Simplified Workflow
For bug fixes, small-to-medium changes, local refactorings (≤1000 LOC). One agent runs the full mini-lifecycle itself — no multi-agent phase handoffs, no separate QA pass.
Brief: WHAT, not HOW
Give the goal, acceptance criteria, and any relevant Prior Knowledge (MemCan) — never a file list or approach (see grand-admiral § Development-Work Delegation). The agent investigates, plans, and executes.
Model
Route to a high-capability model — Codex Astra (--effort high), opus, or fable. Follow any standing dev-routing preference for code work if one is configured. Not for lightweight tiers: the point is trusting one agent through the whole cycle unattended.
Single-Agent Loop
- Plan — investigate, decide the approach, note it briefly before writing code.
- TDD — write tests first; they must fail before implementation (see
coding-best-practicesTDD rules). - Implement — until tests pass.
- Self-review — review the full diff: correctness, edge cases, duplication, comment discipline, formatting/linting (see
coding-best-practices). General review, notgrumpy-review. - Self-fix, then repeat 4–5 until a review pass finds nothing new. Cap at 5 passes — if still finding issues at the cap, stop and report the remainder rather than loop indefinitely.
Out-of-scope findings during self-review: note them in the Final Report, don't fix inline.
Commit Discipline
Commit all changes before exiting — uncommitted work cannot be merged.
Code-mutating agents work in an isolated git worktree, coordinator pre-created (see grand-admiral § Worktree Isolation, Option A default). Push per grand-admiral § Worktree Isolation → Post-wave push.
Stuck?
If the agent can't proceed (ambiguous requirement, can't reproduce, self-fix loop caps out) — surface to the coordinator/user rather than guessing.
Final Report
Approach taken, tests added, self-review passes run, fixes applied, anything left out-of-scope.