Skip to content

Mingbird v3.1.6

Choose a tag to compare

@Gustor-Wang Gustor-Wang released this 05 Nov 05:49
· 20 commits to main since this release

🇨🇳 中文

鸣鸟 v3.1.6

生态兼容 + 还债版。压缩窗口护栏、MCP 常驻 broker、工具参数归一化、AGENTS.md/SKILL.md 生态兼容。桌面 UI 不变。672 项测试全绿(642 + 30 新增)。

  • 压缩请求不再超自己的窗:L2/L3 摘要载荷总量封顶(估算 token ≤ 窗口×0.75),截尾保近——最近的决策与产物名保留;长会话的摘要请求此前会被服务端拒绝且完全不可见(611 格实验日志实锤 8~13K token 超窗请求)。
  • 压缩失败降级而非装作压过:摘要调用失败时退回 L1 机械截断,上下文必然实际下降、运行可继续,不再"压完还是原样"。
  • MCP 常驻 broker:stdio 服务器每任务 spawn+握手一次(此前每调用一次,每次 1.7~5.7s);tools/list 会话内缓存,manifest 刷新复用活会话;任何失败自动回退旧单次路径,AGENT_MCP_BROKER=0 可关闭。
  • 工具参数归一化:坏 JSON 修复(剥 think 泄漏/提取平衡块/修尾逗号与 Python 字面量与单引号),修不好就拒绝执行并回喂原因——告别"静默空参硬跑报 KeyError"。
  • 生态兼容:工作目录 AGENTS.md/CLAUDE.md 项目指令注入;agentskills.io 的 SKILL.md 文件夹技能直接可读(渐进披露:列表只露 name+description);winget 清单工具就位(提交 winget-pkgs 人工执行)。

安装(Install)

  • Windows:下载 Mingbird-v3.1.6-CN-Setup.exe(中文)或 Mingbird-v3.1.6-EN-Setup.exe(英文),双击安装。安装包未签名,Windows 可能弹出 SmartScreen——点"更多信息"→"仍要运行"。
  • Linux(实验性):下载 mingbird-v3.1.6-linux-x64.tar.gz,解压后 sh install-unix.sh,启动 ~/.local/share/Mingbird/LocalAgent。
  • macOS(实验性):下载 mingbird-v3.1.6-macos-arm64.tar.gz,同上。
  • 需本机已装 Ollama 或任何 OpenAI 兼容本地服务(LM Studio、llama.app、llama-server、vLLM 均可)。

完整变更:CHANGELOG_zh.md


🇬🇧 English

Mingbird v3.1.6

Ecosystem-compatibility + debt-repayment release: compaction window guard, MCP resident broker, tool-argument normalization, AGENTS.md/SKILL.md ecosystem compat. Desktop UI unchanged. 672 tests green (642 + 30 new).

  • Compaction can no longer outgrow its own window: the L2/L3 summary payload is now capped at an estimated ≤75% of the window with tail-keeping — recent decisions and artifact names survive. Previously a long session's own compaction request was rejected by the server and the failure was invisible (8–13K-token requests in campaign logs).
  • Compaction failure degrades instead of pretending: on summary failure the harness falls back to L1 mechanical truncation — context always actually shrinks and the run continues.
  • MCP resident broker: stdio servers spawn and handshake once per task instead of once per call (1.7–5.7s saved each); tools/list cached in-session; manifest refreshes reuse live sessions; any failure falls back to the per-call path; AGENT_MCP_BROKER=0 disables.
  • Tool-argument normalization: malformed JSON is repaired (think-leak stripping, balanced-block extraction, trailing commas / Python literals / single quotes); unrepairable calls are rejected with the reason fed back — no more silent empty-args runs ending in confusing KeyErrors.
  • Ecosystem compat: AGENTS.md/CLAUDE.md project instructions injected; agentskills.io SKILL.md folder skills readable directly (progressive disclosure — the list shows name+description only); winget manifest tooling included (winget-pkgs submission is manual).

Install

  • Windows: download Mingbird-v3.1.6-CN-Setup.exe (Chinese) or Mingbird-v3.1.6-EN-Setup.exe (English) and run it. The installer is unsigned — Windows SmartScreen may warn; choose "More info" → "Run anyway".
  • Linux (experimental): download mingbird-v3.1.6-linux-x64.tar.gz, extract, sh install-unix.sh, launch ~/.local/share/Mingbird/LocalAgent.
  • macOS (experimental): download mingbird-v3.1.6-macos-arm64.tar.gz, same steps.
  • Requires Ollama or any OpenAI-compatible local server (LM Studio, llama.app, llama-server, vLLM all work) installed on that machine.

Full changelog: CHANGELOG.md