handoff

Portable context transfer and the phase-boundary decision tree (continue / clear / compact / handoff / subagent).

Smart-zone budget

Dumb-zone onset is model-dependent; working figure ≈ 150,000 tokens (a slope, not a cliff). Bail/compact/handoff as you approach it. At ≥~150k mid-feature: stop hard work; prefer compact+one-line focus to finish a thin slice, or handoff if crossing a boundary — never "just continue."

Tiebreaker: the tree outranks any single skill's default. If a skill suggests Continue but the tree's first question (still in smart zone?) is already NO, the tree wins — compact or handoff.

Ordered decision tree (ask in order)

Continue? Still in smart zone and you still need the primary grilled/implementation context. Typical mid-feature at ~90k: Continue — handoff early throws away primary context. Grill→implement usually Continues.
Else if context is irrelevant to what's next → Clear (cheapest; most room).
Else if work must cross agent / directory / person / side-quest → Handoff.
Else if the next chunk can run AFK (no human), e.g. automated review → Subagent.
Else → Compact with a one-sentence focus (residual default). Implement→QA is a cast-iron compaction case.

"Review" means walk the tree: same-agent QA → Compact; AFK automated review → Subagent; other agent/person/dir → Handoff. Do not answer compact-vs-subagent without asking which of those three. If the user won't disambiguate, assume same-session QA → Compact, and say that's what you assumed.

Do not outsource mid-phase boundaries to auto-compaction.