Codex Web GPT v4.0.7

V4 turns ChatGPT Web into a continuous Codex task runtime: multiple Codex messages stay in one task-bound Temporary Chat, and a full context rolls into a clean chat without losing the agent that did the work.

V4 highlights

• Native context compaction. At the context boundary, Codex asks the exact retained Web agent for a checkpoint through the gRPC-backed MCP tunnel, then starts a clean chat from that summary. If the private chat was closed, canonical Codex history supplies a safe fallback.
• Continuous Web tasks. Sequential Codex messages reuse the same Temporary Chat until compaction, with a fresh turn-bound MCP capability for each message and native tool rounds inside the same Web response.
• Bigger Context. The experimental mode adapts large contexts across two or three verified messages and can expose up to 3× the normal context and compaction limits. Extra requests can increase the chance of rate limits or temporary cooldowns.
• More resilient browser execution. Streaming survives React DOM remounts, reconnects rejoin the existing turn, hidden tabs keep an operational viewport, and active MCP calls count as real progress.
• Current ChatGPT UI support. Personalization, effort sliders, connector rows, cooldown dialogs, login/session races, privacy-safe diagnostics, and cross-platform packaging are handled structurally.

v4.0.1–v4.0.7 hotfixes

• Long-running and staged turns. Adapter liveness now covers every wait, system sleep pauses turn budgets, Bigger Context stages and acknowledgements receive a 180-second budget, completed streams tolerate post-DONE transport resets, and stale launcher pages disconnect before rebind. Includes #226 and #233.
• Think mode for Luna-only accounts. Free and Go Luna-only catalogs expose both Luna and Think. Think presses and verifies the exact ChatGPT control before sending, while normal Luna clears it.
• Launcher visibility and Windows 10 bootstrap. Hidden or minimized parent windows keep running tabs on an offscreen emulated viewport without reload or resend. The primary browser now commits a real inert idle document in an offscreen visible viewport before publishing its descriptor. Fixes #240 and #244.
• Explicit MCP tool deadlines. An abandoned Codex tool call now returns a structured codex_tool_timeout after 90 seconds, retires its binding, and avoids a generic 502. Fixes #237.
• Cross-backend compaction safety. A bridge-owned ocx1: checkpoint is converted into a readable summary before entering native Codex, while native encrypted compaction remains untouched.
• Clear Codex restart guidance. Setup now explains that Codex must be fully quit, including its background process; signing out or closing the window is not a restart, and the launcher must remain open.
• Localized UI and exact connector discovery. Personalized and the current menuitemradio UI are handled structurally across locales; verified chats avoid extra clicks, exact @codex discovery finds Codex Native2, and the current Allow approval action is recognized without matching persistent approval choices.
• Bounded browser recovery. A stalled launcher page is reattached without resending an accepted prompt, and partial diagnostics survive an unresponsive renderer.
• Reliable hidden turns. Background tabs receive a renderer viewport after every full navigation, so fresh turns work with the launcher hidden or on Setup. The repeat action is Repair Codex setup. Fixes #191.
• Append-only streaming through DOM virtualization. ChatGPT source ranges preserve streamed text when React removes a completed prefix, while semantic rewrites still fail closed. Fixes #189.
• Real MCP and tunnel recovery. Interrupted turns revoke MCP capability before the deadline. Internal 502 recovery proves stop → stopped → connect plus fresh diagnostics; Windows cancellation is bounded and deterministic.
• Full-mode MCP liveness. Launcher helper processes now receive exact MCP progress from the daemon. Active tool calls veto premature DOM failure and completion, stale progress is bounded, and answer text emitted between tool calls is no longer discarded. Fixes #221.
• Compaction send readiness. Retained compaction waits a bounded five seconds for ChatGPT to enable Send after the complete verified prompt is attached, while session, rate-limit, cancellation, and closed-page failures remain explicit.
• Native Responses parity. Pro subagent delegation, history metadata, text controls, and strict structured output survive the Web route. agent_message keeps its author, recipient, role, and content instead of becoming a human user message. Fixes #203.
• Correct browser control identity. Connector identity is sent only with its required conversation key, preventing launcher control HTTP 400 failures on ordinary Luna turns. Fixes #204.
• Native model discovery. A missing /v1/models client_version is derived only from an exact first-party Codex User-Agent and never invented from an arbitrary browser or proxy header. Fixes #224.
• Pro continuity. Compaction stays on Pro with the exact retained chat and MCP handoff; completed trailing commentary no longer waits for another status row.
• Markdown fidelity. Literal Obsidian [[wiki links]] remain text instead of being rendered as broken LaTeX.
• Japanese launcher and documentation. Japanese is complete across onboarding, Settings, native dialogs, tray actions, diagnostics timestamps, persisted state, and the full README.
• Launcher and Windows installation. Electron 41.10.7 is audited throughout CI. The installer supports Windows PowerShell 5.1 without IsPathFullyQualified and validates drive and UNC paths. Fixes #190.

Addresses #146, #149, #170, #179, #181, #185, #186, #187, #189, #190, #191, #203, #204, #221, #224, #237, #240, and #244. Includes #171, #172, #176, #197, #200, #201, #211, #223, #226, #232, and #233.

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