ClaudeX

Claude Code skills that put a second model — and an enterprise-grade research pipeline — in your terminal.

ClaudeX ships these Claude Code skills:

/claudex:think — Codex second-opinion brainstorming

/claudex:think arms a lightweight gate, then runs normal brainstorming. A PreToolUse hook intercepts each AskUserQuestion call, sends the drafted question to Codex (read-only, via the Codex plugin's runtime), blocks the question, and feeds Codex's critique back to Claude. Claude then presents:

Codex says: _(Codex's critique, verbatim)_ My take: _(1–3 sentences — agree/disagree, folding in any fixes)_

…and re-issues the (possibly revised) question, which passes through. Claude never has to remember to ask Codex — the hook enforces it.

Design before implementation is also guarded. While a think session is armed, extra PreToolUse hooks block edits to code and to protected instruction files (CLAUDE.md/SKILL.md/AGENTS.md/GEMINI.md), block implementation skills (TDD, executing-plans, …), and refuse Bash commands that write code or tamper with the gate marker — so Claude can't quietly skip brainstorming and start coding. Design/spec/plan markdown stays writable, and the gate opens automatically once brainstorming has run and writing-plans is invoked. Turn it off with enforceImplementationGate: false or the kill switch.

This is a cooperative-agent guardrail, not a security sandbox. It fails open and is heuristic on the shell path: read-only exploration passes, common write patterns are blocked, but a determined bypass (exotic shell, or write-capable MCP tools not covered by the matchers) is still possible. Its job is to interrupt the "just edit the file" reflex, not to contain an adversary.

Requires the Codex plugin (openai/codex-plugin-cc) for the Codex runtime and superpowers (anthropics/claude-plugins-official) for the brainstorming flow. The quickest way to get both is /claudex:setup (see below).

Configuration

Runtime config lives at ~/.claude/claudex/config.json (created on first use; see config.example.json):

Kill switch: rm -f ~/.claude/claudex/armed (this session) disarms both the Codex question gate and the implementation gate; set {"enabled": false} in the config (globally); set {"enforceImplementationGate": false} to keep the Codex second opinion but stop gating edits/skills; or just tell Claude "solo" / "no codex".

deep-research — enterprise-grade deep research, fully on your subscription

A port of the deep-research core of NVIDIA's AI-Q Research Assistant Blueprint (an enterprise multi-agent framework: planner → parallel researchers → orchestrator, with hard citation integrity), re-implemented so it runs entirely on a Claude subscription — no AI-Q server, no external API keys, no shim.

AI-Q's agents drive web search through OpenAI-style function-calling, which is exactly why the server can't run on a subscription CLI. deep-research flips the architecture: Claude Code itself is the researcher.

Flow: route/disambiguate → plan a TOC + acceptance criteria → fan out up to 6 parallel `Task` researchers (native WebSearch/WebFetch, ≤8 searches each) → bounded gap-fill for weak sections → synthesize a long-form report with inline [n] citations → verify.

The core merit — deterministic citation integrity (not model-judged). A bundled verifier (scripts/verify_citations.py, ported from AI-Q's citation_verification.py) builds a registry of the URLs that actually appeared in search results and keeps an inline [n] only if its URL string-matches a captured one (with fuzzy repair and reject-on-ambiguity). Fabricated citations are dropped; URLs are renumbered; if zero real sources were captured it refuses to emit a normal report and labels the output UNVERIFIED — model knowledge only.

deep research on <topic>

The skill is self-documenting — see skills/deep-research/SKILL.md and the prompts under skills/deep-research/prompts/.

/claudex:email — broadcast a question to your panel

When you're hosting a screen-shared brainstorm and scrolling makes it hard for others to follow, /claudex:email emails the current question (plus a 2–4 line context snapshot) to a panel you define in ~/.claude/claudex/panel.json. It previews and asks you to confirm, then sends one group email and logs the sent message id (and thread id when available) so you can find replies later.

• Send-only — it does not collect or summarize replies.
• Filter recipients: /claudex:email professors, /claudex:email florescu bozdog, or an

ad-hoc address like /claudex:email [email protected] (smoke test / one-off).

• Sends from the connected Gmail account; replies arrive there and are not monitored live.

panel.json stores members under the top-level panel key:

{
  "panel": {
    "members": [
      { "name": "Name", "email": "[email protected]", "role": "student", "include": true }
    ]
  }
}

Set include: true per person to put them on the broadcast.

/claudex:build — turn an approved spec into a PR, with Codex as the builder

Where /claudex:think ends (a written spec), /claudex:build begins. Claude conducts; Codex does the writing — both the implementation plan and the real code. Every artifact is independently attacked by a Codex reviewer and a Claude reviewer, then a fresh Codex agent reconciles the two reviews into a single verdict. The whole run happens in an isolated git worktree and ends at an opened PR — never a direct push to main.

Invoke it after a spec exists. The spec is an input.

Phase 0 — Spec readiness gate   (Codex checks the spec is actionable)
Phase 1 — Plan                  ┐ unified verdict loop
Phase 2 — Build                 ┘ (plan doc, then implementation diff)
Phase 3 — Integrate → branch + PR

The unified verdict loop (same for plan and build):

1. Codex builds the artifact via codex:rescue — one builder for the plan;

for the build, parallel builders for independent units (Claude integrates) or a single builder for tightly-coupled work.

1. Dual adversarial review, run independently: the Codex adversarial-review

runtime and an adaptive Claude reviewer (a plan-attacking subagent, or /code-review on the diff). Both return blocking findings, requirement/test gaps, and a merge-readiness signal.

1. Reconcile — a fresh Codex agent weighs both reviews and returns one verdict:

• go — done.
• fix-go — builder applies low-risk fixes; no re-review; done.
• fix-review — builder fixes; loop back to step 2.

1. Loop guard — a second fix-review in a phase stops and asks you. Any

material scope change also pauses for your approval.

Failure handling: a Codex gate that errors is retried once, then degrades (proceeds with a logged warning) by default — set failClosed: true in ~/.claude/claudex/config.json to halt instead. Tests/build must pass before the PR is opened.

It's model-driven (no enforcement hook): the verdict loop is the control structure. See skills/build/SKILL.md for the full contract.

ultra-research — dual-model adversarial research

ultra-research runs independent deep research with Claude and Codex, then performs an adversarial blind cross-audit of the other model's citations. Each cited claim is graded PASS, WEAK, or FAIL; those grades roll up into a trust score and verdict for each draft.

Codex then writes the trust-weighted final report, including a disagreement ledger that preserves material model conflicts instead of smoothing them away. Each run is saved to an Obsidian-style vault with five notes: two model reports, two cross-audits, and one final synthesis, plus a 00-index MOC for the run.

Configure the vault and Codex budgets with vaultPath and codexTimeoutMs in ~/.claude/claudex/config.json (see config.example.json). Prefer ultra-research over deep-research only for high-stakes work, dual-model verification, or cases where prior research had mixed citation integrity.

Install

/plugin marketplace add WillInvest/ClaudeX
/plugin install claudex@willinvest

Then bootstrap the dependencies (the Codex plugin + superpowers) in one command:

/claudex:setup

It installs both plugins, asks you to /reload-plugins once, then verifies the Codex CLI. Re-running it is safe — already-installed pieces are skipped.

License

MIT — see LICENSE. The deep-research skill bundles a port of NVIDIA AI-Q (Apache-2.0); attribution is in skills/deep-research/NOTICE. Built to complement obra/superpowers and openai/codex-plugin-cc.