Wield's two Claude Code dynamic workflows and their Codex skill equivalents for deep reviews and codebase sweeps.

Workflows

Claude Code ships two dynamic workflows — scripts that orchestrate many subagents in parallel and then verify their own output. Where a skill routes a single conversation, a workflow fans out across independent agents, applies codified quality patterns (parallel lenses, adversarial verification, majority voting), and returns a merged, deduplicated result.

Dynamic workflows require Claude Code v2.1.154+ with workflows enabled in /config. They are user-invoked slash commands — Claude only authors or runs one on explicit opt-in.

The workflows

Claude Code	Codex	What it does
`/ultra-review [scope]`	`$wield:ultra-review [scope]`	Persona-diverse review of a diff: parallel lenses → confidence gate → merge/deduplicate → adversarial verification.
`/sweep [glob] [--bugs\|--security]`	`$wield:sweep [glob] [--bugs\|--security]`	Codebase-wide hunt. `--bugs` uses defect-focused finders; `--security` uses attacker personas and STRIDE/OWASP. Run the mode that matches your goal.

/ultra-review

Review a diff, PR, commit, or branch through 2–6 parallel review lenses, then verify each finding.

How workflows differ from skills

Skills are portable: invoke them as /wield:<name> in Claude Code or $wield:<name> in Codex. Codex uses native subagents for the bundled deep-review and sweep behaviors.

A /<name> workflow runs a script that spawns many subagents in parallel, gates their findings by confidence, merges and deduplicates, and verifies adversarially before returning. It's installed separately and invoked explicitly.

Codex needs no separate workflow installation. $wield:ultra-review and $wield:sweep are installed with Wield; start a new Codex session after an install or update so the skills are discovered.

Installation

Claude Code plugins can't bundle workflows — a plugin ships skills/agents/hooks/MCP/LSP/monitors, and there's no workflows component. So the scripts must be copied into a .claude/workflows/ directory Claude Code reads. Wield includes a script to do this:

bash scripts/install-workflows.sh            # -> ./.claude/workflows/ (this project)
bash scripts/install-workflows.sh --global   # -> ~/.claude/workflows/ (all projects)
bash scripts/install-workflows.sh --list     # list bundled workflows, install nothing

Each workflow then appears as /<name> in / autocomplete. The script is safe to re-run — it overwrites only the workflow files it ships.

Enable workflows

In Claude Code, run /config and enable workflows. They require v2.1.154+.

Install the scripts

Run bash scripts/install-workflows.sh (or --global) from the Wield plugin directory to copy the workflows into a .claude/workflows/ folder.

Invoke a workflow

Type /ultra-review or /sweep from the / autocomplete. Each accepts an optional scope.

Workflows fan out across many agents and can consume significant tokens. They're built for high-value, broad-coverage passes — a full pre-merge review or a codebase-wide audit — not quick one-off checks. For those, reach for the equivalent skill (/wield:code-audit, /wield:security-audit, /wield:error-fixing).

Skill vs. workflow: which to use

You want…	Use
A quick review of your pending changes	`/wield:code-audit`
An exhaustive, multi-lens review before merge	`/ultra-review`
A targeted security pass on a change	`/wield:security-audit`
A codebase-wide red-team audit	`/sweep --security` (or `$wield:sweep --security` in Codex)
To debug one specific failure	`/wield:error-fixing --diagnose`
To hunt latent bugs across the whole codebase	`/sweep --bugs` (or `$wield:sweep --bugs` in Codex)

Workflows

/ultra-review

On this page