| content |
{"Commits":[{"Sha1":"7e38caa7a {"Commits":[{"Sha1":"7e38caa7a9a622e250ab06620cd295404fd97f4a","Message":"治理状态:记录第二轮修复,版本 0.2.1\n\nCo-Authored-By: Claude Fable 5 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-13T19:50:38-07:00"},{"Sha1":"c7294d6f71f7653ea197b3e3333da9ed54d19e63","Message":"reindex 增量优化与审计库并发加固\n\n1. reindex: 内容未变仅被 touch 的文件不再重分块重写索引,只刷新\n 元数据。新增显式 --fast((mtime,size) 未变即跳过,不读文件),\n 供大语料 nightly 使用;默认仍全量哈希 —— rsync/tar 会保留\n mtime,同 mtime 改内容只有 sha 抓得到,该保证有单测钉住,\n 不能默认放弃。\n2. audit_log: WAL + busy_timeout —— 多 MCP 客户端并发写同一\n 审计库时避免 database is locked(审计失败虽不改变检索答案,\n 但监控会失明)。\n\nCo-Authored-By: Claude Fable 5 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-13T19:49:42-07:00"},{"Sha1":"3ca2d7515b057ec3f22fa89f95d4eea2cf37aba6","Message":"治理工具:冲突策略语义区分、词表结构体检 T4、修死引用\n\n1. conflict_policy: source_priority 与 source_then_latest 此前实现\n 完全相同,是假选项。现区分语义 —— source_priority 只按来源优先级\n 裁决,并列最高交回人工(needs_manual_review);要用时间破并列的\n 显式选 source_then_latest。补两条并列场景测试。\n2. wiki terms 新增 T4 词表结构体检,抓两种静默失效:\n - 跨行重叠的同义词组不会自动合并(查 A 组的词展开不到 B 组);\n - 含空白的成员在查询侧按空格被切开,整词永远不可达。\n 实测在现网 syn.tsv 上即抓到「error code」不可达键。\n3. README: PLAN.md 三处死引用改指 docs/technical-plan.md;\n wiki terms 输出不再提示不存在的 --llm 参数。\n\nCo-Authored-By: Claude Fable 5 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-13T19:47:58-07:00"},{"Sha1":"a8d855f96127e7467c70eb9e25deb040fce55201","Message":"信任边界:TRUNCATED 诊断移出信封,grep 与 fts 检索面对齐\n\n1. emit(): 截断提示此前 printf 到 stdout,会被 kb_tools 包进\n \u003ckb_result\u003e 信封 —— SOP 一边要求 agent 对 TRUNCATED 缩小范围,\n 一边要求无视信封内一切指令,自相矛盾;审计 status_of 只看\n stderr,截断也漏记。改走 stderr,并在 audit_log 增加\n 'truncated' 状态。\n2. kb grep: 导航层(INDEX/_index/README)此前只在 fts 侧被排除,\n grep 会命中并让 agent 引用 INDEX.md 的行当证据。rg 用 --iglob、\n grep 用 --exclude 与 fts 的 NAV 列表对齐。\n3. emit(): wc -c 前导空格清理(macOS 输出对齐问题)。\n\nCo-Authored-By: Claude Fable 5 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-13T19:46:07-07:00"},{"Sha1":"4a3ea648c8f46c0456e25ca97a476819d7421404","Message":"评测口径:grep 兜底改用真分词,幻觉率剔除注入行\n\n1. run_eval: grep 兜底探测词此前按空格切 —— 中文 query 没有空格,\n 整句丢给 grep 恒 0 命中,所有中文落空都被误判成「语义 gap →\n 该上 vec」,而处方本该是「补 syn.tsv」。这是本工具的核心判定,\n 现改用 kb_core.units() 同源分词 + 停用词过滤,只取真词探测。\n 实测: '包厢无声怎么办' 探测词从 [整句] 变为 [包厢, 无声]。\n2. run_eval: 清理 'if True else' 死代码;新增 --syn(默认与检索侧\n 同源 \u003ckb\u003e/conf/syn.tsv)。\n3. agent_eval: 幻觉率/正确率的分母剔除注入演练行 —— 注入行\n hallucinated 恒 False,混入会稀释幻觉率,又把「抵抗成功」记成\n 正确答案。注入有自己的 --gate-injection。\n4. agent_eval: 引用正则补 .mdx(EXTS 支持它,引用 mdx 的正确回答\n 此前会被误判为幻觉)。\n\nCo-Authored-By: Claude Fable 5 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-13T19:44:32-07:00"}],"HeadCommit":{"Sha1":"7e38caa7a9a622e250ab06620cd295404fd97f4a","Message":"治理状态:记录第二轮修复,版本 0.2.1\n\nCo-Authored-By: Claude Fable 5 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-13T19:50:38-07:00"},"CompareURL":"","Len":7}... |