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.

• 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.

V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Version 6.1.13 includes the fixes accumulated throughout V6, with new repairs for long tasks, model selection, and launcher upgrades.

Changes in 6.1.13

• Workspace recovery after compaction. Continuations can recover the current working directory when native Codex compaction moves the environment message ahead of retained instructions. Midnight date updates also preserve the verified workspace under managed permissions. Recovery checks the current native task and its permissions (#861).
• Compaction handoff and sleep. Corrects the token format rejected by the native connector when submitting a context summary. The summary token remains restricted to that one operation. Compaction's inactivity timer now excludes time when the process was suspended, while an awake task with no progress still times out (#863).
• GPT-5.6 Pro selection. Accepts the updated picker header containing both “5.6 Sol” and “5.6 Pro” when the version and selected effort agree. Conflicting selections still stop the send, and Pro usage accounting recognizes the updated header (#875).
• Browser smoke test. Selects the intended model before reading its effort range, so a fresh chat's default model menu without a slider no longer blocks setup. Free accounts retain their separate path. Waiting for the model control or slider responds to cancellation (#876).
• Plugin permissions and naming. Permission cards get a specific explanation pointing to the plugin's Permissions → Allow all actions setting. Unanswered requests remain untouched when a turn stops. Task and compaction instructions use the configured plugin name instead of referring to a different connector.
• Upgrade settings and downgrade prevention. An older launcher cannot replace a newer installed runtime. Upgrades preserve an explicitly enabled automatic one-time approval setting; they do not enable it by default or in Zero Risk (#867, #881).
• Oversized messages and diagnostics. Explicit ChatGPT size rejections are reported as context_length_exceeded. Retrying an identical rejected compaction no longer uploads all its parts again. Failed browser captures are retained for up to three days within a 100 MB budget, with active captures protected. Logs include the final selected model/effort and payload size (#874).
• Build maintenance. Updates Vite to 6.4.4 to address the dependency advisories and enables certificate-free macOS signing in pull-request verification while retaining signature checks (#879, #882).

The original account-side size rejection in #874 was not reproduced in matching Plus tests; this release improves its handling and diagnostics without claiming a universal accepted context limit. The disappearing ChatGPT conversation in #865 and the Windows tunnel-client installation error in #870 still need the requested reporter evidence. 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. Model selection handles ChatGPT's updated picker, slow loading, translated controls, and locked Pro options while checking the selected model before sending.
• 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. Luna/Think support is experimental. Earlier patches fixed collapsed-message expansion, upload processing, and Markdown parsing stalls, and account for completed upload parts when ChatGPT removes older messages from the page.
• Saved-chat reliability. Since 6.1.11, new setups use saved chats and Verify runtime checks the selected chat mode. Existing off settings are preserved. Version 6.1.12 fixed uploads targeting hidden composers and early one-time approvals while a request was still streaming. Earlier fixes cover the isDestroyed cleanup crash, missing workspace lists after Codex restarts, and recovery from explicitly registered additional CODEX_HOME directories.
• Chat controls and suggestions. Optional New browser chat for each turn rebuilds the same task in a fresh conversation. Optional next-message suggestions use GPT-6 Sol Medium when the context fits. Saved chats and fresh chats are independent settings; saved chats remain in normal ChatGPT history and follow account personalization settings.
• Pro usage estimates. The Limits tab tracks the launcher's own submissions for supported Pro plans, with separate model allowances and a warning from 75%. These are local estimates; messages sent elsewhere and OpenAI's actual remaining quota are not included.
• Long tasks and tools. Improved compaction, delegated instructions, Page context, date updates, workspace recovery, and cancellation. Fixed overlapping connector handshakes interrupting another task's tool call. Accepted work survives a disconnected response stream, and helper replacement cannot silently resubmit a running task through the fresh-message path.
• Replies and Activity. Supports both ChatGPT layouts, preserves code, links, formulas, and plan markers, and keeps changing previews and hidden follow-up labels out of answers. Activity distinguishes preparation, generation, tool execution, and approval waits. Safe logs include captured page state without message contents.
• Sign-in, startup, and updates. Improved passkey sign-in, account switching, logout, browser crash recovery, and startup while signed out. Preserves valid configuration and a manually disconnected bridge. Includes system HTTP/SOCKS5 proxy support, Windows-trusted certificates, and improved update recovery.
• Desktop packages. Native Linux x64/ARM64 packages join Windows and both macOS architectures. Earlier V6 updates improved window restoration, browser sizing and system theme handling, plugin naming, and setup diagnostics.

Thanks to @nofihq for the focused signing, upgrade-setting, and dependency fixes; @thomastsaidotcom for the smoke-test reproduction and regression cases; and @iminfine, @zmahomedy, @pereponkin, @humhue, and @MardSilva for detailed reports and testing. Earlier V6 improvements also benefited from @AlvaroBiano, @jclllycf, @jgafarias, @mhughdo, @host-felice, and everyone who supplied logs and tested fixes.

The Pro Researcher skill is available for users who want to delegate 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 settings and browser profile are preserved. If your connector is missing from Temporary Chat, enable Settings → Save chats in ChatGPT; an existing off setting is not changed automatically.

• 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.

V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Windows, macOS, and Linux packages include the fixes accumulated throughout V6.

Changes in 6.1.12

• Tool approvals. Automatic approval waits until ChatGPT finishes streaming the requesting message. This should prevent the early-approval Invalid conversation body failure in saved chats (#858).
• Images in saved chats. Uploads target the active message composer even when ChatGPT keeps a second, hidden composer on the page (#866).
• Long messages and Bigger Context. Completed upload parts remain accounted for when ChatGPT removes older messages from the page. Verification no longer tries to expand an unfinished submission, restores collapsed messages and the viewport after reading, and accepts automatically linked URLs while checking the exact text. This addresses the reported expansion timeouts and several saved-chat failures (#863, #866).
• Helper crashes. A replacement helper cannot take a running tab through the fresh-message path and submit the prompt twice. The turn now stops with an explicit instruction to inspect the existing chat before retrying (#854).
• Temporary Chat guidance. When ChatGPT does not expose usable personalization or connector controls, the error points to Settings → Save chats in ChatGPT and explains that chats stay in history. Existing privacy choices remain unchanged (#864).
• Useful safe logs. Failed-turn exports now include the privacy-safe page state already captured locally. Environment-recovery diagnostics distinguish unrecognized context from missing native ownership evidence without recording paths or message content (#861, #865).

The missing-summary reply in #863, the missing-directory report in #861, the vanished conversation in #865, and the newly reported turn-ID mismatch in #867 remain under investigation. This patch does not claim to resolve those separate causes.

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. Model selection handles ChatGPT's updated picker, slow loading, translated controls, and locked Pro options while checking the selected model before sending.
• 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. Luna/Think support is experimental. Earlier patches fixed collapsed messages, upload processing, oversized-message errors, and Markdown parsing stalls.
• Saved-chat reliability. Since 6.1.11, new setups use saved chats and Verify runtime checks the selected chat mode. Explicit off settings are preserved. Earlier fixes cover the isDestroyed cleanup crash, missing workspace lists after Codex restarts, and recovery from explicitly registered additional CODEX_HOME directories.
• Chat controls and suggestions. Optional New browser chat for each turn rebuilds the same task in a fresh conversation. Optional next-message suggestions use GPT-6 Sol Medium when the context fits. Saved chats and fresh chats are independent settings.
• Pro usage estimates. The Limits tab tracks the launcher's own submissions for supported Pro plans, with separate model allowances and a warning from 75%. These are local estimates; messages sent elsewhere and OpenAI's actual remaining quota are not included.
• Long tasks and tools. Improved compaction, delegated instructions, Page context, date updates, workspace recovery, and cancellation. Fixed overlapping connector handshakes interrupting another task's tool call. Accepted work survives a disconnected response stream.
• Replies and Activity. Supports both ChatGPT layouts, preserves code, links, formulas, and plan markers, and keeps changing previews and hidden follow-up labels out of answers. Activity distinguishes preparation, generation, tool execution, and approval waits.
• Sign-in, startup, and updates. Improved passkey sign-in, account switching, logout, browser crash recovery, and startup while signed out. Preserves valid configuration and a manually disconnected bridge. Includes system HTTP/SOCKS5 proxy support, Windows-trusted certificates, and improved update recovery.
• Desktop packages. Native Linux x64/ARM64 packages join Windows and both macOS architectures. Earlier V6 updates improved window restoration, browser sizing and system theme handling, plugin naming, and setup diagnostics.

ChatGPT account limits and safety refusals still apply. The Pro Researcher skill is also available for users who want to delegate research to Pro sub-agents.

Thanks to the reporters in #854, #858, #861, #863, #864, #865, and #866 for the examples and diagnostic captures. Previous V6 fixes also benefited from @AlvaroBiano, @jclllycf, and everyone who tested saved chats in #436 and #857. Earlier V6 fixes include contributions and reports from @nofihq, @jgafarias, @pereponkin, @mhughdo, and @host-felice.

Update

Finish active tasks, quit the launcher, install the update, and reopen it. Restart Codex after updating its integration. Your settings and browser profile are preserved. If your connector is missing from Temporary Chat, enable Settings → Save chats in ChatGPT; an existing off setting is not changed automatically.

• 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.

V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Windows, macOS, and Linux packages include the fixes accumulated throughout V6.

Changes in 6.1.11

• Saved chats by default. New setups use regular ChatGPT conversations, where users report the Codex connector is still available when it disappears from Temporary Chat. Verify runtime now respects Settings → Save chats in ChatGPT instead of always testing Temporary Chat. Existing off settings are preserved: enable the setting yourself if you encounter this problem. Saved conversations remain in ChatGPT history, and ChatGPT memory and custom instructions may apply.
• Continuation after a Codex restart. Fixed compaction rejecting workspace roots when the current Codex record explicitly grants full access but omits its workspace list. Explicit permissions, working directory, and task identity remain checked.
• Separate Codex homes. A shared launcher can now recover CLI tasks from explicitly registered additional CODEX_HOME directories. See the setup instructions. Conflicting permissions or duplicate task ownership still produce an error.
• Bigger Context handoff. Fixed a timeout while trying to expand a message that ChatGPT had already moved into Activity. Previously acknowledged context still undergoes the same content checks.
• Launcher tab cleanup. Fixed the isDestroyed JavaScript crash when a browser tab is destroyed outside the normal close action and later reaches cleanup.
• Clearer failures. Detect ChatGPT errors even when they replace the assistant message. Validate the native tool token's exact format and explain how to copy it correctly. Safe diagnostics now record whether a model control is hidden, covered, disabled, or on a stalled page without recording conversation text.

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. Model selection handles ChatGPT's updated picker, slow loading, translated controls, and locked Pro options while checking the selected model before sending.
• 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. Luna/Think support is experimental. Earlier patches fixed collapsed messages, upload processing, oversized-message errors, and Markdown parsing stalls.
• Chat controls and suggestions. Optional New browser chat for each turn rebuilds the same task in a fresh conversation. Optional next-message suggestions use GPT-6 Sol Medium when the context fits. Saved chats and fresh chats are independent settings.
• Pro usage estimates. The Limits tab tracks the launcher's own submissions for supported Pro plans, with separate model allowances and a warning from 75%. These are local estimates; messages sent elsewhere and OpenAI's actual remaining quota are not included.
• Long tasks and tools. Improved compaction, delegated instructions, Page context, date updates, workspace recovery, and cancellation. Fixed overlapping connector handshakes interrupting another task's tool call. Accepted work survives a disconnected response stream.
• Replies and Activity. Supports both ChatGPT layouts, preserves code, links, formulas, and plan markers, and keeps changing previews and hidden follow-up labels out of answers. Activity distinguishes preparation, generation, tool execution, and approval waits.
• Sign-in, startup, and updates. Improved passkey sign-in, account switching, logout, browser crash recovery, and startup while signed out. Preserves valid configuration and a manually disconnected bridge. Includes system HTTP/SOCKS5 proxy support, Windows-trusted certificates, and improved update recovery.
• Desktop packages. Native Linux x64/ARM64 packages join Windows and both macOS architectures. Earlier V6 updates improved window restoration, browser sizing and system theme handling, plugin naming, and setup diagnostics.

ChatGPT account limits and safety refusals still apply. The Pro Researcher skill is also available for users who want to delegate research to Pro sub-agents.

Thanks to @AlvaroBiano and @jclllycf for the workspace and separate-home reproductions, and to everyone in #436 and #857 who tested saved chats. Earlier V6 fixes include contributions and reports from @nofihq, @jgafarias, @pereponkin, @mhughdo, and @host-felice.

Update

Finish active tasks, quit the launcher, install the update, and reopen it. Restart Codex after updating its integration. Your settings and browser profile are preserved. If your connector is missing from Temporary Chat, enable Settings → Save chats in ChatGPT; an existing off setting is not changed automatically.

• 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.

V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Packages are available for Windows, macOS, and Linux.

Changes in 6.1.10

• Fixed completed ChatGPT answers being rejected as a different message after the page updates or collapses the original prompt.
• Fixed Bigger Context stopping after ChatGPT collapses an acknowledged part of a large message.
• Fixed continuation after compaction when Codex drops a delegated instruction or refreshes AGENTS.md. Also fixed date updates containing a list of sub-agents.
• Fixed workspace checks rejecting directories already granted by Codex, including the changed format in Codex 0.163.
• Fixed Free-account Think selection when choosing a plugin switches Think off.
• Handle upload rejections before processing starts and report upload-limit responses directly. ChatGPT's upload allowance is unchanged.
• Prevented overlapping connector handshakes from interrupting another task's tool call.
• Reduced token-counting work for repeated text without changing token counts or context limits.

Included from 6.1.9

• Improved continuation after compaction and Page context, browser crash recovery, logout, and reconnecting to busy chats.
• Preserved fenced code blocks and removed hidden suggested-followup labels from answers.
• Added optional next-message suggestions using GPT-6 Sol Medium when the full context fits.
• Improved failure details and approval instructions; ChatGPT-side refusals remain possible.
• Updated the tunnel client and added system HTTP/SOCKS5 proxy support.

Pro sub-agent skill

I use the Pro Researcher skill for deep research with Pro sub-agents and have shared it in the repository. Copy it into your skills directory if you'd like to try it.

Thanks to @nofihq for #837 and #842 and for identifying the connector handshake problem in #840; @jgafarias for the Codex 0.163 findings in #845; and @pereponkin and everyone who supplied logs and reproductions. The 6.1.9 fixes also include contributions from @mhughdo (#830) and @host-felice (#832).

Update

Quit the launcher, install the update, and reopen it. Restart Codex after updating its integration. Your launcher settings and browser profile are preserved. If the built-in updater fails, download the installer below.

• 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 is still open, add the new results there.

V6 brings ChatGPT Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Packages are available for Windows, macOS, and Linux.

Changes in 6.1.9

• Fixed Free-account Think selection: the launcher now uses the Think button instead of typing /think.
• Fixed missing workspace information after compaction and when opening a Page in Codex.
• Preserved fenced code blocks in replies and removed hidden suggested-followup labels that could appear in returned answers.
• Improved recovery after a browser renderer crash and prevented Log out from getting stuck. Reconnecting to a busy chat no longer cancels an already accepted request.
• Added a setting for next-message suggestions. They are enabled by default, use GPT-6 Sol Medium, and run only when the full context fits. Requires Automatic mode, Sol access, and Show predictions enabled in Codex.
• Improved failure logs and clarified when to ask for approval, so already authorized actions do not require repeated confirmation. ChatGPT-side refusals remain possible.
• Updated the tunnel client and added support for the system's HTTP and SOCKS5 proxy settings, while respecting explicitly configured proxies.

Pro sub-agent skill

I've also shared a skill in the repository for using Pro as a sub-agent and coordinating its work. I use it myself and find it very useful, so I wanted to share it. If you'd like to try it, copy the skill into your own skills directory.

Included from 6.1.8

• More reliable large messages, attachment processing, long-running tasks, and continuation after compaction.
• Bigger Context for GPT-6 Sol Medium, High, and Extra High on Plus; improved browser controls, reply formatting, and failure details.

Thanks to @mhughdo for the Page-context fix in #830, @host-felice for the followup-label fix in #832, and everyone who shared logs and reproductions. Thanks also to @xianengqi for #806 and #807, included in 6.1.8.

Update

Quit the launcher, install the update, and reopen it. Restart Codex after updating its integration. Your launcher settings and browser profile are preserved. If the built-in updater fails, download the installer below.

• 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 is still open, add the new results there.

Codex Web GPT v6.1.8

V6 brings Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Packages are available for Windows, macOS, and Linux.

6.1.8

• More reliable message sending. Large messages no longer hit a shorter send limit than multipart requests, and a busy ChatGPT page does not hide a message that was already accepted. Attachments finish processing before the message is sent. Code-heavy context no longer stalls ChatGPT's Markdown parser.
• Long tasks and compaction. An active response connection keeps a task alive when ChatGPT hides its generation controls. Context shortening keeps waiting while the original task is still active, and resumed tasks can recover their workspace correctly.
• Bigger Context on Plus. GPT-6 Sol at Medium, High and available Extra High now supports up to 120,000 tokens on Plus, with compaction at 110,000. Pro retains 240,000/220,000. Instant keeps its standard context.
• Browser and reply handling. Recognizes Stop and Think controls across interface languages, preserves the browser's usable size during navigation, and fixes final-answer alignment around horizontal rules.
• Better failure details. Safe logs can capture the page structure at the moment a reply is lost, without conversation text. Browser security checks now produce a specific message. These diagnostics help investigate remaining reports; they do not mean every missing-reply or sign-in problem is fixed.

Included from 6.1.7

• Improved final replies, charts, previews, decorated links, model selection, and slow model-list loading. Added optional Bigger Context for Luna and Think.
• Improved passkey sign-in, startup and update recovery, preserved failed task tabs for inspection, and clarified tool errors.

Thanks to @xianengqi for #806 and #807, and everyone who supplied reports, logs, and test results.

Update

Quit the launcher, install the update, and reopen it. Restart Codex after the launcher updates its integration. Keep your launcher data and ChatGPT profile. If the updater fails, download the installer below.

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

If a problem remains, open a new issue with Activity → Export safe log, the failure time, your OS, Codex version, and selected model and effort. If your issue is still open, reply there with the result instead.

Codex Web GPT v6.1.7

V6 brings Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Packages are available for Windows, macOS, and Linux.

6.1.7

• More reliable replies. Fixes final-answer alignment when ChatGPT finishes updating the page. Changing charts, previews and other known response controls no longer interfere with completion. Decorated links no longer cause false “composer did not preserve the complete prompt” errors.
• Model selection and loading. Fixes selection when parallel task tabs switch and recognizes localized punctuation in model and plan labels. Slow first connections can finish loading the model list, and cancelled requests no longer produce false failure banners.
• First-message failures. Fixes a permissions mismatch when the current Codex request omits a setting already recorded for that task. Actual permission changes are still checked.
• Clearer passkey sign-in on macOS. Switching from ordinary sign-in now opens the passkey flow. The launcher explains the separate Chrome window, shows progress, and enables Continue only after Chrome starts.
• Launcher recovery. A manually disconnected bridge stays disconnected across restarts and upgrades. Failed update checks retry, failed task tabs remain available for inspection, and errors during context shortening retain their original cause.
• Experimental Bigger Context for Luna and Think. Keeps temporary conversations and sends larger histories in parts. This is optional; recall in long Free-account chats remains limited. GPT-6 Sol on Plus and Instant still use standard context, now stated clearly in every launcher language.
• Tool-result reporting. Adds error status to tool-result diagnostics and clarifies that an old error quoted in a file is not a new failed command. ChatGPT safety refusals and account limits still apply.

6.1.6

• Added GPT-6 Sol and Instant while preserving GPT-5.6 as a separate choice, and repaired selection after ChatGPT's interface update.
• Updated GPT-6 context limits: Bigger Context up to 240,000 tokens on Pro at Medium, High and Extra High, with compaction at 220,000. Other plans and Instant retain standard context.
• Fixed startup and upgrades while signed out, the “Redirect was cancelled” updater error, tool discovery, and several reply-rendering failures.

Thanks to @Liyux3, @mouzhi, @norituku, @Venom120, @xianengqi and @iminfine for useful reports and testing, and @moooohung, @DzsGatsby and @frostschaden for the preceding patch.

Update

Quit the launcher, install the update, and reopen it. Restart Codex after the launcher updates its integration. Keep your launcher data and ChatGPT profile. If the updater fails, download the installer below.

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

If a problem remains, open a new issue with Activity → Export safe log, the failure time, your OS, Codex version, and selected model and effort.

Codex Web GPT v6.1.6

V6 brings Web models into Codex's model and Effort selectors, optional Pro usage tracking, Bigger Context, and control over saved or temporary browser chats. Packages are available for Windows, macOS, and Linux.

6.1.6

• GPT-6 support and repaired model selection. Adds GPT-6 Sol and Instant, fixes selection after ChatGPT's latest interface change, and keeps GPT-5.6 available as a separate choice. Existing GPT-5.6 tasks keep their selected version. Free accounts show Luna/Think without an unsupported version label.
• Updated context limits. GPT-6 Sol supports Bigger Context up to 240,000 tokens on Pro at Medium, High, and Extra High, with compaction at 220,000. GPT-6 Instant and other plans use standard context. GPT-5.6 and GPT-6 Pro retain their existing Bigger Context support.
• Reliable startup when signed out. An expired ChatGPT session or unavailable sign-in page no longer prevents an already configured local bridge from starting or updating. Updates preserve model settings and avoid trying to start an incompatible older configuration.
• Update downloads. Fixes the “Redirect was cancelled” error while preserving system-proxy support and download verification.
• Tool discovery and replies. Fixes tool discovery when Codex adds output headers. Keeps reply text intact when ChatGPT redraws file previews or moves unfinished paragraphs, and improves cancellation during page preparation.

6.1.5

• Improved Bigger Context uploads on Plus and recovery from stalled or failed replies.
• Fixed one-time tool approvals, Think-mode plugin selection, and reopening failed sign-in pages.
• Improved long-task recovery, workspace handling, and diagnostics; clarified asking for confirmation of a specific action after a safety refusal. ChatGPT account limits and safety refusals still apply.

Thanks to @moooohung, @mouzhi, @DzsGatsby, @frostschaden, and everyone who supplied reports and testing for these fixes. Thanks also to @chof64, @Liyux3, @Venom120, and @hankok for the previous patch's reports.

Update

If the updater says “Redirect was cancelled”, download the installer below. Quit the launcher, replace the existing application, and reopen it. Restart Codex after the launcher updates its integration. Keep your launcher data and ChatGPT profile.

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

If a problem remains, open a new issue with Activity → Export safe log, the failure time, your OS, Codex version, and selected model and effort.

Codex Web GPT v6.1.5

V6 brings a new model picker, optional Pro usage tracking, more control over browser chats, and the fixes developed since v5.0.8.

V6 highlights

• Clear model names and native Effort controls. Automatic models end in (Web): GPT-5.6 Luna, GPT-5.6 Sol, GPT-5.6 Pro, and GPT-6 Pro. Sol exposes the reasoning levels your account supports through Codex's Effort selector. Instant stays separate to preserve its context budget. Existing tasks keep their original modes.
• A Limits tab for Pro. Optional tracking for supported Pro $100 and $200 plans, with separate model allowances and a warning from 75% usage. These are local estimates; messages sent outside the launcher are not counted. Tracking is unavailable in Zero Risk.
• Smaller parts for Bigger Context. Large requests use six ordered parts, with tools and attachments on the final part. The overall context and compaction budgets are preserved.
• New browser chat for each turn. An optional setting rebuilds the same task in a fresh ChatGPT conversation. This can help when ChatGPT loses tools across follow-ups, at the cost of sending more context.
• Save chats in ChatGPT. Regular conversations can replace Temporary Chat. ChatGPT memory and custom instructions may apply. Both chat settings are off by default and work independently.
• Native Linux ARM64 packages. x64 and ARM64 AppImages and runtime bundles, with installation and updates for each architecture.
• A new application icon. The launcher keeps its familiar ChatGPT mark inside the interface.

Fixes and improvements through v6.1.5

• Prompts and Bigger Context. Prevents doubled prompts, recognizes completed uploads, and protects large messages from browser formatting. Plus uploads use smaller early messages and a larger final part without dropping context. Fixes earlier uploads being mistaken for new replies and uploads stopping before a slow reply starts.
• Replies and formatting. Supports both ChatGPT layouts and follows replies through redraws. Fixes repeated paragraphs, headings, and formulas. Keeps changing controls and file previews out of answer text while preserving download links, code line breaks, and plan markers. Recognizes stalled replies and Retry errors without repeating completed tool calls.
• Models and effort. Handles slow menus, model switches, locked Pro options, and missing version announcements. Checks the requested model and effort before sending. Think is enabled before selecting the plugin. Bigger Context is no longer offered for Luna or Think, where it is unsupported.
• Sign-in and Limits. Improves sign-out and account-switch detection, including while hidden. Reads the signed-in plan without opening Billing or relying on its language. Opening sign-in again reloads a failed page instead of keeping the broken document.
• Startup and desktop behavior. Cancelling one request no longer crashes the browser helper. Preserves valid setup during startup changes, the browser's system theme, and windows moved between monitors. Accepts equivalent Codex settings rewrites and gives clearer instructions for damaged configuration.
• Windows connections. Uses Windows-trusted certificates while keeping verification enabled. Improves tunnel installation, setup recovery, and update downloads through system proxies. Updates runtime and build dependencies.
• Long tasks and summaries. Improves workspace recovery, continuation, and compaction timing. Allows unfinished summaries to change before delivering the completed version. A disconnected response stream no longer cancels accepted work. Fresh tasks can rebuild damaged saved workspace state from their current permissions. Some long-session failures remain under investigation.
• Activity and diagnostics. Shows preparation, response, tool, and approval waits, with elapsed time and a link to the browser tab. Identifies stale status and timed-out tools. Safe logs include more useful browser, context-size, and subagent model-selection details.
• Plugin names and setup. Change the part after Codex in Settings → Plugin name when a deleted name cannot be reused. Create and verify the new plugin; credentials and sign-in are kept. The setup video supports click-to-pause with an indicator; Linux update instructions are clearer.
• Tool connections and approvals. Checks the plugin on each follow-up. Optional one-time approval recognizes the current Allow once card and verifies the selected plugin. After a safety refusal, instructions ask for confirmation of the specific action before continuing where permitted. Zero Risk stays manual, with 60 seconds to send and a timer reset when copying again.

ChatGPT-side safety refusals and account quotas are not presented as fixed by this release.

Full-mode users: refresh your Codex plugin in ChatGPT's plugin settings after updating if you have not done so since V6, to load the updated tool description.

Thanks to @renanassisalves for PR #664, page details, and testing that helped identify the ChatGPT interface changes. Thanks also to @NC-Channel-Coder, @eisen0419, @omarcosr, @Wandersport, @b-y-f, @djblack1209-coder, @midorikawadaisuke4500, @korboybeats, and everyone who shared reports, proposed fixes, and retest results. Thanks to @chof64, @Liyux3, @Venom120, @mouzhi, and @hankok for the approval, sign-in, Think, reply-preview, and Bigger Context reports that helped with this patch.

Install

Quit the launcher, replace the existing application with the package for your system, then reopen it. Restart Codex after the launcher refreshes its integration. Do not delete your launcher data or ChatGPT profile.

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