v0.9.1 / v0.12.2

New features and improvements

1. 01

   Coding CLIs became first-class bash harnesses

   The kernel now owns reusable bash-manual harness pages for coding CLIs, while the TUI trims Swiss Knife back to small utilities and routes operator guidance to the new home.

   • Kernel #277 adds bash-manual subskills for Claude Code, OpenAI Codex, OpenCode, Cursor Agent, MiMo Code, Qwen Code, and Oh-My-Pi.
   • Docs-first candidate pages also reserve space for Gemini CLI, Aider, Goose, OpenHands, Crush, and Zed/ACP without pretending those integrations are release-ready.
   • TUI #335 removes the old coding-CLI guidance from Swiss Knife and replaces it with a bash-cli-harnesses redirect, keeping Swiss Knife focused on MiniMax, vision, listen, academic research, HTML reports, Xiaomi MiMo, and Zhipu coding-plan discovery.

   Why it matters:Operator guidance now lives next to the shell tool that actually runs long-lived CLI work, reducing duplicated docs and making future harnesses easier to add safely.

2. 02

   Daemon coding backend coverage expanded

   Kernel v0.12.2 adds three more provider-family daemon backends while keeping the higher-level daemon workflow unchanged for agents.

   • Kernel #275 adds MiMo Code as mimocode / mimo and Qwen Code as qwen-code / qwen through the OpenCode-family JSON runner.
   • Kernel #276 adds Oh-My-Pi as oh-my-pi / omp, giving agents another coding-agent route when that CLI is the right fit.
   • The daemon documentation and i18n strings were updated so the new backends appear in the same operational vocabulary as existing CLI backends.

   Why it matters:LingTai agents can route isolated coding work to more external agent families without changing the parent workflow or treating each CLI as a one-off exception.

3. 03

   GLM-5.2 and a larger default context window landed in the TUI

   The TUI side of this patch is small but user-visible: Zhipu users can pick GLM-5.2, and new/default project agents start with a larger context budget.

   • Commit 9e6ae53 adds GLM-5.2 to the Zhipu model picker ahead of the older GLM models and registers its vision capability.
   • TUI #339 raises the default agent context window to 300k for new/default project agents.
   • Both changes are small release-train polish: they make stronger models and larger work windows available without changing the release workflow.

   Why it matters:Model choice and context budget are day-to-day operator ergonomics. This patch makes the stronger preset and safer default capacity available before the next larger release train.

4. 04

   Same-window reliability work remains context, not double-counted change

   This page intentionally separates strict post-tag delta from the broader release window, so the small patch does not over-claim work already documented in 20260613-1.

   • ToolCallGuard (#270), idle-care/watchdog guidance (#271), and structured tool-error recovery metadata (#273) remain the kernel governance/recovery foundation covered by the previous release log.
   • The TUI Ctrl+O tool-noise ladder (#325, #330, #331, #332, #333, #334) and /setup keep-current fix (#327) shipped in v0.9.0 and remain context for the current operator-facing polish.
   • Telegram polling disconnect recovery (#35) and README product-positioning work (#219/#220) are referenced as release-window context rather than new v0.9.1/v0.12.2 delta.
   • PR #340 is still open and is not included in this release.

   Why it matters:Small patch notes are most useful when they are honest about scope. The context explains why this patch matters without blurring what actually changed after the last tags.