[ranxianglei]

Ships since v0.1.37

• fix: recommend only viable compressible ranges (>=200 tokens) on all surfaces #148: recommend only viable compressible ranges (≥200 tokens) on all surfaces — prevents the atomic-batch rejection failure mode.
• adopt billion-context-kit (shared panel + viableRanges + topicFallback) #149 (supersedes fix(/acp): separate session accounting from sent-to-LLM view in the panel #147): adopt billion-context-kit 0.2.0 — shared /acp panel with dual-scale accounting (session-tree vs sent-view), topic fallback on block rows, kit-provided viableRanges.
• fix: sent-view nudge arbitration + kit 0.2.0 panel scales #150: nudge arbitration on the sent-view estimate — stops permanent false EMERGENCY nudges when getContextUsage falls back to session-tree accounting (compressed originals included, never shrinks; the omp issue fix: inline typebox into dist (ACP tags/nudge silently disappeared) #18 report was 366K tree vs 180K window = 204% while the real sent view was ~5%). e2e scenario 03 hardened accordingly.
• acp-kernel 0.0.23 → 0.0.24: nudge quality batch — pressure-band restore + effective-pending gating (feat: preserve decompress ops meta-memory in block summaries #70), small-range merge to minCompressRange + tier-aware emergency arbitration (release v0.1.18 #57), tier-stamp rebaseline reset (docs: rebrand ACP to billion-context in README prose (EN + ZH) #71).

278/278 tests (node runner), typecheck clean, build green.

Note: pi's realUsage stays provider-anchored when providers report usage (≈ sent view, correct); the sent-view estimate only takes over arbitration, and the tree-scale number remains in logs + the panel's footer-scale line.