Four kinds of page, after the Divio system. Pick by what you need right now.
Follow one from start to end. Each takes a small project through a fixed path and shows what you see at every step.
- Your first feature, when you have installed the plugin and want to see a request become merged code.
- Recover a task, when you want to see what a stopped task looks like and how to pick it back up.
- An epic with a human review, when you want tasks to land on an epic branch and to review the whole epic before it merges.
Each answers the question in its title, for someone who has done a tutorial.
- Choose an integration mode, before the first dispatch of a repository or an epic.
- Choose a review level, when you want more or fewer eyes on a task before it merges.
- Write a good task, when you create tasks by hand or judge what split produced.
- Add project hooks, when workers should install dependencies or run your checks before a merge.
- Handle a stopped or failed task, when dispatch says a task needs you.
- Handle a crashed worker, after a reboot or a killed process.
- Review a task by hand, when a task or an epic waits on a human review gate.
- Clean up leftovers, when worktrees, branches or a merge queue are left behind.
- Run unattended, for headless sessions, cron and CI.
- Run on several sessions, when two people or two machines share one repository.
- Lower the cost, when a run costs more than it should.
- Run workers on another model, for local models through Ollama.
- Keep code under the limits, to let the branch analysis file refactor tasks.
- Make prior-art search find more, when design misses documents you know exist.
- Develop the plugin, to change sdlc itself: layout, conventions, tests.
Complete and terse. One home per fact.
- Skills: every
/sdlc:*skill, its arguments, what it does and prints, and what it does headless. - Settings: every setting, local, shared through beads, per epic and per task, with defaults.
- Task fields: what sdlc stores in beads: fields,
dispatch_*metadata, event beads, merge queues, gates. - Messages: every message a script or skill prints to a person, and what to do about it.
- Scripts: each script, who calls it, its inputs, exit codes and files.
- Hooks: the plugin's Claude Code hooks and the beads hooks that wake the supervisor.
- Requirements: tools, versions, platform limits, what is optional.
Read once, to make good decisions.
- Task lifecycle: the states a task goes through and what moves it.
- Dispatch: the supervisor's contract, work order, slots, takeover.
- Integration: modes, the integration task, merge queues, verify at merge, project hooks.
- Reviews: the three levels, why review comes before the queue, re-review on change.
- Code quality: why metrics gate a merge and how branch analysis files refactor work.
- Prior art: what design and split look at before proposing work.
- Trust model: who can make the dispatching machine run what.
- Harness neutrality: what depends on Claude Code, and what Codex, OpenCode or Ollama would need.
- Planning: from a request to a task graph, and why every task carries a scope and a verify command.
- Rules: why this repository has rule files.
- Roadmap: what comes next.
- Glossary: the terms the other pages use.