Codex Web GPT v5.0.4

V5 adds Zero Risk Mode: the full Codex MCP harness still runs through ChatGPT, but the launcher never reads or changes the ChatGPT page and never sends a prompt for you.

V5 highlights

• Zero Risk Mode. The launcher prepares and copies the complete Codex prompt; you paste it into ChatGPT, add any images yourself, select the Codex Zero Risk connector plus the model and effort you want, send it, and confirm Sent. Codex tools, patches, terminal calls, structured output, streaming, and native progress stay available through the same turn-bound harness without browser automation.
• A separate Zero Risk MCP setup. Automatic and Zero Risk modes have independent connector identities, Tunnel IDs, API keys, runtime state, and recovery. Switching modes is transactional and never silently reuses the other mode's private credentials.
• Optional Zero Risk Pro profile. The install menu can add a second ChatGPT Web — Zero Risk Pro row with roughly 250,000 usable tokens before compaction. It is intentionally explicit: select ChatGPT Pro manually for every turn; a Pro account is required, and the larger limit may not work with every effort.
• Manual turns that remain real Codex turns. A neutral request ID binds each ChatGPT response to the exact Codex turn through MCP metadata. Concurrent turns stay isolated; reconnects rejoin the same request; cancellation, completion, retained-chat continuation, and compaction are explicit and fail closed.
• Zero Risk compaction. The active ChatGPT agent produces the checkpoint through the bound MCP control, the old chat is retired, and the compacted Codex continuation starts from a fresh manual chat without exposing private control values in the pasted prompt.
• A clearer launcher flow. First-run onboarding now offers With Automation and Zero Risk, defaults to automation, and keeps the same choice in Settings. The MCP guide includes trimmed, accelerated video walkthroughs with an in-launcher fullscreen viewer.
• Three-language product parity. The new setup, mode selection, manual-turn guidance, diagnostics, and model-profile controls are complete in English, Chinese, and Japanese.

v5.0.4 hotfixes

• Steering continues instead of deadlocking. A new instruction arriving with an outstanding tool result now retires the previous browser turn and resumes from the complete Codex history. Late requests from the superseded instruction cannot restart it.
• Goals survive repeated compaction. Fixed a checkpoint mismatch caused by Codex removing its internal goal context. Continuations and reopened tasks keep the correct user-message identity and working directory instead of failing with turn_id conflicts or missing cwd in this scenario.
• Zero Risk allows slow model startup after Sent. Confirming Sent now ends the manual submission countdown. The turn stays cancellable while ChatGPT thinks before its first MCP call, instead of failing after another 30 seconds. Fixes #325.
• Retained chats recheck the requested effort. A High request no longer silently inherits Medium from the saved ChatGPT tab. Addresses the effort mismatch in #308.
• Browser recovery keeps the correct owner. A delayed browser wake gets a fresh response check before the missing-response deadline is evaluated. A failed viewport check now releases its newly acquired connection correctly. DOM timeouts have not been increased; the remaining freeze reports in #278, #308, and #321 remain under investigation.
• Normal Codex settings edits preserve the integration. Native TOML edits and Windows-style line-ending normalization no longer make an unchanged managed interrupt hook fail verification. Unrelated settings survive disconnect and repair; changed hook commands or permissions are still rejected.
• Preparation errors keep their real cause. The MCP capability is published only after prompt preparation succeeds, preventing a preparation failure from being replaced by a misleading turn token is invalid or expired error.
• Clearer setup and diagnostics. The final verified MCP setup step shows its checkmark. Connector diagnostics inspect the current composer, and #323 stream-conflict diagnostics now distinguish the failure without logging response contents. The Luna stream-conflict cause remains under investigation.

Thanks to @yzx-deepseek for the preparation and config-editing analysis in #327.

Please update through the launcher, restart Codex, and retry the affected workflow. If the same failure remains, reply in its issue with your versions and a fresh Activity safe-log export from that attempt.

Install

macOS or Linux:

curl -fsSL https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install-launcher.sh | sh

Windows PowerShell:

irm https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install-launcher.ps1 | iex