Release notes: LingTai TUI/Portal v0.10.0 and kernel v0.15.0

This release is about one thing: making the cockpit honest about where a session spends its tokens.

The previous window made the runtime plural — many Codex identities, one fleet — and turned the raw call log into trend reports. This one brings that legibility into the default view, so an operator notices a session quietly burning budget without first opening a report.

What changed

TUI / Portal v0.10.0

• Token telemetry where the calls are: API-call group footers show current-round usage — input, cache miss, output, cache rate — and the home view bottom row shows current-session token telemetry plus a compact context-percentage bar (#441).
• Less noisy replay: Ctrl+O hides the bulky _meta envelope by default and points to /notification when full metadata is wanted (#440).
• Telemetry copy and layout, made legible (#442–#447): the footer is preserved while session telemetry shows (#442); session telemetry no longer needs Ctrl+O (#443); the home telemetry context row layout is fixed (#444); the home status hint and zh telemetry copy are cleaned up (#445); the status hint is corrected to say “expand” (#446) and then localized (#447, the candidate head).
• Richer session observability: session/kanban detail panels show more cached per-session stats — API call stats, expanded token breakdowns, agent paths — keyed to molt windows so refreshes stop drifting (#428, #430, #432, #433, #438, #439).
• Faster session loading: session events load from the SQLite index instead of re-scanning event files (#435).
• Markdown front matter: the markdown viewer renders YAML front matter (name / description / version) instead of silently stripping it (#426).
• A list command surfaces companions, the beginner work manual was rewritten, and a dev-guide check confirms a rebuilt TUI binary actually lands on PATH (#421, #423, #425, #427).

Kernel v0.15.0

• Bash result fidelity: completed commands carry additive, model-visible ok / command_status / warning fields, so a nonzero inner exit or a Python traceback no longer hides under status: ok; status keeps its old meaning for compatibility (#493).
• Secret-redaction hardening: the relaunch/refresh watcher subprocess routes events.jsonl writes through the kernel redactor (key-aware, fail-open with a diagnosable redaction_unavailable marker), closing a separate-process leak path (#491, #492 series).
• Token ledger in SQLite, with scope: sum_token_ledger gains an explicit scope — main_agent excludes daemon rows, all includes them — with parent/child double-count semantics documented and tested (#503, #492).
• Daemon artifact manifest: each run writes a metadata-only artifacts.json, so daemon(action="check") surfaces path, size, mtime, and role without scanning the run directory, with safe fallback for older runs (#499).
• Source-drift nudge: startup runtime fingerprints (git HEAD + source digest) land in .status.json; agents emit source_drift nudges when on-disk source diverges, and lingtai-doctor reports them — skipped in dev runtimes (closes #178).
• Codex summarize guidance: adapter_comment now reports full/incremental counts, a full_to_incremental_ratio, a 1:10 target, and dynamic summarize_economy hints (#518); the guidance thresholds were then tuned (#519) — the newest functional change before the version bump.
• Codex transport and MCP/messaging: continuation is decoupled from transport, session identity is molt-aware, the endpoint pool rotates and can be overridden at molt boundaries, and the WebSocket cache ledger is surfaced (#486, #495, #498, #502, #504, #517); Telegram rich-formatting and empty chat_action / parse_mode handling are fixed and WeChat inbound replay is suppressed (#488, #496, #501, #508).

Why this matters

You cannot trim what you cannot see. Putting per-round token cost and session context in the default view — and making that row read cleanly in both languages — is what turns “we log tokens somewhere” into “the operator notices the expensive session.” On the kernel side, the same theme runs deeper: a bash result that admits its inner command failed, a token total that says whether it counts daemons, an event log that cannot persist a secret, and an agent that knows when it is running stale source are all the runtime refusing to quietly mislead a long, unattended session.

Validation

Kernel v0.15.0 was validated from a clean release worktree at the bump commit c365eec (on base 0797e93):

• python -m compileall over src and tests (clean);
• full pytest: 2935 passed, 4 skipped;
• python -m build for the sdist and wheel, and twine check on both (PASSED);
• the only delta vs origin/main is the pyproject.toml version bump to 0.15.0.

TUI/Portal v0.10.0 was validated at candidate head 37be28b (= origin/main; the build version is injected via make ... VERSION=v0.10.0, with no source bump):

• git diff --check against v0.9.6, clean apart from the known generated docs/stars/stars.csv CRLF caveat;
• full TUI and Portal Go tests passed;
• portal/web npm ci && npm run build passed (dev-only npm audit advisories noted, not blocking);
• make build produced lingtai-tui v0.10.0 and lingtai-portal v0.10.0.

Note on artifacts: the locally built kernel wheel is macOS-arm64 platform-tagged, so the portable sdist is the artifact for PyPI unless per-platform wheels are intentionally published.

Release links

These are the intended tags for v0.10.0 / v0.15.0; the GitHub Releases, PyPI upload, and Homebrew tap bump are cut as part of the publish step.

• Kernel release: https://github.com/Lingtai-AI/lingtai-kernel/releases/tag/v0.15.0
• Runtime package source: https://pypi.org/project/lingtai/0.15.0/
• TUI/Portal release: https://github.com/Lingtai-AI/lingtai/releases/tag/v0.10.0
• 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.

Contributors

@huangzesen (lead, scope and validation owner), @ktwu01 (TUI preset refactor), @9s5bz2jvd2-lang (TUI beginner work-manual rewrite), and @TZZheng (kernel runtime source-drift nudge and MCP inbox latency diagnostics).

Issue reporters whose reports drove shipped or completed work this window: @lin-du (kernel #301, implemented and shipped as PR #488 “Expose Telegram rich formatting options”; #300 closed completed) and @888yzbt888 (TUI #401 preset bugs, with Bug 1 fixed by in-window kernel PR #479; issue closed completed).

Thanks also to @BrianLiubr (TUI #429/#431) and @xczics (TUI #437) for in-window bug reports that are still under triage — credited here as reports, not yet as shipped fixes.

Direction

The theme holds: make the runtime honest, measurable, and now legible by default. With per-session token cost in front of operators and a token ledger they can scope, the next step is to act on what the numbers show — trim the turns that miss cache, and let the cockpit point at the session that is spending the most.