Skip to content

LoopX v2.1.1

Latest
Latest

Choose a tag to compare

View all tags
@loopx-agent loopx-agent released this 23 Oct 00:38
v2.1.1
87e5b4f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

LoopX v2.1.1

Published: The Python package on PyPI, GitHub packages, desktop build assets and signed macOS updater have been independently downloaded and verified. The native desktop remains a preview; Windows execution and Apple notarization are not claimed.
已发布： PyPI 与 GitHub 的 Python 包、桌面构建资产及 macOS 签名更新已独立下载核验。原生桌面继续为预览；不声明已运行 Windows 安装程序或已获 Apple 公证。

At a Glance

Illustrated release guide: 个人飞书图解指南 (existing document access applies). The versioned public guide covers the same upgrade and capability boundaries without a Feishu account.

Recover work through its original conversation, keep Goal evidence visible, and produce content through explicitly selected providers. An acknowledgement or completed Todo still does not prove that the original requester received a result.

Upgrade

Let active Bot/Agent Turns finish before applying the update: runtime activation restarts managed status/chat services. Inspect original results and effect evidence before continuing any interrupted request.

loopx update check
loopx update plan
loopx update apply
loopx --version
loopx doctor --deep

This maintenance release starts from the 2.1 preview source; it is not a rollup of later main snapshots. Preserve the existing source channel and inspect source identity before replacing a development snapshot. Keep the existing installation owner. PyPI publication has been independently verified; new users can install loopx==2.1.1; see Install / Update below. CLI and native App versions are separate readbacks.

Highlights

  • Task context recovery: Back restores the same task and selected row; Forward closes the drawer. Opaque outcome/requirement IDs preserve the three-row preview without losing the full list. Installation-only evidence cannot qualify an untested page. #7101.
  • Original conversation recovery: resume the saved native profile and source-bound request; uncertain effects retain their recovery path. Check the original result rather than redispatching the receiver. #6934, #6974, #7015.
  • Visible work: use Quick navigation (Cmd+K / Ctrl+K) to find Goals, tasks and previews. Opening an item navigates without executing it. #6996.
  • Content production: prepare a composition in the Goal conversation, then use Results → Content artifacts. Local rendering and cloud generation require the optional provider and explicit effect approval. #6876, #6979.
  • Less repeated browser confirmation: normal skill refresh delivers bounded recovery of a known missing Agent-owned Ego workspace; human control and unknown-effect stops remain. #7034.
  • Optional native memory: an explicitly selected Nano Turn can restore scoped memory through the separate MemT host. This does not change default models or add memory to Codex/Claude. #7028, #7037.

Replan Fixes

  • Planning and recovery can settle without fabricating a delivery Todo; material path changes retain their evidence. #6999, #6940.
  • Required objective revisions retain acceptance guidance and require each Agent's own adoption. #6987, #7023.
  • Unsafe redispatch after provider failures stops instead of treating an unknown outcome as an empty attempt. #6932, #6885.

Contributors

Release Decision

Who should upgrade: Operators using Bots, delegation, native Goals, the workspace or desktop preview. Read-only historical installations can use their normal maintenance window.

What this release solves: Lost native execution context, unsafe recovery assumptions, incomplete visible handoffs, navigation friction and missing optional content-production paths.

Breaking changes: Yes, for release-qualification receipt producers: the current installation summary must use release_install_upgrade_host_summary_v1 and include the completed packaged frontend result. Historical v0 receipts remain evidence and cannot qualify this new candidate. For ordinary runtime users, no new state migration is introduced. There is no additional persisted-state migration from 2.0.0. Its canonical-authority cutoff remains: unpromoted Markdown Goals and retired legacy ownership require the reviewed migration. Optional providers remain explicit and disabled until configured.

How to verify: Expect loopx 2.1.1, a successful installation doctor and the original Goal/request readback. A running process or receipt alone is insufficient to prove business recovery.

Contributors: Maintained by @huangruiteng, with @cuipengcx90, @mikamikasuki, @hhyykk, @JasonBuildAI, @Twiink, @Zhouyang66639, @songoow and @steven-kid; concrete contributions appear above and below.

loopx --version
loopx doctor --deep
loopx extension doctor --all-enabled --execute

State Kernel & Control Plane

Bot recovery preserves original native profiles, verified text-only rich posts and source-bound handoffs. Busy lanes are visible before dispatch, and the steward work surface exposes unfinished incoming/outgoing requests. Read, adoption, execution and final return remain separate. #6934, #6941, #7015, #7018, #7019.

Native account handoff exposes quota observation, optional switching, restart and force quit. Force quit targets the verified dedicated GPT client after normal quit fails; unrelated owners continue to block switching. Unknown quota is not exhaustion. #7008.

Known limitation: worker source checks are not atomic with final Task completion. Revocation after the last check can still allow a Task commit before the operation reports revocation. Preserve Task/operation/Host evidence; an authorization error does not establish absence of effects or safe replay. #7016, independent race review.

Capabilities & Workflows

Goal requirements and evidence survive individual Todo completion; objective revision and Agent adoption remain separate decisions. The workbench exposes this state and keyboard navigation without granting execution authority. #6784, #6987, #6996, #7023.

Content artifacts connect Agent-authored composition, local rendering, verified downloads and explicit cloud intents. Full-frame composition retains the complete picture and a source strip. Provider installation, credentials, paid submission and publication remain separate opt-ins. #6876, #6979.

Quality & Testing

Exact qualified source: 87e5b4f88d207adf062eea57002cda6fe04fd535, tree d0288d0ad4372082eb58af8f1a0963b9a2eaf078, clean package 2.1.1. All eight native qualification categories passed on this exact source.

  • Full Python suite: 26,881 passed, 2,242 subtests passed, 686 skipped under the original skip rules; owned processes exited and drained.
  • Full public suite: 514/514 checks executed, no failure, timeout, unsafe command, tracked side effect or Git-required skip. Ruff, mypy and TypeScript checks passed.
  • Complete installed-wheel frontend: all 72 catalog scenarios plus team-evidence-return, team-feedback-recovery and Goal workspace craft passed with unchanged assertions/deadlines and independently drained processes.
  • Real 2.0.0 wheel upgrade preserved a saved Todo and artifact bytes; fresh default installation, doctor, optional installed-host stdio/EOF cleanup, legacy unknown-effect refusal and the complete optional MemT test workflow passed. These are local correctness/upgrade checks, not GPU memory-quality evidence.
  • Risk canary selected and direct checks passed. The three raw PNG privacy holds remain recorded; exact-source image provenance, pixels and metadata were reviewed explicitly rather than called an automatic text-scan pass.
  • The authorized DeepSeek Flash/high substitution ran the unchanged actual-default one-arm portfolio: 22 scenarios × 2 repeats = 44 actor calls, 6 contrasts, no failure or skip. The qualification transport omits the vendor JSON mode, retaining native JSON instructions, strict parsing and serial tool calls; no scenario, oracle, repetition or deadline was weakened. No production model profile was changed.

The first model attempt on this final source stopped after an empty provider JSON-mode response, with 18/44 actors recorded. Its incomplete cases and contrasts were not accepted. The provider documents this JSON-mode limitation; after removing only that request option, this complete portfolio was executed afresh without merging partial model results. This is a qualification transport adaptation, not a production model-profile change.

The immutable 2.1.0 preview and every failed publication check remain retained. The first 2.1.1 candidate passed its Python and complete page checks but failed two public documentation metadata checks. Five documentation files were corrected through #7107, then every final-source lane was executed again. Earlier interrupted suites, the original 12-worker three-second Stop deadline miss, failed public checks and parallel-tool model failure remain historical failures. A separate prompt-only probe reached 19/20; it is not an installed native-model or steward golden-pack qualification. Native live Goal (required executable unavailable), GPU inference, Windows strict tool-tree execution and complete community/steward golden packs remain unqualified. No long-horizon or benchmark improvement is claimed.

Benchmarks & Integrations

Research history retains episodes, source citations and quantitative results independently of model judgment. Explicit source/account selection and the original analysis directory own execution, status and cancellation. #6961, #6966, #6968, #7011.

The MemT package supplies a separately selected Nano CLI/provider prerequisite through the public Turn-host seam. It does not install upstream code/weights, grant tools or add native memory to another model. Upstream client/inference code is PolyForm Noncommercial 1.0.0; weights have separate terms. Review licensing before installation or commercial use. #7028, #7037.

Documentation & Compatibility

Versioned upgrade, provider lifecycle and RFC status documentation distinguish accepted design from installed or independently accepted results. Ego skill refresh permits one recorded replacement only after successful inventory confirms a known Agent-owned space is absent; login, human intervention and uncertain-effect stops remain. Desktop preview startup/update errors and owned-service cleanup are clearer. #7034, #7037, #7024, #7029.

No new persisted-state migration is introduced relative to 2.0.0. Follow the 2.0 migration boundary for older authority. This release does not silently promote a Markdown Goal or authorize a second writer.

Community Contributors

Attribution is verified against the v2.0.0 to release-source Git range and merged PR metadata, excluding automated dependency and maintainer/bot accounts. No first-time-contributor claim is made.

Optional Capability Activation & Use

Content media

Activation: From a checkout of v2.1.1, install the separate provider and register its manifest. Existing providers require an explicit package upgrade and loopx extension upgrade --manifest packages/loopx-media/extension.toml --execute.
Validation: Run provider doctor, inspect the admitted composition and verify the original render/download result. Doctor alone does not prove generation.
Disable / rollback: Disable new work with the command below; retain existing operation IDs. Remove framing in a new composition revision to restore editorial framing. Observation stop is not remote cancellation.
Authority boundary: Local render, optional Blender and paid Ark submission have separate prerequisites. Cloud submission requires credential/cost/upload review; ambiguous submissions must not be repeated. Artifacts grant no acceptance or publication permission.
Docs: Versioned provider guide.

git checkout v2.1.1
uv tool install ./packages/loopx-media
loopx extension install --manifest packages/loopx-media/extension.toml --execute
loopx extension doctor loopx-media --execute
loopx extension disable loopx-media --execute

MemT Nano Turn host

Activation: Install the separate package and reviewed upstream client; provision Nano independently. Set enabled=true only in an operator-owned profile outside the task workspace, and explicitly select loopx-memt-turn for an admitted generic-cli Turn with an independent validator.
Validation: The command below is local preflight only; it reports network/inference as unchecked. Real GPU and memory quality remain unqualified.
Disable / rollback: Set enabled=false, remove the explicit host selection and restore the previous host. Uninstalling the adapter does not stop the external server or delete snapshots.
Authority boundary: Memory is advisory, scoped to workspace/Goal lifetime/Agent; current authority and evidence still come from LoopX. MCP scope inputs are not authentication. Unknown effects require reconciliation; no automatic replay. Upstream Noncommercial client/inference and separate weight terms apply.
Docs: Versioned Nano install, licensing and recovery.

Legacy recovery: Before provider or MCP IO, the 2.1.1 host holds the old alias cursor lock and rejects pending, unfinished, orphaned or unreadable legacy state. A new Goal lifetime cannot waive unknown effects. Completed legacy records stay preserved without importing memory. Reconcile the exact old session and effects, stop its worker, then explicitly select a new state directory under the existing recovery procedure; keep the evidence and avoid blind replay.

python3 -m venv .venv-memt
. .venv-memt/bin/activate
python -m pip install -e . -e packages/loopx-memt-turn
loopx-memt-turn --config /srv/operator/memt.json --doctor
# After stopping active work and restoring the prior host:
python -m pip uninstall loopx-memt-turn

Ego browser continuation and optional source reader

Activation: Refresh normal workflow skills for an already installed Ego skill. The source reader separately needs an explicit entry in the actual host's MCP configuration; skill refresh does not enable it.
Validation: Inspect the ego_browser refresh/readback result. A replacement needs successful complete inventory, known identity and a recorded attempt; unknown cause stays unknown.
Disable / rollback: Remove the optional reader's MCP entry and reload an idle host. loopx workflow-skills --uninstall restores an unchanged managed Ego copy to the current upstream link while preserving custom edits; it also removes other unchanged LoopX workflow skills.
Authority boundary: No Ego installation, login, takeover, publishing, source-scope expansion or unknown-effect replay is granted. Human control and refused claims remain stops.
Docs: Versioned lifecycle and origin configuration.

loopx workflow-skills --format json
loopx workflow-skills --install --format json
loopx workflow-skills --format json

Native Codex account handoff

Activation: In Workspace settings → Native Codex accounts, inspect accounts and observed quotas. Explicitly enable automatic switching and, separately, client restart/force quit when wanted.
Validation: Use Check quota and inspect the account/status readback; unknown reads are not exhaustion. Read the running service's identity independently.
Disable / rollback: Turn off Switch automatically when quota runs out; restart and force-quit options can be disabled independently. Retain existing manual account/model selection.
Authority boundary: Force quit targets only the verified dedicated GPT client after normal quit fails. Other CLI/file owners block switching. No credit purchase, reset coupon or submitted-business replay is granted.
Docs: Versioned workspace/account guide.

loopx --version
loopx doctor --deep
loopx dashboard

Research trajectory analysis

Activation: Inspect an authorized source pack and the selected host/account; the command below previews. Add --execute only after reviewing source, model and cost consequences.
Validation: Read status from the original analysis directory and inspect evidence coverage, citations and result boundaries.
Disable / rollback: Leave the preview unexecuted or cancel the original active analysis directory. A failed/uncertain attempt requires original-result readback before explicit retry.
Authority boundary: History/status reads remain model-free. Analysis does not widen source grants, silently switch identity or establish outcome improvement.
Docs: Versioned assessment guide.

loopx benchmark trajectory-assess --input-json evidence-pack.json --output-dir .local/trajectory-analysis --format json
loopx benchmark trajectory-assessment-status --directory .local/trajectory-analysis/KEY --format json
loopx benchmark trajectory-assessment-cancel --directory .local/trajectory-analysis/KEY --format json

Published Assets

Python wheel and source archive on GitHub match PyPI. GitHub attestations bind Python and desktop build assets to the immutable tag. macOS version/source archive and changed source files match this release; the downloaded App passes deep ad-hoc signature checks and the DMG integrity check. The updater signature verifies with the shipped public key, and a corrupted payload is rejected. The signed stable feed and Git stable source identify this release. Windows installers are build/attestation verified, not locally executed; Apple notarization is not claimed.

Install / Update

Python 3.11+ and Node.js 22.22.3+ are required; Node 24 LTS is recommended. The published PyPI wheel and source archive match the verified GitHub package hashes. The following pinned install is available.

python3 -m pip install --upgrade 'loopx==2.1.1'
loopx workflow-skills --install
loopx --version
loopx doctor --deep

Restart the Agent host after first installing skills. Existing pipx, uv-tool and archive users should keep their installation owner and inspect loopx update check and loopx update plan before loopx update apply. Desktop preview assets and signed-updater artifacts have separate identities. Package checksums and GitHub attestations are acquisition evidence, not a vulnerability-free claim.

中文摘要

图解指南： 个人飞书发布指南沿用当前文档访问权限；无需飞书账号可读公开版本指南。

升级决策

执行升级前，先等活动中的 Bot/Agent Turn 结束：运行时采用会重启受管 status/chat 服务。中断请求须先核对原结果与效果证据，再决定如何继续。

谁需要升级： 使用双 Bot、委派、原生 Goal、工作台或桌面预览的用户优先升级；只读旧历史可按维护窗口安排。
解决了什么： 保留原生执行上下文、收紧未知效果恢复、显示未回传交办，补齐快速导航和可选内容制作路径。
是否有破坏性变更： 有，限发行资格摘要的生产者：当前安装摘要须使用 release_install_upgrade_host_summary_v1 并包含完整页面结果；历史 v0 证据保留，但不能用于放行新候选。普通运行时用户无新增状态迁移。从 2.0.0 到本版没有新增持久状态迁移；2.0 的 canonical authority 迁移边界继续有效。可选 provider 不随升级自动启用。
如何验证： 读回 loopx 2.1.1、doctor 和原业务结果；运行中、已读或 Todo done 均不等于原受众收到答复。
贡献者： @huangruiteng 维护；社区贡献来自 @cuipengcx90、@mikamikasuki、@hhyykk、@JasonBuildAI、@Twiink、@Zhouyang66639、@songoow 和 @steven-kid，具体 PR 见社区贡献者。

loopx --version
loopx doctor --deep
loopx extension doctor --all-enabled --execute

状态内核与控制面

原生 profile、富文本原消息和 source-bound 委派保留原上下文；忙碌 lane 与未回传交办可见。自动账号切换、重启、专用客户端强退分别显式启用；额度读取失败不当成耗尽。 #6934、#6941、#7008、#7015、#7018、#7019。

已知限制：最后一次 source 复核后的撤权与 Task 完成不是原子事务，可能先提交 Task，再报告撤权。必须保留 Task、操作与 Host 证据；权限错误不证明没有效果或可安全重放。 #7016、独立竞态审查。

能力与工作流

Goal 要求、证据、目标修订与各 Agent 采用分别保留；无 Todo 的规划可正常结算。快速导航只导航，不执行。Agent 在会话准备分镜，工作台「成果 → 内容作品」制作与查看；大画面保留完整素材和来源栏。provider、付费生成及发布仍分开授权。 #6784、#6987、#6999、#6996、#6876、#6979。

质量与测试

精确发行源码 87e5b4f88d207adf062eea57002cda6fe04fd535、tree d0288d0ad4372082eb58af8f1a0963b9a2eaf078，clean 2.1.1；八项原生资格均在本源码完成。完整 Python 为 26,881 项通过、2,242 项子测试通过、686 项按原规则跳过；完整公开检查 514/514，无失败、超时、危险命令、tracked side effect 或 Git 前置跳过。Ruff、mypy、TypeScript 通过。

完整已安装 wheel 页面脚本的 72 个 catalog 场景及三条 companion flow 均通过，断言和时限未改，进程独立排空。真实 2.0.0 wheel 升级保留原 Todo 和产物字节；默认安装、doctor、可选 Host 的 stdio/EOF 清理、旧未知效果调用前拦截与完整 MemT 工作流通过。风险检查的三张 PNG 原始人工 hold 保留，由精确源码、像素及元数据审阅闭合，未伪称文本扫描自动通过。

经授权的 DeepSeek Flash/high 替代完成原 actual-default one-arm 组合：22 场景 × 2 次 = 44 actor 调用、6 项对照，零失败、零跳过；资格化请求停用供应商 JSON 模式，保留原 JSON 指令、严格解析和串行工具调用，不改场景、oracle、重复次数或时限。生产模型配置未改。

本最终源码第一次模型验收因供应商 JSON 模式返回空内容而结束，只记录18/44次 actor；未运行项和未完成对照均未接受。该限制有官方说明，停用这一请求选项后完整重跑全部组合，不拼接先前部分结果，也不改变生产模型配置。

2.1.0 不可变预览及全部失败保留。首个 2.1.1 候选虽通过 Python 和完整页面，仍在全公开检查暴露两项文档元数据失败；五个文档经 #7107 修正后，最终源码全部资格重新执行。原中断、12 worker 三秒 Stop 超时、公开检查与并行工具模型失败不改称通过；19/20 的提示词探测也不计作已安装原生模型或完整管家 golden pack。原生 live Goal 所需 executable 不可用，GPU、Windows 严格工具树和完整社区/管家 golden packs 继续未资格化；不声明历史积压全闭、生产模型切换或长程/benchmark 提升。

基准与集成

研究历程展示原始 episode、引用和定量结果；模型判断不替代原证据。执行/取消沿原分析目录及已选 source/account。MemT 只交付显式选择 Nano 的独立 CLI/provider 前置层，不给 Codex/Claude 加原生记忆，不自动安装上游代码/权重。上游客户端/推理代码采用 PolyForm Noncommercial 1.0.0，权重另有许可，使用前需审阅。 #6961、#6966、#6968、#7028、#7037。

文档与兼容性

Ego 的正常 skill 刷新沉淀一次性丢失恢复：成功 inventory 确认已知自有工作区不存在后可记录一次替代；真人接管、登录、拒绝 claim、未知效果仍停止。桌面更新传输/签名失败和自有服务清理分别可见。2.0 到本版无新增持久状态迁移，旧 authority 仍需原迁移路径，不启动第二 writer。 #7034、#7024、#7029、#7037。

社区贡献者

名单基于 v2.0.0 到本次发布源码及已合并 PR 核验，不声明首次贡献，不把依赖机器人或维护者列为社区贡献者。

可选能力启用与使用

Content media

启用： 在 v2.1.1 checkout 安装独立 provider 并注册 manifest；旧 provider 先升级包，再执行 extension upgrade。
验证： 运行 doctor，读回原分镜、制作结果与校验下载；doctor 不证明生成成功。
停用 / 回退： disable 阻止新工作，旧操作保留原 ID；新修订移除 framing 可恢复原构图。停止观察不等于云端取消。
权限边界： Blender extra 与云凭据独立安装；付费生成先看来源上传、价格和用量，不重放未知提交。产物不授予验收或发布权。
文档： 版本化 provider 指南。

uv tool install ./packages/loopx-media
loopx extension install --manifest packages/loopx-media/extension.toml --execute
loopx extension doctor loopx-media --execute
loopx extension disable loopx-media --execute

MemT Nano Turn host

启用： 独立安装适配包、审阅许可后安装上游客户端，并另行部署 Nano。workspace 外 operator profile 设 enabled=true；只有显式 generic-cli host 选择与独立 validator 才进入该 Turn。
验证： doctor 仅本地前置，明确不检查网络/推理；真实 GPU、记忆质量与收益未资格化。
停用 / 回退： 设 enabled=false，移除后续 host 选择并恢复原 host；卸载不停止外部服务或删除快照。
权限边界： 记忆按 workspace、Goal lifetime、Agent 隔离，仍属未验建议；当前权和证据从 LoopX 读回。MCP 字段不当认证，未知效果先对账。上游客户端/推理 Noncommercial、权重单独许可。
文档： 版本化 Nano 安装、许可和恢复。

旧状态恢复： 2.1.1 在 provider/MCP 调用前持有旧 alias 游标锁，拒绝 pending、未完成、孤立或不可读的旧状态；换 Goal 生命周期也不能绕过未知效果。已完成旧记录保留但不导入记忆。先核对原会话与真实效果、停止原 worker，再按原恢复流程显式选择新状态目录；保留证据，不盲目重放。

python3 -m venv .venv-memt
. .venv-memt/bin/activate
python -m pip install -e . -e packages/loopx-memt-turn
loopx-memt-turn --config /srv/operator/memt.json --doctor

Ego browser continuation and optional source reader

启用： 正常刷新已有 Ego skill；可选 source reader 需在真实 host 的 MCP 配置另行显式添加，技能刷新不会启用 reader。
验证： 读回 ego_browser 状态；替代需完整成功 inventory、已知身份和持久尝试记录，未知原因保持未知。
停用 / 回退： 移除 reader MCP 项并在空闲时重载。workflow-skills --uninstall 会移除未修改的 LoopX skills，并将未修改的受管 Ego 副本恢复到当前上游链接，保留定制内容。
权限边界： 不安装 Ego、不登录/接管、不扩来源或发布权、不重放未知效果；真人控制和 claim 拒绝仍停止。
文档： 版本化生命周期和 origin 配置。

loopx workflow-skills --format json
loopx workflow-skills --install --format json
loopx workflow-skills --format json

Native Codex account handoff

启用： 工作区设置 → 原生 Codex 账号，先检查账号与额度；自动切换、重启、强退分别显式启用。
验证： 点击检查额度并读回账号/状态；读取失败不是耗尽，运行服务身份独立核验。
停用 / 回退： 关闭额度耗尽时自动切换；重启和强退选项可分别关闭，保留原手动账号/模型选择。
权限边界： 正常退出失败后仅强退已核验的专用 GPT 实例，其他 CLI/文件占用仍阻断。不能购买额度、消费重置券或重放已提交业务。
文档： 版本化工作区/账号指南。

loopx --version
loopx doctor --deep
loopx dashboard

Research trajectory analysis

启用： 核验 source pack、所选 host/account 与成本后，先 preview；明确执行才加 --execute。
验证： 沿原分析目录读 status、引用、覆盖与结果边界。
停用 / 回退： 不执行 preview，或取消原目录的活动分析；失败/未知结果先读回，不能静默切身份重试。
权限边界： 历史和状态读取不调用模型，分析不扩大 source grant，也不证明长程收益。
文档： 版本化轨迹分析指南。

loopx benchmark trajectory-assess --input-json evidence-pack.json --output-dir .local/trajectory-analysis --format json
loopx benchmark trajectory-assessment-status --directory .local/trajectory-analysis/KEY --format json
loopx benchmark trajectory-assessment-cancel --directory .local/trajectory-analysis/KEY --format json

发布资产与验证

最终源码和测试身份见上方 Quality & Testing；PyPI、桌面构建资产、更新签名与 stable 已分别独立读回。GitHub/PyPI 包哈希一致，Python 与桌面构建 attestation 绑定不可变 tag；macOS 包版本与源码匹配，ad-hoc 深层签名及 DMG 完整性通过。更新签名经随包公钥验证，篡改负例被拒绝，稳定 feed 与 Git stable 均指向本版。Windows 仅验证构建与来源，未本机执行；不声明 Apple 公证。固定版本 v2.1.1；更新后分别核验 CLI、宿主、原服务和启用扩展。原生桌面仍属预览，构建资产不等于真实 Windows 执行、GPU 或公证资格。历史积压是否自动完成，必须以原 Bot 执行与原消息实际回传为准。

Contributors

hhyykk, huangruiteng, and 7 other contributors
Assets 16
Loading