LCU

Keep computer-use behavior, API, tool schemas and instructions in the original locally installed Codex runtime. Do not add another accessibility, input, screenshot, browser, JavaScript, or MCP implementation. Standalone adapters may manage installation, desktop-session selection, and transport/lifetime wiring for original host services. Use the official MCP SDK for harness transport. Reuse original implementations and record each adaptation; missing host services remain parity blockers.

Runtime changes require documented official-source evidence and real desktop integration tests for each claimed platform and architecture. Ship thin platform/architecture-specific LCU archives. Do not use repository app-version, runtime-version, or component-hash allowlists to select a locally installed official app. Report the version and runtime observed from the selected app, and say whether that pair is in the informational tested-versions.json record (a warning only; it never selects or refuses an app), then validate platform, supported architecture, official identity/signature where applicable, required files, and recognized original host layout. The official OpenAI desktop app is a prerequisite on every platform; LCU setup must never download or install it for the user. Linux uses the locally installed app in place; never copy it into the LCU prefix. runtime.lock.json source URLs and checksums are retained only for independently managed development/test package inputs; they do not authorize installer acquisition or select app compatibility. Linux apps are selected by runtime manifest, architecture, required structure, and the version observed from the app or its dpkg ownership. macOS reuses the compatible signed app and native helper in place; never copy, modify or re-sign them. Windows selects the current registered official Store app by identity, publisher/signature and architecture, checks required structure, and records a source-derived inventory for its intact private copy; do not claim new Windows live verification from structural checks. Keep prior platform support and evidence limits unchanged. Never put OpenAI application binaries, copied instructions, or generated fragments in Git, releases, exports, or public images. LCU's MIT license does not relicense installed dependencies.

Run tests/run.sh in disposable containers. Unit tests alone cannot prove desktop behavior. Preserve caller sandbox and approval settings; lcu setup --approval auto is an explicit, reversible opt-in that adds or removes only LCU's own entries. Never test against the user's personal desktop or copy agent credentials into fixtures.

Keep the installer compatible with the documented Luda-style options. Prefer upstream agent installers to new configuration adapters. Keep generated release archives in dist/ and out of Git. Test offline installation with the app already installed locally; development tests may prepare a verified package fixture before invoking LCU, then run setup with acquisition network access disabled. Browser/model services can still require network during use. Do not claim browser actions from discovery or an expected authentication failure.

For the task-owned macOS verification guest, follow docs/verification/macos-test-guest-access.md. Consult .verification/private/README.md before asking the user for guest credentials; keep the credential private.

The port of LCU from Python to the app's Node follows docs/REMOVE-PYTHON.md: its goal, what must and may change, and how it is tested.

Never add caching to a GitHub workflow: no actions/cache, no cache: or package-manager cache on setup actions, and no root package.json with a packageManager field, which turns on setup-node's automatic cache. A job that runs an agent on main must be assumed able to write cache entries, and any later job that restores one would run the agent's content with its own secrets and permissions. Keep no-cache and package-manager-cache: false where they are set.