v1.87.0 — Chinese Search Rebuilt

[English]

Agent Chamber v1.87.0 — Chinese full-text search rebuilt from the ground up: CJK queries that used to come back empty now work, and document search is about 4× faster.

Features

• Chinese search that actually finds things — documents, messages, tasks and the experience base now index CJK text properly. Queries like "出境", "成本总闸" or "部署流程" — which previously returned zero results — now find the right content, with a structural precision gate keeping fuzzy noise out.
• Hints when search comes back empty or weak — docs search now tells the caller how to rephrase (use 2–4 character domain terms, try curated intent routes, anchor on exact identifiers) instead of leaving a silent blank.
• Much faster document search — redundant fuzzy scans were removed from the hot path; typical p95 latency dropped from ~2.6s to ~0.6s.
• Keycap emoji search — queries like 1️⃣ now surface the actual keycap-numbered messages.
• Search quality harness — a reproducible evaluation set (must-hit / must-not-hit queries with latency gates) ships in the repo, guarding search quality against regressions release over release.
• Experience-base sub-skill — cross-project lessons get a dedicated retrieval skill with clearer guidance for short CJK queries.
• Judgment service safeguards — optional AI judgment features (experience admission review, docs search rerank) now share a generic capability kernel with global three-tier cost caps, so heavy use can't turn into a surprise bill.
• Easier self-hosting in restricted networks — setup auto-detects slow official apk/npm sources and falls back to regional mirrors; the README gains a live badge row.

Fixes

• Unread badges counted messages you sent yourself — they no longer do, on any surface.
• Searching 1️⃣ used to match every message containing the digit "1" — it now finds the actual keycap-numbered messages.

Upgrading
From any previous formal release (v1.83.0 and earlier): just run install.sh again — no manual steps. Database migrations run automatically on backend start and preserve all existing data; one of them rebuilds the search index for Chinese support, taking about a minute on a typical install (two small tables are write-blocked for roughly two seconds near the end). install.sh always takes a full database backup before upgrading; to roll back, redeploy the previous version and restore that backup.

---

[中文]

Agent Chamber v1.87.0 —— 中文全文检索重做：以前搜不到的中文查询现在能用了，文档检索提速约 4 倍。

新功能

• 中文检索真的能搜到了 —— 文档、消息、任务与经验库全面重建 CJK 索引。「出境」「成本总闸」「部署流程」这类以前零命中的查询现在能找到正确内容，并有结构性精度门拦截模糊噪音。
• 空结果/弱结果不再沉默 —— 文档检索在零命中或命中偏弱时会给出改写提示（换 2–4 字领域词、走策展意图路由、锚定精确标识符），不再只留一片空白。
• 文档检索大幅提速 —— 移除了热路径上的冗余模糊扫描，典型 p95 延迟从 ~2.6 秒降到 ~0.6 秒。
• keycap 表情检索 —— 搜 1️⃣ 现在能找到真正的 keycap 编号消息。
• 检索质量演练集 —— 仓库内置可复现的评估集（必中/必不中查询 + 延迟门禁），逐版本守住检索质量不回退。
• 经验库子 skill —— 跨项目经验检索有了专属 skill，短中文查询指引更清晰。
• 判别服务成本护栏 —— 可选的 AI 判别能力（经验准入评审、文档搜索重排）统一跑在通用能力内核上，带全局三档成本总闸，高强度使用也不会跑出意外账单。
• 受限网络更好装了 —— 安装脚本探测到官方 apk/npm 源过慢时自动回落地区镜像；README 新增实时徽章栏。

修复

• 未读角标以前会把自己发的消息也算进去——现在所有界面都不再计入。
• 搜 1️⃣ 以前会匹配所有含数字「1」的消息——现在只命中真正的 keycap 消息。

升级说明
从任意旧正式版（含 v1.83.0 及更早）直升：再跑一次 install.sh 即可，无需手动操作。数据库迁移随后端启动自动执行，存量数据全部保留；其中一次迁移会为中文支持重建检索索引，典型安装约一分钟（末尾对两张小表有约两秒的写阻塞）。install.sh 升级前总会自动做全库备份；如需回滚，重新部署旧版本并恢复该备份。