teamai-cli 0.20.0-beta.7 → 0.21.0-beta.0

package/README.md

@@ -16,7 +16,7 @@
 
 Make every AI coding agent work by the same harness.
 
-Git-native management of skills, rules, and docs across Claude Code / Codex / CodeBuddy / WorkBuddy and more.
+Git-native management of skills, rules, and docs across Claude Code / Codex / CodeBuddy / WorkBuddy / OpenCode and more.
 
 For you or your whole team.
 
@@ -30,10 +30,12 @@ npm install -g teamai-cli
 
 ### Team admin / solo user
 
-Create a shared-experience repo on your git host (GitHub, TGit, or CNB), **grant write access to team members**, then have them run `teamai init https://github.com/yourorg/yourrepo`.
+Create a shared-experience repo on your git host (GitHub, TGit, CNB, or a self-hosted Git service), **grant write access to team members**, then have them run `teamai init https://github.com/yourorg/yourrepo`.
 
 > Solo use needs no separate repo setup: `teamai init` checks the target repo and creates it automatically if it doesn't exist.
 
+> Self-hosted GitLab/Gitea-style services use your HTTPS credential helper or SSH key. Create the repository first; merge requests are created manually for now. See [Git Provider docs](docs/providers.md).
+
 > **No team repo yet?** Start from a template pre-loaded with production-ready skills, rules, and review agents. Browse the [teamai-hub](https://github.com/teamai-hub) org, click **Use this template**, then `teamai init` against your new repo.
 
 ### Team members

package/README.zh-CN.md

@@ -16,7 +16,7 @@
 
 面向 AI 智能体的团队 Harness 分发工具。
 
-通过 Git 统一管理 skills、rules、docs，驾驭 Claude Code / Codex / CodeBuddy / WorkBuddy 等多种 AI 工具。
+通过 Git 统一管理 skills、rules、docs，驾驭 Claude Code / Codex / CodeBuddy / WorkBuddy / OpenCode 等多种 AI 工具。
 
 一个人也能用，团队用更强。
 
@@ -30,10 +30,12 @@ npm install -g teamai-cli
 
 ### 团队管理员 / 个人使用者
 
-在 Git 托管平台（GitHub、TGit 或 CNB）创建共享经验仓库，**授予团队成员写权限**，然后让他们运行 `teamai init https://github.com/yourorg/yourrepo`。
+在 Git 托管平台（GitHub、TGit、CNB 或自建 Git 服务）创建共享经验仓库，**授予团队成员写权限**，然后让他们运行 `teamai init https://github.com/yourorg/yourrepo`。
 
 > 个人使用无需单独建仓：`teamai init` 会检查目标仓库，不存在时自动创建。
 
+> 自建 GitLab/Gitea 等服务支持 HTTPS Credential Helper 或 SSH Key；需预先创建仓库，MR 暂时手动创建。详见 [Git Provider 说明](docs/providers.md)。
+
 > **还没有团队仓库？** 可以从内置了成套 skills、rules、review agents 的模板起步。浏览 [teamai-hub](https://github.com/teamai-hub) org，点 **Use this template** 生成自己的仓库，再对它执行 `teamai init`。
 
 ### 团队成员

package/agents/teamai-recall.md

@@ -124,7 +124,29 @@ none exists, silently skip.
 ### Step 2 — Extract keywords from the task description
 
 Pick 3–6 high-signal keywords from the user query. Strip filler words
-("the", "how", "please"). Mix English and Chinese terms when both appear.
+("the", "how", "please").
+
+**Always search bilingually (cross-language recall).** The knowledge base
+mixes Chinese and English: a learning may be titled in English while your
+task is described in Chinese, or vice versa. The search is purely lexical
+(BM25 token matching), so a Chinese-only query never matches an
+English-only document. To bridge this, expand every domain term into
+**both languages** and include both in your keyword set:
+
+- Chinese task → add the English term for each concept
+  (e.g. "重启推理集群" → also add `restart`, `inference`, `cluster`).
+- English task → add the Chinese term for each concept
+  (e.g. "restart inference cluster" → also add `重启`, `推理`, `集群`).
+- Only translate terms that genuinely differ across languages. Keep any
+  **language-neutral token** as-is — anything that is written the same way
+  in both languages, or has no meaningful translation: proper nouns, brand
+  and product names, acronyms and initialisms, code identifiers and symbols,
+  error codes, version strings, file paths, CLI commands, and API names
+  (e.g. `AppID`, `Kubernetes`, `GPU`, `RuntimeError`, `v2.1.3`, `oauth`). Translating
+  these only hurts matching.
+
+Only translate the meaningful domain terms; do not pad the query with
+low-signal words. Aim for the same 3–6 concepts, each in both languages.
 
 **Lead with the terms that pin down *this* task**: proper nouns, customer or
 product names, service IDs, error codes, versions, symbol names
@@ -208,9 +230,10 @@ Use judgement rather than counting, in both directions:
   discriminating term as a lead to verify by opening the file, not as a
   conclusion.
 - **Fewer than 5 results does not mean the corpus is thin.** Results are
-  deduplicated when title, date, author and content all match, so a learning
-  shared twice appears once. Body-
-  only matches are also dropped. Both are intentional.
+  deduplicated when type, domain, title, date, author, tags and content all
+  match, so a learning shared twice appears once — including when the two copies
+  have collected different vote counts. Body-only matches are also dropped. Both
+  are intentional.
 
 For each hit you keep, read the source file directly (use `Read`) and
 condense each into **one or two sentences**.