subagent-driven-developmentlisted

# Subagent-Driven Development (workflow-driven) Execute a written plan by handing it to a deterministic Workflow. **You** (the controller, on the most capable model) do only what needs human judgment or Opus reasoning — plan parsing, conflict scan, model tiering, the go/no-go gate, and the irreversible finish. The background Workflow (`workflows/sdd.mjs`) runs the per-task loop as code so it cannot drift, skip a review, or rubber-stamp. **Why a workflow, not a flowchart:** the implement → review → fix loop has known structure. Encoding it as control flow (not prose the model walks by hand) is the recommended pattern for known-structure tasks and removes the dominant agentic-SWE failure — verification theater / skipped steps. The model still fills every node's judgment; only the loop is fixed. ## When to use - A written plan exists with `# Task N` / `## Task N` headings (what `task-brief` parses) and the work has test coverage. - **Not** for large, ambiguous, or brownfield work where the plan can't be decomposed into independently testable tasks — elaborate task ceremony there creates review overload without guaranteeing compliance. Recommend smaller tasks or manual execution instead. ## Controller process ### 1. Read the plan; capture Global Constraints verbatim Note the project-wide requirements (exact values, formats, naming rules). They bind every task and are passed to every agent. ### 2. Pre-flight conflict scan Scan the plan once for tasks that contradict