← Releases
中文 English

id 20260627-2 · 2026-06-27 · kernel v0.15.2

Kernel v0.15.2

LingTai kernel v0.15.2: roomier result reads and leaner runtime guidance

A compact kernel patch after v0.15.1: Codex summarize/cache behavior now waits for the real reconstruction boundary before resetting websocket continuation, summarize/molt guidance is clearer about local compaction versus provider context, `_meta` exposes current-session token efficiency while shedding duplicated notification prose, and `read` now defaults to a 100k-character page budget while retaining the 200k hard cap for short-lived long-result inspection.

Upgrade / runtime

This is a kernel-only patch release. The lingtai PyPI package is the runtime package source used by LingTai-managed environments; existing projects should follow their normal TUI-managed refresh or setup path instead of treating a bare global pip command as the everyday upgrade story.

New features and improvements

  1. 01

    Summarize waits for the reconstruction boundary

    The Codex path no longer treats every local summarize as an immediate websocket fresh-epoch event. It waits for the runtime point where provider context is actually rebuilt.

    • Codex summarize epoch resets are delayed until provider-side reconstruction is required (#534).
    • Summarize/molt guidance now explains that local result compaction is immediate, while provider reconstruction is delayed and task-boundary molt is the stronger completed-work boundary (#535).

    Why it matters:Keeping the incremental Codex websocket chain alive until the real reconstruction point protects cache continuity without hiding the moment when a full context reset becomes necessary.

  2. 02

    Token efficiency is visible without repeating guidance

    The runtime now gives agents the compact numbers they need to reason about current-session token economy while moving static instruction text back into resident guidance.

    • _meta.agent_meta.token_efficiency exposes API calls, input/cached tokens, cache rate, average input tokens, and current context size/window (#537).
    • Notification guidance is now a meta_guidance.notification_handling hook with source names, per-channel duplicate guidance is removed, and tool-result char leaders are capped to the top five (#538).
    • The post-reconstruction reminder now points at 0.6 * context_window as the summarize-then-molt threshold instead of carrying a separate guiding-average field (#538).

    Why it matters:A long-running agent needs the live budget numbers, not repeated static prose. This patch keeps the signal and cuts the duplicated context weight.

  3. 03

    A 100k read default with a 200k hard ceiling

    With summarize and molt carrying the cleanup path, short-lived long-result inspection can be roomier by default.

    • read now defaults to a 100k-character per-call page budget while keeping the 200k runtime hard cap.
    • The read manual, English/Chinese/Wen tool descriptions, continuation tests, and stale tool-result-cap comments were aligned with the new budget.
    • Release validation covered the full kernel suite, build, twine check, archive inspection, PyPI upload, GitHub release, and no-cache pip index verification.

    Why it matters:The everyday page size should match the new operating model: inspect enough context to decide, then summarize or molt deliberately instead of losing useful evidence to an overly tight cap.

Contributors

Thanks to everyone whose commits, co-authored changes, PRs, reviews, assignments, or reported/closed issues are represented in this release window.

Validation

Final release validation was run from a clean release worktree at commit 603e5d854ed7.

Links