Skip to content

Mingbird v3.1.4

Choose a tag to compare

@Gustor-Wang Gustor-Wang released this 22 Oct 13:54
· 23 commits to main since this release

鸣鸟 Mingbird v3.1.4

🇨🇳 中文在上 / English below.

会话卫生版——源自一次真实 6 分钟会话的事后解剖,全部按"类"修不按"例"修。桌面 UI 不变。642 项测试全绿(631 + 11 新增)。

🔧 修复

  • 追问消息不再从转录消失:续跑游标此前取 len(messages),每次 --append 的新用户消息被静默吞掉(事故会话上下文 33 条 vs 转录 30 条)。游标现在数转录文件实际行数;压缩后续跑的完整历史也被保住。
  • 交付自查对照正确的题面:回注"本次交付所响应的指令"(最近一条真实用户输入),不再把会话第一句当题面;⏱/提示等全部注入消息补进前缀表,不会再被误当原文。
  • 假完成门两则豁免:问答会话豁免;实质性 summary 算交付(讲解型 finish 不再被逼去建文件)。基准注记:主臂数据在旧门禁语义下产生。
  • 拒绝不再携带成功语义:六道 finish 门拒绝时一律回填中性结果(旧版回填含 [TASK_COMPLETE] 的结果,给模型矛盾信号、中途标 done)。
  • 测试门豁免问答:工作目录恰好有 test_*.py 时,聊天收尾不再先跑测试套件。
  • 工具话术如实:不在装配列表的工具(幻觉名/未装配)得到"不可用"话术,不再误标"连续重复调用保护"。
  • MAX ITERATIONS 只在 200 轮真耗尽时打印。

English

Conversation-hygiene release from a real 6-minute session post-mortem; every fix is class-level, not scenario-level. Desktop UI unchanged. 642 tests green (631 + 11 new).

🔧 Fixes

  • Follow-up messages no longer vanish from the transcript: the resume cursor now counts the transcript file's actual lines instead of len(messages), which silently swallowed every --append user message (incident: 33 context vs 30 transcript). Full history survives compacted resumes.
  • Delivery self-check reads the right instruction: the last real user input is re-injected — not the conversation's first line; budget nudges and teaching notes joined the harness-prefix list so they can never pose as the original task.
  • Fake-finish gate exemptions: QA sessions are exempt, and a substantive summary counts as delivery (explanatory finishes are no longer pushed into creating files). Benchmark note: main arms ran under the old gate semantics.
  • Rejections carry no success semantics: all six finish gates append neutral results instead of the [TASK_COMPLETE]-bearing one (which fed the model contradictory signals and marked the meta done mid-run).
  • Test-run gate exempts QA: a chat ending in a workdir that happens to contain test_*.py no longer runs the suite first.
  • Truthful tool wording: a tool outside the loaded list (hallucinated, or real but unloaded) reads "unavailable", not "disabled by repeat-call protection".
  • MAX ITERATIONS only prints on true exhaustion.

Install / 安装

Windows:下载 Mingbird-v3.1.4-CN-Setup.exe(中文)或 Mingbird-v3.1.4-EN-Setup.exe(English),双击安装。需要 Ollama 或任何 OpenAI 兼容本地服务;2~9B 小模型核显即可跑,全程数据不出本机。
Windows (EN): download Mingbird-v3.1.4-CN-Setup.exe (Chinese UI) or Mingbird-v3.1.4-EN-Setup.exe (English UI) and run the installer. Requires Ollama or any OpenAI-compatible local server; 2–9B models run on integrated graphics, and your data never leaves the machine.

macOS / Linux(实验性):下载 mingbird-v3.1.4-macos-arm64.tar.gz / mingbird-v3.1.4-macos-x64.tar.gz / mingbird-v3.1.4-linux-x64.tar.gz,解压后运行 LocalAgent。系统语言自动识别(中文系统直接显示中文;英文系统输出自动跟随英文。也可在可执行文件旁建 app_lang.txt 写 zh/en 强制)。
macOS / Linux (experimental, EN): download the tar.gz for your arch, extract, run LocalAgent. System language is auto-detected; you can also force it by placing an app_lang.txt (zh/en) next to the executable.

From source:pip install -r requirements.txt → python agent_gui.py(Python 3.12 推荐)。 / From source (EN): pip install -r requirements.txt → python agent_gui.py (Python 3.12 recommended).