Generated from docs/roles.md at df1dbe1. Edit the repository file, not the wiki.

Roles

Role	Who	Does	Does not
Lead	Claude (the session the maintainer talks to)	Plans, splits work into cards, sets the Agent field, writes briefs, reviews results, keeps the board current	Merge, or decide what only the maintainer can
Implementer and reviewer	Claude subagents, Sonnet by default	Implement a task from a complete brief; review another agent's work	Pick their own scope
Mechanical coder	Claude Haiku	Mechanical code from a complete spec, transcribed exactly (a plan's own bugs come along, so it is reviewed)	Anything that needs judgment
Deep reviewer	Claude Opus	A genuinely hard judgment problem, or a whole-branch design review of something risky	Run by default: use only with the maintainer's OK and the reason stated
Research, planning and attack	Codex	Research, plans, and the adversarial review of security-sensitive code	Resume across git worktrees (see Lessons)
Small fixes	Copilot coding agent	Small, fully specified issues and deferred minor items	Anything with open design questions
Maintainer	A person	Decisions, bench and hardware work, and every merge

Defaults and why

• Sonnet is the default for implementation and review. Opus burns the plan's allowance several times faster, so it is a deliberate choice: the lead names the reason and gets a yes first, and never upgrades silently.
• Haiku only for work with no judgment in it. If a brief needs the word "decide", it is not a Haiku task.
• Every card carries an Agent. An unassigned card is unowned work; agents-status shows it as unassigned.
• The maintainer merges. Agents open pull requests and report; a person decides. That is the one gate that never moves.

See Personas for the Claude subagent roles, which are defined in the Department of Internal Arguments, not here.