v0.10.10

New features and improvements

1. 01

   Daemon backend expansion: OpenCode is now a runnable backend

   This release adds backend="opencode", allowing LingTai daemons to run work through OpenCode as an ephemeral parallel execution backend.

   • daemon(action="emanate", backend="opencode", ...) is now part of the daemon backend schema.
   • The backend invokes opencode run --format json and parses JSON events for output, errors, and session ids.
   • When a resumable OpenCode session id is available, daemon(action="ask") resumes with opencode run --session <id>.
   • Session-id extraction is defensive: ordinary event/message ids are no longer mistaken for OpenCode session ids.

   Why it matters:LingTai can integrate multi-provider coding agents more easily instead of tying all external execution to a single CLI.

2. 02

   Configurable CLI daemons and smoother follow-up interaction

   CLI daemon backends now support safe backend_options passthrough and avoid blocking the parent agent during follow-up asks.

   • backend_options can be passed to CLI backends such as Claude Code, Codex, and OpenCode for model, agent, permission, or backend-specific flags.
   • Options are converted safely: booleans become flags, strings/numbers become --flag value, lists expand, and unsafe keys or nested objects are rejected.
   • CLI daemon ask now runs in the background instead of blocking the parent agent.
   • The parent can keep responding to humans or handling tool results while an external CLI daemon continues running.

   Why it matters:Daemons behave more like a real asynchronous execution layer — configurable without freezing the primary agent.

3. 03

   Agent recovery and long-session stability fixes

   This release fixes several failure modes that could make agents go silent, fail recovery, or repeatedly wake under complex sessions.

   • Fixed poll-backoff correlation using the wrong tool result id, preventing agents from appearing idle after repeated inbound reads.
   • Fixed OpenAI/Codex Responses recovery when history contained a function_call without a matching function_call_output.
   • Fixed a livelock where MCP notifications could repeatedly fail to wake an ASLEEP agent.
   • Fixed refresh-time env-file reload behavior so refreshed agents pick up the intended configuration.

   Why it matters:These are runtime reliability foundations; the longer the task and the richer the tool chain, the more visible the improvement.

4. 04

   File reading, cross-platform behavior, and intrinsic tool manuals

   Text reading now pins UTF-8, and built-in file tools now have first-class manual guidance.

   • Fixed UnicodeDecodeError risks on Windows or non-UTF-8 locale hosts by reading text as UTF-8.
   • Added an intrinsic file manual skill for core file tools such as read, write, edit, glob, and grep.
   • This brings intrinsic tools closer to the same “read the manual before operating” discipline used for MCP/addon tools.

   Why it matters:Fewer cross-platform read/startup crashes and less chance of agents misusing core file operations.

5. 05

   MCP/addon communication polish

   In the same release window, IMAP, Feishu, WeChat, Telegram and related addons received practical communication improvements.

   • IMAP: stale socket handling no longer requires restarting the MCP; tool calls can use bounded reconnect.
   • IMAP: bare search tokens now fall back to TEXT search more compatibly for services like Gmail.
   • Feishu: fixed orphan typing indicators after p2p send failures, deduplicated replayed incoming events, and aligned conversation context with Telegram.
   • WeChat: aligned conversation context with Telegram so agents see fuller conversation history.
   • Telegram / Feishu: added Cleanup / Footprint audit guidance — read-only reports first, destructive cleanup only with user consent.

   Why it matters:External communication addons become more reliable long-lived entry points rather than fragile connections needing manual repair.

6. 06

   TUI state visibility and community contributions

   This release log also includes same-window TUI and community-facing contributions, especially TZZheng’s network activity badge.

   • The project-level network activity badge can show active / daemon-active / idle / asleep / suspended states.
   • It addresses the gap where a single agent mind state is not enough to understand whether the whole project network is busy.
   • The same window also includes local mailbox time display, API-key edit locking, and secondary human update guidance refinements.

   Why it matters:LingTai’s collaborative network state is more visible, and community contributions are represented in the official release narrative.