Skip to content

Releases

v1.2.10

Choose a tag to compare

@github-actions github-actions released this 17 Nov 03:27

1.2.10

中文

  • 修「红队模式被重复注册挤掉」:1.2.9 把预设写进 rc.3 声明行后,宿主 loader 会在启动时注册 redteam;插件运行时又调用 agentPresets.register(),得到 Duplicate agent preset: redteam,模式下拉因此被过滤。preset-declare.json 会停在 register-failed。
  • 新版分支改为只验证、不重复注册:检测到 profile 已有 preset-redteam 声明行时,等待 loader 的注册结果并写回 registered / registered-broken,不再调用 registry.register();rc.2 的运行时注册路径保留。
  • 成功状态会清掉上一轮遗留的 error 字段,避免面板显示旧错误。
  • 真机复验:preset-declare.json = ok:true / phase:registered / plugins:19 / broken:null / error:null;/redteam/api version 1.2.9、164 技能、blockers 0。全套 node --test 233 通过 / 0 失败;compat:check 一致。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fixes the red-team preset being squeezed out by duplicate registration: 1.2.9 wrote the rc.3 declaration row, so the host loader already registered redteam at startup; the plugin then called agentPresets.register() again and hit Duplicate agent preset: redteam, which filtered the mode out of the picker. preset-declare.json stayed at register-failed.
  • The modern branch now verifies instead of registering again: when the profile already carries the preset-redteam declaration row, it waits for the loader's registration and writes registered / registered-broken without calling registry.register(); the rc.2 runtime path is unchanged.
  • Successful status clears the stale error field from the previous attempt.
  • Live check: preset-declare.json = ok:true / phase:registered / plugins:19 / broken:null / error:null; /redteam/api reports version 1.2.9, 164 skills and 0 blockers. node --test 233 passing / 0 failing; compat:check clean.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.9

中文

  • 修「红队模式加载失败」:1.2.8 把预设写成 rc.3 声明行后,模式卡显示了但带红标「加载失败」。逐行核对插件名后定位到两处 rc.3 改名/删包:
    • @deepseek-ai/dsh-tool-subagent-control/list-agents 子路径在 rc.3 已删除,list_agents 现在由主入口注册;继续写这一行 → 整份预设激活失败。
    • @deepseek-ai/dsh-workflow-worker-thread 包在 rc.3 不存在,即便 disabled: true 也会拖垮激活。
  • 预设模板 presets/redteam/agent.cordis.yml 删掉这两行,其余 16 行逐一在宿主里校验存在且导出匹配。
  • 真机 profile 的 preset-redteam 声明行已用 force 重写并复验(宿主自带 YAML 解析器:id: redteam、name: 红队模式、无失效插件名)。
  • 全套 node --test 233 通过 / 0 失败;compat:check 一致。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fixes "red-team mode failed to load": 1.2.8 moved the preset to an rc.3 declaration row, and the card appeared with a red "failed to activate" badge. Auditing every plugin name found two rc.3 changes:
    • @deepseek-ai/dsh-tool-subagent-control/list-agents no longer exists as a subpath; list_agents is registered by the main entry, and keeping the old row fails the whole preset.
    • @deepseek-ai/dsh-workflow-worker-thread has no package in rc.3, and even disabled: true still breaks activation.
  • Both rows are removed from presets/redteam/agent.cordis.yml; the remaining 16 rows are checked against the host for presence and matching exports.
  • The live profile's preset-redteam row was rewritten with force and re-validated with the host's YAML parser (id: redteam, name: 红队模式, no dead plugin names).
  • node --test 233 passing / 0 failing; compat:check clean.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.8

中文

  • 红队模式在 rc.3 上装不出来的根因:0.2.2 起宿主不再读取 $DSH_HOME/.agent-presets/<id>/(官方文档原话 "Nothing reads that directory any more"),预设改为 bundle 补丁里的 @deepseek-ai/dsh-agent-preset 声明行。本插件一直只写旧目录,于是模式下拉里没有红队模式,面板仍在、/redteam/api 也通——只剩入口消失。
  • installPreset 增加新版分支:宿主能解析到 @deepseek-ai/dsh-agent-preset 时,改为在挂着本插件的 profile cordis.patch.yml 追加 preset-redteam 声明行(id/name/description/order/plugins),插件列表从包内 presets/redteam/agent.cordis.yml 生成、跳过锚点门行(它是性能项,不是功能项);幂等,已存在就保留,force 才重写。旧目录安装保留给 rc.2 及更早。
  • 声明行里显示的预设名是「红队模式」;插件逐项校验:本轮只缺 dsh-workflow-worker-thread(模板里本就 disabled: true,不影响激活)。
  • 真机已按此格式写入 ~/.dsh/profiles/desktop/cordis.patch.yml(YAML 用宿主自带解析器验证:preset-redteam / plugins 18 行 / 首行 persona),重启后模式下拉应出现红队模式。
  • 全套 node --test 233 通过 / 0 失败;compat:check 一致。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Why the red-team mode never showed up on rc.3: 0.2.2 no longer reads $DSH_HOME/.agent-presets/<id>/ ("Nothing reads that directory any more" in the official docs); presets are now @deepseek-ai/dsh-agent-preset declaration rows inside bundle patches. This plugin kept writing only the legacy directory, so the console and /redteam/api stayed up while the mode vanished from the picker.
  • installPreset gained a modern branch: when the host resolves @deepseek-ai/dsh-agent-preset, it appends a preset-redteam declaration row (id/name/description/order/plugins) to the profile cordis.patch.yml that carries this plugin, generating the plugin list from the bundled presets/redteam/agent.cordis.yml and skipping the anchor-gate row (a performance knob, not a feature). Idempotent: kept when present, rewritten only with force. Legacy install remains for rc.2 and older.
  • The row shows the preset as 红队模式. Every plugin name was checked; only dsh-workflow-worker-thread is absent, and the template already ships it disabled: true, so activation is unaffected.
  • The live profile has the row now (validated with the host's own YAML parser: preset-redteam, 18 plugins, first row persona); restart makes it appear in the picker.
  • node --test 233 passing / 0 failing; compat:check clean.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.7

中文

  • 修 1.2.6 的复发口子:宿主重启会走「过期基线上刷新」自愈(refreshStaleBaksForUpgrade),原判据是「正文里没有 dsh-purge 字面量且与 bak 不同 → 认定成官方新版,把 bak 刷成正文」。但 #1/#22/#33/#48 这类补丁只改文案、不写 marker 注释,于是刚重建好的基线被下一次启动原样打回已打态——真机实测重启一次,8 个刚修好的文件全部复发成 baseline_lost(dsh-agent-instructions/lib/index.js 从 63281 字节的官方原件又变回 63970 字节的已打态)。
  • 判据换成补丁自己的:逐文件按「marker 命中 + 原 pattern 已不命中」判定正文是否已被我们写过,命中就不动 bak;只有正文确实回到官方新版(原 pattern 还在)才刷新基线。拿不到补丁索引时退回旧的字面量判据,不改变无补丁文件的行为。
  • 新增 test/upgrade-refresh.test.js 3 条:已打态无字面量不刷基线、官方新版照常刷新、applyPatches 内部刷新后再 revertAll 回到官方新版而非旧版。全套 node --test 233 通过 / 0 失败。
  • npm run cycle:isolated 在 rc.3 隔离树 OK 6/6,与 1.2.6 同口径。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Closes the 1.2.6 regression window: every host boot runs the "stale baseline refresh" self-heal (refreshStaleBaksForUpgrade), which used to decide "no dsh-purge literal + differs from bak ⇒ this is the official new version, overwrite bak with the current bytes". Patches #1/#22/#33/#48 only rewrite sentences and never write marker comments, so every freshly rebuilt baseline was reverted to the patched state on the next boot — on the live host all 8 repaired files relapsed to baseline_lost after one restart (dsh-agent-instructions/lib/index.js went from the 63281-byte official copy back to the 63970-byte patched one).
  • The refresh now uses the patch's own verdict: per file it checks "marker present + original pattern gone"; when that holds the backup is left alone, and only a file that really is the official new version gets its baseline refreshed. Without a patch index it falls back to the old literal check, so files no patch touches keep the previous behaviour.
  • 3 new cases in test/upgrade-refresh.test.js; node --test 233 passing / 0 failing. npm run cycle:isolated remains OK 6/6 on the rc.3 tree.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.6

中文

  • 修掉「假还原」:.dshpurge.bak 只在缺失时才创建,遇到「先打补丁、后才有备份」的历史状态会把已打态存成基线。此时 revertAll 把打过补丁的文件覆盖回自己:不报错、0 差异、界面显示成功,用户以为退回了官方原版,其实补丁全在。真机实测 54 个备份里 10 个是这种,其中 dsh-agent-instructions/lib/index.js、dsh-tool-fs/lib/index.js 能直接验到补丁痕迹(强句 / READ_MAX_LINE_LENGTH = 1e4)。
  • 新判定 baseline_lost:备份与本体逐字节相同、且本体确实带补丁痕迹(marker 命中 + 原 pattern 已不在)时,不再计入 reverted,单独列出并保留 bak 留证;--revert、面板「还原」都会提示「需从官方安装包/asar 重建基线」。只同字节、但没补丁痕迹的文件仍按 skipped 处理,不误报。
  • --apply 预先告警:应用前先跑一次只读检测,若存在基线丢失文件就先说明「还原回不到官方原文」,避免用户事后才发现。
  • 新增 test/baseline-lost.test.js 5 条(判丢失、真原文不丢、未打不丢、revert 不谎报、same-bytes 无痕迹仍 skipped),全套 node --test 230 通过 / 0 失败。
  • 真机已按官方 rc.3 原包重建 8 个文件的基线(旧备份留档 *.dshpurge.bak.lost-20261116),剩余 2 个属于已退役的 #50 残留,官方原件里没有对应锚点,按「无法重建」如实列出。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fake reverts are gone: .dshpurge.bak was only created when missing, so a history of "patch first, backup later" stored the already-patched bytes as the baseline. revertAll then copied a patched file over itself — no error, zero diff, a success message, and every patch still in place. On the live host 10 of 54 backups were in this state; two of them prove it outright (dsh-agent-instructions/lib/index.js carries the strong sentence, dsh-tool-fs/lib/index.js carries READ_MAX_LINE_LENGTH = 1e4).
  • New baseline_lost verdict: when the backup is byte-identical to the body and the body really is patched (marker present, original pattern gone), the file is no longer counted as reverted; it is listed separately, the backup is kept as evidence, and both --revert and the panel revert say "rebuild the baseline from the official package/asar". A same-bytes backup without patch traces is still just skipped, so untouched files are not misreported.
  • --apply warns first: a read-only pre-check reports baseline-lost files before patching, instead of letting users discover it later.
  • 5 new cases in test/baseline-lost.test.js; node --test 230 passing / 0 failing.
  • On the live host 8 baselines were rebuilt from the official rc.3 package (old copies kept as *.dshpurge.bak.lost-20261116); the remaining 2 are leftovers of the retired #50 whose anchors do not exist in the official files, so they are listed as unrecoverable rather than quietly cleared.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.5

中文

  • **...
Read more

v1.2.9

Choose a tag to compare

@github-actions github-actions released this 16 Nov 12:52

1.2.9

中文

  • 修「红队模式加载失败」:1.2.8 把预设写成 rc.3 声明行后,模式卡显示了但带红标「加载失败」。逐行核对插件名后定位到两处 rc.3 改名/删包:
    • @deepseek-ai/dsh-tool-subagent-control/list-agents 子路径在 rc.3 已删除,list_agents 现在由主入口注册;继续写这一行 → 整份预设激活失败。
    • @deepseek-ai/dsh-workflow-worker-thread 包在 rc.3 不存在,即便 disabled: true 也会拖垮激活。
  • 预设模板 presets/redteam/agent.cordis.yml 删掉这两行,其余 16 行逐一在宿主里校验存在且导出匹配。
  • 真机 profile 的 preset-redteam 声明行已用 force 重写并复验(宿主自带 YAML 解析器:id: redteam、name: 红队模式、无失效插件名)。
  • 全套 node --test 233 通过 / 0 失败;compat:check 一致。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fixes "red-team mode failed to load": 1.2.8 moved the preset to an rc.3 declaration row, and the card appeared with a red "failed to activate" badge. Auditing every plugin name found two rc.3 changes:
    • @deepseek-ai/dsh-tool-subagent-control/list-agents no longer exists as a subpath; list_agents is registered by the main entry, and keeping the old row fails the whole preset.
    • @deepseek-ai/dsh-workflow-worker-thread has no package in rc.3, and even disabled: true still breaks activation.
  • Both rows are removed from presets/redteam/agent.cordis.yml; the remaining 16 rows are checked against the host for presence and matching exports.
  • The live profile's preset-redteam row was rewritten with force and re-validated with the host's YAML parser (id: redteam, name: 红队模式, no dead plugin names).
  • node --test 233 passing / 0 failing; compat:check clean.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.8

中文

  • 红队模式在 rc.3 上装不出来的根因:0.2.2 起宿主不再读取 $DSH_HOME/.agent-presets/<id>/(官方文档原话 "Nothing reads that directory any more"),预设改为 bundle 补丁里的 @deepseek-ai/dsh-agent-preset 声明行。本插件一直只写旧目录,于是模式下拉里没有红队模式,面板仍在、/redteam/api 也通——只剩入口消失。
  • installPreset 增加新版分支:宿主能解析到 @deepseek-ai/dsh-agent-preset 时,改为在挂着本插件的 profile cordis.patch.yml 追加 preset-redteam 声明行(id/name/description/order/plugins),插件列表从包内 presets/redteam/agent.cordis.yml 生成、跳过锚点门行(它是性能项,不是功能项);幂等,已存在就保留,force 才重写。旧目录安装保留给 rc.2 及更早。
  • 声明行里显示的预设名是「红队模式」;插件逐项校验:本轮只缺 dsh-workflow-worker-thread(模板里本就 disabled: true,不影响激活)。
  • 真机已按此格式写入 ~/.dsh/profiles/desktop/cordis.patch.yml(YAML 用宿主自带解析器验证:preset-redteam / plugins 18 行 / 首行 persona),重启后模式下拉应出现红队模式。
  • 全套 node --test 233 通过 / 0 失败;compat:check 一致。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Why the red-team mode never showed up on rc.3: 0.2.2 no longer reads $DSH_HOME/.agent-presets/<id>/ ("Nothing reads that directory any more" in the official docs); presets are now @deepseek-ai/dsh-agent-preset declaration rows inside bundle patches. This plugin kept writing only the legacy directory, so the console and /redteam/api stayed up while the mode vanished from the picker.
  • installPreset gained a modern branch: when the host resolves @deepseek-ai/dsh-agent-preset, it appends a preset-redteam declaration row (id/name/description/order/plugins) to the profile cordis.patch.yml that carries this plugin, generating the plugin list from the bundled presets/redteam/agent.cordis.yml and skipping the anchor-gate row (a performance knob, not a feature). Idempotent: kept when present, rewritten only with force. Legacy install remains for rc.2 and older.
  • The row shows the preset as 红队模式. Every plugin name was checked; only dsh-workflow-worker-thread is absent, and the template already ships it disabled: true, so activation is unaffected.
  • The live profile has the row now (validated with the host's own YAML parser: preset-redteam, 18 plugins, first row persona); restart makes it appear in the picker.
  • node --test 233 passing / 0 failing; compat:check clean.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.7

中文

  • 修 1.2.6 的复发口子:宿主重启会走「过期基线上刷新」自愈(refreshStaleBaksForUpgrade),原判据是「正文里没有 dsh-purge 字面量且与 bak 不同 → 认定成官方新版,把 bak 刷成正文」。但 #1/#22/#33/#48 这类补丁只改文案、不写 marker 注释,于是刚重建好的基线被下一次启动原样打回已打态——真机实测重启一次,8 个刚修好的文件全部复发成 baseline_lost(dsh-agent-instructions/lib/index.js 从 63281 字节的官方原件又变回 63970 字节的已打态)。
  • 判据换成补丁自己的:逐文件按「marker 命中 + 原 pattern 已不命中」判定正文是否已被我们写过,命中就不动 bak;只有正文确实回到官方新版(原 pattern 还在)才刷新基线。拿不到补丁索引时退回旧的字面量判据,不改变无补丁文件的行为。
  • 新增 test/upgrade-refresh.test.js 3 条:已打态无字面量不刷基线、官方新版照常刷新、applyPatches 内部刷新后再 revertAll 回到官方新版而非旧版。全套 node --test 233 通过 / 0 失败。
  • npm run cycle:isolated 在 rc.3 隔离树 OK 6/6,与 1.2.6 同口径。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Closes the 1.2.6 regression window: every host boot runs the "stale baseline refresh" self-heal (refreshStaleBaksForUpgrade), which used to decide "no dsh-purge literal + differs from bak ⇒ this is the official new version, overwrite bak with the current bytes". Patches #1/#22/#33/#48 only rewrite sentences and never write marker comments, so every freshly rebuilt baseline was reverted to the patched state on the next boot — on the live host all 8 repaired files relapsed to baseline_lost after one restart (dsh-agent-instructions/lib/index.js went from the 63281-byte official copy back to the 63970-byte patched one).
  • The refresh now uses the patch's own verdict: per file it checks "marker present + original pattern gone"; when that holds the backup is left alone, and only a file that really is the official new version gets its baseline refreshed. Without a patch index it falls back to the old literal check, so files no patch touches keep the previous behaviour.
  • 3 new cases in test/upgrade-refresh.test.js; node --test 233 passing / 0 failing. npm run cycle:isolated remains OK 6/6 on the rc.3 tree.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.6

中文

  • 修掉「假还原」:.dshpurge.bak 只在缺失时才创建,遇到「先打补丁、后才有备份」的历史状态会把已打态存成基线。此时 revertAll 把打过补丁的文件覆盖回自己:不报错、0 差异、界面显示成功,用户以为退回了官方原版,其实补丁全在。真机实测 54 个备份里 10 个是这种,其中 dsh-agent-instructions/lib/index.js、dsh-tool-fs/lib/index.js 能直接验到补丁痕迹(强句 / READ_MAX_LINE_LENGTH = 1e4)。
  • 新判定 baseline_lost:备份与本体逐字节相同、且本体确实带补丁痕迹(marker 命中 + 原 pattern 已不在)时,不再计入 reverted,单独列出并保留 bak 留证;--revert、面板「还原」都会提示「需从官方安装包/asar 重建基线」。只同字节、但没补丁痕迹的文件仍按 skipped 处理,不误报。
  • --apply 预先告警:应用前先跑一次只读检测,若存在基线丢失文件就先说明「还原回不到官方原文」,避免用户事后才发现。
  • 新增 test/baseline-lost.test.js 5 条(判丢失、真原文不丢、未打不丢、revert 不谎报、same-bytes 无痕迹仍 skipped),全套 node --test 230 通过 / 0 失败。
  • 真机已按官方 rc.3 原包重建 8 个文件的基线(旧备份留档 *.dshpurge.bak.lost-20261116),剩余 2 个属于已退役的 #50 残留,官方原件里没有对应锚点,按「无法重建」如实列出。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fake reverts are gone: .dshpurge.bak was only created when missing, so a history of "patch first, backup later" stored the already-patched bytes as the baseline. revertAll then copied a patched file over itself — no error, zero diff, a success message, and every patch still in place. On the live host 10 of 54 backups were in this state; two of them prove it outright (dsh-agent-instructions/lib/index.js carries the strong sentence, dsh-tool-fs/lib/index.js carries READ_MAX_LINE_LENGTH = 1e4).
  • New baseline_lost verdict: when the backup is byte-identical to the body and the body really is patched (marker present, original pattern gone), the file is no longer counted as reverted; it is listed separately, the backup is kept as evidence, and both --revert and the panel revert say "rebuild the baseline from the official package/asar". A same-bytes backup without patch traces is still just skipped, so untouched files are not misreported.
  • --apply warns first: a read-only pre-check reports baseline-lost files before patching, instead of letting users discover it later.
  • 5 new cases in test/baseline-lost.test.js; node --test 230 passing / 0 failing.
  • On the live host 8 baselines were rebuilt from the official rc.3 package (old copies kept as *.dshpurge.bak.lost-20261116); the remaining 2 are leftovers of the retired #50 whose anchors do not exist in the official files, so they are listed as unrecoverable rather than quietly cleared.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.5

中文

  • 修掉一个比 1.2.1 更大的诊断盲区:diagnose 只认数组形态的 file,而 29 条补丁(#1/#2/#3/#21/#22/#45/#46/#82-#85…,占 35%)用字符串声明载体,整批被塞进「无文件字段 29 条」——等于这三分之一补丁从未参与漂移诊断。现在 file 字符串与数组同口径展开;真机 --diagnose 从「42 命中 + 29 无文件字段」变为 71 条可命中 / 无文件字段 0,与 --status 的 70/71 完全对齐。
  • 补齐四类特殊补丁的判定语义(否则它们一旦可见就会被误判成「目标已消失」):① 校验型(官方原文 marker,如 #2/#3)→ verified;② 带编号自标记(#88 校验 #87 的 [dsh-purge-87])→ 未出现时 verify_pending(待配套),不是漂移;③ 占位/撤销型(#14/#80 的 [dsh-purge] 自标记永不写入)→ noop;④ 迁移/恢复默认型(softSettle,如 #21)→ satisfied。新增独立桶并在诊断行里分开计数。
  • 两条被基线丢失误伤的补丁:#1 WORKSPACE_CONTEXT_INTRO、#22 READ_CAPS_RAISED 的 marker 是补丁自己写进去的人造文本(强句 / 1e4、1MB、2e4),但基线备份曾被覆盖成「已打态」,diff 为空导致诊断判成「目标已消失」。声明 markersStrict 后按「marker 在 + 原 pattern 不在」判定为已生效。
  • 多候选载体不再被一个不存在的键整条判死:#24 的 preset-unrestricted/preset-liangshen、#46 的 system-prompt-nested 在 rc.3 里根本没有,旧口径只要有一个键解析不到就整条进「载体缺失」。现在只要有一个候选能探到就按实际结论归类,缺的键记进 unresolvedKeys;#83 DESKTOP_HOST_CLOSE_NO_FATAL 的载体是桌面壳 main.js(纯 npm 树不存在),改标 optional,隔离门禁不再因此判红。
  • #24 PRESET_FETCH_ENABLED 判为默认态满足:rc.3 的 preset 原生就是 fetch: true,fetch: false 块已不存在;声明 softSettle 且要求 marker(fetch: true)同时命中,避免上游换写法时被误报成已满足。
  • status 口径:#88 在净树上由 drifted 改为 pending(待配套),revert 后不再出现「已还原却仍报漂移」的假告警。
  • 新增 test/diagnose-caliber.test.js 8 条(字符串 file、校验型两向、待配套、占位、softSettle、softSettle+markers、多候选载体、#88 净树/已打);全套 node --test 225 通过 / 0 失败;npm run cycle:isolated 六判据 OK 6/6;真机 --status 70/71、0 待清洗,--diagnose 71 命中、漂移 0、目标已消失 0、待配套 0、载体缺失 0、无文件字段 0。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • A diagnostic blind spot larger than the 1.2.1 one: diagnose only accepted array-form file, while 29 of 82 patches (35%) declared their carrier as a string (#1/#2/#3/...
Read more

v1.2.8

Choose a tag to compare

@github-actions github-actions released this 16 Nov 12:44

1.2.8

中文

  • 红队模式在 rc.3 上装不出来的根因:0.2.2 起宿主不再读取 $DSH_HOME/.agent-presets/<id>/(官方文档原话 "Nothing reads that directory any more"),预设改为 bundle 补丁里的 @deepseek-ai/dsh-agent-preset 声明行。本插件一直只写旧目录,于是模式下拉里没有红队模式,面板仍在、/redteam/api 也通——只剩入口消失。
  • installPreset 增加新版分支:宿主能解析到 @deepseek-ai/dsh-agent-preset 时,改为在挂着本插件的 profile cordis.patch.yml 追加 preset-redteam 声明行(id/name/description/order/plugins),插件列表从包内 presets/redteam/agent.cordis.yml 生成、跳过锚点门行(它是性能项,不是功能项);幂等,已存在就保留,force 才重写。旧目录安装保留给 rc.2 及更早。
  • 声明行里显示的预设名是「红队模式」;插件逐项校验:本轮只缺 dsh-workflow-worker-thread(模板里本就 disabled: true,不影响激活)。
  • 真机已按此格式写入 ~/.dsh/profiles/desktop/cordis.patch.yml(YAML 用宿主自带解析器验证:preset-redteam / plugins 18 行 / 首行 persona),重启后模式下拉应出现红队模式。
  • 全套 node --test 233 通过 / 0 失败;compat:check 一致。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Why the red-team mode never showed up on rc.3: 0.2.2 no longer reads $DSH_HOME/.agent-presets/<id>/ ("Nothing reads that directory any more" in the official docs); presets are now @deepseek-ai/dsh-agent-preset declaration rows inside bundle patches. This plugin kept writing only the legacy directory, so the console and /redteam/api stayed up while the mode vanished from the picker.
  • installPreset gained a modern branch: when the host resolves @deepseek-ai/dsh-agent-preset, it appends a preset-redteam declaration row (id/name/description/order/plugins) to the profile cordis.patch.yml that carries this plugin, generating the plugin list from the bundled presets/redteam/agent.cordis.yml and skipping the anchor-gate row (a performance knob, not a feature). Idempotent: kept when present, rewritten only with force. Legacy install remains for rc.2 and older.
  • The row shows the preset as 红队模式. Every plugin name was checked; only dsh-workflow-worker-thread is absent, and the template already ships it disabled: true, so activation is unaffected.
  • The live profile has the row now (validated with the host's own YAML parser: preset-redteam, 18 plugins, first row persona); restart makes it appear in the picker.
  • node --test 233 passing / 0 failing; compat:check clean.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.7

中文

  • 修 1.2.6 的复发口子:宿主重启会走「过期基线上刷新」自愈(refreshStaleBaksForUpgrade),原判据是「正文里没有 dsh-purge 字面量且与 bak 不同 → 认定成官方新版,把 bak 刷成正文」。但 #1/#22/#33/#48 这类补丁只改文案、不写 marker 注释,于是刚重建好的基线被下一次启动原样打回已打态——真机实测重启一次,8 个刚修好的文件全部复发成 baseline_lost(dsh-agent-instructions/lib/index.js 从 63281 字节的官方原件又变回 63970 字节的已打态)。
  • 判据换成补丁自己的:逐文件按「marker 命中 + 原 pattern 已不命中」判定正文是否已被我们写过,命中就不动 bak;只有正文确实回到官方新版(原 pattern 还在)才刷新基线。拿不到补丁索引时退回旧的字面量判据,不改变无补丁文件的行为。
  • 新增 test/upgrade-refresh.test.js 3 条:已打态无字面量不刷基线、官方新版照常刷新、applyPatches 内部刷新后再 revertAll 回到官方新版而非旧版。全套 node --test 233 通过 / 0 失败。
  • npm run cycle:isolated 在 rc.3 隔离树 OK 6/6,与 1.2.6 同口径。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Closes the 1.2.6 regression window: every host boot runs the "stale baseline refresh" self-heal (refreshStaleBaksForUpgrade), which used to decide "no dsh-purge literal + differs from bak ⇒ this is the official new version, overwrite bak with the current bytes". Patches #1/#22/#33/#48 only rewrite sentences and never write marker comments, so every freshly rebuilt baseline was reverted to the patched state on the next boot — on the live host all 8 repaired files relapsed to baseline_lost after one restart (dsh-agent-instructions/lib/index.js went from the 63281-byte official copy back to the 63970-byte patched one).
  • The refresh now uses the patch's own verdict: per file it checks "marker present + original pattern gone"; when that holds the backup is left alone, and only a file that really is the official new version gets its baseline refreshed. Without a patch index it falls back to the old literal check, so files no patch touches keep the previous behaviour.
  • 3 new cases in test/upgrade-refresh.test.js; node --test 233 passing / 0 failing. npm run cycle:isolated remains OK 6/6 on the rc.3 tree.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.6

中文

  • 修掉「假还原」:.dshpurge.bak 只在缺失时才创建,遇到「先打补丁、后才有备份」的历史状态会把已打态存成基线。此时 revertAll 把打过补丁的文件覆盖回自己:不报错、0 差异、界面显示成功,用户以为退回了官方原版,其实补丁全在。真机实测 54 个备份里 10 个是这种,其中 dsh-agent-instructions/lib/index.js、dsh-tool-fs/lib/index.js 能直接验到补丁痕迹(强句 / READ_MAX_LINE_LENGTH = 1e4)。
  • 新判定 baseline_lost:备份与本体逐字节相同、且本体确实带补丁痕迹(marker 命中 + 原 pattern 已不在)时,不再计入 reverted,单独列出并保留 bak 留证;--revert、面板「还原」都会提示「需从官方安装包/asar 重建基线」。只同字节、但没补丁痕迹的文件仍按 skipped 处理,不误报。
  • --apply 预先告警:应用前先跑一次只读检测,若存在基线丢失文件就先说明「还原回不到官方原文」,避免用户事后才发现。
  • 新增 test/baseline-lost.test.js 5 条(判丢失、真原文不丢、未打不丢、revert 不谎报、same-bytes 无痕迹仍 skipped),全套 node --test 230 通过 / 0 失败。
  • 真机已按官方 rc.3 原包重建 8 个文件的基线(旧备份留档 *.dshpurge.bak.lost-20261116),剩余 2 个属于已退役的 #50 残留,官方原件里没有对应锚点,按「无法重建」如实列出。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fake reverts are gone: .dshpurge.bak was only created when missing, so a history of "patch first, backup later" stored the already-patched bytes as the baseline. revertAll then copied a patched file over itself — no error, zero diff, a success message, and every patch still in place. On the live host 10 of 54 backups were in this state; two of them prove it outright (dsh-agent-instructions/lib/index.js carries the strong sentence, dsh-tool-fs/lib/index.js carries READ_MAX_LINE_LENGTH = 1e4).
  • New baseline_lost verdict: when the backup is byte-identical to the body and the body really is patched (marker present, original pattern gone), the file is no longer counted as reverted; it is listed separately, the backup is kept as evidence, and both --revert and the panel revert say "rebuild the baseline from the official package/asar". A same-bytes backup without patch traces is still just skipped, so untouched files are not misreported.
  • --apply warns first: a read-only pre-check reports baseline-lost files before patching, instead of letting users discover it later.
  • 5 new cases in test/baseline-lost.test.js; node --test 230 passing / 0 failing.
  • On the live host 8 baselines were rebuilt from the official rc.3 package (old copies kept as *.dshpurge.bak.lost-20261116); the remaining 2 are leftovers of the retired #50 whose anchors do not exist in the official files, so they are listed as unrecoverable rather than quietly cleared.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.5

中文

  • 修掉一个比 1.2.1 更大的诊断盲区:diagnose 只认数组形态的 file,而 29 条补丁(#1/#2/#3/#21/#22/#45/#46/#82-#85…,占 35%)用字符串声明载体,整批被塞进「无文件字段 29 条」——等于这三分之一补丁从未参与漂移诊断。现在 file 字符串与数组同口径展开;真机 --diagnose 从「42 命中 + 29 无文件字段」变为 71 条可命中 / 无文件字段 0,与 --status 的 70/71 完全对齐。
  • 补齐四类特殊补丁的判定语义(否则它们一旦可见就会被误判成「目标已消失」):① 校验型(官方原文 marker,如 #2/#3)→ verified;② 带编号自标记(#88 校验 #87 的 [dsh-purge-87])→ 未出现时 verify_pending(待配套),不是漂移;③ 占位/撤销型(#14/#80 的 [dsh-purge] 自标记永不写入)→ noop;④ 迁移/恢复默认型(softSettle,如 #21)→ satisfied。新增独立桶并在诊断行里分开计数。
  • 两条被基线丢失误伤的补丁:#1 WORKSPACE_CONTEXT_INTRO、#22 READ_CAPS_RAISED 的 marker 是补丁自己写进去的人造文本(强句 / 1e4、1MB、2e4),但基线备份曾被覆盖成「已打态」,diff 为空导致诊断判成「目标已消失」。声明 markersStrict 后按「marker 在 + 原 pattern 不在」判定为已生效。
  • 多候选载体不再被一个不存在的键整条判死:#24 的 preset-unrestricted/preset-liangshen、#46 的 system-prompt-nested 在 rc.3 里根本没有,旧口径只要有一个键解析不到就整条进「载体缺失」。现在只要有一个候选能探到就按实际结论归类,缺的键记进 unresolvedKeys;#83 DESKTOP_HOST_CLOSE_NO_FATAL 的载体是桌面壳 main.js(纯 npm 树不存在),改标 optional,隔离门禁不再因此判红。
  • #24 PRESET_FETCH_ENABLED 判为默认态满足:rc.3 的 preset 原生就是 fetch: true,fetch: false 块已不存在;声明 softSettle 且要求 marker(fetch: true)同时命中,避免上游换写法时被误报成已满足。
  • status 口径:#88 在净树上由 drifted 改为 pending(待配套),revert 后不再出现「已还原却仍报漂移」的假告警。
  • 新增 test/diagnose-caliber.test.js 8 条(字符串 file、校验型两向、待配套、占位、softSettle、softSettle+markers、多候选载体、#88 净树/已打);全套 node --test 225 通过 / 0 失败;npm run cycle:isolated 六判据 OK 6/6;真机 --status 70/71、0 待清洗,--diagnose 71 命中、漂移 0、目标已消失 0、待配套 0、载体缺失 0、无文件字段 0。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • A diagnostic blind spot larger than the 1.2.1 one: diagnose only accepted array-form file, while 29 of 82 patches (35%) declared their carrier as a string (#1/#2/#3/#21/#22/#45/#46/#82–#85, …). All of them fell into a bucket mislabelled "no file field", so a third of the patch set had never been drift-diagnosed. String and array forms are now expanded identically; live --diagnose moves from "42 matched + 29 no-file-field" to 71 matched / 0 no-file-field, exactly aligning with --status 70/71.
  • Four special patch classes now have explicit semantics (otherwise they would be misread as "target gone" the moment they became visible): verify-only official markers → verified; numbered self-markers (#88 checks #87's [dsh-purge-87]) → verify_pending when absent, not drift; placeholder/undo patches (#14/#80 [dsh-purge] never written) → noop; migration/restore-default (softSettle, #21) → satisfied. Each has its own bucket and count in the diagnostic line.
  • Two patches wrongly hit by a lost baseline: #1 WORKSPACE_CONTEXT_INTRO and #22 READ_CAPS_RAISED carry markers the patch itself writes (strong sentences / 1e4, 1MB, 2e4), but the backup had been overwritten with the applied state, leaving an empty diff and a "target gone" verdict. Declaring markersStrict makes "marker present + original pattern absent" count as applied.
  • One unresolvable candidate key no longer fails a whole patch: #24's preset-unrestricted/preset-liangshen and #46's system-prompt-nested simply do not exist in rc.3, yet the old rule sent the entire patch to "carrier missing" as soon as any key failed to resolve. Now one evaluable candidate is enough, unresolvable keys are recorded in unresolvedKeys, a...
Read more

v1.2.7

Choose a tag to compare

@github-actions github-actions released this 16 Nov 12:28

1.2.7

中文

  • 修 1.2.6 的复发口子:宿主重启会走「过期基线上刷新」自愈(refreshStaleBaksForUpgrade),原判据是「正文里没有 dsh-purge 字面量且与 bak 不同 → 认定成官方新版,把 bak 刷成正文」。但 #1/#22/#33/#48 这类补丁只改文案、不写 marker 注释,于是刚重建好的基线被下一次启动原样打回已打态——真机实测重启一次,8 个刚修好的文件全部复发成 baseline_lost(dsh-agent-instructions/lib/index.js 从 63281 字节的官方原件又变回 63970 字节的已打态)。
  • 判据换成补丁自己的:逐文件按「marker 命中 + 原 pattern 已不命中」判定正文是否已被我们写过,命中就不动 bak;只有正文确实回到官方新版(原 pattern 还在)才刷新基线。拿不到补丁索引时退回旧的字面量判据,不改变无补丁文件的行为。
  • 新增 test/upgrade-refresh.test.js 3 条:已打态无字面量不刷基线、官方新版照常刷新、applyPatches 内部刷新后再 revertAll 回到官方新版而非旧版。全套 node --test 233 通过 / 0 失败。
  • npm run cycle:isolated 在 rc.3 隔离树 OK 6/6,与 1.2.6 同口径。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Closes the 1.2.6 regression window: every host boot runs the "stale baseline refresh" self-heal (refreshStaleBaksForUpgrade), which used to decide "no dsh-purge literal + differs from bak ⇒ this is the official new version, overwrite bak with the current bytes". Patches #1/#22/#33/#48 only rewrite sentences and never write marker comments, so every freshly rebuilt baseline was reverted to the patched state on the next boot — on the live host all 8 repaired files relapsed to baseline_lost after one restart (dsh-agent-instructions/lib/index.js went from the 63281-byte official copy back to the 63970-byte patched one).
  • The refresh now uses the patch's own verdict: per file it checks "marker present + original pattern gone"; when that holds the backup is left alone, and only a file that really is the official new version gets its baseline refreshed. Without a patch index it falls back to the old literal check, so files no patch touches keep the previous behaviour.
  • 3 new cases in test/upgrade-refresh.test.js; node --test 233 passing / 0 failing. npm run cycle:isolated remains OK 6/6 on the rc.3 tree.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.6

中文

  • 修掉「假还原」:.dshpurge.bak 只在缺失时才创建,遇到「先打补丁、后才有备份」的历史状态会把已打态存成基线。此时 revertAll 把打过补丁的文件覆盖回自己:不报错、0 差异、界面显示成功,用户以为退回了官方原版,其实补丁全在。真机实测 54 个备份里 10 个是这种,其中 dsh-agent-instructions/lib/index.js、dsh-tool-fs/lib/index.js 能直接验到补丁痕迹(强句 / READ_MAX_LINE_LENGTH = 1e4)。
  • 新判定 baseline_lost:备份与本体逐字节相同、且本体确实带补丁痕迹(marker 命中 + 原 pattern 已不在)时,不再计入 reverted,单独列出并保留 bak 留证;--revert、面板「还原」都会提示「需从官方安装包/asar 重建基线」。只同字节、但没补丁痕迹的文件仍按 skipped 处理,不误报。
  • --apply 预先告警:应用前先跑一次只读检测,若存在基线丢失文件就先说明「还原回不到官方原文」,避免用户事后才发现。
  • 新增 test/baseline-lost.test.js 5 条(判丢失、真原文不丢、未打不丢、revert 不谎报、same-bytes 无痕迹仍 skipped),全套 node --test 230 通过 / 0 失败。
  • 真机已按官方 rc.3 原包重建 8 个文件的基线(旧备份留档 *.dshpurge.bak.lost-20261116),剩余 2 个属于已退役的 #50 残留,官方原件里没有对应锚点,按「无法重建」如实列出。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fake reverts are gone: .dshpurge.bak was only created when missing, so a history of "patch first, backup later" stored the already-patched bytes as the baseline. revertAll then copied a patched file over itself — no error, zero diff, a success message, and every patch still in place. On the live host 10 of 54 backups were in this state; two of them prove it outright (dsh-agent-instructions/lib/index.js carries the strong sentence, dsh-tool-fs/lib/index.js carries READ_MAX_LINE_LENGTH = 1e4).
  • New baseline_lost verdict: when the backup is byte-identical to the body and the body really is patched (marker present, original pattern gone), the file is no longer counted as reverted; it is listed separately, the backup is kept as evidence, and both --revert and the panel revert say "rebuild the baseline from the official package/asar". A same-bytes backup without patch traces is still just skipped, so untouched files are not misreported.
  • --apply warns first: a read-only pre-check reports baseline-lost files before patching, instead of letting users discover it later.
  • 5 new cases in test/baseline-lost.test.js; node --test 230 passing / 0 failing.
  • On the live host 8 baselines were rebuilt from the official rc.3 package (old copies kept as *.dshpurge.bak.lost-20261116); the remaining 2 are leftovers of the retired #50 whose anchors do not exist in the official files, so they are listed as unrecoverable rather than quietly cleared.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.5

中文

  • 修掉一个比 1.2.1 更大的诊断盲区:diagnose 只认数组形态的 file,而 29 条补丁(#1/#2/#3/#21/#22/#45/#46/#82-#85…,占 35%)用字符串声明载体,整批被塞进「无文件字段 29 条」——等于这三分之一补丁从未参与漂移诊断。现在 file 字符串与数组同口径展开;真机 --diagnose 从「42 命中 + 29 无文件字段」变为 71 条可命中 / 无文件字段 0,与 --status 的 70/71 完全对齐。
  • 补齐四类特殊补丁的判定语义(否则它们一旦可见就会被误判成「目标已消失」):① 校验型(官方原文 marker,如 #2/#3)→ verified;② 带编号自标记(#88 校验 #87 的 [dsh-purge-87])→ 未出现时 verify_pending(待配套),不是漂移;③ 占位/撤销型(#14/#80 的 [dsh-purge] 自标记永不写入)→ noop;④ 迁移/恢复默认型(softSettle,如 #21)→ satisfied。新增独立桶并在诊断行里分开计数。
  • 两条被基线丢失误伤的补丁:#1 WORKSPACE_CONTEXT_INTRO、#22 READ_CAPS_RAISED 的 marker 是补丁自己写进去的人造文本(强句 / 1e4、1MB、2e4),但基线备份曾被覆盖成「已打态」,diff 为空导致诊断判成「目标已消失」。声明 markersStrict 后按「marker 在 + 原 pattern 不在」判定为已生效。
  • 多候选载体不再被一个不存在的键整条判死:#24 的 preset-unrestricted/preset-liangshen、#46 的 system-prompt-nested 在 rc.3 里根本没有,旧口径只要有一个键解析不到就整条进「载体缺失」。现在只要有一个候选能探到就按实际结论归类,缺的键记进 unresolvedKeys;#83 DESKTOP_HOST_CLOSE_NO_FATAL 的载体是桌面壳 main.js(纯 npm 树不存在),改标 optional,隔离门禁不再因此判红。
  • #24 PRESET_FETCH_ENABLED 判为默认态满足:rc.3 的 preset 原生就是 fetch: true,fetch: false 块已不存在;声明 softSettle 且要求 marker(fetch: true)同时命中,避免上游换写法时被误报成已满足。
  • status 口径:#88 在净树上由 drifted 改为 pending(待配套),revert 后不再出现「已还原却仍报漂移」的假告警。
  • 新增 test/diagnose-caliber.test.js 8 条(字符串 file、校验型两向、待配套、占位、softSettle、softSettle+markers、多候选载体、#88 净树/已打);全套 node --test 225 通过 / 0 失败;npm run cycle:isolated 六判据 OK 6/6;真机 --status 70/71、0 待清洗,--diagnose 71 命中、漂移 0、目标已消失 0、待配套 0、载体缺失 0、无文件字段 0。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • A diagnostic blind spot larger than the 1.2.1 one: diagnose only accepted array-form file, while 29 of 82 patches (35%) declared their carrier as a string (#1/#2/#3/#21/#22/#45/#46/#82–#85, …). All of them fell into a bucket mislabelled "no file field", so a third of the patch set had never been drift-diagnosed. String and array forms are now expanded identically; live --diagnose moves from "42 matched + 29 no-file-field" to 71 matched / 0 no-file-field, exactly aligning with --status 70/71.
  • Four special patch classes now have explicit semantics (otherwise they would be misread as "target gone" the moment they became visible): verify-only official markers → verified; numbered self-markers (#88 checks #87's [dsh-purge-87]) → verify_pending when absent, not drift; placeholder/undo patches (#14/#80 [dsh-purge] never written) → noop; migration/restore-default (softSettle, #21) → satisfied. Each has its own bucket and count in the diagnostic line.
  • Two patches wrongly hit by a lost baseline: #1 WORKSPACE_CONTEXT_INTRO and #22 READ_CAPS_RAISED carry markers the patch itself writes (strong sentences / 1e4, 1MB, 2e4), but the backup had been overwritten with the applied state, leaving an empty diff and a "target gone" verdict. Declaring markersStrict makes "marker present + original pattern absent" count as applied.
  • One unresolvable candidate key no longer fails a whole patch: #24's preset-unrestricted/preset-liangshen and #46's system-prompt-nested simply do not exist in rc.3, yet the old rule sent the entire patch to "carrier missing" as soon as any key failed to resolve. Now one evaluable candidate is enough, unresolvable keys are recorded in unresolvedKeys, and #83 DESKTOP_HOST_CLOSE_NO_FATAL (carrier = desktop-shell main.js, absent from a pure npm tree) is marked optional so the isolated gate stops failing on it.
  • #24 PRESET_FETCH_ENABLED classified as default-state satisfied: rc.3 presets already ship fetch: true and the fetch: false block is gone; softSettle now requires its marker (fetch: true) to hit as well, so a rewritten upstream config cannot be misread as satisfied.
  • Status caliber: #88 flips from drifted to pending on a pristine tree (companion patch not yet run), so a completed revert no longer reports a phantom drift.
  • 8 cases in test/diagnose-caliber.test.js; node --test 225 passing / 0 failing; npm run cycle:isolated OK 6/6; live host --status 70/71 with 0 pending, --diagnose 71 matched, 0 drift, 0 dead target, 0 verify-pending, 0 missing carrier, 0 no-file-field.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.4

中文

  • 隔离全链路固化成门禁脚本:新增 npm run cycle:isolated(scripts/isolated-cycle.mjs)——在干净官方包树上跑 backupAll → apply → diagnose → revert,六条判据任一失败即退出码 1:G1 干净树(有 .dshpurge.bak/tx 残留直接拒跑)、G2 所有解析出的载体路径必须都在隔离树内(跨树副作用门禁,1.2.2 那类事故从此机器可拦)、G3 apply 0 失败且确实改写了文件、G4 drifted/partial/carrier_missing 全 0、G5 改写 .js 语法全过、G6 revert 后与原始树逐字节 0 差异。rc.3 实测 OK 6/6。
  • CLI 行为回归:树缺失 / 树有残留两种情形退出码 2 且给出铺树命令(test/isolated-cycle-cli.test.js 2 条)。
  • .tmp/ 工作区加入 .gitignore;README / README.en / llms.txt 增加门禁命令并把「离体实打」数字更新为本轮实测(58 applied / 49 改写 / 39 .js 语法全过)。
  • 全套 node --test 217 通过 / 0 失败。维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Isolated full-cycle promoted to a gate script: npm run cycle:isolated (scripts/isolated-cycle.mjs) runs backupAll → apply → diagnose → revert on a clean official package tree and exits 1 when any of six criteria fails: G1 clean tree (refuses to run on leftover .dshpurge.bak/tx files), G2 every resolved carrier path must stay inside the isolation tree (the cross-tree-side-effect gate — the 1.2.2 class of bug is now machine-catchable), G3 apply with 0 failures and real changes, G4 drifted/partial/carrier_missing all 0, G5 syntax-clean on every modified .js, G6 byte-for-byte 0 diff after revert. Measured OK 6/6 on rc.3.
  • CLI behaviou...
Read more

v1.2.6

Choose a tag to compare

@github-actions github-actions released this 16 Nov 11:52

1.2.6

中文

  • 修掉「假还原」:.dshpurge.bak 只在缺失时才创建,遇到「先打补丁、后才有备份」的历史状态会把已打态存成基线。此时 revertAll 把打过补丁的文件覆盖回自己:不报错、0 差异、界面显示成功,用户以为退回了官方原版,其实补丁全在。真机实测 54 个备份里 10 个是这种,其中 dsh-agent-instructions/lib/index.js、dsh-tool-fs/lib/index.js 能直接验到补丁痕迹(强句 / READ_MAX_LINE_LENGTH = 1e4)。
  • 新判定 baseline_lost:备份与本体逐字节相同、且本体确实带补丁痕迹(marker 命中 + 原 pattern 已不在)时,不再计入 reverted,单独列出并保留 bak 留证;--revert、面板「还原」都会提示「需从官方安装包/asar 重建基线」。只同字节、但没补丁痕迹的文件仍按 skipped 处理,不误报。
  • --apply 预先告警:应用前先跑一次只读检测,若存在基线丢失文件就先说明「还原回不到官方原文」,避免用户事后才发现。
  • 新增 test/baseline-lost.test.js 5 条(判丢失、真原文不丢、未打不丢、revert 不谎报、same-bytes 无痕迹仍 skipped),全套 node --test 230 通过 / 0 失败。
  • 真机已按官方 rc.3 原包重建 8 个文件的基线(旧备份留档 *.dshpurge.bak.lost-20261116),剩余 2 个属于已退役的 #50 残留,官方原件里没有对应锚点,按「无法重建」如实列出。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Fake reverts are gone: .dshpurge.bak was only created when missing, so a history of "patch first, backup later" stored the already-patched bytes as the baseline. revertAll then copied a patched file over itself — no error, zero diff, a success message, and every patch still in place. On the live host 10 of 54 backups were in this state; two of them prove it outright (dsh-agent-instructions/lib/index.js carries the strong sentence, dsh-tool-fs/lib/index.js carries READ_MAX_LINE_LENGTH = 1e4).
  • New baseline_lost verdict: when the backup is byte-identical to the body and the body really is patched (marker present, original pattern gone), the file is no longer counted as reverted; it is listed separately, the backup is kept as evidence, and both --revert and the panel revert say "rebuild the baseline from the official package/asar". A same-bytes backup without patch traces is still just skipped, so untouched files are not misreported.
  • --apply warns first: a read-only pre-check reports baseline-lost files before patching, instead of letting users discover it later.
  • 5 new cases in test/baseline-lost.test.js; node --test 230 passing / 0 failing.
  • On the live host 8 baselines were rebuilt from the official rc.3 package (old copies kept as *.dshpurge.bak.lost-20261116); the remaining 2 are leftovers of the retired #50 whose anchors do not exist in the official files, so they are listed as unrecoverable rather than quietly cleared.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.5

中文

  • 修掉一个比 1.2.1 更大的诊断盲区:diagnose 只认数组形态的 file,而 29 条补丁(#1/#2/#3/#21/#22/#45/#46/#82-#85…,占 35%)用字符串声明载体,整批被塞进「无文件字段 29 条」——等于这三分之一补丁从未参与漂移诊断。现在 file 字符串与数组同口径展开;真机 --diagnose 从「42 命中 + 29 无文件字段」变为 71 条可命中 / 无文件字段 0,与 --status 的 70/71 完全对齐。
  • 补齐四类特殊补丁的判定语义(否则它们一旦可见就会被误判成「目标已消失」):① 校验型(官方原文 marker,如 #2/#3)→ verified;② 带编号自标记(#88 校验 #87 的 [dsh-purge-87])→ 未出现时 verify_pending(待配套),不是漂移;③ 占位/撤销型(#14/#80 的 [dsh-purge] 自标记永不写入)→ noop;④ 迁移/恢复默认型(softSettle,如 #21)→ satisfied。新增独立桶并在诊断行里分开计数。
  • 两条被基线丢失误伤的补丁:#1 WORKSPACE_CONTEXT_INTRO、#22 READ_CAPS_RAISED 的 marker 是补丁自己写进去的人造文本(强句 / 1e4、1MB、2e4),但基线备份曾被覆盖成「已打态」,diff 为空导致诊断判成「目标已消失」。声明 markersStrict 后按「marker 在 + 原 pattern 不在」判定为已生效。
  • 多候选载体不再被一个不存在的键整条判死:#24 的 preset-unrestricted/preset-liangshen、#46 的 system-prompt-nested 在 rc.3 里根本没有,旧口径只要有一个键解析不到就整条进「载体缺失」。现在只要有一个候选能探到就按实际结论归类,缺的键记进 unresolvedKeys;#83 DESKTOP_HOST_CLOSE_NO_FATAL 的载体是桌面壳 main.js(纯 npm 树不存在),改标 optional,隔离门禁不再因此判红。
  • #24 PRESET_FETCH_ENABLED 判为默认态满足:rc.3 的 preset 原生就是 fetch: true,fetch: false 块已不存在;声明 softSettle 且要求 marker(fetch: true)同时命中,避免上游换写法时被误报成已满足。
  • status 口径:#88 在净树上由 drifted 改为 pending(待配套),revert 后不再出现「已还原却仍报漂移」的假告警。
  • 新增 test/diagnose-caliber.test.js 8 条(字符串 file、校验型两向、待配套、占位、softSettle、softSettle+markers、多候选载体、#88 净树/已打);全套 node --test 225 通过 / 0 失败;npm run cycle:isolated 六判据 OK 6/6;真机 --status 70/71、0 待清洗,--diagnose 71 命中、漂移 0、目标已消失 0、待配套 0、载体缺失 0、无文件字段 0。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • A diagnostic blind spot larger than the 1.2.1 one: diagnose only accepted array-form file, while 29 of 82 patches (35%) declared their carrier as a string (#1/#2/#3/#21/#22/#45/#46/#82–#85, …). All of them fell into a bucket mislabelled "no file field", so a third of the patch set had never been drift-diagnosed. String and array forms are now expanded identically; live --diagnose moves from "42 matched + 29 no-file-field" to 71 matched / 0 no-file-field, exactly aligning with --status 70/71.
  • Four special patch classes now have explicit semantics (otherwise they would be misread as "target gone" the moment they became visible): verify-only official markers → verified; numbered self-markers (#88 checks #87's [dsh-purge-87]) → verify_pending when absent, not drift; placeholder/undo patches (#14/#80 [dsh-purge] never written) → noop; migration/restore-default (softSettle, #21) → satisfied. Each has its own bucket and count in the diagnostic line.
  • Two patches wrongly hit by a lost baseline: #1 WORKSPACE_CONTEXT_INTRO and #22 READ_CAPS_RAISED carry markers the patch itself writes (strong sentences / 1e4, 1MB, 2e4), but the backup had been overwritten with the applied state, leaving an empty diff and a "target gone" verdict. Declaring markersStrict makes "marker present + original pattern absent" count as applied.
  • One unresolvable candidate key no longer fails a whole patch: #24's preset-unrestricted/preset-liangshen and #46's system-prompt-nested simply do not exist in rc.3, yet the old rule sent the entire patch to "carrier missing" as soon as any key failed to resolve. Now one evaluable candidate is enough, unresolvable keys are recorded in unresolvedKeys, and #83 DESKTOP_HOST_CLOSE_NO_FATAL (carrier = desktop-shell main.js, absent from a pure npm tree) is marked optional so the isolated gate stops failing on it.
  • #24 PRESET_FETCH_ENABLED classified as default-state satisfied: rc.3 presets already ship fetch: true and the fetch: false block is gone; softSettle now requires its marker (fetch: true) to hit as well, so a rewritten upstream config cannot be misread as satisfied.
  • Status caliber: #88 flips from drifted to pending on a pristine tree (companion patch not yet run), so a completed revert no longer reports a phantom drift.
  • 8 cases in test/diagnose-caliber.test.js; node --test 225 passing / 0 failing; npm run cycle:isolated OK 6/6; live host --status 70/71 with 0 pending, --diagnose 71 matched, 0 drift, 0 dead target, 0 verify-pending, 0 missing carrier, 0 no-file-field.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.4

中文

  • 隔离全链路固化成门禁脚本:新增 npm run cycle:isolated(scripts/isolated-cycle.mjs)——在干净官方包树上跑 backupAll → apply → diagnose → revert,六条判据任一失败即退出码 1:G1 干净树(有 .dshpurge.bak/tx 残留直接拒跑)、G2 所有解析出的载体路径必须都在隔离树内(跨树副作用门禁,1.2.2 那类事故从此机器可拦)、G3 apply 0 失败且确实改写了文件、G4 drifted/partial/carrier_missing 全 0、G5 改写 .js 语法全过、G6 revert 后与原始树逐字节 0 差异。rc.3 实测 OK 6/6。
  • CLI 行为回归:树缺失 / 树有残留两种情形退出码 2 且给出铺树命令(test/isolated-cycle-cli.test.js 2 条)。
  • .tmp/ 工作区加入 .gitignore;README / README.en / llms.txt 增加门禁命令并把「离体实打」数字更新为本轮实测(58 applied / 49 改写 / 39 .js 语法全过)。
  • 全套 node --test 217 通过 / 0 失败。维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Isolated full-cycle promoted to a gate script: npm run cycle:isolated (scripts/isolated-cycle.mjs) runs backupAll → apply → diagnose → revert on a clean official package tree and exits 1 when any of six criteria fails: G1 clean tree (refuses to run on leftover .dshpurge.bak/tx files), G2 every resolved carrier path must stay inside the isolation tree (the cross-tree-side-effect gate — the 1.2.2 class of bug is now machine-catchable), G3 apply with 0 failures and real changes, G4 drifted/partial/carrier_missing all 0, G5 syntax-clean on every modified .js, G6 byte-for-byte 0 diff after revert. Measured OK 6/6 on rc.3.
  • CLI behaviour regression: missing tree / dirty tree exit 2 with the staging command (test/isolated-cycle-cli.test.js, 2 cases).
  • .tmp/ workspace gitignored; README / README.en / llms.txt document the gate command and refresh the isolated-apply numbers to this round's measurements.
  • node --test 217 passing / 0 failing. Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.3

中文

  • 1.2.2 的桌面壳作用域修复补上 macOS 布局:shellMainForAiBase 改为按「@deepseek-ai 上四级目录名 = resources(大小写不敏感)」定位,Windows resources 与 macOS Contents/Resources 都只跟随本树,仍不会跨树。
  • 新增第 3 条范围回归(macOS Contents/Resources 布局);全套 node --test 215 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Extends the 1.2.2 desktop-shell scope fix to the macOS layout: shellMainForAiBase now derives the resources directory four levels above @deepseek-ai and accepts either resources or Contents/Resources (case-insensitive), still never crossing trees.
  • Adds a third scope regression (macOS Contents/Resources); node --test 215 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.2

中文

  • 修掉跨树副作用:targetFiles 里的 desktop-shell 以前无条件取「当前官方桌面安装」的 resources/app/lib/main.js,于对隔离树 / 归档 base 跑 revert 时,会把真实宿主的桌面壳补丁一起摘掉。现在只跟随 aiBase 自己所属的 resources/app,不命中就留空。
  • 离体全周期验证(这次补上的证据):在干净 rc.3 npm 隔离树里 backupAll → apply → status → diagnose → revert;apply 58 条、0 失败、49 个文件改写、diagnose 42 命中且漂移 0、语法 0 错;revert 49 个文件后与原始树逐字节比对 0 差异,真实宿主桌面壳 marker 前后均为 9(未被触碰)。
  • 回归:新增 test/target-scope.test.js 2 条(桌面壳只跟随自己的 resources/app;隔离包布局不挂真实桌面壳);全套 node --test 214 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Cross-tree side effect fixed: desktop-shell in targetFiles used to resolve the currently installed official desktop resources/app/lib/main.js regardless of aiBase, so a revert on an isolation tree or...
Read more

v1.2.5

Choose a tag to compare

@github-actions github-actions released this 16 Nov 08:23

1.2.5

中文

  • 修掉一个比 1.2.1 更大的诊断盲区:diagnose 只认数组形态的 file,而 29 条补丁(#1/#2/#3/#21/#22/#45/#46/#82-#85…,占 35%)用字符串声明载体,整批被塞进「无文件字段 29 条」——等于这三分之一补丁从未参与漂移诊断。现在 file 字符串与数组同口径展开;真机 --diagnose 从「42 命中 + 29 无文件字段」变为 71 条可命中 / 无文件字段 0,与 --status 的 70/71 完全对齐。
  • 补齐四类特殊补丁的判定语义(否则它们一旦可见就会被误判成「目标已消失」):① 校验型(官方原文 marker,如 #2/#3)→ verified;② 带编号自标记(#88 校验 #87 的 [dsh-purge-87])→ 未出现时 verify_pending(待配套),不是漂移;③ 占位/撤销型(#14/#80 的 [dsh-purge] 自标记永不写入)→ noop;④ 迁移/恢复默认型(softSettle,如 #21)→ satisfied。新增独立桶并在诊断行里分开计数。
  • 两条被基线丢失误伤的补丁:#1 WORKSPACE_CONTEXT_INTRO、#22 READ_CAPS_RAISED 的 marker 是补丁自己写进去的人造文本(强句 / 1e4、1MB、2e4),但基线备份曾被覆盖成「已打态」,diff 为空导致诊断判成「目标已消失」。声明 markersStrict 后按「marker 在 + 原 pattern 不在」判定为已生效。
  • 多候选载体不再被一个不存在的键整条判死:#24 的 preset-unrestricted/preset-liangshen、#46 的 system-prompt-nested 在 rc.3 里根本没有,旧口径只要有一个键解析不到就整条进「载体缺失」。现在只要有一个候选能探到就按实际结论归类,缺的键记进 unresolvedKeys;#83 DESKTOP_HOST_CLOSE_NO_FATAL 的载体是桌面壳 main.js(纯 npm 树不存在),改标 optional,隔离门禁不再因此判红。
  • #24 PRESET_FETCH_ENABLED 判为默认态满足:rc.3 的 preset 原生就是 fetch: true,fetch: false 块已不存在;声明 softSettle 且要求 marker(fetch: true)同时命中,避免上游换写法时被误报成已满足。
  • status 口径:#88 在净树上由 drifted 改为 pending(待配套),revert 后不再出现「已还原却仍报漂移」的假告警。
  • 新增 test/diagnose-caliber.test.js 8 条(字符串 file、校验型两向、待配套、占位、softSettle、softSettle+markers、多候选载体、#88 净树/已打);全套 node --test 225 通过 / 0 失败;npm run cycle:isolated 六判据 OK 6/6;真机 --status 70/71、0 待清洗,--diagnose 71 命中、漂移 0、目标已消失 0、待配套 0、载体缺失 0、无文件字段 0。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • A diagnostic blind spot larger than the 1.2.1 one: diagnose only accepted array-form file, while 29 of 82 patches (35%) declared their carrier as a string (#1/#2/#3/#21/#22/#45/#46/#82–#85, …). All of them fell into a bucket mislabelled "no file field", so a third of the patch set had never been drift-diagnosed. String and array forms are now expanded identically; live --diagnose moves from "42 matched + 29 no-file-field" to 71 matched / 0 no-file-field, exactly aligning with --status 70/71.
  • Four special patch classes now have explicit semantics (otherwise they would be misread as "target gone" the moment they became visible): verify-only official markers → verified; numbered self-markers (#88 checks #87's [dsh-purge-87]) → verify_pending when absent, not drift; placeholder/undo patches (#14/#80 [dsh-purge] never written) → noop; migration/restore-default (softSettle, #21) → satisfied. Each has its own bucket and count in the diagnostic line.
  • Two patches wrongly hit by a lost baseline: #1 WORKSPACE_CONTEXT_INTRO and #22 READ_CAPS_RAISED carry markers the patch itself writes (strong sentences / 1e4, 1MB, 2e4), but the backup had been overwritten with the applied state, leaving an empty diff and a "target gone" verdict. Declaring markersStrict makes "marker present + original pattern absent" count as applied.
  • One unresolvable candidate key no longer fails a whole patch: #24's preset-unrestricted/preset-liangshen and #46's system-prompt-nested simply do not exist in rc.3, yet the old rule sent the entire patch to "carrier missing" as soon as any key failed to resolve. Now one evaluable candidate is enough, unresolvable keys are recorded in unresolvedKeys, and #83 DESKTOP_HOST_CLOSE_NO_FATAL (carrier = desktop-shell main.js, absent from a pure npm tree) is marked optional so the isolated gate stops failing on it.
  • #24 PRESET_FETCH_ENABLED classified as default-state satisfied: rc.3 presets already ship fetch: true and the fetch: false block is gone; softSettle now requires its marker (fetch: true) to hit as well, so a rewritten upstream config cannot be misread as satisfied.
  • Status caliber: #88 flips from drifted to pending on a pristine tree (companion patch not yet run), so a completed revert no longer reports a phantom drift.
  • 8 cases in test/diagnose-caliber.test.js; node --test 225 passing / 0 failing; npm run cycle:isolated OK 6/6; live host --status 70/71 with 0 pending, --diagnose 71 matched, 0 drift, 0 dead target, 0 verify-pending, 0 missing carrier, 0 no-file-field.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.4

中文

  • 隔离全链路固化成门禁脚本:新增 npm run cycle:isolated(scripts/isolated-cycle.mjs)——在干净官方包树上跑 backupAll → apply → diagnose → revert,六条判据任一失败即退出码 1:G1 干净树(有 .dshpurge.bak/tx 残留直接拒跑)、G2 所有解析出的载体路径必须都在隔离树内(跨树副作用门禁,1.2.2 那类事故从此机器可拦)、G3 apply 0 失败且确实改写了文件、G4 drifted/partial/carrier_missing 全 0、G5 改写 .js 语法全过、G6 revert 后与原始树逐字节 0 差异。rc.3 实测 OK 6/6。
  • CLI 行为回归:树缺失 / 树有残留两种情形退出码 2 且给出铺树命令(test/isolated-cycle-cli.test.js 2 条)。
  • .tmp/ 工作区加入 .gitignore;README / README.en / llms.txt 增加门禁命令并把「离体实打」数字更新为本轮实测(58 applied / 49 改写 / 39 .js 语法全过)。
  • 全套 node --test 217 通过 / 0 失败。维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Isolated full-cycle promoted to a gate script: npm run cycle:isolated (scripts/isolated-cycle.mjs) runs backupAll → apply → diagnose → revert on a clean official package tree and exits 1 when any of six criteria fails: G1 clean tree (refuses to run on leftover .dshpurge.bak/tx files), G2 every resolved carrier path must stay inside the isolation tree (the cross-tree-side-effect gate — the 1.2.2 class of bug is now machine-catchable), G3 apply with 0 failures and real changes, G4 drifted/partial/carrier_missing all 0, G5 syntax-clean on every modified .js, G6 byte-for-byte 0 diff after revert. Measured OK 6/6 on rc.3.
  • CLI behaviour regression: missing tree / dirty tree exit 2 with the staging command (test/isolated-cycle-cli.test.js, 2 cases).
  • .tmp/ workspace gitignored; README / README.en / llms.txt document the gate command and refresh the isolated-apply numbers to this round's measurements.
  • node --test 217 passing / 0 failing. Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.3

中文

  • 1.2.2 的桌面壳作用域修复补上 macOS 布局:shellMainForAiBase 改为按「@deepseek-ai 上四级目录名 = resources(大小写不敏感)」定位,Windows resources 与 macOS Contents/Resources 都只跟随本树,仍不会跨树。
  • 新增第 3 条范围回归(macOS Contents/Resources 布局);全套 node --test 215 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Extends the 1.2.2 desktop-shell scope fix to the macOS layout: shellMainForAiBase now derives the resources directory four levels above @deepseek-ai and accepts either resources or Contents/Resources (case-insensitive), still never crossing trees.
  • Adds a third scope regression (macOS Contents/Resources); node --test 215 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.2

中文

  • 修掉跨树副作用:targetFiles 里的 desktop-shell 以前无条件取「当前官方桌面安装」的 resources/app/lib/main.js,于对隔离树 / 归档 base 跑 revert 时,会把真实宿主的桌面壳补丁一起摘掉。现在只跟随 aiBase 自己所属的 resources/app,不命中就留空。
  • 离体全周期验证(这次补上的证据):在干净 rc.3 npm 隔离树里 backupAll → apply → status → diagnose → revert;apply 58 条、0 失败、49 个文件改写、diagnose 42 命中且漂移 0、语法 0 错;revert 49 个文件后与原始树逐字节比对 0 差异,真实宿主桌面壳 marker 前后均为 9(未被触碰)。
  • 回归:新增 test/target-scope.test.js 2 条(桌面壳只跟随自己的 resources/app;隔离包布局不挂真实桌面壳);全套 node --test 214 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Cross-tree side effect fixed: desktop-shell in targetFiles used to resolve the currently installed official desktop resources/app/lib/main.js regardless of aiBase, so a revert on an isolation tree or archived base could strip the live host's shell patches. It now follows only the resources/app that owns the given aiBase.
  • Full isolated cycle: on a clean rc.3 npm tree, backupAll → apply → status → diagnose → revert: 58 applied / 0 failed / 49 files changed, diagnosis 42 hits with 0 drift, 0 syntax errors; after revert, byte-for-byte comparison against the pristine tree showed 0 differences, and the live desktop shell markers stayed at 9.
  • Tests: 2 new scope cases; node --test 214 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.1

中文

  • 修掉一处静默漏打:#57 DEEPSEEK_SYSTEM_ROLE 在 rc.3 的 openai-responses-shared.js 里,变量名是 instructionRole 且带可选链(compat?.supportsDeveloperRole !== false),旧 pattern 只认 role,导致这个文件一直没打上、诊断却因另一份文件已打而整条报 clean。现已补上该候选锚点,真机两份副本(宿主 + profile)都已改写并带 allowlist marker,node --check 通过。
  • 诊断补上老式补丁的命中数:probeCarrier 此前只读 replacements,7 条用 patterns + replace 老式声明的补丁在诊断里永远 0 命中、只能靠 dsh-purge 字样蒙混。现在与新式补丁同口径展开,上游再改这几条能直接报「可命中/漂移」。
  • 事务临时文件清理:写盘失败时不再把 .dshpurge.tx-* 临时文件留在宿主目录。
  • 验证:新增 3 条回归(instructionRole 锚点命中、重复执行不再改写、老式 patterns 计数);全套 node --test 212 通过 / 0 失败,compat:check 一致;真机连打两次 apply,62 个载体文件哈希零变化(第二次 applied 0),确认幂等。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • One silent miss fixed: patch #57 DEEPSEEK_SYSTEM_ROLE uses instructionRole with optional chaining in rc.3's openai-responses-shared.js; the old pattern only matched role, so that file was never patched while diagnosis called the patch clean because its sibling file was. A candidate anchor is now included; both the host and profile copies carry the allowlist marker and pass node --check.
  • Legacy patches now report hit counts: probeCarrier only read replacements, so the 7 patches declared with patterns + replace always showed zero hits. It now expands them the same way as modern patches.
  • Transaction temp cleanup: failed writes no longer leave .dshpurge.tx-* files in the host tree.
  • Verification: 3 new regression tests; node --test 212 passing / 0 failing; two consecutive live applies changed zero bytes across 62 carrier files (second apply: applied 0).
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.0

中文

  • 本地重锚覆盖(用户可撤销):$DSH_HOME/dsh-purge/anchor-overrides.json 允许给某条补丁钉一个新锚点,`app...
Read more

v1.2.4

Choose a tag to compare

@github-actions github-actions released this 16 Nov 07:56

1.2.4

中文

  • 隔离全链路固化成门禁脚本:新增 npm run cycle:isolated(scripts/isolated-cycle.mjs)——在干净官方包树上跑 backupAll → apply → diagnose → revert,六条判据任一失败即退出码 1:G1 干净树(有 .dshpurge.bak/tx 残留直接拒跑)、G2 所有解析出的载体路径必须都在隔离树内(跨树副作用门禁,1.2.2 那类事故从此机器可拦)、G3 apply 0 失败且确实改写了文件、G4 drifted/partial/carrier_missing 全 0、G5 改写 .js 语法全过、G6 revert 后与原始树逐字节 0 差异。rc.3 实测 OK 6/6。
  • CLI 行为回归:树缺失 / 树有残留两种情形退出码 2 且给出铺树命令(test/isolated-cycle-cli.test.js 2 条)。
  • .tmp/ 工作区加入 .gitignore;README / README.en / llms.txt 增加门禁命令并把「离体实打」数字更新为本轮实测(58 applied / 49 改写 / 39 .js 语法全过)。
  • 全套 node --test 217 通过 / 0 失败。维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Isolated full-cycle promoted to a gate script: npm run cycle:isolated (scripts/isolated-cycle.mjs) runs backupAll → apply → diagnose → revert on a clean official package tree and exits 1 when any of six criteria fails: G1 clean tree (refuses to run on leftover .dshpurge.bak/tx files), G2 every resolved carrier path must stay inside the isolation tree (the cross-tree-side-effect gate — the 1.2.2 class of bug is now machine-catchable), G3 apply with 0 failures and real changes, G4 drifted/partial/carrier_missing all 0, G5 syntax-clean on every modified .js, G6 byte-for-byte 0 diff after revert. Measured OK 6/6 on rc.3.
  • CLI behaviour regression: missing tree / dirty tree exit 2 with the staging command (test/isolated-cycle-cli.test.js, 2 cases).
  • .tmp/ workspace gitignored; README / README.en / llms.txt document the gate command and refresh the isolated-apply numbers to this round's measurements.
  • node --test 217 passing / 0 failing. Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.3

中文

  • 1.2.2 的桌面壳作用域修复补上 macOS 布局:shellMainForAiBase 改为按「@deepseek-ai 上四级目录名 = resources(大小写不敏感)」定位,Windows resources 与 macOS Contents/Resources 都只跟随本树,仍不会跨树。
  • 新增第 3 条范围回归(macOS Contents/Resources 布局);全套 node --test 215 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Extends the 1.2.2 desktop-shell scope fix to the macOS layout: shellMainForAiBase now derives the resources directory four levels above @deepseek-ai and accepts either resources or Contents/Resources (case-insensitive), still never crossing trees.
  • Adds a third scope regression (macOS Contents/Resources); node --test 215 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.2

中文

  • 修掉跨树副作用:targetFiles 里的 desktop-shell 以前无条件取「当前官方桌面安装」的 resources/app/lib/main.js,于对隔离树 / 归档 base 跑 revert 时,会把真实宿主的桌面壳补丁一起摘掉。现在只跟随 aiBase 自己所属的 resources/app,不命中就留空。
  • 离体全周期验证(这次补上的证据):在干净 rc.3 npm 隔离树里 backupAll → apply → status → diagnose → revert;apply 58 条、0 失败、49 个文件改写、diagnose 42 命中且漂移 0、语法 0 错;revert 49 个文件后与原始树逐字节比对 0 差异,真实宿主桌面壳 marker 前后均为 9(未被触碰)。
  • 回归:新增 test/target-scope.test.js 2 条(桌面壳只跟随自己的 resources/app;隔离包布局不挂真实桌面壳);全套 node --test 214 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Cross-tree side effect fixed: desktop-shell in targetFiles used to resolve the currently installed official desktop resources/app/lib/main.js regardless of aiBase, so a revert on an isolation tree or archived base could strip the live host's shell patches. It now follows only the resources/app that owns the given aiBase.
  • Full isolated cycle: on a clean rc.3 npm tree, backupAll → apply → status → diagnose → revert: 58 applied / 0 failed / 49 files changed, diagnosis 42 hits with 0 drift, 0 syntax errors; after revert, byte-for-byte comparison against the pristine tree showed 0 differences, and the live desktop shell markers stayed at 9.
  • Tests: 2 new scope cases; node --test 214 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.1

中文

  • 修掉一处静默漏打:#57 DEEPSEEK_SYSTEM_ROLE 在 rc.3 的 openai-responses-shared.js 里,变量名是 instructionRole 且带可选链(compat?.supportsDeveloperRole !== false),旧 pattern 只认 role,导致这个文件一直没打上、诊断却因另一份文件已打而整条报 clean。现已补上该候选锚点,真机两份副本(宿主 + profile)都已改写并带 allowlist marker,node --check 通过。
  • 诊断补上老式补丁的命中数:probeCarrier 此前只读 replacements,7 条用 patterns + replace 老式声明的补丁在诊断里永远 0 命中、只能靠 dsh-purge 字样蒙混。现在与新式补丁同口径展开,上游再改这几条能直接报「可命中/漂移」。
  • 事务临时文件清理:写盘失败时不再把 .dshpurge.tx-* 临时文件留在宿主目录。
  • 验证:新增 3 条回归(instructionRole 锚点命中、重复执行不再改写、老式 patterns 计数);全套 node --test 212 通过 / 0 失败,compat:check 一致;真机连打两次 apply,62 个载体文件哈希零变化(第二次 applied 0),确认幂等。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • One silent miss fixed: patch #57 DEEPSEEK_SYSTEM_ROLE uses instructionRole with optional chaining in rc.3's openai-responses-shared.js; the old pattern only matched role, so that file was never patched while diagnosis called the patch clean because its sibling file was. A candidate anchor is now included; both the host and profile copies carry the allowlist marker and pass node --check.
  • Legacy patches now report hit counts: probeCarrier only read replacements, so the 7 patches declared with patterns + replace always showed zero hits. It now expands them the same way as modern patches.
  • Transaction temp cleanup: failed writes no longer leave .dshpurge.tx-* files in the host tree.
  • Verification: 3 new regression tests; node --test 212 passing / 0 failing; two consecutive live applies changed zero bytes across 62 carrier files (second apply: applied 0).
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.0

中文

  • 本地重锚覆盖(用户可撤销):$DSH_HOME/dsh-purge/anchor-overrides.json 允许给某条补丁钉一个新锚点,apply / status / diagnose 共用同一份覆盖;覆盖项追加为候选,旧锚点 0 命中不再按命中数契约拒写,候选自带 expect 时仍严格校验。坏 JSON 只报错不生效,删文件即恢复;--anchors 只读查看,--diagnose 出 ⚓ 本地重锚 N 条生效。
  • fail-closed 补丁写盘:一条补丁的全部目标文件先在内存算完,提交前逐文件校验内容指纹,再原子写入;任一文件失败即把已写文件回滚到原文,报 write_error:…:rolled_back_N,不再出现「第 1 个文件改了、第 2 个失败」的半打状态。
  • 候选锚点表与多版本声明:replacements[].pattern 支持数组(旧版/新版原文二选一,唯一命中才替换,两版同时存在则整条拒写);package.json 新增 dshTargets,可声明多个已验证宿主版本,命中其一即适配;诊断与探测同口径展开候选。
  • 回归测试:test/anchor-overrides.test.js 6 条 + test/transaction.test.js 6 条(事务提交/指纹拦改/失败回滚、候选命中/歧义拒写、多版本命中),全套 node --test 209 通过 / 0 失败,npm run compat:check 三者一致。
  • 真机:无覆盖文件时行为不变(--status 70/71、--diagnose 42 命中、partial 0);--apply 走事务路径后数字不变、无 write_error。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Local re-anchor overrides: $DSH_HOME/dsh-purge/anchor-overrides.json lets you pin a new anchor for one patch; apply, status and diagnose read the same file. Overrides are appended as candidates, a zero-hit original no longer trips the hit-count contract, and an override that declares expect is still checked strictly. Broken JSON reports an error without taking effect; delete the file to restore. --anchors lists entries, --diagnose prints ⚓ 本地重锚 N 条生效.
  • Fail-closed patch writes: all target files of one patch are computed in memory first, fingerprinted before commit, then written atomically. Any failure rolls the already-written files back to their original text and reports write_error:…:rolled_back_N, so a patch is never half-applied.
  • Candidate anchor table and multi-version declaration: pattern in replacements accepts an array (old/new upstream text, replace only when exactly one is present, refuse when both are); package.json gains dshTargets so several verified host versions can be declared, any match counts as supported.
  • Tests: 6 anchor-override cases + 6 transaction cases (commit, fingerprint guard, rollback; candidate hit/ambiguity; multi-version match); node --test 209 passing / 0 failing, npm run compat:check consistent.
  • Maintainer qiuqianyi66. Derivative work of the original project by YuJunZhiXue, Apache-2.0, LICENSE/NOTICE retained.

1.1.91

中文

  • 修掉「status 说待清洗、diagnose 说没问题」的口径打架:同一个载体键落到多份副本时(宿主自带 pi-ai 与 profile 里装的 pi-ai 各一份 anthropic-messages.js / openai-completions.js),--status 按文件计数报 pending,--diagnose 按「任一文件已生效」把整条判成 clean。现在诊断新增「部分副本未清洗」桶与每文件证据(partial / pending_files),两边口径归一;建议文案直接说“跑一次 apply 补齐,pattern 不用改”。
  • 真机复现与修复:#47 ANTHROPIC_OAUTH_NO_CLAUDE_IDENTITY、#49 ANTHROPIC_OAUTH_CLI_NEUTRAL、#57 DEEPSEEK_SYSTEM_ROLE 三条因 profile 侧 pi-ai 升级到 1.5.2、原文又匹配旧 pattern 而挂 pending;--apply 补齐后 --status = 70/71 已清洗、0 待清洗,--diagnose = 42 命中、partial 0、漂移 0。两份 pi-ai 文件 node --check 通过,原文件留 .dshpurge.bak。
  • 回归测试:新增 test/diagnose-partial.test.js 三条用例(一打一原文 → partial;两份都打 → clean;单文件 → 旧行为不变),全套 node --test 197 通过 / 0 失败。
  • 文档:中英 README 与 llms.txt 的证据数字同步为 42 命中 + 197 测试 + partial 0,FAQ 补上「部分副本未清洗」的判读方式。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Fixes the status-vs-diagnose contradiction: when one carrier key maps to several copies (the host-bundled pi-ai and the pi-ai plugin installed under a profile each ship anthropic-messages.js / openai-completions.js), --status counted per file and reported pending while --diagnose took the strongest evidence and called the whole patch clean. Diagnosis now carries a dedicated partially-patched copies bucket with per-file evidence (partial / pending_files) so both commands agree, and the advice says exactly what to do: run apply once, no pattern change needed.
  • Reproduced and fixed on a real host: #47, #49 and #57 sat in pending because the profile-side pi-ai had been upgraded to 1.5.2 and matched the old patterns again; one --apply caught it up — --status now reports 70/71 applied with 0 pending, --diagnose reports 42 matched, 0 partially-patched, 0 drift. Both patched pi-ai files pass node --check and keep their .dshpurge.bak originals.
  • Regression tests: test/diagnose-partial.test.js covers half-patched → partial, fully patched → clean, and single-file → unchange...
Read more

v1.2.3

Choose a tag to compare

@github-actions github-actions released this 16 Nov 07:41

1.2.3

中文

  • 1.2.2 的桌面壳作用域修复补上 macOS 布局:shellMainForAiBase 改为按「@deepseek-ai 上四级目录名 = resources(大小写不敏感)」定位,Windows resources 与 macOS Contents/Resources 都只跟随本树,仍不会跨树。
  • 新增第 3 条范围回归(macOS Contents/Resources 布局);全套 node --test 215 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Extends the 1.2.2 desktop-shell scope fix to the macOS layout: shellMainForAiBase now derives the resources directory four levels above @deepseek-ai and accepts either resources or Contents/Resources (case-insensitive), still never crossing trees.
  • Adds a third scope regression (macOS Contents/Resources); node --test 215 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.2

中文

  • 修掉跨树副作用:targetFiles 里的 desktop-shell 以前无条件取「当前官方桌面安装」的 resources/app/lib/main.js,于对隔离树 / 归档 base 跑 revert 时,会把真实宿主的桌面壳补丁一起摘掉。现在只跟随 aiBase 自己所属的 resources/app,不命中就留空。
  • 离体全周期验证(这次补上的证据):在干净 rc.3 npm 隔离树里 backupAll → apply → status → diagnose → revert;apply 58 条、0 失败、49 个文件改写、diagnose 42 命中且漂移 0、语法 0 错;revert 49 个文件后与原始树逐字节比对 0 差异,真实宿主桌面壳 marker 前后均为 9(未被触碰)。
  • 回归:新增 test/target-scope.test.js 2 条(桌面壳只跟随自己的 resources/app;隔离包布局不挂真实桌面壳);全套 node --test 214 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Cross-tree side effect fixed: desktop-shell in targetFiles used to resolve the currently installed official desktop resources/app/lib/main.js regardless of aiBase, so a revert on an isolation tree or archived base could strip the live host's shell patches. It now follows only the resources/app that owns the given aiBase.
  • Full isolated cycle: on a clean rc.3 npm tree, backupAll → apply → status → diagnose → revert: 58 applied / 0 failed / 49 files changed, diagnosis 42 hits with 0 drift, 0 syntax errors; after revert, byte-for-byte comparison against the pristine tree showed 0 differences, and the live desktop shell markers stayed at 9.
  • Tests: 2 new scope cases; node --test 214 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.1

中文

  • 修掉一处静默漏打:#57 DEEPSEEK_SYSTEM_ROLE 在 rc.3 的 openai-responses-shared.js 里,变量名是 instructionRole 且带可选链(compat?.supportsDeveloperRole !== false),旧 pattern 只认 role,导致这个文件一直没打上、诊断却因另一份文件已打而整条报 clean。现已补上该候选锚点,真机两份副本(宿主 + profile)都已改写并带 allowlist marker,node --check 通过。
  • 诊断补上老式补丁的命中数:probeCarrier 此前只读 replacements,7 条用 patterns + replace 老式声明的补丁在诊断里永远 0 命中、只能靠 dsh-purge 字样蒙混。现在与新式补丁同口径展开,上游再改这几条能直接报「可命中/漂移」。
  • 事务临时文件清理:写盘失败时不再把 .dshpurge.tx-* 临时文件留在宿主目录。
  • 验证:新增 3 条回归(instructionRole 锚点命中、重复执行不再改写、老式 patterns 计数);全套 node --test 212 通过 / 0 失败,compat:check 一致;真机连打两次 apply,62 个载体文件哈希零变化(第二次 applied 0),确认幂等。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • One silent miss fixed: patch #57 DEEPSEEK_SYSTEM_ROLE uses instructionRole with optional chaining in rc.3's openai-responses-shared.js; the old pattern only matched role, so that file was never patched while diagnosis called the patch clean because its sibling file was. A candidate anchor is now included; both the host and profile copies carry the allowlist marker and pass node --check.
  • Legacy patches now report hit counts: probeCarrier only read replacements, so the 7 patches declared with patterns + replace always showed zero hits. It now expands them the same way as modern patches.
  • Transaction temp cleanup: failed writes no longer leave .dshpurge.tx-* files in the host tree.
  • Verification: 3 new regression tests; node --test 212 passing / 0 failing; two consecutive live applies changed zero bytes across 62 carrier files (second apply: applied 0).
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.0

中文

  • 本地重锚覆盖(用户可撤销):$DSH_HOME/dsh-purge/anchor-overrides.json 允许给某条补丁钉一个新锚点,apply / status / diagnose 共用同一份覆盖;覆盖项追加为候选,旧锚点 0 命中不再按命中数契约拒写,候选自带 expect 时仍严格校验。坏 JSON 只报错不生效,删文件即恢复;--anchors 只读查看,--diagnose 出 ⚓ 本地重锚 N 条生效。
  • fail-closed 补丁写盘:一条补丁的全部目标文件先在内存算完,提交前逐文件校验内容指纹,再原子写入;任一文件失败即把已写文件回滚到原文,报 write_error:…:rolled_back_N,不再出现「第 1 个文件改了、第 2 个失败」的半打状态。
  • 候选锚点表与多版本声明:replacements[].pattern 支持数组(旧版/新版原文二选一,唯一命中才替换,两版同时存在则整条拒写);package.json 新增 dshTargets,可声明多个已验证宿主版本,命中其一即适配;诊断与探测同口径展开候选。
  • 回归测试:test/anchor-overrides.test.js 6 条 + test/transaction.test.js 6 条(事务提交/指纹拦改/失败回滚、候选命中/歧义拒写、多版本命中),全套 node --test 209 通过 / 0 失败,npm run compat:check 三者一致。
  • 真机:无覆盖文件时行为不变(--status 70/71、--diagnose 42 命中、partial 0);--apply 走事务路径后数字不变、无 write_error。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Local re-anchor overrides: $DSH_HOME/dsh-purge/anchor-overrides.json lets you pin a new anchor for one patch; apply, status and diagnose read the same file. Overrides are appended as candidates, a zero-hit original no longer trips the hit-count contract, and an override that declares expect is still checked strictly. Broken JSON reports an error without taking effect; delete the file to restore. --anchors lists entries, --diagnose prints ⚓ 本地重锚 N 条生效.
  • Fail-closed patch writes: all target files of one patch are computed in memory first, fingerprinted before commit, then written atomically. Any failure rolls the already-written files back to their original text and reports write_error:…:rolled_back_N, so a patch is never half-applied.
  • Candidate anchor table and multi-version declaration: pattern in replacements accepts an array (old/new upstream text, replace only when exactly one is present, refuse when both are); package.json gains dshTargets so several verified host versions can be declared, any match counts as supported.
  • Tests: 6 anchor-override cases + 6 transaction cases (commit, fingerprint guard, rollback; candidate hit/ambiguity; multi-version match); node --test 209 passing / 0 failing, npm run compat:check consistent.
  • Maintainer qiuqianyi66. Derivative work of the original project by YuJunZhiXue, Apache-2.0, LICENSE/NOTICE retained.

1.1.91

中文

  • 修掉「status 说待清洗、diagnose 说没问题」的口径打架:同一个载体键落到多份副本时(宿主自带 pi-ai 与 profile 里装的 pi-ai 各一份 anthropic-messages.js / openai-completions.js),--status 按文件计数报 pending,--diagnose 按「任一文件已生效」把整条判成 clean。现在诊断新增「部分副本未清洗」桶与每文件证据(partial / pending_files),两边口径归一;建议文案直接说“跑一次 apply 补齐,pattern 不用改”。
  • 真机复现与修复:#47 ANTHROPIC_OAUTH_NO_CLAUDE_IDENTITY、#49 ANTHROPIC_OAUTH_CLI_NEUTRAL、#57 DEEPSEEK_SYSTEM_ROLE 三条因 profile 侧 pi-ai 升级到 1.5.2、原文又匹配旧 pattern 而挂 pending;--apply 补齐后 --status = 70/71 已清洗、0 待清洗,--diagnose = 42 命中、partial 0、漂移 0。两份 pi-ai 文件 node --check 通过,原文件留 .dshpurge.bak。
  • 回归测试:新增 test/diagnose-partial.test.js 三条用例(一打一原文 → partial;两份都打 → clean;单文件 → 旧行为不变),全套 node --test 197 通过 / 0 失败。
  • 文档:中英 README 与 llms.txt 的证据数字同步为 42 命中 + 197 测试 + partial 0,FAQ 补上「部分副本未清洗」的判读方式。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Fixes the status-vs-diagnose contradiction: when one carrier key maps to several copies (the host-bundled pi-ai and the pi-ai plugin installed under a profile each ship anthropic-messages.js / openai-completions.js), --status counted per file and reported pending while --diagnose took the strongest evidence and called the whole patch clean. Diagnosis now carries a dedicated partially-patched copies bucket with per-file evidence (partial / pending_files) so both commands agree, and the advice says exactly what to do: run apply once, no pattern change needed.
  • Reproduced and fixed on a real host: #47, #49 and #57 sat in pending because the profile-side pi-ai had been upgraded to 1.5.2 and matched the old patterns again; one --apply caught it up — --status now reports 70/71 applied with 0 pending, --diagnose reports 42 matched, 0 partially-patched, 0 drift. Both patched pi-ai files pass node --check and keep their .dshpurge.bak originals.
  • Regression tests: test/diagnose-partial.test.js covers half-patched → partial, fully patched → clean, and single-file → unchanged behaviour; node --test is 197 passing / 0 failing.
  • Docs: the Chinese and English README plus llms.txt now carry the same numbers (42 matched, 197 tests, 0 partially-patched), and the FAQ explains the new state.
  • Apache-2.0 derivative of the original project by YuJunZhiXue; LICENSE/NOTICE attribution preserved.

1.1.90

中文

  • 适配 0.2.2-rc.3 桌面版的 asar 密封布局:rc.3 把整个 dsh 运行时打进 resources/app.asar(209 MB),resources/runtime 为空,不再有 rc.1 的 resources/app/dsh 明文目录。面板点「应用」时插件会自动解开 asar 到 resources/app(451 MB、338 个 @deepseek-ai 包)、补上 app.asar.unpacked 的原生件、把 asar 改名 .bak 并重启宿主 —— 用户不用再自己跑脚本。
  • 命令行不再漏掉桌面壳补丁:CLI 跑在 node 里探不到 Electron 的 resourcesPath,过去会把自己判成 web 宿主,整批跳过 surfaces: ["desktop"] 的补丁(#82 DESKTOP_SHELL_PROFILE_FALLBACK、#83、#84 等)。1.1.90 起按 aiBase 是否落在 resources/app/dsh/... 自动判成 desktop,实测这 7 条从「na / pending」变成 applied。
  • --apply 末尾明说 asar 后果:写完补丁后若 app.asar 仍在原位,提示「宿主会优先读它」并给出退出后改名的具体做法,不再让人以为打完就生效。
  • 诊断口径修正(1.1.90 内):--diagnose 在 CLI 下把桌面宿主当成 web,桌面壳补丁全被标成「本表面不适用」;现在按 aiBase 是否落在 resources/app/dsh/... 自动判 desktop(rc.3 真机:本表面不适用 0)。另外 #33/#48 这类「替换文案就是 marker」的补丁,在 pattern 已全被替换、又没有 .bak 差异时被误判为「目标已消失」;新增 markersStrict,人造后置文案命中即算已生效(真机目标已消失 0)。
  • 真机验证(0.2.2-rc.3 桌面版):82 条补丁 = 42 条文件核验命中(40 条靠 markers)+ 29 条无文件字段(运行时注入)+ 7 条本机无对应载体(可选:#28/#30/#31/#64/#65/#71/#76,hooks/preset 包没装)+ 4 条已退役;改过原文 0、目标已消失 0、载体缺失 0;重启宿主后插件自检日志记 sealed:false / boot:host-untouched。兼容矩阵 hostSmoke 由 not-run 改为 verified,verifiedOn 记为 2026-11-16。
  • 安装提醒:rc.3 起注册表项迁到 HKLM(per-machine),静默安装必须提权;升级前请用 dsh-purge --preupdate 看一遍「解包目录会与新版 asar 冲突」的预检结论。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版...
Read more

v1.2.2

Choose a tag to compare

@github-actions github-actions released this 16 Nov 07:37

1.2.2

中文

  • 修掉跨树副作用:targetFiles 里的 desktop-shell 以前无条件取「当前官方桌面安装」的 resources/app/lib/main.js,于对隔离树 / 归档 base 跑 revert 时,会把真实宿主的桌面壳补丁一起摘掉。现在只跟随 aiBase 自己所属的 resources/app,不命中就留空。
  • 离体全周期验证(这次补上的证据):在干净 rc.3 npm 隔离树里 backupAll → apply → status → diagnose → revert;apply 58 条、0 失败、49 个文件改写、diagnose 42 命中且漂移 0、语法 0 错;revert 49 个文件后与原始树逐字节比对 0 差异,真实宿主桌面壳 marker 前后均为 9(未被触碰)。
  • 回归:新增 test/target-scope.test.js 2 条(桌面壳只跟随自己的 resources/app;隔离包布局不挂真实桌面壳);全套 node --test 214 通过 / 0 失败。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • Cross-tree side effect fixed: desktop-shell in targetFiles used to resolve the currently installed official desktop resources/app/lib/main.js regardless of aiBase, so a revert on an isolation tree or archived base could strip the live host's shell patches. It now follows only the resources/app that owns the given aiBase.
  • Full isolated cycle: on a clean rc.3 npm tree, backupAll → apply → status → diagnose → revert: 58 applied / 0 failed / 49 files changed, diagnosis 42 hits with 0 drift, 0 syntax errors; after revert, byte-for-byte comparison against the pristine tree showed 0 differences, and the live desktop shell markers stayed at 9.
  • Tests: 2 new scope cases; node --test 214 passing / 0 failing.
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.1

中文

  • 修掉一处静默漏打:#57 DEEPSEEK_SYSTEM_ROLE 在 rc.3 的 openai-responses-shared.js 里,变量名是 instructionRole 且带可选链(compat?.supportsDeveloperRole !== false),旧 pattern 只认 role,导致这个文件一直没打上、诊断却因另一份文件已打而整条报 clean。现已补上该候选锚点,真机两份副本(宿主 + profile)都已改写并带 allowlist marker,node --check 通过。
  • 诊断补上老式补丁的命中数:probeCarrier 此前只读 replacements,7 条用 patterns + replace 老式声明的补丁在诊断里永远 0 命中、只能靠 dsh-purge 字样蒙混。现在与新式补丁同口径展开,上游再改这几条能直接报「可命中/漂移」。
  • 事务临时文件清理:写盘失败时不再把 .dshpurge.tx-* 临时文件留在宿主目录。
  • 验证:新增 3 条回归(instructionRole 锚点命中、重复执行不再改写、老式 patterns 计数);全套 node --test 212 通过 / 0 失败,compat:check 一致;真机连打两次 apply,62 个载体文件哈希零变化(第二次 applied 0),确认幂等。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • One silent miss fixed: patch #57 DEEPSEEK_SYSTEM_ROLE uses instructionRole with optional chaining in rc.3's openai-responses-shared.js; the old pattern only matched role, so that file was never patched while diagnosis called the patch clean because its sibling file was. A candidate anchor is now included; both the host and profile copies carry the allowlist marker and pass node --check.
  • Legacy patches now report hit counts: probeCarrier only read replacements, so the 7 patches declared with patterns + replace always showed zero hits. It now expands them the same way as modern patches.
  • Transaction temp cleanup: failed writes no longer leave .dshpurge.tx-* files in the host tree.
  • Verification: 3 new regression tests; node --test 212 passing / 0 failing; two consecutive live applies changed zero bytes across 62 carrier files (second apply: applied 0).
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.0

中文

  • 本地重锚覆盖(用户可撤销):$DSH_HOME/dsh-purge/anchor-overrides.json 允许给某条补丁钉一个新锚点,apply / status / diagnose 共用同一份覆盖;覆盖项追加为候选,旧锚点 0 命中不再按命中数契约拒写,候选自带 expect 时仍严格校验。坏 JSON 只报错不生效,删文件即恢复;--anchors 只读查看,--diagnose 出 ⚓ 本地重锚 N 条生效。
  • fail-closed 补丁写盘:一条补丁的全部目标文件先在内存算完,提交前逐文件校验内容指纹,再原子写入;任一文件失败即把已写文件回滚到原文,报 write_error:…:rolled_back_N,不再出现「第 1 个文件改了、第 2 个失败」的半打状态。
  • 候选锚点表与多版本声明:replacements[].pattern 支持数组(旧版/新版原文二选一,唯一命中才替换,两版同时存在则整条拒写);package.json 新增 dshTargets,可声明多个已验证宿主版本,命中其一即适配;诊断与探测同口径展开候选。
  • 回归测试:test/anchor-overrides.test.js 6 条 + test/transaction.test.js 6 条(事务提交/指纹拦改/失败回滚、候选命中/歧义拒写、多版本命中),全套 node --test 209 通过 / 0 失败,npm run compat:check 三者一致。
  • 真机:无覆盖文件时行为不变(--status 70/71、--diagnose 42 命中、partial 0);--apply 走事务路径后数字不变、无 write_error。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Local re-anchor overrides: $DSH_HOME/dsh-purge/anchor-overrides.json lets you pin a new anchor for one patch; apply, status and diagnose read the same file. Overrides are appended as candidates, a zero-hit original no longer trips the hit-count contract, and an override that declares expect is still checked strictly. Broken JSON reports an error without taking effect; delete the file to restore. --anchors lists entries, --diagnose prints ⚓ 本地重锚 N 条生效.
  • Fail-closed patch writes: all target files of one patch are computed in memory first, fingerprinted before commit, then written atomically. Any failure rolls the already-written files back to their original text and reports write_error:…:rolled_back_N, so a patch is never half-applied.
  • Candidate anchor table and multi-version declaration: pattern in replacements accepts an array (old/new upstream text, replace only when exactly one is present, refuse when both are); package.json gains dshTargets so several verified host versions can be declared, any match counts as supported.
  • Tests: 6 anchor-override cases + 6 transaction cases (commit, fingerprint guard, rollback; candidate hit/ambiguity; multi-version match); node --test 209 passing / 0 failing, npm run compat:check consistent.
  • Maintainer qiuqianyi66. Derivative work of the original project by YuJunZhiXue, Apache-2.0, LICENSE/NOTICE retained.

1.1.91

中文

  • 修掉「status 说待清洗、diagnose 说没问题」的口径打架:同一个载体键落到多份副本时(宿主自带 pi-ai 与 profile 里装的 pi-ai 各一份 anthropic-messages.js / openai-completions.js),--status 按文件计数报 pending,--diagnose 按「任一文件已生效」把整条判成 clean。现在诊断新增「部分副本未清洗」桶与每文件证据(partial / pending_files),两边口径归一;建议文案直接说“跑一次 apply 补齐,pattern 不用改”。
  • 真机复现与修复:#47 ANTHROPIC_OAUTH_NO_CLAUDE_IDENTITY、#49 ANTHROPIC_OAUTH_CLI_NEUTRAL、#57 DEEPSEEK_SYSTEM_ROLE 三条因 profile 侧 pi-ai 升级到 1.5.2、原文又匹配旧 pattern 而挂 pending;--apply 补齐后 --status = 70/71 已清洗、0 待清洗,--diagnose = 42 命中、partial 0、漂移 0。两份 pi-ai 文件 node --check 通过,原文件留 .dshpurge.bak。
  • 回归测试:新增 test/diagnose-partial.test.js 三条用例(一打一原文 → partial;两份都打 → clean;单文件 → 旧行为不变),全套 node --test 197 通过 / 0 失败。
  • 文档:中英 README 与 llms.txt 的证据数字同步为 42 命中 + 197 测试 + partial 0,FAQ 补上「部分副本未清洗」的判读方式。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Fixes the status-vs-diagnose contradiction: when one carrier key maps to several copies (the host-bundled pi-ai and the pi-ai plugin installed under a profile each ship anthropic-messages.js / openai-completions.js), --status counted per file and reported pending while --diagnose took the strongest evidence and called the whole patch clean. Diagnosis now carries a dedicated partially-patched copies bucket with per-file evidence (partial / pending_files) so both commands agree, and the advice says exactly what to do: run apply once, no pattern change needed.
  • Reproduced and fixed on a real host: #47, #49 and #57 sat in pending because the profile-side pi-ai had been upgraded to 1.5.2 and matched the old patterns again; one --apply caught it up — --status now reports 70/71 applied with 0 pending, --diagnose reports 42 matched, 0 partially-patched, 0 drift. Both patched pi-ai files pass node --check and keep their .dshpurge.bak originals.
  • Regression tests: test/diagnose-partial.test.js covers half-patched → partial, fully patched → clean, and single-file → unchanged behaviour; node --test is 197 passing / 0 failing.
  • Docs: the Chinese and English README plus llms.txt now carry the same numbers (42 matched, 197 tests, 0 partially-patched), and the FAQ explains the new state.
  • Apache-2.0 derivative of the original project by YuJunZhiXue; LICENSE/NOTICE attribution preserved.

1.1.90

中文

  • 适配 0.2.2-rc.3 桌面版的 asar 密封布局:rc.3 把整个 dsh 运行时打进 resources/app.asar(209 MB),resources/runtime 为空,不再有 rc.1 的 resources/app/dsh 明文目录。面板点「应用」时插件会自动解开 asar 到 resources/app(451 MB、338 个 @deepseek-ai 包)、补上 app.asar.unpacked 的原生件、把 asar 改名 .bak 并重启宿主 —— 用户不用再自己跑脚本。
  • 命令行不再漏掉桌面壳补丁:CLI 跑在 node 里探不到 Electron 的 resourcesPath,过去会把自己判成 web 宿主,整批跳过 surfaces: ["desktop"] 的补丁(#82 DESKTOP_SHELL_PROFILE_FALLBACK、#83、#84 等)。1.1.90 起按 aiBase 是否落在 resources/app/dsh/... 自动判成 desktop,实测这 7 条从「na / pending」变成 applied。
  • --apply 末尾明说 asar 后果:写完补丁后若 app.asar 仍在原位,提示「宿主会优先读它」并给出退出后改名的具体做法,不再让人以为打完就生效。
  • 诊断口径修正(1.1.90 内):--diagnose 在 CLI 下把桌面宿主当成 web,桌面壳补丁全被标成「本表面不适用」;现在按 aiBase 是否落在 resources/app/dsh/... 自动判 desktop(rc.3 真机:本表面不适用 0)。另外 #33/#48 这类「替换文案就是 marker」的补丁,在 pattern 已全被替换、又没有 .bak 差异时被误判为「目标已消失」;新增 markersStrict,人造后置文案命中即算已生效(真机目标已消失 0)。
  • 真机验证(0.2.2-rc.3 桌面版):82 条补丁 = 42 条文件核验命中(40 条靠 markers)+ 29 条无文件字段(运行时注入)+ 7 条本机无对应载体(可选:#28/#30/#31/#64/#65/#71/#76,hooks/preset 包没装)+ 4 条已退役;改过原文 0、目标已消失 0、载体缺失 0;重启宿主后插件自检日志记 sealed:false / boot:host-untouched。兼容矩阵 hostSmoke 由 not-run 改为 verified,verifiedOn 记为 2026-11-16。
  • 安装提醒:rc.3 起注册表项迁到 HKLM(per-machine),静默安装必须提权;升级前请用 dsh-purge --preupdate 看一遍「解包目录会与新版 asar 冲突」的预检结论。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Handles the sealed-app.asar layout of the 0.2.2-rc.3 desktop build: rc.3 packs the whole dsh runtime into resources/app.asar (209 MB) and leaves resources/runtime empty, so the rc.1-era plain resources/app/dsh directory no longer exists. Pressing Apply in the panel now extracts the asar into resources/app (451 MB, 338 @deepseek-ai packages), overlays app.asar.unpacked, renames the asar to .bak and restarts the host — no manual script needed.
  • The CLI no longer drops desktop-shell patches: running under plain node it cannot see Electron's resourcesPath and used to classify itself as a web host, silently skipping every surfaces: ["desktop"] patch (#82, #83, #84, …). 1.1.90 infers desktop from an `ai...
Read more

v1.2.1

Choose a tag to compare

@github-actions github-actions released this 16 Nov 07:27

1.2.1

中文

  • 修掉一处静默漏打:#57 DEEPSEEK_SYSTEM_ROLE 在 rc.3 的 openai-responses-shared.js 里,变量名是 instructionRole 且带可选链(compat?.supportsDeveloperRole !== false),旧 pattern 只认 role,导致这个文件一直没打上、诊断却因另一份文件已打而整条报 clean。现已补上该候选锚点,真机两份副本(宿主 + profile)都已改写并带 allowlist marker,node --check 通过。
  • 诊断补上老式补丁的命中数:probeCarrier 此前只读 replacements,7 条用 patterns + replace 老式声明的补丁在诊断里永远 0 命中、只能靠 dsh-purge 字样蒙混。现在与新式补丁同口径展开,上游再改这几条能直接报「可命中/漂移」。
  • 事务临时文件清理:写盘失败时不再把 .dshpurge.tx-* 临时文件留在宿主目录。
  • 验证:新增 3 条回归(instructionRole 锚点命中、重复执行不再改写、老式 patterns 计数);全套 node --test 212 通过 / 0 失败,compat:check 一致;真机连打两次 apply,62 个载体文件哈希零变化(第二次 applied 0),确认幂等。
  • 维护者 qiuqianyi66;衍生作品,Apache-2.0,保留 LICENSE/NOTICE。

English

  • One silent miss fixed: patch #57 DEEPSEEK_SYSTEM_ROLE uses instructionRole with optional chaining in rc.3's openai-responses-shared.js; the old pattern only matched role, so that file was never patched while diagnosis called the patch clean because its sibling file was. A candidate anchor is now included; both the host and profile copies carry the allowlist marker and pass node --check.
  • Legacy patches now report hit counts: probeCarrier only read replacements, so the 7 patches declared with patterns + replace always showed zero hits. It now expands them the same way as modern patches.
  • Transaction temp cleanup: failed writes no longer leave .dshpurge.tx-* files in the host tree.
  • Verification: 3 new regression tests; node --test 212 passing / 0 failing; two consecutive live applies changed zero bytes across 62 carrier files (second apply: applied 0).
  • Maintainer qiuqianyi66. Derivative work, Apache-2.0, LICENSE/NOTICE retained.

1.2.0

中文

  • 本地重锚覆盖(用户可撤销):$DSH_HOME/dsh-purge/anchor-overrides.json 允许给某条补丁钉一个新锚点,apply / status / diagnose 共用同一份覆盖;覆盖项追加为候选,旧锚点 0 命中不再按命中数契约拒写,候选自带 expect 时仍严格校验。坏 JSON 只报错不生效,删文件即恢复;--anchors 只读查看,--diagnose 出 ⚓ 本地重锚 N 条生效。
  • fail-closed 补丁写盘:一条补丁的全部目标文件先在内存算完,提交前逐文件校验内容指纹,再原子写入;任一文件失败即把已写文件回滚到原文,报 write_error:…:rolled_back_N,不再出现「第 1 个文件改了、第 2 个失败」的半打状态。
  • 候选锚点表与多版本声明:replacements[].pattern 支持数组(旧版/新版原文二选一,唯一命中才替换,两版同时存在则整条拒写);package.json 新增 dshTargets,可声明多个已验证宿主版本,命中其一即适配;诊断与探测同口径展开候选。
  • 回归测试:test/anchor-overrides.test.js 6 条 + test/transaction.test.js 6 条(事务提交/指纹拦改/失败回滚、候选命中/歧义拒写、多版本命中),全套 node --test 209 通过 / 0 失败,npm run compat:check 三者一致。
  • 真机:无覆盖文件时行为不变(--status 70/71、--diagnose 42 命中、partial 0);--apply 走事务路径后数字不变、无 write_error。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Local re-anchor overrides: $DSH_HOME/dsh-purge/anchor-overrides.json lets you pin a new anchor for one patch; apply, status and diagnose read the same file. Overrides are appended as candidates, a zero-hit original no longer trips the hit-count contract, and an override that declares expect is still checked strictly. Broken JSON reports an error without taking effect; delete the file to restore. --anchors lists entries, --diagnose prints ⚓ 本地重锚 N 条生效.
  • Fail-closed patch writes: all target files of one patch are computed in memory first, fingerprinted before commit, then written atomically. Any failure rolls the already-written files back to their original text and reports write_error:…:rolled_back_N, so a patch is never half-applied.
  • Candidate anchor table and multi-version declaration: pattern in replacements accepts an array (old/new upstream text, replace only when exactly one is present, refuse when both are); package.json gains dshTargets so several verified host versions can be declared, any match counts as supported.
  • Tests: 6 anchor-override cases + 6 transaction cases (commit, fingerprint guard, rollback; candidate hit/ambiguity; multi-version match); node --test 209 passing / 0 failing, npm run compat:check consistent.
  • Maintainer qiuqianyi66. Derivative work of the original project by YuJunZhiXue, Apache-2.0, LICENSE/NOTICE retained.

1.1.91

中文

  • 修掉「status 说待清洗、diagnose 说没问题」的口径打架:同一个载体键落到多份副本时(宿主自带 pi-ai 与 profile 里装的 pi-ai 各一份 anthropic-messages.js / openai-completions.js),--status 按文件计数报 pending,--diagnose 按「任一文件已生效」把整条判成 clean。现在诊断新增「部分副本未清洗」桶与每文件证据(partial / pending_files),两边口径归一;建议文案直接说“跑一次 apply 补齐,pattern 不用改”。
  • 真机复现与修复:#47 ANTHROPIC_OAUTH_NO_CLAUDE_IDENTITY、#49 ANTHROPIC_OAUTH_CLI_NEUTRAL、#57 DEEPSEEK_SYSTEM_ROLE 三条因 profile 侧 pi-ai 升级到 1.5.2、原文又匹配旧 pattern 而挂 pending;--apply 补齐后 --status = 70/71 已清洗、0 待清洗,--diagnose = 42 命中、partial 0、漂移 0。两份 pi-ai 文件 node --check 通过,原文件留 .dshpurge.bak。
  • 回归测试:新增 test/diagnose-partial.test.js 三条用例(一打一原文 → partial;两份都打 → clean;单文件 → 旧行为不变),全套 node --test 197 通过 / 0 失败。
  • 文档:中英 README 与 llms.txt 的证据数字同步为 42 命中 + 197 测试 + partial 0,FAQ 补上「部分副本未清洗」的判读方式。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Fixes the status-vs-diagnose contradiction: when one carrier key maps to several copies (the host-bundled pi-ai and the pi-ai plugin installed under a profile each ship anthropic-messages.js / openai-completions.js), --status counted per file and reported pending while --diagnose took the strongest evidence and called the whole patch clean. Diagnosis now carries a dedicated partially-patched copies bucket with per-file evidence (partial / pending_files) so both commands agree, and the advice says exactly what to do: run apply once, no pattern change needed.
  • Reproduced and fixed on a real host: #47, #49 and #57 sat in pending because the profile-side pi-ai had been upgraded to 1.5.2 and matched the old patterns again; one --apply caught it up — --status now reports 70/71 applied with 0 pending, --diagnose reports 42 matched, 0 partially-patched, 0 drift. Both patched pi-ai files pass node --check and keep their .dshpurge.bak originals.
  • Regression tests: test/diagnose-partial.test.js covers half-patched → partial, fully patched → clean, and single-file → unchanged behaviour; node --test is 197 passing / 0 failing.
  • Docs: the Chinese and English README plus llms.txt now carry the same numbers (42 matched, 197 tests, 0 partially-patched), and the FAQ explains the new state.
  • Apache-2.0 derivative of the original project by YuJunZhiXue; LICENSE/NOTICE attribution preserved.

1.1.90

中文

  • 适配 0.2.2-rc.3 桌面版的 asar 密封布局:rc.3 把整个 dsh 运行时打进 resources/app.asar(209 MB),resources/runtime 为空,不再有 rc.1 的 resources/app/dsh 明文目录。面板点「应用」时插件会自动解开 asar 到 resources/app(451 MB、338 个 @deepseek-ai 包)、补上 app.asar.unpacked 的原生件、把 asar 改名 .bak 并重启宿主 —— 用户不用再自己跑脚本。
  • 命令行不再漏掉桌面壳补丁:CLI 跑在 node 里探不到 Electron 的 resourcesPath,过去会把自己判成 web 宿主,整批跳过 surfaces: ["desktop"] 的补丁(#82 DESKTOP_SHELL_PROFILE_FALLBACK、#83、#84 等)。1.1.90 起按 aiBase 是否落在 resources/app/dsh/... 自动判成 desktop,实测这 7 条从「na / pending」变成 applied。
  • --apply 末尾明说 asar 后果:写完补丁后若 app.asar 仍在原位,提示「宿主会优先读它」并给出退出后改名的具体做法,不再让人以为打完就生效。
  • 诊断口径修正(1.1.90 内):--diagnose 在 CLI 下把桌面宿主当成 web,桌面壳补丁全被标成「本表面不适用」;现在按 aiBase 是否落在 resources/app/dsh/... 自动判 desktop(rc.3 真机:本表面不适用 0)。另外 #33/#48 这类「替换文案就是 marker」的补丁,在 pattern 已全被替换、又没有 .bak 差异时被误判为「目标已消失」;新增 markersStrict,人造后置文案命中即算已生效(真机目标已消失 0)。
  • 真机验证(0.2.2-rc.3 桌面版):82 条补丁 = 42 条文件核验命中(40 条靠 markers)+ 29 条无文件字段(运行时注入)+ 7 条本机无对应载体(可选:#28/#30/#31/#64/#65/#71/#76,hooks/preset 包没装)+ 4 条已退役;改过原文 0、目标已消失 0、载体缺失 0;重启宿主后插件自检日志记 sealed:false / boot:host-untouched。兼容矩阵 hostSmoke 由 not-run 改为 verified,verifiedOn 记为 2026-11-16。
  • 安装提醒:rc.3 起注册表项迁到 HKLM(per-machine),静默安装必须提权;升级前请用 dsh-purge --preupdate 看一遍「解包目录会与新版 asar 冲突」的预检结论。
  • 维护者 qiuqianyi66;本版本系原始项目(作者 YuJunZhiXue)的衍生作品,Apache-2.0,保留 LICENSE/NOTICE 版权声明。

English

  • Handles the sealed-app.asar layout of the 0.2.2-rc.3 desktop build: rc.3 packs the whole dsh runtime into resources/app.asar (209 MB) and leaves resources/runtime empty, so the rc.1-era plain resources/app/dsh directory no longer exists. Pressing Apply in the panel now extracts the asar into resources/app (451 MB, 338 @deepseek-ai packages), overlays app.asar.unpacked, renames the asar to .bak and restarts the host — no manual script needed.
  • The CLI no longer drops desktop-shell patches: running under plain node it cannot see Electron's resourcesPath and used to classify itself as a web host, silently skipping every surfaces: ["desktop"] patch (#82, #83, #84, …). 1.1.90 infers desktop from an aiBase under resources/app/dsh/...; those 7 patches went from na / pending to applied on a real rc.3 install.
  • --apply states the asar consequence: if app.asar is still in place after writing, it now says the host reads it first and tells you to rename it after quitting.
  • Verified on a real 0.2.2-rc.3 desktop host: 71 of 82 patches applied directly, 7 not applicable to this surface, 4 skipped, 0 drift, 4 retired; the plugin self-check log records sealed:false and boot:host-untouched after restart (6 processes, port 65088). The matrix hostSmoke cell moves from not-run to verified, verifiedOn is 2026-11-16.
  • Apache-2.0 derivative of the original project by YuJunZhiXue; LICENSE/NOTICE attribution preserved.