[TanChuping]

Why

A Python toolkit for agents that hand ChatGPT web Pro several tasks over the same repository through Google Drive: a persistent repository mirror on Drive, a request file for changing a running task without sending another message, and a delivery marker the local agent can wait on. Offered as a self-contained contrib/ directory, not wired into the CLI.

How it relates to the CLI

• It builds on --timebox, --approve-apps and --drive rather than reimplementing them; the suggested task header is submitted with voyager ask --timebox 70.
• --drive uploads specified files per question with rclone; the mirror keeps one committed snapshot of a repository at a stable Drive location across questions (text layer with file/symbol index and redaction, optional binary volumes, MIRROR.json with commit and hashes).
• Suggested merge: keep it as is under contrib/, outside the npm package. If something moves into the CLI later, the mirror and the versioned request file are the candidates; the sandbox reminder hooks and prompt examples fit better as optional recipes. contrib/web-pro-kit/README.md says the same for an agent reading it cold.

Outline

contrib/web-pro-kit/
├── README.md            what it is, how to merge, platform note
├── RUNBOOK.md           setup, what a project shares, publish, steer, collect, adapt
├── config.example.json  workspace path and redaction terms (~/.voyager/web-pro-kit.json)
├── scripts/             drive_mirror, code_index, secret_patterns, kit_config, drive_request, drive_wait
├── request-channel/     channel.py (sandbox reminder), README.txt, VERSION, PROMPT-BLOCK.md
└── templates/           suggested task header, blank project header

Verification

• End to end on Windows 11 + Google Drive for desktop + Python 3.12, through voyager ask --model extra-high --timebox 25 with a paired extension: the mirror was fetched and its commit and hashes verified; the task wrote test.txt first; channel.py passed its selftest and installed in the ChatGPT sandbox; a second request appended mid-task was picked up within three minutes and applied; the ZIP, NOTES.md and DONE.txt landed on Drive in that order and drive_wait.py returned on the marker. The approval card was answered by hand.
• That run surfaced one bug, fixed here: an unpaired BEGIN … PRIVATE KEY line made the redactor drop the rest of the file. A selftest now mirrors the kit itself, extracts it and runs every script's selftest from the copy.
• All seven selftests pass on the Windows host and from an extracted mirror. In the ChatGPT Linux sandbox five passed and two failed on the bug above; the fix has not been re-run there. oxfmt --check passes on the new directory; Python is checked by the selftests.
• macOS hosts and Pro subagent runs were not tested.

🤖 Generated with Claude Code