Development workflow

After Setup, try a small change: add a retry limit to outbound calls.

Specify

Use code-spec to start a new code development. For example:

“I need to limit outbound calls to three retries and add a test that proves we stop at the cap.”

The agent reads the documentation, then writes a document with the requirements and how to verify them. You then review and approve the spec before implementation.

Plan

Ask for code-plan along with a clear implementation description or a spec written with code-spec.

The agent identifies the files to change, the tests to add, and the checks to run. Then presents you with a task plan that covers each implementation wave and the verification gates. You approve the plan before work starts.

Execute

Ask for code-execute and the agent makes a code change and runs the tests and harness checks. If you pass along a plan written with code-plan, it can orchestrate subagents to execute it in one go, stopping at the defined human gates.

You review the diff and results, then proceed.

Commits and pull requests

By default, the work stays uncommitted until you ask for code-commit.

You can enable automatic commits in .agents/code-commit.config.yml with autocommit: true. Set autocommit-rule to unit to commit after each planned work unit, or wave to commit after a group of units passes its checks. The defaults are false and wave; setup does not create this config.

Automatic commits require a plan, the code-commit skill, and passing validation. Invalid config values stop execution. Without a plan or the commit skill, the work stays uncommitted.

Opening a pull request (code-pr), reviewing it (code-review), and resolving CI failures (code-ci) each require a separate request. None merges the PR.