@auron-labs/opencode-improve

OpenCode plugin that injects an improve agent for auditing codebases and writing self-contained implementation plans.

The idea: use your most capable model for the part where intelligence compounds — understanding the codebase, judging what's worth doing, writing the spec — and hand execution to cheaper models. The plugin never implements anything itself. The plan is the product.

you          →  /improve                    (expensive model, advises)
plans/       →  001-fix-n-plus-one.md       (self-contained specs)
other agent  →  implements, tests, ships    (cheap model, executes)

Install

bun add @auron-labs/opencode-improve

Requirements

• OpenCode installed and loading plugins from your config.

Usage

Add to your OpenCode config:

{
  "plugin": [
    ["@auron-labs/opencode-improve", {}]
  ]
}

Restart OpenCode. The plugin adds:

• An improve agent (invoked via /improve or selected from the agent picker)
• A /improve slash command

What /improve does

Invoke in an OpenCode session:

/improve                        full audit → prioritized findings → plans
/improve quick                  cheap pass: hotspots, top findings only
/improve deep                   exhaustive: every package, every category
/improve security               focused audit (also: perf, tests, bugs, …)
/improve branch                 audit only what the current branch changes
/improve next                   feature suggestions — where to take the project
/improve plan <description>     skip the audit, spec one thing
/improve review-plan <file>     critique and tighten an existing plan
/improve execute <plan>         dispatch a cheaper subagent, review its work
/improve reconcile              refresh the backlog: verify, unblock, retire
/improve … --issues             also publish plans as GitHub issues
/improve help                   show the usage guide above in-session

The advisor audits across nine categories (correctness, security, performance, tests, tech debt, dependencies, DX, docs, direction), vets every finding before showing it, and writes plans for the ones you select.

Plans land in plans/ (or advisor-plans/ if plans/ already exists for another purpose). Each plan is self-contained — a different agent with zero context can execute it.

Hard constraints

• Never edits source code. The advisor only writes to plans/ and advisor-plans/. Permission rules enforce this: all edits outside those paths are denied.
• Never mutates the working tree. Read, search, and analysis commands only. No installs, builds, or commits.
• Never reproduces secret values. Credential locations are cited by file:line with a rotation recommendation.

execute limitation

The execute <plan> variant dispatches a subagent to implement a plan in an isolated worktree. OpenCode subagent worktree isolation is not guaranteed. If true worktree isolation is unavailable, the advisor will tell you and hand the plan over for manual execution instead. The advisor never merges, pushes, or commits to your branch.

Notes

• The improve agent is a primary agent — you can select it in the agent picker or invoke it via /improve.
• The upstream shadcn/improve skill markdown is bundled in the package and read at runtime, along with the reference files it points to.
• Refresh the bundled files from upstream with bun run sync:improve at the repo root, or run bun run sync:improve:check to fail CI when local copies drift. The script discovers SKILL.md and references/*.md from shadcn/improve@main:skills/improve/ (via the GitHub tree API) and writes them flat into this package's references/ directory, so renames and additions are picked up automatically. Local .md files in references/ that no longer exist upstream are removed on sync and reported as failures by check.