V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage estimates, Bigger Context, and control over when a task starts a fresh browser conversation. Version 6.1.14 improves context transfer, compaction, tool polling, and setup, and makes ordinary ChatGPT chats the only supported chat mode.
Changes in 6.1.14
- Reliable multipart context transfer. Rich formatting in ChatGPT's user-message bubbles no longer makes correctly submitted JSON appear changed after acknowledgement (#886).
- Workspace recovery when Codex omits the working directory. Tasks with several workspace roots recover their working directory from the current native Codex turn. The selected directory must belong to the verified roots (#894).
- Compaction keeps the task and goal. Preserves the accepted task and active goal through repeated compaction, supports child-agent conversations without inventing a human message, and retains attribution for native desktop delegations. Date changes after compaction keep the verified environment (#884, #889, #892, #896).
- Tool polling fits the MCP request limit. Command polls longer than 60 seconds are rejected with guidance to use shorter polls. This prevents a requested 300-second wait from being cut off by the transport's 90-second limit (#904).
- More time for a delayed final answer. After visible generation or tool activity stops, the bridge allows up to ten minutes for ChatGPT's final answer instead of failing after one minute. This addresses the delayed-final case reported in #885; other missing-final cases remain open.
- Simplified Chinese effort controls. Model verification recognizes the observed Chinese labels, including GPT-6 High (
高), while continuing to reject a conflicting selection (#911). - MCP setup and model diagnostics. Completing setup no longer requires a fresh model-catalog check when the core installation is already complete (#893). Doctor reports model-catalog failures separately from local proxy health; this does not resolve the connection refusal reported in #895.
- Ordinary chats throughout. Removes Temporary Chat mode and its setting. Existing configurations remain readable, and other settings are preserved. Task chats appear in ChatGPT history; account memory and custom instructions may apply. The optional New browser chat for each turn setting remains available.
- Security and diagnostics. Updates Electron to 42.11.16 and affected dependencies. Logs include a numerical breakdown of context token costs to help investigate repeated compaction without recording message contents.
The repeated-compaction case in #863, disappearing conversation in #865, Windows tunnel installation failure in #870, and other open account or transport reports remain under investigation. ChatGPT account limits and safety refusals still apply.
V6 features and earlier fixes
- Models and effort. GPT-5.6 and GPT-6 choices remain separate, with account-supported reasoning levels in Codex. Free accounts expose Luna/Think. Earlier patches improved the updated model picker, GPT-5.6 Pro verification, translated controls, locked Pro options, and cancellation during Browser Smoke.
- Bigger Context. Large tasks use one, two, or six messages. GPT-6 Sol supports up to 120,000 tokens on Plus and 240,000 on Pro at supported reasoning levels; Instant keeps its standard context. Earlier fixes cover collapsed messages, upload processing, hidden composers, oversized-message errors, and repeated rejected uploads. Luna/Think support remains experimental.
- Chat controls and Pro estimates. Optional fresh chats rebuild the same Codex task in a new conversation. Optional next-message suggestions use GPT-6 Sol Medium when the context fits. The Limits tab estimates the launcher's own submissions for supported Pro plans, with separate model allowances and a warning from 75%; it does not report OpenAI's remaining quota or messages sent elsewhere.
- Long tasks and tools. Earlier updates corrected compaction summary tokens, excluded computer sleep from inactivity timers, preserved accepted work after stream disconnection, and prevented overlapping connector handshakes from interrupting other tasks. Permission cards point to the plugin's Permissions → Allow all actions setting, and prompts use the configured plugin name.
- Replies and diagnostics. Preserves code, links, formulas, and plan markers across supported ChatGPT layouts. Activity distinguishes preparation, generation, tool execution, and approval waits. Failed browser captures are retained for up to three days within a 100 MB budget, with active captures protected.
- Startup, updates, and desktop packages. Earlier fixes improved passkey sign-in, account switching, browser recovery, system proxy support, Windows certificate handling, and update rollback. Upgrades preserve explicit settings, and older launchers cannot replace a newer runtime. Packages cover Windows x64, both macOS architectures, and Linux x64/ARM64.
Thanks to @nofihq for the compaction, delegation, date-handling, and dependency contributions; @thomastsaidotcom for earlier smoke-test reproductions; and everyone who supplied logs and tested fixes, including @iminfine, @zmahomedy, @alexzwy51, and @martjay.
The Pro Researcher skill is available for delegating research to Pro sub-agents.
Update
Finish active tasks, quit the launcher, install the update, and reopen it. Restart Codex after updating its integration. Your other settings and browser profile are preserved; the old Temporary Chat preference is no longer used.
If a closed issue still happens, open a new issue with Activity → Export safe log, the failure time, launcher/Codex versions, and model/effort. If your issue remains open, add the new results there.