Easy Kit

• The line is a statement, not a question: Track: minimal — C0, Risk Low, one file or Track: extended — C2, Risk Medium.
• Extended track if any of these is true: Complexity C2 or above; Risk Medium or High; a word such as refactor or migrate; a change that deletes, moves, renames or converts user data (for example CSV to JSON); more than one session or agent; a UI component.
• Minimal track otherwise. The Supervisor writes the code itself: test first, change, a /simplify pass, real run. It saves a Decision record in tasks/MINIMAL_LOG.md and adds a board row marked Track: minimal. It commits on a new branch without asking. It writes no task guide. It starts no sub-agent, except the read-only reviewers of /simplify (and of a Debate in auto mode).
• Your override: reply full or minimal at any time and the Supervisor switches. It may move a task up to extended on its own. It never moves a task down.

Put the exact tag [auto] in your request. The Supervisor runs either track up to the point where it waits for your review and asks you no question. It pauses only for a few stops, such as a safety-hook block. It writes a final report that lists every decision and how it made each one. The task Risk sets how it decides:

• Low: the Supervisor decides and logs the reason.
• Medium: a Debate with 2 Advocates.
• High: a Debate with 3 Advocates, and at least one decision is always debated.

In a Debate, read-only Advocate agents each frame the problem and propose their own option. resolve-question runs it in auto mode only. The Supervisor decides and takes the safer option on a tie. It marks the line ⚠ decided for you — please review. A task has at most 3 Debates. The 4th question goes to you.

A Stop hook checks the final report before the run may end. Safety hooks still block.

Auto verify runs for any task, with or without [auto]. When a task waits for your review, the Supervisor starts a fresh headless /verify run with python3 .claude/hooks/stop_auto_verify.py --start <ID>. A Stop hook starts it as a backstop. That run, not the Supervisor's own session, writes the verify row.

Close-out: after a pass, the hook moves the row to Done only if the run record proves the pass, the tests row and the UI rows hold, and the report is complete. With CLAUDE_AUTO_VERIFY=off, /verify is yours again for every task without [auto]. With [auto], merge, push and ship stay with you. With [admin] instead, the run is the same, and the Supervisor may also merge and push the task's own branch, and then push main, only after its auto verify passed. A task that changes only docs, task files, memory or tests has no runtime surface: when the run answers skip, the hook checks the changed paths in git and records a pass, so you need not type /verify. One path outside that list keeps the result blocked. A force push or a push of any other branch stays blocked. With [admin], ship runs as a plan only, never a deploy. The [admin] merge and push also refuse when the repo config holds a key that can run code or reach another host, such as a merge driver, a filter or core.hooksPath; the refusal names the key.

Installing into an existing project never deletes what is already there: a file or folder the kit would replace is first moved to a backup, listed in the install output. Identical files are left alone.

The command takes no options — you choose from numbered menus, and Enter takes the default each time:

• What to do — 1) Install 2) Cancel (or Update / Reinstall / Cancel where Easy Kit is already installed).
• Which CLIs — 1) Claude Code 2) Codex. Pick one or both (1 2 or 1,2); the ones found on your PATH are pre-selected, and Claude Code when neither is.
• New or existing project — 1) New project 2) Existing / legacy project. An existing / legacy project gets CLAUDE_LEGACY.md's rules as its CLAUDE.md.

A plan then lists the CLIs, the project type and anything that will be backed up, and asks Proceed? [Y/n] before a single file changes.

• No terminal (CI, containers) — no menus: the defaults are printed and taken.
• Anything on the command line — it stops with nothing written.

To pull in newer Easy Kit changes later, run the same command again from your repo root and choose Update (see below).

There is one command: run the same install line again inside your installed project (still a git repository). Easy Kit sees it is installed and asks what to do:

Easy Kit is already installed in /path/to/your-project.
  1) Update (keeps your edits)
  2) Reinstall (backs up your edits)
  3) Cancel
Choose [1]:

Before changing anything it shows a plan — files to ask about, files to remove because the kit no longer ships them, anything to back up — and ends with Proceed? [Y/n].

If the plan lists files you edited, one question about all of them replaces Proceed? [Y/n]:

How should Update treat your edited files?
  1) Ask me about each one
  2) Overwrite all of them (each backed up first)
  3) Keep all of them
  4) Cancel
Choice [1]:

• 1 asks about each file, as in the table below. Enter picks it.
• 2 moves each file to <file>.bak-<commit>, then overwrites it. A file it cannot back up is kept.
• 3 keeps every edit. 4 changes nothing.

• Answer n to go back to the menu. Cancel changes nothing.
• No terminal (CI, containers, ssh without -t) — no menu: it prints the safe default — Update, keeping every edit — and takes it. Reinstall never runs without a terminal.
• An existing update.sh in a checkout still works: it runs the same Update action.

Update re-fetches the framework fresh, then compares each MANIFEST file — plus CLAUDE.md — against the content hash recorded at the last install/update:

Every Update run adds a dated entry to the top of .claude/update-log.md: how your edited files were decided, what happened to each one, and counts for the rest.

## 2026-10-26 14:05 — Update 75040d9 → b122017
Decided by: auto-approve (overwrite all)
- CLAUDE.md — overwrote (your edits; backup CLAUDE.md.bak-75040d9)
- agents/qa.md — overwrote (your edits; backup agents/qa.md.bak-75040d9)
- 41 kit files you never edited — refreshed
- 2 files added

CLAUDE.md is delivered from the same source the project was installed with — a brownfield/existing project keeps receiving CLAUDE_LEGACY.md's rules, never the greenfield CLAUDE.md — recorded per project in .claude/harness-lock.json.

Which CLIs get updated

Easy Kit stores no list of CLIs. Update asks no CLI question: it refreshes every CLI whose directory is already present. Reinstall shows the CLI menu, pre-selecting those. That means:

• A Codex-only project refreshes .codex/skills/ and is not given .claude/{skills,agents} it never asked for.
• A Claude project keeps its .claude/{skills,agents} links repaired — including when a link has gone missing, stale, absolute, or been replaced by a real directory.
• Adding a CLI later: run the same command, choose Reinstall, and pick the CLI. A CLI already set up that you do not pick is kept and refreshed, and the plan says so — to drop one, delete its folder.

One edge worth knowing: if both .claude/skills and .claude/agents are deleted outright, Update leaves them absent. Restore them: run the command, choose Reinstall, and pick Claude Code.

At Stage 1 you list each CLI and the model it spawns at each Complexity level (C0–C3) in the ## CLI Model Table of PROJECT_SPEC.md, and the effort at each level (low, medium, high, xhigh or max) in the ## CLI Effort Table beside it. craft-spawn-prompt reads both and recommends the model and the effort for each spawn. An empty model cell makes it stop and ask; an empty effort cell passes no effort.

When a task moves up to the extended track, or verify returns fail, the Supervisor first raises the effort of the next spawn by one step, before it changes the model or splits the task. It never lowers effort on its own.

code-review (Stage 4): its security reviewer follows one short rule file, skills/code-review/references/security-reviewer.md. A finding must name who attacks, the boundary they cross and the harm. A lead that cannot be proven yet stays as needs_validation with the one fact still missing, and never blocks a merge by itself.

The wiring lives in .claude/settings.json. Both setup.sh and update.sh merge every entry below into that file. No manual merge step.

Codex executes the kit's skills directly, by name, from a projected .codex/skills/ directory (choose Codex in the installer's CLI menu) — it is not left with doctrine alone.

• Each skill's body is capped at 8 KB per Codex's own limit.
• A skill over the cap is skipped with a named, loud warning rather than truncated.
• Four of this kit's skills currently exceed it and are skipped for Codex: bugfix, craft-spawn-prompt, diagnose, and write-better-skill.

What a non-Claude provider still cannot enforce: Claude Code's PreToolUse/PostToolUse hooks, the Agent spawn tool, and everything built on them:

• code-review, security-review, verify, ship, migration-safety, and the git-guardrails hook.
• Those stay on the Claude supervisor: an adapter carries the doctrine, and Codex can now run the skills it has room for, but the deterministic enforcement layer does not follow.