Codex Web GPT v3.0.0

V3 is the largest architecture update to Codex Web GPT so far. It adds a production-shaped
development harness, enables the complete Codex MCP harness for ChatGPT Pro, and hardens the
browser/runtime lifecycle around real failures reported in Issues.

Isolated DEV harness

The new DEV mode runs the current working tree through the same launcher browser, Temporary Chat,
prompt compiler, Responses parser, MCP broker, and compaction handlers used in production—without
rerouting or stopping the maintainer's live Codex installation.

• A separate DEV home isolates launcher state, browser cookies, ChatGPT login, CODEX_HOME, named
  chats, diagnostics, runtime state, broker, and tunnel configuration.
• You can sign in with a different ChatGPT account and keep production Codex Native2 separate from
  the dedicated Codex Native2 DEV connector.
• DEV control endpoints are isolated and the production Responses port remains untouched, so an
  agent can drive test chats while the normal Codex session stays available.
• Named CLI chats are persistent: reuse a name to continue its canonical history, or pass a message
  directly for a one-shot run.
• Optional Full mode exercises the real MCP connector and broker. Every dispatched action returns an
  explicit simulation receipt with simulated: true and side_effects_performed: false; DEV never
  pretends that a command, patch, image read, or external mutation actually happened.
• /status, /fill, /send-fill, /compact, /model, and /reset make context pressure,
  compaction, transport boundaries, and model routing reproducible from the terminal.

bun run dev:launcher
bun run src/cli.ts dev status
bun run dev:chat compaction-lab "Reply with exactly: DEV READY"
bun run dev:chat compaction-lab

ChatGPT Pro now has the full Codex harness

ChatGPT Pro previously exposed no usable custom MCP connector in our browser flow. That changed.
Over the last several days, Pro was exercised across multiple real development tasks and
successfully called the same turn-bound MCP tools as the other ChatGPT Web efforts.

Pro can now work directly with the active Codex task's filesystem, shell, patches, images,
approvals, installed tools, and apps. Calls and real results stay in the same browser response; Pro
is the selected model, not a tool called by another host model. To our knowledge, this is currently
the only open-source bridge that exposes ChatGPT Pro as a native Codex model with its local harness
attached this way.

Normal Pro work still runs on Pro. Only Pro's isolated history-compaction turn uses Extra High,
which has the same measured 95,000-token pre-compaction budget and completes summarization faster.

Reliability and issue fixes

• Upgraded the embedded runtime, source toolchain, CI, and release packages from Bun 1.3.14 to
  stable Bun 1.4.0, addressing the Windows stream-disconnecting segfault tracked in #86.
• Fixed Lexical prompt corruption at 16k insertion boundaries and added a bounded clean pre-submit
  compaction retry for the complete-composer-loss failure tracked in #122 and #135. A retry is
  allowed only before any user turn, assistant turn, or generation exists, preventing duplicates.
• Closing a running launcher tab now cancels its owned browser turn instead of recreating a tab.
• Command-Q and explicit shutdown cancel active HTTP/browser turns and force-stop only the
  launcher-owned runtime when graceful shutdown cannot complete.
• Command-Plus/Minus now scales the launcher shell while the embedded browser keeps its independent
  zoom and correctly tracks the resized native bounds.
• Hardened MCP connector verification, tunnel supervision, route restoration, config journaling,
  session persistence, DOM commentary streaming, and launcher diagnostics.
• Linux package smoke now uses isolated extraction state and a bounded larger output buffer (#114).
• The Windows installer now shows a normal completion screen and offers to launch Codex Web GPT.
• Added a documented account-bound release-validation gate instead of treating package smoke as
  proof that sign-in, MCP, compaction, and real model turns work.

Validated release

The complete test suite and native package smoke pass on macOS arm64/x64, Windows x64, and Linux
x64. The published V3 candidate was also installed and exercised manually on Windows 11 with a
real ChatGPT Pro account: authenticated browser state, Codex model routing, Full-mode MCP actions,
Pro development turns, compaction, cancellation, and preserved launcher/connector state all passed.

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