本語	自動再開アシスタント	続けて
한국어	자동 재개 도우미	계속 진행해
Deutsch	Auto-Fortsetzen	Weiter
Français	Reprise automatique	Continue
Italiano	Ripresa automatica	Continua

How it decides

State it reads	What it does
Running	Waits. Continues when the task ends.
Quota exhausted, reset time known	Sleeps until the reset time, then continues.
Quota exhausted, no reset time	Checks every minute until quota returns.
Reset time passed, banner still says exhausted	Still tries to continue, then checks whether the task actually started. The banner is known to lag.
Waiting for permission	Does nothing. Sends a notification instead.
Waiting for your reply	Does nothing. Sends a notification instead.
Idle	Waits.
Goal paused, or stalled by the quota	Continues when the limit resets or the task ends.
Goal completed	Records the time, notifies you, and turns itself off.
Goal blocked	Confirmed over 3 checks before it turns itself off. If a reset time is still pending it keeps waiting and tries Resume once at that time.

The app never approves a permission request for you, and never sends a message when the composer already has a draft.

Requirements

• Windows 10 or 11
• The ChatGPT desktop app (package OpenAI.Codex) installed and signed in
• No administrator rights, unless the client itself runs elevated

Tested against client build 26.1013.4894.0.

Run it

From source

pip install uiautomation psutil pystray pillow
python main.py

Build a single exe

pip install pyinstaller
python -m PyInstaller --noconfirm --clean --onefile --windowed ^
  --name AutoResume --collect-all comtypes --collect-all pystray ^
  --hidden-import comtypes.gen main.py

The result is dist\AutoResume.exe, about 37 MB, no Python required.

Using it

1. Start the app. It shows the conversation that is currently open.
2. Pick a target conversation, either from the dropdown or with Use current.
3. Turn on Auto continue. A dialog explains what will happen: the app must bring the window to the front and briefly switch conversations to force the client to refresh. Confirm only if that is acceptable.
4. Leave it running. It checks every minute.

Closing the window sends it to the tray. Use the tray menu to reopen it or to quit.

What it shows

• Target conversation and the project it belongs to.
• Current state in plain words.
• Limit reset time, taken from the client banner.
• Quota: 5-hour and weekly, with remaining percentage, reset time, and how old the data is.
• Last 20 log lines, with timestamps.

Honest limitations

• The client does not expose a live quota percentage. Percentages come from the client's own local session records, which are only written when a turn finishes. When the recorded window no longer matches the banner, the app says "percentage pending" rather than showing a stale number.
• UI automation depends on the client's wording. A client update that renames a button can break detection. The strings are centralised in app/constants.py and app/i18n.py.
• One conversation at a time. Watching several would mean switching threads every minute, which is disruptive.
• Not affiliated with OpenAI. This is a third-party automation tool.

How it works

Layer	What it uses
Finding the client	Win32 EnumWindows + process name
Reading state	Windows UI Automation on the Chromium accessibility tree
Sending	UI Automation invoke, clipboard paste, and keystrokes
Quota numbers	Read-only parsing of %USERPROFILE%\.codex\sessions\**\rollout-*.jsonl
Scheduling	A background thread with a state machine

No network access. No telemetry. Configuration and logs live in %LOCALAPPDATA%\AutoResume\.

Development

main.py              entry point: single instance, tray, UI, scheduler
app/discovery.py     find the window (Win32), bring it to the front
app/reader.py        read a state snapshot via UIA
app/planner.py       pure decision functions
app/actor.py         refresh ladder, switch thread, resume goal, send message
app/engine.py        background scheduler and state machine
app/ratelimit.py     read 5-hour and weekly quota from local session records
app/i18n.py          7 languages
app/tray.py          tray icon and menu
app/autostart.py     HKCU Run entry
app/notify.py        tray balloon notifications
app/ui.py            tkinter dashboard

Tests (40 cases, no GUI required):

python tests\test_planner.py
python tests\test_engine.py
python tests\test_i18n.py

Design notes live in specs/auto-resume.md.