Slash Commands ate My Workflow

In my last post, I described how git worktrees let you run multiple coding agents in parallel. The idea was sound, but the workflow still had friction: creating worktrees, running setup, writing commit messages, pushing PRs, merging back, cleaning up. Each of those steps required me to stay in the loop or (more likely) context switch back to sheppard a change along.

Since then I've automated almost all of that away. My coding session logs now look like this:

> /build-worktree implement @feature-spec.md
> [review code and make some minor edits]
> /merge-worktree

That's it. Two slash commands. I estimate this makes me 3-4X faster compared to the non-worktree agent flow, because I'm no longer the bottleneck for all the operational overhead.

The Tools

Worktrunk

I've switched from wtp to Worktrunk. Worktrunk (wt) is a git worktree manager built specifically for parallel AI agent workflows. Three commands cover 90% of what I need:

wt switch -- create and switch between worktrees by branch name (no path juggling)
wt list -- see all worktrees with branch status, diff stats, and CI state at a glance
wt merge -- squash, rebase, fast-forward merge, and clean up in one shot

The killer feature for my use case is wt merge. It handles committing uncommitted changes, squashing, rebasing onto the target branch, fast-forward merging, and removing the worktree -- all in a single command. Where my old workflow required five separate git commands and a mental checklist, wt merge is one line.

Worktrunk also supports hooks (run scripts on create, pre-merge, post-merge), LLM-generated commit messages, wt list --full with CI status per branch, and configurable path templates so worktree directories follow a consistent convention.

Install it with brew install worktrunk.

OpenCode

OpenCode is an open-source CLI coding agent. I use it over alternatives because it supports custom slash commands -- markdown files that define multi-step workflows. This is the glue that turns "build a feature in a worktree" from a multi-step process into a single command.

OpenCode also supports custom agents, MCP tools, and a skill system for extending its capabilities. It runs in your terminal, works with any git repo, and keeps all context local.

The Slash Commands

I wrote two custom OpenCode commands that handle the full lifecycle of a worktree-based task. Both are just markdown files in ~/.config/opencode/commands/.

`/build-worktree`

Given a task description (or a file reference), this command:

Infers a branch name from the task -- feat/, fix/, chore/, etc.
Creates a worktree using wt switch -c $BRANCH_NAME --base $BASE_BRANCH --no-cd
Implements the feature -- reads the codebase, writes code, follows existing conventions
Validates locally -- discovers and runs lint, typecheck, and test commands
Pushes a PR -- creates the branch, pushes, and opens a pull request with gh
Monitors CI -- polls until CI passes, and if it fails, automatically fixes issues and retries (up to 5 times)

The entire pipeline runs autonomously. I can kick off a /build-worktree and switch to another task while it works.

`/merge-worktree`

Once I'm happy with the result, this command:

Finds the worktree path from the branch name
Runs wt merge to squash, rebase, fast-forward merge, and clean up
Handles conflicts -- if rebase fails, it launches a sub-agent to resolve conflicts automatically

One command, no fiddling.

The Result

My workflow went from managing worktrees manually to a two-command loop. The efficiency gains come from eliminating all the small decisions and typing that used to interrupt my flow:

No more typing branch names three times
No more remembering which validation commands to run
No more writing commit messages
No more manually pushing and creating PRs
No more multi-step merge and cleanup rituals

The agent handles all of it. I review the code, make edits if needed, and merge. That's my job now.

The full commands are open-source if you want to try them. You'll need OpenCode and Worktrunk installed.