Repository navigation
Meta_Kim v3.9.0
Meta_Kim 3.9.0
English
Release date: 2026-11-04.
This release improves the runtime connection from user intent and product decisions to research, design context and actual task dispatch.
- Select the product solution before production. Options describe real user flows and tradeoffs separately from execution tools. Delivery with an option list requires a selected mechanism and reason; exploration can remain undecided, and clear bounded changes need no artificial alternatives.
- Research what changes the decision. Official documentation, user flows, source, cases and papers can resolve the question. Generic repository counts, star floors and paired-channel quotas are removed, while explicit user research constraints remain enforced.
- Pass the design that workers need. Related complete journeys, decisions, shared terms, state, interfaces and the selected solution enter the actual work order. Context, source bytes and current upstream artifacts are checked before invocation. Necessary dependency text is retained separately from a labelled preview.
- Choose depth per task and subtask. Design uncertainty, coupling, operational risk and verification needs drive the current route. Worker/file counts and internal provider scores alone do not create risk or a user decision; missing facts stay explicit, and permission rules still apply.
- Continue independent ready work. The existing DAG runs eligible tasks while missing-context work and its real dependents wait. Recovery retains completed results, the overall run remains partial, and observed calls remain distinct from whole-batch completion.
- Repair valid measured failures. Numerical target failures can use the existing focused revision path. Invalid samples, forged artifact evidence and stale task/version bindings remain blocked.
- Preserve history through reviewed migration. Legacy v1 research and governance facts remain read-only. New dispatch requires current reviewed v2 records and design context; old work orders need re-evaluation instead of silently continuing. Completed history is retained, and design inputs without new optional fields keep their digest. Injected callbacks still cannot certify native execution.
This release publishes Meta_Kim only. Unpublished prompt and method changes in the separate Kim_Service repository are not included.
Verification scope
The clean 3.9.0 candidate passed 55/55 targeted tests, and normal packaging produced 614 tracked files. The separate existing integration comparison remains 28/32: the same four first-assertion failures were reproduced against clean v3.8.0. Assertions after those failures were not reached; overlapping runs are not added together.
The full release gate did not pass: dependency Git acquisition failed in preflight, so none of its 17 standard stages ran. Isolated graph rebuilding lacked the Graphify Python module. A minimal offline installation of the same package failed with ENOTCACHED while resolving which; the CLI entry was not tested. This is not a standard-release-verified or live-certified claim. No comparative model-quality, model-latency or real-business improvement benchmark was performed, and comprehensive macOS, Linux and native-host acceptance remains unverified.
简体中文
发布日期：2026-11-04。
这一版改善从用户意图、产品选择到研究、设计上下文和实际派工的运行链路。
- 先选定产品解法，再进入制作。 方案比较真实用户流程与取舍，和执行工具分开。有方案列表的定稿必须记录所选机制和理由；探索可以未决，明确的小改动无需凑替代方案。
- 研究能改变决定的问题。 官方文档、用户流程、源码、案例和论文都可用于回答关键问题。取消通用仓库数、star 下限和双渠道检索配额，保留用户明确提出的研究条件。
- 执行者拿到所需设计。 完整相关旅程、决定、共享术语、状态、接口及选中方案进入实际工单。调用前核对上下文、源文件和当前上游产物；必要的依赖正文完整保留，导航预览另外标明。
- 整体与子任务分别判断深度。 当前设计未知、耦合、操作风险和验证条件决定路线。文件数、worker 数和内部能力评分不会单独制造风险或要求用户决定；缺少事实如实保留，权限规则继续独立生效。
- 独立就绪工作可以继续。 现有依赖图先执行合格任务，缺上下文的工作及其真实依赖等待。恢复保留完成结果，整批仍是部分完成；已观察到的调用与全批完成分别记录。
- 有效数值失败能够返修。 实测未达标可进入原有局部返修路径，缺样本、伪造产物证据以及错任务、旧版本仍受阻断。
- 复核后迁移，保留历史。 旧 v1 研究和治理事实保持只读，新派工需当前复核后的 v2 记录与设计上下文；旧工单须重新评估，不能静默继续。完成历史仍保留，没有新增可选字段的旧设计输入保持原摘要。注入回调仍不能证明原生执行。
本次仅发布 Meta_Kim，Kim_Service 独立仓库中未发布的提示词和方法修改不包含在此版本中。
验证范围
干净的 3.9.0 候选定向测试为 55/55 通过，正常打包包含 614 个受版本管理的文件。另一次旧集成对照仍为 28/32：同样四处首次断言失败已在干净的 v3.8.0 上复现，失败之后的断言没有运行到；重叠检查不重复累计。
完整发布检查未通过：前置步骤取得 Git 依赖失败，17 个标准阶段均未运行。隔离环境的图谱重建缺少 Graphify Python 模块。同一安装包的最小离线安装在解析 which 依赖时报告 ENOTCACHED，命令行入口未验。本说明不作 standard-release-verified 或 live-certified 声明；没有进行可比的模型质量、模型耗时或真实业务改进评测，macOS、Linux 和全部原生宿主也未全面验收。