Skip to content

Releases

v0.2.0 — evidence-driven gates

Choose a tag to compare

@hyj28 hyj28 released this 27 Aug 21:26
9ff01be

v0.2.0 — evidence-driven gates

This release makes the pair workflow prove more than “the tests passed.” Claude still
plans and reviews, Codex still implements in a pinned workspace sandbox, and the human
still decides what merges — but the green gate now demands stronger evidence.

Highlights

  • Baseline-aware testing. Test counts are captured before the handoff, so a suite that
    collects nothing or loses coverage cannot quietly read as green.
  • Test-diff integrity. Deleted or loosened assertions, new skips, tautological expected
    values, and mocks that replace the behavior under test are explicit red gates.
  • Run it like a user. A fresh agent that has not read the implementation drives the
    pre-agreed runtime scenarios and records commands, verbatim output, expectations, and
    match or mismatch.
  • Safer installation and CI. Empty or root install destinations are refused; CI checks
    the installer, skill frontmatter, opt-in guarantees, and repository-relative links.
git clone https://github.com/hyj28/claude-codex-pair.git
cd claude-codex-pair
./install.sh

Then invoke /pair, /pair fast, /pair strict, or /pair-review explicitly inside
Claude Code. Installing the skills does not change ordinary Claude Code or Codex behavior.

Read the complete workflow design
and changelog.

v0.1.0 — first release

Choose a tag to compare

@hyj28 hyj28 released this 27 Jul 21:09

Opt-in dual-agent development workflow: Claude Code plans and reviews, Codex implements headlessly, git is automated, and merging stays a human decision.

Nothing runs until you type the command — installing the skills changes nothing about ordinary Claude Code usage.

What is in this release

  • /pair [fast|strict] <task> — the main workflow. Claude writes a handoff (goal, constraints, acceptance criteria — deliberately not the implementation approach), branches, hands off to Codex via stdin, runs the project's real quality gates, then reviews the diff with a fresh read-only subagent. Fix and review loops are capped at 2 rounds each.
  • /pair-review [codex] [base <branch>] — standalone second pair of eyes on current changes, optionally merging Codex's native reviewer.
  • Deterministic gates — format, lint, typecheck, test. Exit codes decide; an agent claiming "tests pass" is never trusted.
  • Sandbox pinned explicitly on every codex exec, never inherited from ~/.codex/config.toml — headless runs disable approvals, so the sandbox is the only remaining boundary.
  • A commit at every gated checkpoint — one for the implementation, one per review round, so the branch carries a readable trail and can be merged as-is.
  • Tiered dependency policy — project-local installs are self-service, package-manager caches are granted explicitly via --add-dir, and anything system-level (brew, global installs, sudo, runtimes, containers) stops and asks.

Install

git clone https://github.com/hyj28/claude-codex-pair.git
cd claude-codex-pair && ./install.sh

Requires the Claude Code CLI, the Codex CLI, and git. Verified against Codex CLI 0.145.0 — the -s, --add-dir, and resume --last flags are version-sensitive.

See WORKFLOW.md for the design doc.