Flow — research

1. Codebase recon first (inline): what already exists for this task — relevant files, conventions, constraints, prior art (existing specs/plans, git history). Small task area → do it all inline.
2. Fan out only for breadth. Up to 3 parallel flow-researcher agents, each with ONE scoped question (e.g. one per subsystem, or one for an unfamiliar library). Cap at 3; if you want more, the task is too big — say so. Web research only when the task genuinely needs outside knowledge. Dispatch synchronously when you'd only wait anyway; if you background them, end each prompt with "deliver your findings via SendMessage to main" — background agents otherwise idle silently with the results stranded in their own context.
3. Write research.md in the task folder: Findings (with file:line), Constraints & prior art, Open questions. Dense; no pasted file bodies.
4. Triage the unknowns, then ask. Sort every open question: code-resolvable → go read the code, never ask; conventional default → record it, don't ask; deferred or tunable → record it with an owner; shape-changing intent → ask now. Only that last class interrupts the user: ONE AskUserQuestion round, ≤4 questions, recommended option first and marked. Write each answer to the workflow.md decision log immediately. Everything you recorded rather than asked carries into plan.md's Assumptions section — silently guessing is the failure this step exists to prevent. Guard: more than 3 shape-changing questions means the task was never shaped. Stop and route the user to /flow shape instead of planning a guess.
5. Advance. Update updated + Next action, set phase: planning, announce (reference research.md by path), invoke the flow-plan skill.

Idempotent resume: if research.md already has findings, extend — don't redo.