Release notes: LingTai TUI v0.9.5 and kernel v0.14.0

TL;DR. LingTai TUI v0.9.5 and LingTai kernel v0.14.0 are out. The kernel makes the Codex integration more honest and observable — accurate request metadata, the operator's own ChatGPT account header, and end-to-end LLM latency telemetry — while hardening turn recovery. The TUI gets a faster launch path and richer, more readable notifications.

This release is about telling the truth and measuring the cost.

The previous window made tool-result metadata visible. This one turns that lens on the runtime’s own behavior: what it tells the backend about itself, and how long each phase of a turn actually takes. Less guessing, more measurement.

What changed

TUI / Portal v0.9.5

Launch is faster: the authoritative session rebuild is deferred off the launch path, and the session cache is now concurrency-safe so a deferred rebuild can’t race the UI.
/daemons lazy-loads run detail, and token-ledger reads now stream instead of loading whole JSONL files into memory.
Notification snapshots render their full meta envelope and markdown blocks, are scrollable, and persist across restarts.
The initial mail view shows a loading… banner while the first rebuild is pending instead of looking empty.
Codex thinking defaults to xhigh, with a reasoning-effort preset option.
Setup preserves agent identity and legacy init-preset fields; portal meta hygiene and repo ignore rules were tightened.

Kernel v0.14.0

Agents send an honest metadata envelope to the Codex backend, and forward the operator’s own ChatGPT-Account-ID header on Codex requests, so attribution and account routing are accurate.
New LLM latency telemetry records provider-wait and per-phase timings, so operators can see where a turn’s time is actually spent.
Long tool results are surfaced in agent meta with a top-10 preview list and char counts; metadata is nested under _meta and excluded from summarize sizing, so it no longer distorts the context budget.
Durable tool results are replayed before heal, and turns poisoned by a stuck worker (WorkerStillRunning) now recover instead of wedging.
Legacy molt-pressure configuration is ignored; molt pressure moved into agent meta.

Why this matters

When an agent runs for hours, two questions decide whether you trust it: what is it telling the world about itself, and where is the time going. This release answers both. The honest metadata envelope and forwarded account header mean Codex sees the truth about who is calling; the latency telemetry means a slow turn is now a measurable fact instead of a hunch. The recovery work — durable replay, poison recovery — keeps a single stuck turn from taking down a long session.

Validation

Kernel v0.14.0 was validated with:

full pytest suite: 2635 passed, 4 skipped;
python -m build for the sdist and macOS arm64 wheel;
twine check on both artifacts (PASSED);
a clean-venv install from the published wheel (import lingtai → 0.14.0).

TUI/Portal v0.9.5 was validated with:

focused TUI tests for the changed areas;
full TUI Go tests;
Portal web install/build (vite);
Portal Go tests.

Release links

Kernel release: https://github.com/Lingtai-AI/lingtai-kernel/releases/tag/v0.14.0
Runtime package source: https://pypi.org/project/lingtai/0.14.0/
TUI release: https://github.com/Lingtai-AI/lingtai/releases/tag/v0.9.5
Homebrew tap: https://github.com/Lingtai-AI/homebrew-lingtai

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

The theme stays steady: make the runtime honest about itself, and measurable. With latency now instrumented, the next step is to act on what it shows — trim the slow phases, and keep long-running work both truthful and recoverable.