Kernel v1.0.0 · TUI/Portal v1.0.0

New features and improvements

1. 01

   Daemon supervision and honest terminal notifications

   Daemon work gets a detached supervisor lifecycle, exactly-once terminal notifications, and truthful failure receipts instead of silent or partial completion.

   • The detached supervisor publishes done/failed/cancelled/timeout terminal notifications exactly once, with the result path attached for follow-up.
   • Partial-finish receipts, per-batch timeouts, and provider_defaults survive detached-child reconstruction, so fleets report effort truthfully.
   • Daemon tasks can carry Agent Plugins and per-task plugin paths, and task cards add proactive nudges for expired watches and fleets.

   Why it matters:Long-running agent work becomes observable and attributable: when a fleet finishes, fails, or stalls, operators get one clear terminal signal and a path to the evidence.

2. 02

   Dependable channels: Telegram, IMAP, mail, Feishu, WeChat, WhatsApp

   Channel adapters converge on one protocol while each surface gains real reliability fixes: native rich rendering, correct account routing, and replay-safe webhooks.

   • Telegram gains native Rich Messages rendering, configurable Task Card locale, three-boundary read/seen/replied reactions, and device-footer + latency visibility.
   • IMAP responses preserve the requested account; mail stops dropping unreadable or corrupt messages, prunes unbounded seen-state, and guards against double-start listeners.
   • Feishu gets the full channel/task-card integration; WeChat and WhatsApp add personal-account bridging, webhook replay dedup, and credential redaction.

   Why it matters:Messages arrive, read receipts are honest, and credentials stay out of logs — the transport surfaces operators actually depend on stop failing silently.

3. 03

   A hardened shell contract for Windows, macOS, and POSIX

   The shell adapter learns real platform semantics: Windows PowerShell discovery and tree-kill, macOS detection, benign non-zero exits, and a hard timeout ceiling.

   • Windows: pwsh discovery from well-known paths, UTF-8 console output with OEM fallback, Job Object tree-kill with identity re-check, quote-aware metachar safety, and trusted cmd.exe shims.
   • POSIX/macOS: ShellKind classifier drives spawn args and model-facing descriptions, macOS login-shell detection, and process-group kill with timeout fallbacks.
   • Everywhere: benign non-zero exits (grep/rg no-match), ANSI/control sanitization, ASCII-only cmdline + UTF-8 stdin bootstrap, and LINGTAI_TOOL_TIMEOUT_MAX_SECONDS as a hard ceiling.

   Why it matters:Shell automation is predictable across every host the kernel runs on, with failures that are bounded in time and safe to reason about.

4. 04

   LLM, Agent Plugins, and the TUI/Portal 1.0

   The runtime grows a seven-tier thinking model and OpenAI Responses reasoning semantics; plugins become a first-class capability; the TUI/Portal reaches 1.0 with launch reliability and a GitHub-only kernel check.

   • Seven-tier thinking levels, Responses default allow, generic reasoning_content round-trip fallback, and provider-aware reasoning-effort contracts for Claude Code, Kimi, DeepSeek, and GLM.
   • Agent Plugins become a real capability: plugin.json manifests mount skills and MCP servers for the main agent and daemon tasks, with resident prompt sections.
   • TUI/Portal 1.0: launch heartbeat and agent.log fd/atomic-write fixes, preset expansion, doctor D1–D5 startup checks, recipe picker simplification, install.sh fixes, and a kernel update check that reads GitHub releases only — ending the 0.19.5→0.19.3 PyPI downgrade.

   Why it matters:The system gets smarter about reasoning, more extensible through plugins, and dramatically more reliable to install and keep up to date.