Release LingTai kernel v0.15.2 · Lingtai-AI/lingtai-kernel

Highlights

Codex summarize/cache behavior now waits for the provider-side reconstruction trigger before resetting the websocket epoch, preserving incremental continuation when summarize only changed local visible history (#534).
Summarize/molt guidance now distinguishes local compaction from delayed provider reconstruction, treats task-boundary molt as the stronger completed-work boundary, and avoids summarize-first when a molt is already planned (#535).
_meta.agent_meta.token_efficiency now exposes current-session token economy (API calls, input/cached tokens, cache rate, average input tokens, and context size/window) so agents and UI surfaces can reason about live token efficiency (#537).
_meta payloads are leaner: notification guidance is now a resident meta_guidance.notification_handling hook, per-channel duplicate guidance is removed, tool-result char leaders are capped to top 5, and the post-reconstruction molt reminder is aligned to 0.6 * context_window (#538).
read now defaults to a 100k-character page budget while preserving the 200k runtime hard cap, matching the new assumption that short-lived long-result inspection is acceptable before summarizing or molting.

Release branch is based on 5b29ac963412856a89b75aeb49e58ac4a0ef64d6 (Compact meta block guidance payloads (#538)) on top of v0.15.1, with the release PR adding the 0.15.2 version bump and the 100k read default cap patch.
git diff --check passed.
PYTHONPATH=src python -m pytest -q passed: 3040 passed, 4 skipped in 291.20s.
python -m build succeeded; python -m twine check dist/* passed; archive inspection found no __pycache__ or .pyc files.
SHA256: wheel 2854c7aa7c7630d4f91e86e73395cee3497cf769b3e5e713e66db3a0a3f36aa4; sdist 5bd4976b10b1e4157ea4bb257e5e63932015bea54ef21e382351f4952e8693ae.

Thanks to: huangzesen.