Tandem Skill

Tandem

**Persistent pair-programming between Claude Code and the OpenAI Codex CLI.**

Tandem turns Codex from a one-shot reviewer into a standing partner. One Codex thread per project keeps the whole history: the project summary, every decision, every objection and how it was resolved. Claude sends deltas instead of briefings, Codex answers in strict JSON, and a small Node runner enforces the protocol, the budget and the safety rules.

Tandem is a [Claude Code](https://claude.com/claude-code) skill. It was designed, reviewed and hardened together with Codex itself (see *Credits*).

---

Why

Ad-hoc "ask Codex" workflows have three recurring problems:

1. No memory. Every call starts from zero, so every prompt has to re-explain the project.
2. No protocol. Reviews happen when someone remembers, in whatever format the model picks.
3. No budget control. A review loop can burn through a ChatGPT usage window before anyone notices.

Tandem fixes all three: a persistent thread, a contact protocol with schema-validated verdicts, and a rate-limit guard that reads the remaining quota *before* each call and refuses to spend it below a threshold.

---

What is in this release (Plan A, the core)

Capability	What it does
Companion mode (Begleiter)	Claude implements; Codex reviews at protocol checkpoints (after a plan, after each coherent edit block, before "done"). Verdict OK / CONCERN / BLOCK with up to five prioritized points.
Plan consensus with automode	A plan travels between Claude and Codex for at most three rounds. Codex answers APPROVE / REVISE; the runner computes a hard consensus flag (no open BLOCKER/MAJOR, sources read, test strategy feasible, residual risk named). On consensus the plan is executed without a human stop. No consensus after round three: both positions go to the user.
Two final verdicts	The persistent thread checks goal fidelity and earlier objections; a fresh Co