Skills

Workrex installs six native skills for supported agent CLIs. A skill contains instructions in SKILL.md; it can be selected by the agent or invoked directly. Codex uses $workrex, $coordinator, and other $skill-name invocations. Claude Code uses /workrex, /coordinator, and /skill-name. The sections below use slash names as labels; choose the invocation syntax of your agent.

Using with workrex

You can run workrex commands yourself, but skills let agents handle the complete workflow - committing with context-aware messages, resolving conflicts intelligently, and delegating tasks to parallel worktrees.

• /workrex - Teach the agent how to use workrex
• /merge - Commit, rebase, and merge the current branch
• /rebase - Rebase with flexible target and smart conflict resolution
• /worktree - Delegate tasks to parallel worktree agents
• /coordinator - Orchestrate multiple agents with full lifecycle control
• /open-pr - Write a PR description using conversation context

For a Codex agent, you can trigger $merge from the dashboard using the m keybinding:

dashboard:
  merge: "$merge"

Installation

Run workrex setup to install all skills automatically:

workrex setup --skills

This detects installed Claude Code, Codex, OpenCode, Grok, Pi, and Oh My Pi agents and copies skills to the right location. The setup wizard also offers skill installation on first run.

Skill sources in skills/ are already native Workrex instructions. Setup copies those source bytes into each detected agent’s skill directory. Manual installation uses the same folder names and SKILL.md files, with no command-name translation.

Claude Code: ~/.claude/skills/ (or project .claude/skills/). If CLAUDE_CONFIG_DIR is set, workrex setup --skills installs to $CLAUDE_CONFIG_DIR/skills/ instead.

Codex: ~/.agents/skills/. Invoke skills with $workrex, $merge, and so on. Codex detects new skills automatically; restart Codex if they do not appear. This shared skills directory is independent of CODEX_HOME.

OpenCode: ~/.config/opencode/skills/.

Pi: ~/.pi/agent/skills/.

Oh My Pi: ~/.omp/agent/skills/.

/workrex

Teaches the agent how to use the workrex CLI. Invoke /workrex to give the agent knowledge of workrex commands, configuration, and concepts. The agent can then use workrex to manage worktrees, check status, and interact with other agents.

View skill ->

This is a reference skill, not an action skill. It loads workrex documentation into the agent’s context so it can use workrex commands as needed. For specific workflows, the agent is directed to the dedicated skills below.

/merge

Handles the complete merge workflow:

1. Commit staged changes using a specific commit style
2. Rebase onto the base branch with smart conflict resolution
3. Run workrex merge to merge, clean up, and send a notification when complete

View skill →

Instead of just running workrex merge, this skill:

• Commits staged changes first - the agent has full context on the work done and can write a meaningful commit message
• Reviews base branch changes before resolving conflicts - the agent understands both sides and can merge intelligently
• Asks for guidance on complex conflicts

/rebase

Rebases with flexible target selection and smart conflict resolution.

View skill →

Usage: /rebase, /rebase origin, /rebase origin/develop, /rebase feature-branch

See Resolve merge conflicts with Claude Code for more on this approach.

/worktree

Delegates tasks to parallel worktree agents. A main agent on the main branch can act as a coordinator: planning work and delegating tasks to worktree agents.

View skill →

See the blog post on delegating tasks for a detailed walkthrough.

Usage:

> /worktree Implement user authentication
> /worktree Fix the race condition in handler.go
> /worktree Add dark mode, Implement caching  # multiple tasks

Customization

You can customize the skill to add additional instructions for worktree agents. For example, to have agents review their changes with a subagent before finishing, or run workrex merge after completing their task.

/coordinator

Orchestrates the full lifecycle of multiple worktree agents: spawning, monitoring, communicating, and merging. Unlike /worktree which dispatches tasks and returns, /coordinator turns the agent into a persistent orchestrator that manages agents through completion.

View skill ->

The coordinator agent does not implement tasks itself. It writes prompt files, spawns worktree agents, monitors their status, sends follow-up instructions, and triggers merges.

Key commands used

Command	Purpose
workrex add -b -P <file>	Spawn an agent in the background
workrex status	Check agent statuses
workrex wait	Block until agents reach a target status
workrex capture	Read terminal output from an agent
workrex send	Send instructions or skill commands to an agent
workrex run	Run shell commands in an agent’s worktree

Cross-project agent communication

Agent commands (send, capture, status, wait, run) can target agents in other projects, not just the current git repository. If a worktree name is not found locally, workrex searches all active agents globally.

# From any project, send to an agent in another project
workrex send other-project-worktree "run the tests"

# Use project:handle syntax to disambiguate when names collide
workrex send myproject:docs-update "also add the API reference"

# Check status of agents across projects
workrex status myproject:feature-auth

Lifecycle commands (add, open, merge, remove, close) remain scoped to the current repository.

Fan-out / fan-in pattern

The typical coordinator workflow:

1. Write prompt files with full context for each task
2. Spawn all agents in the background
3. Confirm agents started with workrex wait --status working
4. Wait for completion with workrex wait
5. Review results with workrex capture
6. Merge one at a time by sending /merge to each agent sequentially

When to use /coordinator vs /worktree

• /worktree: fire and forget. Spawn agents and return control to you. Good for delegating tasks you will review later yourself.
• /coordinator: full automation. The agent manages the entire lifecycle, including waiting, reviewing output, sending follow-ups, and merging. Good for multi-step plans where tasks depend on each other.

/open-pr

Writes a PR description using the conversation context and opens the PR creation page in browser. This is the recommended way to finish work in repos that use pull requests.

View skill →

The skill is opinionated: it opens the PR creation page in your browser rather than creating the PR directly. This lets you review and edit the description before submitting.

The agent knows what it built and why, so it can write a PR description that captures that context.