subagent-driven-development

Dispatch a fresh implementer subagent per task with a file-based task brief, then dispatch a task reviewer subagent against the generated diff package. After all tasks pass review, run a final whole-branch review before merging.

Subagent-driven development stays in the same session with no context switch, gives each task a fresh subagent, and iterates faster without human-in-the-loop between tasks. Parallel-session plan execution hands off the whole plan to a separate session instead.

Match model capability to task complexity: cheap models for mechanical single-file tasks with complete specs, standard models for multi-file integration, and the most capable model for architecture work and the final whole-branch review. Always specify the model explicitly when dispatching.

Assess the blocker before retrying: provide missing context and re-dispatch, escalate to a more capable model for reasoning-heavy tasks, break oversized tasks into smaller pieces, or escalate to the human if the plan itself is wrong. Never force the same model to retry unchanged.

A progress ledger stored in the git directory records each completed task with its commit range. After compaction, the controller reads the ledger and git log to resume at the first incomplete task instead of re-dispatching finished work.

Avoid it when tasks are tightly coupled and cannot be implemented independently, when no written implementation plan exists yet, or when you cannot stay in the current session. Brainstorm or write the plan first in those cases.