miuuyy/codex-chatgpt-web v6.1.18

on GitHub

V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage estimates, Bigger Context, and control over browser conversations. 6.1.18 improves long-task continuation, preserves goal changes, and fixes macOS passkey-session import.

Changes in 6.1.18

• Current workspace folders survive partial updates and compaction. The bridge checks the folders recorded in the current native Codex session instead of relying only on the task's initial folders. This addresses the partial visualization-folder update in #939 and improves recovery for the related missing-folder and compaction reports in #936 and #945. Working-directory, permission and task-ownership checks remain enforced; those two related reports still need confirmation on the affected tasks.
• Goal changes remain effective during commands and after compaction. Clearing or editing a goal during a running command no longer creates a false turn-ID conflict. The completed command result is preserved, and later compaction keeps the updated goal instead of reviving the previous task (#957).
• Automatic compaction can run before the next instruction arrives. An exact completed checkpoint can identify the previous task when native Codex starts pre-turn compaction before adding the next instruction. This does not authorize replaying an old task in another ordinary turn (#935).
• Old messages returning into view are not new submissions. Previously observed messages that reappear while scrolling a retained ChatGPT conversation no longer cause false multiple-turn errors. A genuinely competing new submission is still rejected (#948).
• macOS passkey sign-in closes its dedicated Chrome cleanly. Continue requests a normal Chrome shutdown and waits for a clean exit before importing the session, allowing pending cookies to be saved. This addresses the lost-session failure reported with a YubiKey; an unclean exit now stops the import with an explicit error (#973, thanks @zm2231).
• Update checks explain failures and allow a manual retry. Settings exposes update-check errors and a check-again action; GitHub rate-limit reset times are respected. Safe diagnostics also retain recognized browser network error codes without exposing response bodies.
• Build dependencies are patched. Vite and the affected minimatch dependencies are updated, while the legacy archive dependencies retain their required APIs.

Earlier V6 patches

• Model controls and setup. The October inline picker is recognized alongside supported earlier layouts. Setup preserves the previous configuration when an interface is unrecognized, instead of incorrectly replacing paid models with Luna. Existing installations affected by #932 need the model refresh described above.
• Additional workspace folders. Full Access tasks can continue after Codex adds its visualization or output folder. Current projects, working directory and permissions remain checked (#933).
• Large compaction and targeted recovery. Bigger Context compaction can trim older history when rebuilding a lost chat, preserving the latest checkpoint and instruction. Targeted delivery-timeout and multipart-staging recovery avoids resubmitting accepted tool work (#929, #930).

V6 features and earlier fixes

• Models and effort. Account-supported GPT-5.6 and GPT-6 models appear in Codex's native picker. Sol reasoning uses Effort, Instant keeps a separate context budget, and Free accounts expose Luna/Think. Earlier patches improved translated controls, Pro verification, locked options, exact Sol/Astra selection, six-position controls and Browser Smoke cancellation.
• Bigger Context and long tasks. Large histories use one, two or six ordered messages. GPT-6 Sol supports configured windows up to 120,000 tokens on Plus and 240,000 on Pro at supported reasoning levels; Instant keeps standard context. These are bridge context budgets, not guaranteed upstream message limits. Earlier fixes preserve the accepted task, active goal, date and delegation context through compaction, handle formatted user messages, and prevent repeated uploads of the same rejected transaction. One oversized message still must fit its individual stage. Luna/Think Bigger Context remains experimental.
• Chat controls and Pro estimates. Temporary Chat is enabled by default; turn off Settings → Temporary Chat to use saved conversations. An explicit opt-out is preserved. Optional fresh chats rebuild the same Codex task in a new browser 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; it does not report OpenAI's remaining allowance or messages sent elsewhere.
• Tools and replies. Full harness mode provides native files, commands, tools and approvals through MCP. Command polls are bounded to 60 seconds per MCP call. Earlier updates improved permission guidance, worktree access, tool settlement, cancellation, Markdown fidelity and delayed-final handling without resubmitting accepted work. Think recognizes the current icon across languages.
• Setup and desktop reliability. MCP setup and model-catalog loading have separate health checks. Doctor explains conflicting context overrides. Session warnings stay above the browser without shrinking it, and startup preserves settings and owned runtime state on failure. Earlier patches improved passkey sign-in, account switching, system proxy support, Windows certificate handling, update rollback and downgrade prevention.
• Diagnostics and packages. Failed browser captures are retained for up to three days within a 100 MB budget, with active captures protected. Diagnostics record stream completion and observed UI structures without message text or tool arguments. UI contracts have separate owners, dated evidence and regression fixtures. The launcher uses Electron 42.11.16 and pinned Bun 1.4.0, with Windows x64, macOS Apple silicon/Intel and Linux x64/ARM64 packages.

The separate model-selection, superseded-response, Linux Send-timeout and Codex 0.166 environment reports in #956, #961, #963, #969 and #971 remain under investigation. New native-environment records have arrived for #971; this release does not claim to resolve that case or the other reports. ChatGPT account limits and safety refusals still apply; the shared tracker #622 remains open.

Thanks to @zm2231 for the passkey fix, @nofihq for the compaction and retained-chat reports, and @zmahomedy and @renddre2 for the goal and workspace reproductions. Earlier V6 improvements include contributions and evidence from @kaimiller144-web, @lce52147, @thomastsaidotcom, @alexzwy51, @martjay and the reporters credited in previous releases.

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. If your Web models disappeared, run Setup → Install into Codex → Reinstall after updating, then fully quit and reopen Codex. Do not delete your launcher data or ChatGPT profile. Check Settings → Temporary Chat if you prefer ordinary saved conversations.

• Windows x64
• macOS Apple silicon · macOS Intel
• Linux x64 · Linux ARM64

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.