Codewright recipes (agent skill)

When to use this

Before running a terminal command you didn't get from the user verbatim — especially one that's destructive, hard to undo, or you're not 100% sure of the exact flags for. Search first; only fall back to writing the command yourself if nothing matches.

Risk vocabulary

Workflow

1. Step 1

   Search

   Search for a recipe matching the task.

   $pnpm --filter @codewright/agent-skill recipe search "undo last commit"

   Copy

2. Step 2

   Inspect

   Read risk, requires_confirmation, and platforms from the search result before acting, or fetch the full document for parameters, verification steps, and undo instructions.

3. Step 3

   Render

   Render the recipe for the current platform and parameter values — this never executes anything, it only substitutes parameter values into command text and returns it.

   $pnpm --filter @codewright/agent-skill recipe render git.undo_last_commit macos

   Copy

4. Step 4

   Confirm — non-negotiable

   Before running any command from a render result, check whether it requires confirmation. If it does, stop and ask the user — do not run destructive commands on the agent's own judgment.

5. Step 5

   Run, then verify

   After running the rendered commands, run the recipe's own verification commands before reporting success. Don't assume a command worked because it exited 0.

6. Step 6

   Undo if needed

   Use the render result's undo path — either the undo commands, or, if there's no undo path, surface the honest note instead of guessing a rollback.

Configuration

Points at production by default (https://codewright.tools/api/v1). Override with CODEWRIGHT_API_BASE_URL to hit a staging deployment or local dev server.