Releases
Release notes for LingTai. Each entry links to the published artifacts and the validation that accompanied them.
- Kernel v0.18.1 · TUI/Portal v0.11.6 2026-07-23
LingTai 0.18.1 / 0.11.6: fixes for first-run Codex, vision routing, and doctor safety
This coordinated patch release fixes three first-run and reliability issues: generic vision requests now follow the active Codex/Codex-pool route instead of guessing from provider spelling, a genuinely empty production-shaped account-pool snapshot now correctly takes the intended legacy fallback instead of being misclassified as non-empty, and `lingtai-tui doctor --help`/`-h` is now read-only. It also clarifies Codex OAuth preset behavior and fresh zsh installer PATH guidance, without changing existing installation destinations or non-OAuth preset behavior.
Read release log → - Kernel v0.18.0 · TUI/Portal v0.11.5 2026-07-21
LingTai 0.18.0 / 0.11.5: a more resilient runtime and a verified Windows install path
This coordinated release focuses on runtime resilience and a verifiable Windows distribution path: native Windows kernel adapters and process lifecycle support, native multi-account Codex quota/epoch handling, Outlook OAuth2 IMAP, lossless Telegram inbound updates with task-card self-healing, and a PowerShell-validated Windows AMD64 install path that sits alongside the existing source and Homebrew builds.
Read release log → - Kernel v0.17.1 · TUI/Portal v0.11.0 2026-07-17
Source releases, live Task Cards, and a clearer first-run journey
This coordinated release makes delivery and operation more concrete: verified kernel artifacts, a source-only TUI/Portal release with Homebrew source builds, Telegram Task Cards with truthful elapsed progress, a project-first startup journey, more resilient daemon backends, and clearer runtime and model contracts.
Read release log → - TUI/Portal v0.10.7 2026-07-13
LingTai TUI/Portal v0.10.7: an admin network switcher, faster mail history, and four honest platform archives
Published from the independently locked TUI/Portal candidate 5fb554b69f966b39cb9eca32a70c84d628714d18 (tree 05c068dbc5e50832d2d5d425099c104290acd2f4), this TUI-only release makes /projects a live admin network switcher (#620, #639), makes large mail history fast and consistent (#631, #632, #637, #638, #640), and keeps projects working across both runtime import layouts (#617). It also corrects daemon cache accounting (#636), refreshes documentation (#618, #628, #629, #630, #633, #616, #642), and publishes four platform archives. The darwin-amd64 archive was manually recovered after the macos-13 Intel runner failed; its full checksum is recorded below.
Read release log → - Kernel v0.16.3 · TUI/Portal v0.10.6 2026-07-10
LingTai release day: tool knowledge ownership, one-shot installer, and leaner prompt budgets
This paired release reorganizes how the kernel owns tool knowledge — every built-in tool now lives in its own package directory with a CONTRACT.md, glossary files in EN/ZH/Wen, and a centralized glossary validator — while single-sourcing repeated tool guidance in the system prompt so #842 saves roughly 2,975 tokens per call (a different measurement scope; not a universal net). The TUI ships a one-shot GitHub-release installer with hardened uv/venv/portal/Go bootstrapping, a custom OpenAI wire selector, codex-pool model classification, mailbox search, session tool-call stats, loopback-only portal binding, and credential-labeling for Codex OAuth. The kernel adds a codex-pool exact-model auth lane, backend flag-discovery submanuals for 9 backends, IMAP attachment filename sanitization, extended credential redaction, AED manifest recognition, a skill/glossary i18n migration that grows ZH prompt descriptions by +5,224 chars (+221.17%) and Wen by +5,106 chars (+212.66%) while EN adds 0 new chars from glossary-only files (#844), and a final packaging repair (#847) that makes the native lingtai-search-sidecar platform wheels platlib-compliant.
Read release log → - Kernel v0.16.2 · TUI/Portal v0.10.5 2026-07-07
Reliability patch: safer migrations, cleaner summaries, and steadier releases
This paired patch release closes the loop on the July reliability batch: schema-critical migrations now fail loudly instead of advancing half-applied state, `lingtai clean` protects live agents, global config writes are atomic, and the kernel gains cleaner summary/manual/notification contracts for long-running agents.
Read release log → - Kernel v0.16.1 · TUI/Portal v0.10.4 2026-07-03
Summarize rebuilds, portal reliability, and a cleaner update path
This paired release makes LingTai's context-rebuild behavior explicit, ships a safer TUI/Portal update and replay path, and closes another round of runtime reliability work across logging, addons, MCP workbench docs, and security hygiene.
Read release log → - Kernel v0.16.0 · TUI/Portal v0.10.3 2026-07-01
LingTai release day: cache-miss budgets, soul-flow opt-in, and leaner runtime telemetry
A paired release that puts 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 that defaults to a million cache-miss tokens; when it is reached the runtime restamps a `molt now` cue under durable context metadata alongside the running cache-miss count. Soul flow becomes an explicit env opt-in via `LINGTAI_SOUL_FLOW_ENABLED`, returning a stable disabled result unless enabled. Token and meta surfaces get leaner and more inspectable — flattened token diagnostics, sparse agent and notification meta, large-result ranking, and a-priori `summary=true` — while 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. The window also broadens daemon/runtime backends (Kimi Code, claude-p wait guard, explicit daemon email tool, MCP completion, HTTP identity headers, Codex encrypted-reasoning replay self-heal).
Read release log → - Kernel v0.15.3 · TUI/Portal v0.10.2 2026-06-28
LingTai release day: an honest init contract and steadier reconstruction signals
A small paired patch that makes two runtime contracts honest. Generated, template, and example init.json no longer seed legacy `prompt` or empty `lingtai` fields — long-lived character state is owned by `system/lingtai.md` / psyche after creation, and the kernel treats a missing seed as a valid empty seed. The kernel also makes context-pressure signals more truthful: a delayed-summarization reconstruction now reports one-shot evidence while sustained high-context pressure is surfaced separately as a molt warning. The window rounds out with system-prompt resources refactored into Markdown catalogs, a WeChat MCP config-path fix and stronger inbound-media validation, grep and IMAP ergonomics, runtime-identity stamps on event logs, and the usual release hygiene.
Read release log → - Kernel v0.15.2 2026-06-27
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.
Read release log → - Kernel v0.15.1 · TUI/Portal v0.10.1 2026-06-27
LingTai release day: self-updating TUI, Claude Code provider, and steadier runtime state
A paired patch release that finishes the self-update story the previous window started for Homebrew: the source / user-local TUI can now update itself — manually and on startup — for installs that are not Homebrew-managed. The kernel adds a `claude-code` LLM provider, an MCP manual sidecar contract with bundled manuals for curated MCPs, shared filesystem/JSON/JSONL helpers and a session-recovery refactor that de-risk kernel state writes, and honest spill/refresh failure visibility — plus cockpit, doctor, and mail polish.
Read release log → - Kernel v0.15.0 · TUI/Portal v0.10.0 2026-06-26
LingTai release day: a cockpit that shows where the tokens go
A paired kernel and TUI/Portal release built around one idea: make the cockpit honest about where a session spends its tokens. The TUI grows home and session token telemetry — a context-percentage bar and per-round usage — followed by a focused copy and layout pass, bilingual status hints, less noisy replay, richer session/kanban observability, and faster SQLite-indexed session loading. The kernel makes bash results report inner failures, routes refresh-watcher events through the secret redactor, indexes the token ledger in SQLite with an explicit reporting scope, writes a per-run daemon artifact manifest, nudges agents when on-disk source drifts from the running runtime, and tunes Codex summarize guidance after framing it as an investment ratio.
Read release log → - Kernel v0.14.2 · TUI/Portal v0.9.6 2026-06-24
LingTai release day: multi-account Codex, trend reports, and a calmer cockpit
A paired kernel and TUI/Portal release centered on running several Codex identities side by side and seeing how a fleet actually spends its calls: per-agent and per-preset Codex OAuth token files, a multi-OAuth setup UI, swiss-knife tool-call and API-call trend reports, a scrollable `/doctor`, bilingual loading strings, a lower default context cap, cache-miss detail in the kanban call view, and the daemon and bash fixes that keep long-running work honest.
Read release log → - Kernel v0.14.1 2026-06-23
LingTai kernel v0.14.1: Codex Responses state, cache discipline, and honest identity
A Codex-focused runtime release: persistent Responses WebSocket continuation, safer tool-output baselines, explicit fresh-epoch reset, concise summarize/cache guidance, honest LingTai request identity, and refreshed kernel ANATOMY citations.
Read release log → - Kernel v0.14.0 · TUI/Portal v0.9.5 2026-06-22
LingTai release day: honest backends, measured latency, and a faster cockpit
A paired kernel and TUI/Portal release window centered on truth and measurement: the kernel sends an honest metadata envelope and the operator's own ChatGPT account header to Codex, records end-to-end LLM latency, surfaces long tool results in agent meta, and recovers from wedged turns; the TUI starts faster, renders fuller notification snapshots, and streams the token ledger.
Read release log → - Kernel v0.13.0 · TUI/Portal v0.9.3 2026-06-20
LingTai release day: context hygiene, notification history, and honest backends
A paired kernel and TUI/Portal release that makes long-running agents quieter, more inspectable, and easier to recover: progressive disclosure for large tool results, a dedicated notification tool, mandatory session journals before deliberate molts, persisted notification-block history, Codex header fixes, and operator views with timezone and daemon token context.
Read release log → - v0.12.4 2026-06-20
LingTai kernel v0.12.4
A tightly scoped kernel release after v0.12.3 focused entirely on Codex cache identity. Root agents now derive a stable 8-character sha256-prefix hash from the resolved agent/init.json path and present it byte-identically as the Codex session-id, thread-id, and prompt_cache_key, so repeated turns from the same agent look like one continuous conversation to the provider. LingTai-backend daemon runs stop reusing one parent/global identity: each run gets its own daemon-scoped LLMService and a Codex anchor derived from its resolved per-run daemon.json path, and preset daemons keep a dedicated service whose Codex presets also receive daemon anchors. The net effect is a practical improvement in Codex prompt-cache hit opportunity without crossing isolation boundaries between independent runs.
Read release log → - v0.12.3 2026-06-14
LingTai kernel v0.12.3
A focused kernel release after v0.12.2: daemon tasks can now carry explicit one-run MCP registrations, /daemon list gains historical run records with search/filter metadata and lazy data-version migration, and the agent procedures now teach a pad-first daemon workflow that protects the parent context while sending noisy work to disposable workers.
Read release log → - v0.9.1 / v0.12.2 2026-06-13
LingTai TUI/Portal v0.9.1 + Kernel v0.12.2
A focused patch after the v0.9.0 / v0.12.0 release train: coding CLIs move out of Swiss Knife into first-class bash-manual harnesses; daemon execution gains MiMo Code, Qwen Code, and Oh-My-Pi backends; the Zhipu picker exposes GLM-5.2; and new agents default to a 300k context window. The release keeps the previous release-log boundary clear: ToolCallGuard, structured tool-error recovery, Ctrl+O cockpit work, /setup keep-current behavior, and Telegram reconnect recovery remain same-window context already covered by 20260613-1, not newly claimed post-tag work here.
Read release log → - v0.9.0 / v0.12.0 2026-06-13
LingTai TUI/Portal v0.9.0 + Kernel v0.12.0
This release log is written from the previous published release log to the current v0.9.0 / v0.12.0 release, not merely from the immediately previous patch tag. The audited window spans LingTai TUI/Portal, lingtai-kernel, and the Telegram addon work that affected the same operating surface: 129 commits, 370 changed files, +31,717 / -2,047 lines, 109 PRs updated in the window (81 merged, 20 closed unmerged), and 48 issues (41 closed). The main theme is operational quality under real load: the TUI becomes a layered cockpit for long agent runs; setup, presets, manifests, and first-run paths become harder to misread; the kernel turns tool execution into a governed and observable surface; daemon/idle-care practices become explicit operations; knowledge, skills, tutorials, and research workflows become more teachable; MCP/chat/mail integrations become more reliable; and release hygiene records the validation, packaging, Homebrew, PyPI, contributors, and lessons for the next release.
Read release log → - v0.8.15 / v0.11.3 2026-06-09
LingTai TUI/Portal v0.8.15 + Kernel v0.11.3
This release pairs TUI/Portal v0.8.15 with Kernel v0.11.3 and centers on operating the runtime: more visibility and control from the TUI (a skill-backed, multilingual `/help` command, `/notification` for read-only notification inspection, a daemon browser, daemon counts in kanban, tool-call display controls, adaptive mail input, and `/clear` routed through kernel context clear), and a safer agent runtime under real workloads (the nested secondary tool-call channel is removed and residual stale `secondary` args are ignored/logged, notification sync stops leaking diary-looking synthesized text into visible conversation history, top-level tool calls gain lifecycle tracing plus API-call IDs that group tool batches, with enriched error results, plus a strict relay-compatible avatar schema, repeated-tool-error recovery, OpenAI Responses compact-threshold handling, a stale-notification version guard, refresh recovery from stale processes, MCP stale-stdio restart, and daemon limit / dead-parent cleanup). Knowledge and skill surfaces become more teachable — molt history knowledge entries, a strengthened pre-molt checklist, textbook distillation, a read-only preset-health subskill, optional skill repository remotes, a LICC chat tutorial, and Graphify / cyclic-manifold docs — and research and setup paths get more practical with an authorized-publisher PDF tier, in-house publisher extraction, a Zotero institutional handoff, an academic evidence gate, an install.sh writable-bin fallback, an AED max-attempts setting, and a Codex service-tier preset. Candidate window: 189 files changed, +15,541 / -3,428 lines across 44 first-parent merged PRs (25 TUI/Portal + 19 kernel), plus small direct/doc commits and star-count chores. Final release validation is still to be completed before tag, PyPI, and Homebrew publication.
Read release log → - v0.8.14 / v0.11.2 2026-06-02
LingTai TUI/Portal v0.8.14 + Kernel v0.11.2
This paired release cleans up the skill and manual surfaces that agents read every day. TUI/Portal v0.8.14 continues the move from large top-level utility documents toward router + reference skill layouts, consolidates research/media helpers under Swiss Knife, and tightens setup/preset reliability. Kernel v0.11.2 mirrors that work inside intrinsic manuals, splits the consequential-molt scaffold into a psyche-manual asset, adds Cursor daemon/backend and validation fixes, and includes a late rescue fix for Codex/gpt-5.5 agents by scrubbing rejected Responses tool schemas and removing the duplicate email scheduler in favor of host cron. Release window: 49 commits, 31 merged PRs, 172 files changed, +10,990 / -5,476 lines.
Read release log → - v0.8.13 / v0.11.1 2026-05-31
LingTai TUI/Portal v0.8.13 + Kernel v0.11.1
A release window about making long-running LingTai work easier to resume and easier to inspect. Molt continuation now relies on explicit post-molt notifications rather than fragile next-action guessing, while local traces gain a rebuildable SQLite index and optional historical backfill. Soul flow also becomes more exploratory and voice-aware. The same window brings first-party WhatsApp MCP support, a kernel-owned intrinsic doctor, progressive-disclosure guidance for resident prompts, and product/license documentation polish.
Read release log → - v0.8.12 / v0.11.0 2026-05-29
LingTai TUI/Portal v0.8.12 + Kernel v0.11.0
A coordinated release window for the Go TUI/Portal and the Python/Rust kernel. The TUI reaches v0.8.12 with safer first-run recovery, safer Homebrew upgrades, clearer MCP controls, Rust sidecar diagnostics, and dev-runtime protection; the kernel reaches v0.11.0 to mark the Rust-backed rewrite line and ships on PyPI after full release validation.
Read release log → - v0.10.10 2026-05-26
LingTai Kernel v0.10.10
A kernel release focused on runtime reliability, daemon extensibility, and MCP/addon communication polish. It adds an OpenCode daemon backend, improves configurable/asynchronous CLI daemon behavior, fixes several long-session recovery and wake-up issues, and improves cross-platform file reads, intrinsic tool manuals, and external communication addons.
Read release log →