subagent-driven-development

Dispatch a fresh implementer subagent for each task with the complete task text, relevant design sections, and allowed file scope. After implementation, run a specification review against the task BASE_SHA diff, then a code quality review before approving the task.

Specification review verifies the task's requirements are fully implemented without unplanned features or compatibility shims. Code quality review runs afterward and checks correctness, naming, test quality, and file scope against the actual workspace diff.

Run a stage review after all tasks in the current stage are committed. It verifies stage design coverage, cross-task interface consistency, and deliverables for later stages, without requiring the service to be deployable at that point.

Diagnose the cause first: supplement missing context, use a stronger model for reasoning limits, split oversized tasks into fix tasks, or escalate plan gaps back to the planning skill. Never redispatch without changing something.

No. Implementers are forbidden from running git add, commit, amend, switch, merge, rebase, reset, stash, or clean. All Git and progress operations are controlled centrally by the calling smart-exec-plan workflow.

Tasks and stages are not deployment boundaries, so adapters, dual reads/writes, temporary schemas, and feature flags add unplanned complexity. The final review verifies old implementations and transitional code are fully removed before merge.