carefully-crafted-plugins

A small, opinionated set of multi-agent plugins for Claude Code. Two single-agent bridges — codex (OpenAI Codex) and agy (Google Antigravity) — plus contexthub, the multi-agent hub for cross-provider convergence and token-efficient Claude supervision of Codex workers. Three plugins, ten skills, deliberately lightweight by construction.

Superpowers provides the development methodology. Carefully Crafted provides Codex/Agy bridges, explicit cross-provider convergence, and token-efficient Claude supervision of Codex workers.

What you can delegate

codex — OpenAI Codex CLI

Codex CLI fills capability gaps from Claude Code:

• Image generation (gpt-image-2 via Codex's built-in $imagegen)
• Hard reasoning at xhigh effort (GPT-5.6 Sol)
• An independent, read-only code review on explicit request
• Session resume across follow-ups
• Raw codex exec passthrough
• Optional, explicit bridge setup

/codex:imagegen   generate a 256x256 todo app icon
/codex:reason     solve this dynamic programming problem ...
/codex:review     audit src/auth for security bugs
/codex:resume     <follow-up for the most recent Codex session>
/codex:exec       <any raw prompt to codex>
/codex:setup      optional — copy editable defaults into this repo, re-verify the Codex CLI

Only /codex:imagegen and /codex:reason auto-trigger from natural language; the other four are slash-command only. codex is an optional, separately-installed plugin — install it if you want /codex:* commands. /contexthub:supervise talks to Codex directly through its own scripts and does not require this plugin.

agy — Google Antigravity CLI

Antigravity's terminal coding agent (agy) is a narrow two-skill bridge:

• Nano Banana text-to-image generation through the authenticated agy CLI
• Raw agy -p passthrough

/agy:nanobanana  a poster of a lighthouse at dusk, warm palette
/agy:exec        <any raw prompt to agy>

Nano Banana no longer has a wired multi-tool image backend — there is no story/multi-scene generation, natural-language editing, photo restoration, icon-set, pattern, or diagram tooling here anymore. It is plain text-to-image only, generated directly through the authenticated agy CLI, and may require access or billing for whichever image model your Antigravity account is configured with. For a first-choice default image generator, or for any of the retired capabilities, prefer /codex:imagegen.

Both agy skills are slash-command only.

contexthub — the multi-agent hub

The hub ties Claude, Codex, and Antigravity together for two things Superpowers does not do on its own: cross-provider debate, and bounded external-worker execution.

/contexthub:converge  [--full] <question>   # cross-provider debate: Claude + Codex + Antigravity
/contexthub:supervise <task>                # Claude plans/reviews; isolated Codex workers implement

/contexthub:converge stages a short cross-provider debate — independent answers by default, full mutual critique and refinement with --full — and produces consensus, disagreements, and a recommendation. It writes nothing to disk.

/contexthub:supervise is the token-efficient execution engine: Claude plans (via Superpowers), grades the task, and reviews; isolated GPT-5.6 Sol Codex workers implement in their own worktrees and report back compact evidence, never their full transcripts. At most one correction wave runs before Claude accepts the final result and, with your consent, finishes the branch. Every methodology step — brainstorming, planning, TDD, debugging, review, verification, branch-finishing — is delegated to the corresponding Superpowers skill; contexthub never re-implements it.

Both contexthub skills are slash-command only and manual-invocation gated (disable-model-invocation: true) — that defers each skill's larger body cost until you explicitly run it and blocks automatic invocation. It does not remove the skill's name/description from Claude's always-on context; measured manual-only skills carry ~60–100 tok always-on, the same order as model-invocable ones.

Install

In Claude Code:

/plugin marketplace add anthropics/claude-plugins-official
/plugin marketplace add https://github.com/infroid/carefully-crafted-plugins
/plugin install contexthub@carefully-crafted-plugins
/plugin install codex@carefully-crafted-plugins
/plugin install agy@carefully-crafted-plugins

Add claude-plugins-official first. contexthub requires superpowers@claude-plugins-official, and Claude Code resolves that dependency only from a marketplace you have already added. With it present, installing contexthub pulls in and enables Superpowers automatically — no separate install step. Without it, the install still reports success and gives you contexthub alone; the unresolved dependency surfaces only on a later install attempt, as 1 dependency still unresolved: superpowers@claude-plugins-official. If you see that, add the marketplace and re-run the install.

codex and agy are optional, separately-installed bridges: install codex only for /codex:* commands, agy only for /agy:* commands. Neither is required by /contexthub:supervise.

Requirements

• contexthub requires Superpowers (superpowers@claude-plugins-official) — installing contexthub auto-resolves and enables it; you never install it separately.
• codex, the public bridge behind /codex:* commands, is optional and installed separately; /contexthub:supervise talks to Codex directly and does not need it.
• Claude Code 2.1.143 or later is required for dependency enable/disable enforcement (older versions can still install contexthub, but won't reject an attempt to disable its required Superpowers dependency).
• Codex CLI auth (codex login) remains a user prerequisite for every Codex-touching skill.
• Codex-side Superpowers (superpowers@openai-curated, installed and enabled inside Codex, with its worker skills present) is a separately required prerequisite that /contexthub:supervise checks before running and never installs on your behalf.
• Antigravity CLI auth is only needed for /contexthub:converge and the agy bridge skills — /contexthub:supervise never spawns Antigravity workers.
• Node.js ≥ 20 and Git are the only script/runtime dependencies — pure standard library, no npm packages.
• /codex:setup is optional and explicit; nothing auto-scaffolds a repo on first use of any skill.
• Nano Banana (/agy:nanobanana) generates through the authenticated agy CLI directly: a smaller, text-to-image-only guarantee that may require access or billing for whichever image model your Antigravity account is configured with.

Layout

.claude-plugin/marketplace.json   # 3 plugins listed
plugins/
├── codex/
│   ├── .claude-plugin/plugin.json
│   ├── skills/{imagegen,reason,review,exec,resume,setup}/SKILL.md
│   ├── skills/{imagegen,reason}/evals/evals.json   # the two model-invocable skills only
│   ├── reference/critical-evaluation.md
│   ├── reference/defaults/{constraints,output-formats}/   # packaged defaults, copied in by /codex:setup
│   ├── reference/schemas/code-review.schema.json
│   └── scripts/
│       ├── spec-builder.mjs     # writes the 5-section spec
│       ├── codex-invoke.mjs     # wraps `codex exec` (+ resume, model, sandbox)
│       ├── result-handler.mjs   # parses output, places artifacts
│       ├── setup.mjs            # optional, explicit repo scaffolding
│       └── output-schema.json
├── agy/
│   ├── .claude-plugin/plugin.json
│   ├── skills/{exec,nanobanana}/SKILL.md
│   └── scripts/
│       └── agy-invoke.mjs        # wraps `agy -p` (+ --collect artifact retrieval)
└── contexthub/
    ├── .claude-plugin/plugin.json
    ├── skills/converge/{SKILL.md,references/critique-and-refinement-prompts.md}
    ├── skills/supervise/{SKILL.md,references/protocol.md}
    ├── schemas/{complexity,worker-report}.schema.json
    └── scripts/
        ├── agent-availability.mjs # detects Claude/codex/agy for converge
        ├── supervise.mjs          # supervise's phase-machine entry point
        └── supervise/{checkpoint,codex,contracts,git,scheduler,state,verify}.mjs
tests/unit/                      # node --test, no external deps
tools/
├── lint-skill.mjs               # quality-bar enforcer (run in CI via tests/)
├── eval-check.mjs               # evals.json structural validator
└── run-supervise-live-evals.mjs # paid, human-run release-eval harness (not part of ordinary supervision)
quality-bar.md                   # the gates every skill must clear

How the codex bridge works

The codex bridge writes a 5-section spec to docs/carefully-crafted-plugins/handoffs/ for every structured delegation — /codex:imagegen, /codex:reason, and /codex:review. The sections are: what to do (role + task), how to do (numbered steps or open delegation), standard constraints, expected output format, and any pre-flight clarifications Claude resolved with you. Codex receives a tiny prompt pointing at the spec file and reads everything from disk — no 800KB prompt limit, no opaque handoffs.

The constraint and output-format sections link to packaged defaults under ${CLAUDE_PLUGIN_ROOT}/reference/defaults/ unless you have opted in to project-local copies. Because /codex:setup is optional and never runs on its own, those packaged defaults — not files in your repo — are what a fresh install references; run /codex:setup if you want editable copies under docs/carefully-crafted-plugins/ to customize.

/codex:exec and /codex:resume deliberately skip this scaffolding — exec is a raw passthrough and resume continues an existing session, so neither builds a spec. The agy and contexthub bridges skip it by design too: Antigravity reads the repo itself, /contexthub:converge is its own debate protocol, and /contexthub:supervise drives Codex through a compact machine task graph instead of a handoff spec.

Run the tests

node --test tests/unit/*.test.mjs

The suite includes lint-skill.mjs (gates every SKILL.md against quality-bar.md) and eval-check.mjs (validates every evals/evals.json structure). To run them standalone:

node tools/lint-skill.mjs    # SKILL.md quality bar
node tools/eval-check.mjs    # evals.json structural validation

Migrating to 6.0.0

v6 narrows the marketplace to ten high-value skills and adds /contexthub:supervise, a token-efficient engine where Claude plans and reviews while isolated GPT-5.6 Sol Codex workers implement. Twelve commands were removed with no aliases — Superpowers, not a new Carefully Crafted command, now owns every methodology step they used to perform.

Upgrading from 4.x or earlier? See the 5.0.0 release notes for the forge/triage → contexthub renames first, then apply the table below.

Removed methodology commands → the Superpowers skill that replaces them

Removed	Use instead
/contexthub:spec	superpowers:brainstorming — explore intent, requirements, and design before implementation
/contexthub:plan	superpowers:writing-plans
/contexthub:tdd	superpowers:test-driven-development, run via superpowers:executing-plans or superpowers:subagent-driven-development
/contexthub:review	superpowers:requesting-code-review (or superpowers:receiving-code-review on the receiving end)
/contexthub:verify	superpowers:verification-before-completion
/contexthub:debug	superpowers:systematic-debugging
/contexthub:ship	superpowers:finishing-a-development-branch
/contexthub:triage	grading is now a built-in step of /contexthub:supervise, not a standalone command

Removed capabilities — cut, not replaced

Removed	Why
/codex:playwright	Browser automation is out of scope for this marketplace; no replacement command
/agy:longctx	1M-token long-context analysis is out of scope for this marketplace; no replacement command
/agy:veo	Video generation is out of scope for this marketplace; no replacement command
/agy:setup	Nano Banana no longer has a wired multi-tool backend to set up — /agy:nanobanana now generates directly through the authenticated agy CLI with no setup step

/codex:exec, /codex:imagegen, /codex:reason, /codex:resume, /codex:setup, /agy:exec, /agy:nanobanana (now text-to-image only), and /contexthub:converge are unchanged in name; /contexthub:supervise is new.

Everything above this marker is the v6 migration record and is the only place in this README permitted to name removed commands. Anything added below it is held to the current-surface standard by tests/unit/stale-references.test.mjs.