Release notes: LingTai kernel v0.16.0 and TUI/Portal v0.10.3

techdevlog
TL;DR. LingTai kernel v0.16.0 and TUI/Portal v0.10.3 put the runtime's cache economy and context-pressure signals under an explicit budget. The kernel adds a cache_miss_budget — a per-session, per-molt soft cap defaulting to 1,000,000 cache-miss tokens; when it is reached the runtime restamps cache miss budget {budget} reached, molt now under durable context metadata, alongside a running cache-miss count. Soul flow becomes an explicit env opt-in via LINGTAI_SOUL_FLOW_ENABLED. Token and meta surfaces get leaner and more inspectable, and the TUI shows current-session cache-miss tokens inline, teaches the soul-flow opt-in in setup and kanban, exposes model-visible summary previews, and adds clipboard image paste. Both versions were validated from clean release worktrees before publication.

This window is about making the cost of a long run visible and spendable. A cache miss is where a long-running agent quietly burns money and context, and until now it accrued invisibly. v0.16.0 turns it into a budget: a soft cap you can see counting down, with a durable molt now cue when it runs out. Alongside it, soul flow stops running silently by default, the runtime’s token and meta surfaces get leaner, and the cockpit surfaces exactly what the model saw.

What changed

Kernel v0.16.0

TUI / Portal v0.10.3

Why this matters

A cache miss is the quiet cost of a long run — it burns money and context at the same time, and an invisible cost is one nobody manages. Turning it into a budget with a visible running count and a durable molt now cue makes that cost spendable: the operator and the agent both see it accrue, and both know when it is time to molt rather than keep paying.

The same instinct runs through the rest of the window. A powerful behavior that runs silently by default is a surprise waiting to happen, so soul flow becomes an explicit flag with a stable disabled result — on only when someone chose it, with the cockpit showing that choice rather than guessing at it. And an agent reasons better on a lean, ranked view than on a wall of raw meta, so the token and meta surfaces get flattened and thinned while the TUI surfaces exactly what the model saw. Less hidden cost, fewer silent defaults, more of the picture visible.

Contributor scope

After the first website draft, we re-audited contributors using the broader release accounting now required for LingTai release notes. The contributor list on the release card covers the previous two version windows for both repositories: kernel v0.15.2..v0.16.0 and TUI/Portal v0.10.1..v0.10.3. It includes release/commit evidence plus PR authors, PR commenters/reviewers, issue authors, and issue commenters — including closed or unmerged PRs and closed issues that were discussed but not accepted.

The audit scanned 160 PR records and 128 issue records. The public contributor set for this release entry is: huangzesen, TZZheng, ZigongXu, BrianLiubr, 9s5bz2jvd2-lang, BatalloLu, houleixx, Keesan12, LinnkidChen, LuuOW, rawpaper123, wchwawa, xczics, github-actions[bot].

Validation and release hygiene

Both versions were validated from clean release worktrees before publication.

Strict ranges for this window:

Kernel v0.16.0 gates:

TUI/Portal v0.10.3 gates:

For regular users, LingTai’s managed project environments remain the normal way the runtime package is resolved. The PyPI page is the published runtime package source and a useful verification point, not the primary end-user upgrade story.

Direction

With cache misses under a visible budget and soul flow behind an explicit flag, both a long-running agent and its operator start from a clearer picture of what the run is spending and what it is doing. The next steps keep paying down the same debt — turning hidden costs and silent defaults into things you can see, budget, and choose.