Release notes: LingTai TUI v0.9.6 and kernel v0.14.2

TL;DR. LingTai TUI v0.9.6 and LingTai kernel v0.14.2 are out. The headline is multi-account Codex: each agent or preset can point at its own OAuth token file, with a setup UI to match. The cockpit also learns to report where its calls go — tool-call and API-call trends, cache-miss detail in the kanban view — and gets a scrollable /doctor, bilingual loading text, and a more conservative 250k default context cap. The kernel fixes two run-directory edge cases that made long-running daemon work misreport its state.

This release is about running more than one identity, and seeing more than one number.

The previous window made the runtime honest about who it is and how long a turn takes. This one lets a fleet be several whos at once — different agents signing in as different Codex accounts — and turns the call log into something you can actually read.

What changed

TUI / Portal v0.9.6

Multi-OAuth Codex: setup gained UI and preset support so you can choose which Codex identity a preset uses, instead of editing a token path by hand (#415).
Swiss-knife trend reports: a new script and skill produce tool-call and API-call trend reports, with the script and skill wording aligned (#414).
/doctor is scrollable and has section headers, so a long diagnostic readout is navigable instead of a single wall of text (#413).
Loading strings are bilingual and i18n-aware, so the cockpit speaks the operator’s language while it warms up (#412).
The default context/token cap was lowered to 250k — a more conservative out-of-the-box working budget, still raisable per setup (#411).
The kanban call view now shows cache-miss detail, so a turn that paid full price instead of hitting cache is visible where operators already look (#410).

Kernel v0.14.2

Per-agent and per-preset Codex OAuth token files: a new manifest.llm.codex_auth_path lets each agent or preset point at its own Codex OAuth token, so multiple ChatGPT accounts can run side by side; the default falls back to the shared path (#484).
The daemon historical check now resolves run directories correctly after a refresh or molt, so prior-run inspection no longer breaks across a context shed (#483).
An empty working_dir in a bash call is treated as unset and falls back to the default agent directory, instead of running against an ambiguous path (#480).

Alongside the features, the window carried release hygiene: the kernel update command is confirm-gated, a destructive global preset-split migration was neutralized so it can no longer rewrite operator presets, TUI release version comparisons are now classified explicitly, and the release star CSV was normalized.

Why this matters

A fleet that mixes accounts needs each agent to authenticate as itself; a single shared token makes attribution and rate limits everyone’s problem at once. Per-preset Codex auth turns “which account is this agent on” into a deliberate, inspectable choice. And you cannot tune what you cannot see — making tool-call trends and cache misses legible is the first step toward trimming them. The daemon and bash fixes target the exact moments a long session is most likely to lose track of where it is: just after a molt, or when a field is left empty.

Validation

Kernel v0.14.2 was validated with:

python -m compileall over src and tests (clean);
focused and expanded pytest over the daemon, bash, notification, preset, deep-refresh, tool-executor, openai, codex, provider, and auth surfaces: 513 passed;
python -m build for the sdist and wheel, and twine check on both (PASSED);
PyPI confirmed info.version = 0.14.2.

TUI/Portal v0.9.6 was validated with:

git diff --check against v0.9.5 (clean);
full TUI Go tests, and make clean && make build;
Portal web npm ci && npm run build (npm audit warnings noted, build and tests passed);
full Portal Go tests, and make clean && make build;
the Homebrew Release workflow succeeded for tag v0.9.6.

Release links

Kernel release: https://github.com/Lingtai-AI/lingtai-kernel/releases/tag/v0.14.2
Runtime package source: https://pypi.org/project/lingtai/0.14.2/
TUI release: https://github.com/Lingtai-AI/lingtai/releases/tag/v0.9.6
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, measurable, and now plural. With multiple Codex identities and readable call trends in hand, the next step is to act on what they show — route work to the right account, and trim the calls that miss cache.