plan-tango

A Claude Code plugin that runs a Claude ↔ Codex plan-review loop on its own.
Run one command, come back to a polished plan.

---

plan-tango is a Claude Code plugin that runs a plan-review loop between Claude and Codex (gpt-5) with a single command.

More than once I've caught myself sending plans from Claude into Codex by hand. GPT's feedback is usually useful, and one round rarely closes it. So I'd write the plan, paste it into Codex, wait for the findings, bring them back to Claude, ask to apply, send the updated plan to Codex for another round, wait for the next batch of findings, bring those back. And again. Between rounds you wait, sometimes lose track of which terminal has the latest verdict and is waiting on you.

plan-tango runs that loop on its own. Claude and Codex work as a pair; the plugin passes context between them and applies fixes to the plan. Default budget is up to 6 iterations, hard cap 12. The plugin will stop sooner if there's nothing important left to flag. You stay in plan mode and can step away to do something else; when a question comes up that needs your call, the plugin pauses and asks. For time-pressed runs (or just when you have Codex quota to spare), --fast flips on Codex's priority service tier (~1.5× faster).

Install

Prerequisites: Claude Code 2.x, Node.js 18+.

1. Install Codex CLI — in your terminal (macOS/Linux) or PowerShell (Windows), outside Claude Code. One-time setup: plan-tango calls into this Codex install at runtime.

npm install -g @openai/codex
codex login

2. Install the plugin — inside a Claude Code session. The two slash commands below go directly in the chat with the agent, one per message (Claude Code treats each chat message as a single command — pasting both at once jams the second one into the URL of the first):

/plugin marketplace add egsok/plan-tango

Wait for the marketplace to be added, then:

/plugin install plan-tango@plan-tango

Or just ask the agent: "install plan-tango from github.com/egsok/plan-tango". Claude will figure out the commands and walk you through.

3. Reload plugins — in the terminal Claude Code, run /reload-plugins to pick up the new skill, scripts, and agent without leaving the session. In the VS Code extension that slash command isn't available — use Developer: Reload Window from the Command Palette (or restart Claude Code) instead.

Usage

/plan-tango:run                          # use the active plan from plan mode
/plan-tango:run <slug-or-path>           # or pass a specific plan file
/plan-tango:run --fast                   # priority service tier (~1.5× faster)
/plan-tango:run --max-iter 10 --effort medium --lenient --quiet --fresh-each

Persistent defaults live in ~/.claude/plan-tango/config.json. For an interactive wizard instead of hand-editing JSON: /plan-tango:settings.

Full reference (flag list, status codes, architecture notes) — plugins/plan-tango/README.md (English) · plugins/plan-tango/README.ru.md (Russian).

What the plugin handles beyond the loop itself

• Thread mode (continue / fresh). By default all iterations share one Codex thread (continue): cheaper thanks to prompt-cache hits on repeated plan content, faster, and only one thread per run in the Codex panel. At iter ≥ 2 a reset block is injected into the prompt — a short "forget your previous verdicts on this plan and evaluate fresh" instruction — to keep Codex from anchoring on its earlier output. If you want a fully independent audit each round, pass --fresh-each or set thread_mode: fresh in config.
• Severity-aware stop. If a round leaves only cosmetic findings (formatting, wording nitpicks), the plugin doesn't trigger another corrective round — auto-applying cosmetic fixes usually makes them worse. Those findings render as an advisory list in the final report; apply manually if you want. Turn off via severity_aware: false in config.
• Snapshots and hash integrity. Every apply phase writes .iter{N}.bak next to the plan. Before each iteration the plan's sha256 is compared against last_known; if it changed outside the loop (an IDE save, another tool), the skill stops with a clear error rather than clobbering those edits.
• Resume. State is persisted after every apply phase; an interrupted run (Ctrl+C, Claude Code crash, machine reboot) picks up from the same iteration via --resume.

Plus the smaller stuff: per-plan lock file (two parallel runs on the same plan can't clobber each other; lock becomes stale after 30 min), off-plan guard (the plugin only edits the plan file — if Codex asks to change code in the repo, the finding is logged as off_plan_blocked and skipped), and an optional Opus final sanity-check via --final-check for plans that touch runtime contracts (subagents / permissions / hooks / MCP).

Update

Easiest path: run /plan-tango:update. The skill version-checks against GitHub, asks you to confirm, runs git fetch + git reset --hard v<latest> in the marketplace clone, and prints a reload reminder. Pass --check to print the status without updating, or --force to discard any local modifications in the marketplace clone.

Manual path (always works): open /plugin in Claude Code, go to Marketplaces → plan-tango, and pick Update. Claude Code pulls the marketplace and reinstalls if a newer version is published. Auto-update is opt-in via the same menu: Enable auto-update. Third-party marketplaces have auto-update off by default — that's Claude Code policy, not a plan-tango choice.

Independent of both, plan-tango watches the GitHub release channel: at session start (a SessionStart hook) and at the end of each run, it checks for a newer tag — at most once per 7 days, silently on network errors — and prints a one-line notice if one is out. Opt out via update_check: false in ~/.claude/plan-tango/config.json. (The opt-out silences the notices but does NOT block /plan-tango:update — running the skill explicitly is always honored.)

Feedback

Issues and PRs: github.com/egsok/plan-tango/issues.

Telegram channel where I post about AI tooling and Claude Code workflows: @neiroset_ne_vinovata.

Author

Built by Egor Sokolov — 10 years in product (Sberbank, Rolf, Claustrophobia). Writing and experimenting with AI — mostly Claude Code, Codex, and squeezing more out of dev workflows. Deep-dive on plan-tango (the why and what I learned): egorsokolov.ru/ai/plan-tango/.

License

MIT — see LICENSE. Copyright (c) 2026 Egor Sokolov.