Codex Jev 0.4.0-beta.1: Evidence Selection and Recovery

Codex Jev: Public Beta

An evidence-selection and quota-aware recovery layer for Codex. Keep your chosen
coding model; use local discovery and optional bounded Jev ranking to retrieve
source-addressed evidence with exact follow-up reads.

Included

• Five stdio MCP tools for search, diagnostic selection, omissions, exact reads and status.
• Local exclusions/redaction, persistent spending reservations and explicit paid opt-in.
• Verified Git recovery boundaries and ownership-aware install/upgrade/rollback/uninstall.
• Model-neutral defaults, protected request caps and privacy regressions.
• Setup, architecture, recovery, security and measurement documentation.
• Upstream MIT attribution and third-party notices.

Verification

The release pipeline runs 205 Node tests, 71 Python tests, build/lint, package
lifecycle checks, public-content checks, dependency audit and Gitleaks history
scanning. No API key, paid model call or native Codex inference is required by CI.

Install

Use the pinned tag and README quick start, or download the prebuilt .tgz, verify
it against SHA256SUMS, extract it into a permanent directory and run its
runtime/manage.py install command for an existing Git workspace. Node 22.13+,
Python 3.11+, Git and ripgrep are required. Local-only mode is the default.
Native hooks require your own review/trust in Codex; the installer never grants it.

The package includes source and bundled MCP runtime. sbom.cdx.json inventories
the locked development/build dependency tree; optional dependencies for other
platforms may appear in the SBOM. SHA256SUMS covers the package and SBOM.
This release is not published to npm; install from GitHub.

Important limits

This is a tested beta, not a claim of proven universal cost or speed savings.
The native comparison was interrupted and remains incomplete. Evidence-byte
reduction is not an account-quota measurement. Jev hosting is separately billed.
Mac-to-Linux operation has been used; Mac-local execution remains unverified.
Hooks are partial guards, redaction is not perfect, and recovery never replays
external effects automatically. No existing private installation is upgraded by
publishing this release.

Independent Hyper AI Lab project, derived from James Cressler's MIT-licensed
Jev Codex Token Saver. Not an official OpenAI or TypeSafe product.