Skip to content

Developer

Repo maintenance automation. Point it at a GitHub issue and it triages, implements, evaluates, and opens a PR. Point it at a PR and it posts a code review. Point it at a repo and it audits docs, tests, and code quality.

Developer runs agent pipelines powered by Claude Code SDK. The agent reads code and reasons about it. All the actual git and GitHub operations -- commits, pushes, PRs, issue creation -- happen in deterministic Python code that the agent can't touch. This split is the core design decision. The agent is good at understanding code. Python is good at not accidentally force-pushing to main.

Quick look

# Resolve a GitHub issue (triage -> implement -> evaluate -> PR)
developer resolve assistant 25

# Review a pull request
developer review assistant 42

# Address PR feedback
developer address assistant 42

# Audit documentation for drift
developer audit docs assistant

# Validate a repo's config from scratch
developer validate assistant

Commands accept GitHub URLs too:

developer resolve https://github.com/gregology/assistant/issues/25
developer review gregology/assistant/pull/42

How it works

Each command runs a pipeline made up of stages. A stage is an agent call with a specific prompt template, a set of allowed tools, and a structured output model. After each stage, a routing function looks at the output and decides what comes next.

The resolve pipeline has 13 stages. The review pipeline has 3. Audits are single-stage. But they all follow the same pattern: render template, call agent, route based on output, repeat.

Python code handles everything between stages. It creates git worktrees, captures diffs, runs quality tools, commits changes, pushes branches, and opens PRs. The agent never sees a git push command.

What you need

  • Python 3.11+
  • uv for dependency management
  • A GitHub App with issue and PR permissions
  • Claude Code SDK access

Head to Getting started for the full setup walkthrough.