Executes tasks.md via subagents. Does not ask for permissions between tasks — works to the end, reports at the end.
Input
$ARGUMENTS — path to tasks.md or feature-name.
Process
1. Preparation
git rev-parse HEAD # BASE_SHA
Read tasks.md and design.md. Report:
Выполняю: <feature>, задач: N
2. Loop over tasks
For each task:
Implementer
Task(subagent_type="general-purpose", prompt=<implementer-prompt.md>)
Pass: the task text (NOT the path), project context, dependencies, list of skills.
Reviewer (one pass)
Task(subagent_type="general-purpose", prompt=<reviewer-prompt.md>)
Checks BOTH compliance with the specification AND code quality in one pass.
On ❌ from reviewer
Send the remarks back to the implementer. Do not ask the user — fix and continue. Maximum 2 iterations per task. If it is not resolved in 2 — record the problem and move on.
Skipping review
For tasks without BSL code (pure metadata via /meta-compile, /subsystem-edit, etc.) — reviewer is not needed. Validation via the corresponding -validate skill is sufficient.
Marking
В tasks.md: [ ] → [x]
3. Final report
Готово: N/N задач
Коммиты: <список>
Проблемы в ревью: <если были>
Нерешённые замечания: <если есть>
When to ask the user
Only on a blocker: a dependency is missing, the task is contradictory, it is impossible to continue. Everything else — decide yourself.
When NOT to use
- 1-3 tasks → execute directly with skills (without subagents)
- No tasks.md → first
/write-planor/brainstorm