[wusijian007]

… / v4)

Layer 2 of the v4 pipeline: a deterministic stage that drops stale tool SCAFFOLDING. In the stale zone (outside the root task and the recent window) it snips oversized tool_result content to a tiny marker — the dominant token cost of old tool turns — while leaving tool_use blocks (tiny), prose/reasoning text, and user turns alone. Pairing is always safe (blocks are snipped in place, never removed) and the artifactPath is preserved so the full output stays restorable.

The layer separation is the point: L2 reclaims only the cheap, low-value tool apparatus; prose/reasoning is intentionally untouched because that is L5's job (semantic recap). The cascade is now [spill, snip, reclaim].

Because L1 spill and L2 snip both target tool scaffolding, a fixture whose whale is a stale tool_result would get reclaimed before the semantic layer runs — so the semantic-compaction eval task and the M3.2c query integration test now seed an assistant PROSE whale instead, which L1/L2 skip and only L5 can condense. That is the correct layer separation and keeps the eval fingerprint unchanged (prose replaces tool_result; scripted usage is the same).

• core: snipStaleToolScaffolding + SnipScaffoldingOptions (context.ts); query loop inserts the snip stage between spill and reclaim.
• tests: 2 new L2 unit tests (snips stale tool_result, leaves prose/tool_use/ root/recent; preserves artifactPath in the marker); two compaction fixtures retargeted to prose whales.
• docs: CLAUDE.md query-loop step 7 lists L2; v4 roadmap M4.2 delivered.

232 tests pass (+2); all 10 eval tasks green, fingerprint unchanged.