Best Practices Research

1. Scope the recon

Identify 1-3 domains the task touches (a stack, library, or framework). For each domain, distill 1-3 concrete, specific questions and not generic ("Python best practices") but task-scoped ("rate-limiting a FastAPI endpoint with Redis in 2026", "current React Server Component data-fetching patterns"). Vague questions produce vague, unusable answers.

Done when: every domain has its own scoped question set (1-3 questions each).

2. Dispatch the recon subagent(s)

One domain → one subagent with all its questions. 2-3 clearly unrelated domains → one subagent per domain, dispatched together in a single message (parallel Task calls).

Probe Task/Agent enums; route per ../council/references/task-workflow.md (SSOT).

Use portable role general-purpose, write-capable, foreground (you need findings before writing code; read-only mode strips MCP/internet access these subagents need).

Prompt each subagent with: