subagent-driven-development

Dispatch one fresh implementer subagent per task using a file-based task brief extracted by scripts/task-brief, then mark completion in a progress ledger. After all tasks finish, dispatch a final whole-branch code reviewer over the full commit range.

Subagent-driven development stays in the current session, dispatching fresh subagents per task with no context pollution and no human check-ins between tasks. The executing-plans alternative runs in a separate parallel session with human-in-the-loop checkpoints.

The ledger lives in a git-ignored per-plan workspace at .ai/sdd/<plan>/progress.md, recording each completed task with its commit range. After compaction, the controller trusts the ledger and git log instead of conversation memory to resume at the first incomplete task.

Always specify the model explicitly: cheap models for mechanical single-file tasks with complete specs, standard models for multi-file integration work, and the most capable model for architecture and the final whole-branch review. Omitting the model silently inherits the session's most expensive one.

Never batch tightly coupled or judgment-heavy tasks. Batching is only safe when consecutive tasks share the same shape, each is independently a cheap-model task, none depends on another's output, and their plan Interfaces blocks do not collide.

Assess the blocker: provide more context and re-dispatch for context problems, re-dispatch with a more capable model for reasoning-heavy tasks, split oversized tasks, or escalate to the human if the plan itself is wrong. Never force the same model to retry unchanged.