feat(search): fs.search 文件名搜索优先使用 fd/rg,缺失回退 JS 遍历 (fixes #203) by yanzhaohui1999 · Pull Request #303 · omdsh-dev/DSH-better-sidebar

yanzhaohui1999 · 2026-08-21T12:20:11Z

背景

侧边栏文件窗口的全局文件名搜索（fs.search）之前是纯 JS 递归遍历：小目录没毛病，一到大目录（十万级）就是秒级，而且 10 万条访问预算一顶，结果会被静默截断——用户感知就是"又慢、结果还不全"。issue #203 里有实测：28G 目录下 JS 遍历 1.56s，fd 只要 0.013s。

做了什么

启动后探测本机的 fd / rg（进程内一次、带缓存、500ms 超时验证），探到就用引擎，探不到或运行失败自动回退原来的 JS 遍历，功能只增不减：

fd：--fixed-strings 字面量子串匹配，--max-results 天然对应 200 条结果上限；文件名 + 目录名都匹配（与遍历一致）
rg：优先用 DSH 自己捆绑的 @vscode/ripgrep（agent 侧搜索工具同一个二进制），PATH 里没有也能命中；query 双 glob + deriveRgMatches 保持"按名字匹配（文件 + 目录）"语义
噪声目录排除（.git / node_modules / 构建缓存等 18 项）与兜底遍历单源共享（SKIP_DIR_NAMES），两条路径结果形状一致
引擎输出统一换算成 JS 遍历的既有契约（根相对、/ 分隔、大小写不敏感、排序、截断），路由和客户端零改动
运行时失败的引擎进程内禁用，不拖慢后续搜索；超时只降级本次（树太大 ≠ 二进制坏）；abort 正常杀子进程、不误禁引擎

几个取舍，说明一下

~~node_modules 不排除（留待产品决策）~~ → 已排除：main 在 d0f194b 给遍历加了噪声目录跳过，本 PR 跟随并把名单单源化到引擎（SKIP_DIR_NAMES，18 项）——常见词搜索的 200 个名额不再被依赖树占满
~~rg 只报文件、不报目录（documented lossy）~~ → 已对齐：rg 无斜杠 glob 按 gitignore 语义只匹配 basename（真机 rg 15 验证），补路径级 glob **/*<q>*/** + 从文件路径倒推目录命中（deriveRgMatches）；残留差异仅剩空目录不可见（rg 看不到无文件的目录）
探测超时取 500ms（比 issue 建议的 1s 更保守），失败静默降级

2026-09-06 rebase 说明

main 大版本合并（v0.17.1 → v0.18.0，154 commits）后本分支已线性化 rebase（丢弃 merge commit，15 commits / 15 files / +1404−48），上述排除单源化与 rg 目录匹配即为 rebase 期间追加的对齐；期间解决的新 main 语义冲突：

SEARCH_SKIP_DIRS（d0f194b）与本 PR 的 worktree 式 .git 指针文件跳过合并（两种形态都跳）
main 新增的 node_modules 测试与本 PR 的 .git-file 测试都保留，walk 用例钉 searchFilesPlain
fd/rg 文档从旧 AGENTS.md 锚定到 README（v0.12.3 条目）与 docs/external-plugin-guide.md（参考实现 §）

原设想中的 Windows CI lane 曾被 #520（ci-windows）覆盖而暂时撤出；2026-09-12 起改为并进那条 lane（见文末「2026-09-12 更新」），#335 随之关闭。

实测

设计文档 4 组真实目录基准（17 万 / 9 万 / 24 万 / 3.7 万文件）：rg 比 JS 遍历快 2.5~20 倍，大目录下 JS 遍历全部预算截断、rg 结果完整。门控调试插桩（DSH_SEARCH_DEBUG=1 才写 $DSH_HOME/search-debug.log，默认完全静默）真机 UI 过一遍：9 万文件 workspace 常见词 15~200ms、罕见词全量遍历 ~478ms 且不截断、通配符按字面量处理、无 .git 内部路径命中。

验证

pnpm typecheck + pnpm lint + pnpm build 全绿；全量测试与干净 main 基线逐项一致（零回归）
tests/search-engines.spec.ts（探测缓存 / 失败禁用 / abort 不误禁 / argv 对称性 / deriveRgMatches）+ fs-search.spec.ts dispatch 组，共 60+ 用例
真机：macOS + DSH 捆绑 rg 15.0.0，含目录命中的端到端冒烟（util → ['util', 'util.ts']、src → ['src']）
文档同步：README、external-plugin-guide、设计文档（验收记录与取舍更新）

测试计划：hook 注入走单测（引擎路径全部经 setEngineHooks 注入，不依赖机器上真有 fd/rg）；Windows 由 ci-windows lane 承载——该 lane 现在会 choco install fd ripgrep 装真引擎（整条套件因此都在引擎在场的环境下跑），套件结束后跑 node scripts/win-engine-check.cjs --assert 的字节级断言。

2026-09-12 更新

合并 main（普通 merge commit，未改写历史、未 force-push）：分支此前那个红着的 ci-windows 是 main 已修的 agent-pty 超时 flake，同步后转绿。
把 ci(search): add search-windows lane for engine platform regressions #335 的搜索引擎 CI 并进 main 已有的 ci-windows lane（两个步骤：choco install fd ripgrep -y --no-progress、node scripts/win-engine-check.cjs --assert），而不是单开 search-windows job——单开要把「装依赖 + 跑全量」再付一遍（CI 分钟翻倍）却覆盖更少；并进来则让整条套件都在引擎在场的环境下跑。引擎 spec 是插桩式的，所以引擎在不在 PATH 都不改变单测结论，ubuntu lane 继续覆盖「引擎缺失 → JS 回退」。
scripts/win-engine-check.cjs 一并升级为 --assert 版：自建 scratch 目录（不再硬编码 C:\Users\y\wintest）、引擎缺失即跳过、钉死 --path-separator / 的输出不得含 \/CR、中文文件名必须命中、失败非零退出。新增 docs/ci-windows.md（红线清单 / 历史平台修正 / 标准动作），AGENTS.md §2 埋入口并顺手改正该节过期的 --maxWorkers=2（实为 1）。ci(search): add search-windows lane for engine platform regressions #335 已关闭，价值去向记录在其评论里。
真机证据（run 34684285338）：choco 装到 fd 10.5.0 / ripgrep 15.2.0；未钉分隔符的 rg 吐 .\中文文件名.ts\n.\src\a.ts（backslash: true），钉 --path-separator / 后干净，fd/rg 的中文名都命中，断言全过。

fixes #203

Menghuan1918 · 2026-08-22T06:48:37Z

我其实主要担心fd / rg在windows上有什么奇奇怪怪的格式问题）所以当时写的时候直接用的node原生的

… test-pin them Review concern (PR omdsh-dev#303): fd/rg output formats on Windows. rg emits backslash-separated paths with dot-backslash prefixes and CRLF line endings; the normalizer only handled the platform sep case, which Linux CI could not exercise with Windows-shaped lines, and the trailing CR was never explicitly stripped (relied on readline behavior). Make the separator injectable (defaults to the platform separator, behavior unchanged) and strip a trailing CR, then pin both Windows shapes in unit tests.

yanzhaohui1999 · 2026-08-22T08:04:40Z

Windows 的坑主要在输出格式不稳定:同一个 rg,cmd 下输出 `,Git Bash 下输出 /;还有 CRLF 行尾、.前缀,以及 DSH 自带 rg 的探测布局在 Windows 上完全是另一套(没有lib/层)。现在让 fd/rg 都在源头用--path-separator /` 钉死分隔符,输出再归一化成和原 JS 遍历一致的格式,探测布局也按 Windows 实际装法补齐,全有单测守着;引擎挂了就回退原实现,不会比之前差。

Windows 真机我会跑一轮验证,完了把结果补回 PR;另外要不要顺便在 CI 加一个 windows runner 跑这两组搜索测试?仓库现在是纯 ubuntu,Windows 格式这类问题只有真机/CI 才兜得住,加的话我可以一起带上。

Menghuan1918 · 2026-08-22T08:14:31Z

加一下？

yanzhaohui1999 · 2026-08-22T13:23:22Z

Windows 真机验证补完了,结果都写在设计文档里。贴几个关键点:

你担心的 Windows 格式问题是真实的:Windows 上 rg 默认输出 \ 分隔 + . 前缀(同一个 rg,cmd 和 Git Bash 行为还不一样,rg#501),CRLF 也有坑。现在 fd/rg 都在源头用 --path-separator / 钉死分隔符,输出再归一化成和原 JS 遍历一致的格式(CRLF/\ 兜底剥离),全有单测守着。

Windows 上的 DSH 自带 rg 探测也修了:npm 全局装在 %APPDATA%\npm\node_modules(没有 lib/ 层),旧公式在 Windows 必 miss。实测装 DSH 0.1.1-rc.2 后,新公式 probe 到捆绑 rg.exe;PATH 里没有引擎时它照样兜底。

Bench(Windows 11,10 万+ 文件目录):引擎比 JS 遍历快 3.7~34.7 倍;而且原生遍历会撞 100k visited 预算截断——query=1234 时 plain 只找到 5 条,引擎 40 条。"慢+结果不全"确实是 issue #203 的实锤。

完整链路(打包→dsh plugin add→dsh web→fs.search 路由)也跑通了,engine=fd ... 25ms 这种日志在真机可见;repro 脚本在 scripts/win-*.cjs。

另外顺手把 fd 截断和 rg 对齐了(fd 之前 --max-results 钉在哨兵位,满集永不标 truncated、多返 1 条;现在 cap+1,两条引擎行为一致,有单测钉死)。

Windows CI 那条我另开 PR 跟进(装 fd/rg + 跑这两组搜索测试 + 吸收这三个脚本)。

…-dev#203) The editor side panel's global name search walked the tree in plain JS. On large workspaces a single search can take a second or more. Detect a native engine once per process (lazily, cached) and dispatch to it: - fd (PATH + fixed Homebrew/usr/cargo paths) with fixed-string, case-insensitive, hidden/ignore-aware listing capped by max-results - rg next: DeepSeek Harness' own bundled @vscode/ripgrep binary (resolved from process.execPath's global prefix, the same binary the harness's agent-side search tool runs) before any PATH rg, then fixed paths - plain JS walk as the fallback when no engine is available or fails Engine output is normalized to the plain walk's contract (root-relative, / separators, case-insensitive substring, sorted, truncated); .git is excluded in both engine and plain paths; runtime failures disable the engine in-process and degrade to the next one. Probe/health hooks keep the engine selection unit-testable without real binaries.

Debug instrumentation now lives in src/search-debug.ts and is a no-op unless DSH_SEARCH_DEBUG=1 is set: default is zero console output and zero disk writes. The log file resolves like pty-deps does (DSH_HOME, falling back to ~/.dsh); write failures are swallowed so a missing log dir can never break a search. relRoot uses homedir()/DSH_HOME with a separator boundary instead of process.env.HOME.

… verification in design doc README and AGENTS.md now mention the fd/rg engine dispatch for fs.search (DSH-bundled rg preferred, JS-walk fallback). The design doc records the gated instrumentation file, today real-host acceptance run (workspace timings, literal-semantics checks, .git exclusion) and the node_modules tradeoff decision.

… test-pin them Review concern (PR omdsh-dev#303): fd/rg output formats on Windows. rg emits backslash-separated paths with dot-backslash prefixes and CRLF line endings; the normalizer only handled the platform sep case, which Linux CI could not exercise with Windows-shaped lines, and the trailing CR was never explicitly stripped (relied on readline behavior). Make the separator injectable (defaults to the platform separator, behavior unchanged) and strip a trailing CR, then pin both Windows shapes in unit tests.

…robe The bundled-rg candidate derived only from process.execPath with a POSIX npm-global shape (<prefix>/lib/node_modules/...). On Windows npm global there is no lib/ layer — packages land in %APPDATA%\npm\node_modules — so the probe would resolve to a bogus C:\lib\... path and silently fall back to plain-walk (or system rg, usually absent). Also add the launchd-profile layout (~/.dsh/profiles/node_modules), which a real machine check (2026-08-22) shows is where DSH actually lands here, and the Homebrew/pnpm global roots. Wrong guesses stay cheap: every candidate still goes through the 500ms --version verify; tests pin the Windows shape (%APPDATA%, no lib/ layer) and the profile layout.

rg emits a platform-dependent separator (backslash in cmd/PowerShell on Windows, forward slash in Git Bash — rg#501). fd already pinned '/' via --path-separator; rg supports the same flag since 0.8, so pin it too and let the output match the walk contract before normalization even runs. normalizeEnginePaths stays as a safety net for exotic builds.

Runs rg/fd with and without --path-separator / on a scratch tree and reports backslash/CR/LF presence byte-level, pinning the Windows output shapes this PR normalizes (runs on any OS; Windows is where it earns its keep).

Runs bundledRgCandidates + probeEngines against the live machine and prints which candidate EXISTS plus what the real probe resolves. On a npm-global Windows DSH install this proves the %APPDATA% layout derivation hits the bundled rg.exe (verified live 2026-08-22: C:\Users\y\AppData\Roaming\npm\node_modules\@deepseek-ai\dsh\...).

…fs.search) Replicates scripts/e2e-mount.sh on Windows without bash: scratch profile bootstrap files (package.json / pnpm-workspace.yaml), a helper that checks dsh web state and POSTs /sidebar/api/fs.search with curl. The full chain was verified live on Windows (2026-08-22): plugin mounted into a scratch profile via dsh plugin add, dsh web up on an OS-assigned port, fs.search returned ok with 200 matches, and DSH_SEARCH_DEBUG proved the bundled rg.exe under %APPDATA%\npm\node_modules was probed and picked (fd engine ran the search, 222ms).

… by tests) fd capped at maxMatches+1 exactly, so it stopped AT the stream sentinel: a full result set was never marked truncated and up to maxMatches+1 rows leaked past the 200-row contract. Pin fd one line higher (cap+1) so it overflows into the same sentinel rg triggers; both engines now report truncated identically and the caller slices back to maxMatches. Extract fdArgv/rgArgv as exported pure functions so the argv contract (literal query, / separators, sentinel above cap) is pinned by unit tests. Windows bench script added: plain walk vs engines on the same root (D:\\Project real machine, 2026-08-22: engines 3.7-34.7x faster AND plain walk truncates on the 100k visited budget, dropping matches e.g. query=1234 plain 5 vs engines 40).

Windows 11 + npm-global DSH 0.1.1-rc.2 + scoop fd/rg: byte-level output shapes (backslash + .\ prefix reproduced, --path-separator / verified), %APPDATA% probe hit, full mount->dsh web->fs.search chain, bench (plain truncated on visited budget, engines 3.7-34.7x faster), and the fd truncation-symmetry fix.

…ile parity Review follow-ups on the fd/rg engine path (verified against real rg 15.0): - escapeGlob now escapes { }: globset reads {a,b} as alternation, so an unbalanced '{' made rg fail glob parsing (exit 2 -> engine wrongly disabled process-wide) and 'a{b}' silently searched 'ab' - a 15s runtime TIMEOUT no longer adds the engine to the broken set: a timeout means THIS tree is too large, not that the binary is broken; the broken set is per-engine, so one huge directory would otherwise strip every other root of the engine (new EngineTimeoutError) - worktree '.git' FILE parity: rg gains a second exclusion glob next to the directory one (it requires a path segment after .git and leaked the pointer file), and the plain walk skips any entry named .git -- same semantics as fd's --exclude .git

…ew base Rebase follow-ups for the new v0.18.0 main: - Single-source the skip-name list as SKIP_DIR_NAMES (search-engines.ts); the plain walk builds SEARCH_SKIP_DIRS from it. fd (--exclude, one per name) and rg (--iglob directory + entry-only pair, per name) now exclude every name: with --no-ignore active the engines would otherwise re-enter node_modules etc. and regress d0f194b's budget savings. - The plain walk skips a worktree-style .git FILE too (removed the isDirectory gate), keeping parity with the engine excludes; both the node_modules spec and the .git-file spec are kept (walk tests pinned to searchFilesPlain for determinism). - argv symmetry tests: fd excludes == SKIP_DIR_NAMES, rg carries the glob pair per name, query iglob stays last. - escapeGlob keeps escaping '[' (rg globs are gitignore-style: '[d]' is a character class, not a literal) with a reasoned no-useless-escape exemption. - win-* ops scripts: .cjs files are run directly with node on a real host — file-level no-require-imports exemptions; drop an unused os import. - fd/rg engine docs re-anchored to README (v0.12.3 entry) and docs/external-plugin-guide.md (old AGENTS.md is now a slim rules file).

rg --files never emits a directory line, so rg-only machines missed the directory matches fd and the plain walk report. Two fixes, pinned by a real-rg smoke on this machine: - rgArgv gains a second query glob '**/*<q>*/**': the slash-free form follows gitignore semantics and matches BASENAMES only (verified on real rg 15 — 'util/helper.ts' never matched '*util*', so matching directories were invisible to derivation too); the path-level form admits files under matching directories. - deriveRgMatches(paths, query, maxMatches): basename hit stays a file match, a matching directory segment is derived as a directory match; deduped, sorted, capped at maxMatches + 1 with truncated raised when the budget (not the engine) cut the set. Residual lossy: empty directories stay invisible to rg (no file path carries them). - runChild rg branch delegates to deriveRgMatches and ORs the stream truncation flag; unit tests cover derivation, dedupe, case insensitivity, budget capping, and the two-glob argv shape.

yanzhaohui1999 · 2026-09-05T17:29:45Z

Rebase 到 v0.18.0 基线（2026-09-06）

main 大版本合并（v0.17.1 → v0.18.0，154 commits）后原分支已 DIRTY。已线性化 rebase（丢弃 merge commit，现为 15 commits / 15 files / +1404−48），描述已同步更新。除机械冲突外，rebase 期间追加了三笔实质对齐：

1. 引擎排除与 walk 跳过名单单源化（关键修复）

原实现引擎只排除 .git，但 fd/rg 都带 --no-ignore——引擎路径激活时 node_modules 会被重新搜进来，回退掉 main d0f194b 的噪声目录优化。现把跳过名单单源化为 SKIP_DIR_NAMES（18 项，walk 与引擎共用）：fd 每名一个 --exclude；rg 每名一对 --iglob（目录 glob + 条目 glob，后者保住 worktree .git 指针文件语义）。新增 argv 对称性测试钉住。

2. rg 目录名匹配（收窄 documented lossy）

rg --files 不输出目录行，rg-only 机器上目录命中缺失（fd / plain walk 都有）。真机冒烟还暴露了一个 mock 测不出的坑：rg 无斜杠 glob 按 gitignore 语义只匹配 basename（util/helper.ts 不匹配 *util*）。修复 = 双 query glob（basename 形态 + 路径级 **/*<q>*/**）+ deriveRgMatches 从文件路径倒推目录命中；预算超出时 cap + truncated（与流式 sentinel 同语义）。残留 lossy 收窄为：空目录不可见。

本机（DSH 自带 rg 15）端到端：util → ['util', 'util.ts']、src → ['src']，与 walk 逐项一致。

3. 冲突解决与文档重锚定

SEARCH_SKIP_DIRS（d0f194b）与本 PR 的 worktree .git 文件跳过合并（两种形态都跳）
main 新增的 node_modules 测试与本 PR 的 .git-file 测试都保留；walk 用例钉 searchFilesPlain
fd/rg 文档从旧 AGENTS.md 锚定到 README（v0.12.3 条目）与 docs/external-plugin-guide.md（参考实现 §）

验证

typecheck / lint / build 全绿；search 相关 spec 42+9 全绿；全量测试与干净 main 基线逐项一致（零回归）。原设想的 Windows CI lane 已被 #520（ci-windows）覆盖，本 PR 不含 CI 改动。

…lane Fold PR omdsh-dev#335's standalone `search-windows` lane into the existing ci-windows job instead of adding a second Windows job: real fd/ripgrep land on PATH before the suite (so the whole suite runs with engines present) and the byte-level assertions run against them afterwards. A separate job would pay the install + full-suite cost twice for strictly less coverage; the engine specs inject their own probes, so engine presence cannot flip unit-test outcomes and the ubuntu lane keeps covering the missing-engine fallback. scripts/win-engine-check.cjs gains the --assert mode: self-built scratch tree (no hard-coded C:\Users\y\wintest), skip when an engine is missing, pinned (--path-separator /) output must carry no '\'/CR, and a Unicode filename must survive the UTF-8 pipeline. docs/ci-windows.md carries the red-line list / historical fixes / standard procedure; AGENTS.md §2 points at it. That section's stale `--maxWorkers=2` is corrected to 1 (what package.json and the lane's own comment say).

yanzhaohui1999 · 2026-09-12T08:56:43Z

2026-09-12：同步 main + 把 #335 的引擎 CI 并进来

1. 同步 main（670eb0a，普通 merge commit，未改写历史）

分支此前唯一的红灯 ci-windows 是 main 已修的 agent-pty 5000ms 超时 flake（5451f5c 把全局 testTimeout 提到 15s、lane 改单 fork），同步后 ci / ci-windows / plugin-mount 三 lane 全绿。

2. #335 的搜索引擎 CI 并进 ci-windows（b97a003）

没有单开 search-windows job：main 已有 ci-windows（#520），单开要把「装依赖 + 跑全量」再付一遍（CI 分钟翻倍）却覆盖更少；并进去则让整条套件都在引擎在场的环境下跑。落地的两个步骤：

- name: Install fd / ripgrep (choco)      # pnpm install 之前，真二进制进 PATH
  run: choco install fd ripgrep -y --no-progress
- name: Engine output-shape assertions (real fd/rg)   # 套件之后
  run: node scripts/win-engine-check.cjs --assert

引擎 spec 是插桩式的（tests/search-engines.spec.ts 经 setEngineHooks 注入 probe），所以引擎在不在 PATH 上都不改变单测结论——ubuntu lane 继续覆盖「引擎缺失 → JS 遍历回退」这条路径。

scripts/win-engine-check.cjs 同步升级为 --assert 版（自建 scratch 目录，不再硬编码 C:\Users\y\wintest；引擎缺失即跳过；钉死 --path-separator / 的输出不得含 \/CR；中文文件名必须命中）。新增 docs/ci-windows.md（红线清单 / 历史平台修正 / 标准动作），AGENTS.md §2 埋入口并顺手改正该节过期的 --maxWorkers=2（package.json 与 lane 注释都是 1）。

3. 真机证据（ci-windows，run 34684285338，ci-windows 5m01s 绿）

choco: fd v10.5.0 / ripgrep v15.2.0 installed
=== rg default (no path-separator) ===   backslash: true    raw: ".\中文文件名.ts\n.\src\a.ts\n.\README.md\n"
=== rg with --path-separator / ===       backslash: false   raw: "./README.md\n./中文文件名.ts\n./src/a.ts\n"
=== fd with --path-separator / ===       backslash: false
=== rg Chinese filename ===              backslash: false   raw: "./中文文件名.ts\n"
=== fd Chinese filename ===              backslash: false   raw: "./中文文件名.ts\n"

即当年 review 里担心的「Windows 上 rg 的 \ / . 前缀 / 输出格式」在真机上被如实复现，而 PR 的 --path-separator / 归一化确实把它压住了；中文文件名两条引擎都命中。

4. 范围与遗留：#335 已关闭（价值去向记录在其评论里）；原先设想随本 PR 携带的 build 跨平台、smoke.spec.ts 的 autocrlf、pty-deps.spec.ts 平台断言三处均已在 main 独立落地，不再重复。本 PR 现在相对 main 是 18 文件 / +1563−49（原 15 文件 / +1404−48，新增 CI 移植的 3 文件）。

This was referenced Aug 22, 2026

ci(search): add search-windows lane for engine platform regressions yanzhaohui1999/DSH-better-sidebar#1

Closed

ci(search): add search-windows lane for engine platform regressions #335

Closed

HuanLinOTO added enhancement New feature or request P2 中：一般功能/优化 area/explorer File tree / explorer: navigation, context menu, search labels Aug 23, 2026

yanzhaohui1999 added 15 commits September 6, 2026 00:51

docs(search): add Windows verification manual for fs.search engines

c0c42be

test(search): add Windows engine-output shape checker script

aac5305

Runs rg/fd with and without --path-separator / on a scratch tree and reports backslash/CR/LF presence byte-level, pinning the Windows output shapes this PR normalizes (runs on any OS; Windows is where it earns its keep).

yanzhaohui1999 force-pushed the feat/fs-search-engines branch from 8771ccd to 64dafd4 Compare September 5, 2026 17:29

yanzhaohui1999 added 2 commits September 12, 2026 16:40

Merge remote-tracking branch 'origin/main' into sync/pr303

670eb0a

Navigation Menu

Uh oh!

feat(search): fs.search 文件名搜索优先使用 fd/rg,缺失回退 JS 遍历 (fixes #203) - #303

feat(search): fs.search 文件名搜索优先使用 fd/rg,缺失回退 JS 遍历 (fixes #203)#303
yanzhaohui1999 wants to merge 17 commits into
omdsh-dev:mainfrom
yanzhaohui1999:feat/fs-search-engines

yanzhaohui1999 commented Aug 21, 2026 •

edited

Loading

Uh oh!

Menghuan1918 commented Aug 22, 2026

Uh oh!

yanzhaohui1999 commented Aug 22, 2026

Uh oh!

Menghuan1918 commented Aug 22, 2026

Uh oh!

yanzhaohui1999 commented Aug 22, 2026

Uh oh!

yanzhaohui1999 commented Sep 5, 2026

Uh oh!

yanzhaohui1999 commented Sep 12, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

yanzhaohui1999 commented Aug 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Menghuan1918 commented Aug 22, 2026

Uh oh!

yanzhaohui1999 commented Aug 22, 2026

Uh oh!

Menghuan1918 commented Aug 22, 2026

Uh oh!

yanzhaohui1999 commented Aug 22, 2026

Uh oh!

yanzhaohui1999 commented Sep 5, 2026

Rebase 到 v0.18.0 基线（2026-09-06）

1. 引擎排除与 walk 跳过名单单源化（关键修复）

2. rg 目录名匹配（收窄 documented lossy）

3. 冲突解决与文档重锚定

验证

Uh oh!

yanzhaohui1999 commented Sep 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

2026-09-12：同步 main + 把 #335 的引擎 CI 并进来

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

yanzhaohui1999 commented Aug 21, 2026 •

edited

Loading

yanzhaohui1999 commented Sep 12, 2026 •

edited

Loading