ModelRudder 0.1.0-preview.4 — Codex 0.161.0 compatibility

Compatibility update for Codex 0.161.0, based on the latest GitHub main including merged Claude routing work.

• Accept exactly Codex 0.161.0 after native protocol initialization, fresh-thread, model-setting and read-only permission checks. Historical 0.159.3/0.160.1 builds remain accepted; unknown or modified builds remain rejected.
• Regression tests cover the launcher's compatibility gate and doctor readiness.
• The actual local smart-codex launcher opened the native 0.161.0 terminal and exited normally with zero task/model/classifier calls.
• All six CI jobs passed using native Codex 0.161.0 on macOS, Ubuntu and Windows with Node 24 and 26. Type checking, client tests, production dependency audit, protocol probe and packaging are included.
• Local updated client suite: 161 passed, one Windows-only skipped. Dependency audit: zero known vulnerabilities.
• This installer also bundles the previously merged experimental smart-claude adapter. Its native task-quality and platform qualification limits remain documented in docs/CLAUDE_CODE.md. This compatibility update does not certify Claude live execution.

Use Node 24+, an accepted independently installed Codex build and native ChatGPT login. Bring your own Jev key and classification consent for observe/auto; pinned mode does not require Jev. This remains an experimental community preview. Full native TUI qualification and general task-quality/savings claims remain pending.

Installer SHA-256: 0870d0b5dd5ac71324ad322597c77a4d888dca91b3522e88a287ebaf40ed7f16.

Verify the attached installer against SHA256SUMS, then run it to upgrade the existing managed installation. Repeat any custom --root, --bin-dir and --env-file flags used originally. Existing configuration and state are retained.

CI: https://github.com/cybrking/modelrudder/actions/runs/37873563134

ModelRudder 0.1.0-preview.3 — security and readiness fixes

Free local MIT community preview. This replaces preview.2 for new installations.

Changes:

• Withhold configured Jev, OpenAI and gateway secrets regardless of Windows environment-variable capitalization; retain unrelated tool credentials.
• Report immediate editor startup failures during setup and provide a manual-open path.
• Add failing-before-fix regression tests and exercise the actual Windows .cmd launcher from PowerShell in CI.
• Record installed lifecycle and real macOS terminal checks, including interruption, resume and live Jev observe/auto routing.

Validation:

• Final six-job CI: macOS, Ubuntu and Windows with Node 24 and 26; type checks, tests, zero-inference Codex 0.160.1 protocol checks, dependency audit and packaging.
• Local client: 101 passed, one Windows-only skipped; Windows CI executes all 102 tests. Dependency audit: zero known vulnerabilities.
• Independent security/code review found no additional actionable defects in reviewed paths.
• Verified standalone install, upgrade from preview.2, rollback and uninstall preserve private configuration/state. Local, macOS and Windows CI installers are byte-identical.
• Real installed macOS pinned/read/interruption/resume/exit checks passed. Live observe kept Sol; one auto task selected Luna and returned its exact expected answer. These are smoke checks, not general quality/savings evidence.

Requirements: Node 24+ and independently installed Codex exactly 0.159.3 or 0.160.1 with the user's native ChatGPT login. Other versions are rejected. Own Jev key and explicit classification consent are required for observe/auto modes. Provider charges and limits apply.

Remaining qualification: actual Windows/Linux terminal journeys, confirmation-required approval prompts, compaction, automatic effort and representative task quality. Paid hosted service readiness is separate. See docs/RELEASE_READINESS.md at this tag. Keep the preview designation.

Installer SHA-256: 19b0e54491d84a310acc0fc934f207193160759acf9e74af2ffecb5d6ad4b610.

Verify the installer against SHA256SUMS before running it. No repository checkout or npm dependency download is needed. Existing credentials/configuration are retained; setup creates a commented template and requests a local editor. The README contains the single copy/paste Codex installation prompt.

ModelRudder v0.1.0-preview.2 — Windows and agent setup

Free local community preview, MIT licensed. Requires Node.js 24+ and your independently installed native Codex CLI (accepted versions: 0.159.3 and 0.160.1). Use your own native ChatGPT login and Jev key; no ModelRudder subscription is required.

Changes:

• Native Windows per-user installer and launcher, regular-file release pointer, upgrade/rollback/uninstall preservation, and private ACL enforcement. Linux/WSL and macOS remain supported.
• Authenticated loopback relay on Windows; Unix sockets on Linux/macOS.
• smart-codex setup creates a private, fully commented Jev configuration and opens a local editor. --no-open supports agents and headless terminals. Existing configuration is preserved.
• Official Jev signup/key dashboard link in the template, plus agent installation instructions.

Validation: 99 tests pass on native Windows, 98 pass on Linux/macOS (the Windows ACL test is skipped there), on Node 24 and 26. Type checking, dependency audit, packaging and a no-inference installed Codex 0.160.1 protocol probe pass on all platforms. Full interactive TUI certification, including Windows, remains pending. No paid model turns were submitted.

Download the installer and SHA256SUMS from this release. Verify the exact installer checksum, then run node -- ./smart-codex-0.1.0-preview.2-8bc00f2dac0b65c2-install.mjs (PowerShell: node -- .\smart-codex-0.1.0-preview.2-8bc00f2dac0b65c2-install.mjs). Add the printed launcher directory to your user PATH, then run smart-codex setup. All Jev settings remain disabled until you explicitly uncomment them.

Cross-platform CI. Windows and macOS/Linux installer bytes were compared before publication.

ModelRudder 0.1.0 — community preview

ModelRudder's first free MIT community preview adds experimental model routing to the native Codex terminal experience. Use your own Codex/ChatGPT login and TypeSafe Jev key; no ModelRudder account or subscription is required.

Start with observe mode, which recommends a route while Sol executes. Auto mode is experimental and selects Luna, Sol or Astra for eligible turns. Native Codex retains tools, workspace permissions and approvals.

Requirements: macOS, Node.js 24+, and independently installed Codex 0.159.3 or 0.160.1. Full TUI behavior and general version compatibility are not certified. Clef and Claude support are not included.

Download the installer and SHA256SUMS together, verify the checksum, then run:

shasum -a 256 -c SHA256SUMS
node -- ./smart-codex-0.1.0-preview.1-022bfa94a95603c1-install.mjs

See the README for private configuration, data disclosure, observe/pinned/auto modes, rollback and uninstall. Provider charges and limits still apply; this preview does not guarantee task quality or subscription savings.

Validation: 91 offline client tests and type checking passed; dependency audit reported zero known vulnerabilities. GitHub CI passed on macOS with Node 24 and 26, and the CI-built installer matches the locally tested installer byte-for-byte. Native Codex 0.160.1 protocol checks and one pinned read-only terminal turn/exit passed. The installer checksum, disposable installation, help, activation and uninstall passed; offline tests also cover rollback and retained state. A targeted credential-pattern scan and scoped security review found no remaining concrete release blockers.

Passing CI · Release evidence

SHA-256: 20bda1ec4048180d06bb1af736a72b7e04c55b51b11ee95fb870ae4e47469b98