Skip to content

LoopX 1.0 — a workspace for long-running agents

Latest
Latest

Choose a tag to compare

View all tags
@huangruiteng huangruiteng released this 06 Sep 12:52
v1.0.0
14c5f87

LoopX 1.0 — a workspace for long-running agents

At a Glance

See what agents are doing, what is waiting on you, and what has actually been delivered—in one persistent Workspace.

Upgrade

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

Highlights

  • A richer Workspace: all agent lanes, completed work, capability settings and verified reports.
  • IM as an operating surface: reconnect Goal Topics and choose live steering, session queuing or async inbox.
  • Signed desktop maintenance: paired App/runtime updates with stable/main, repair and rollback.
  • Real frontend project stories, and a stronger Doubao evolving behavior gate.

Contributors

Release Decision

Who should upgrade: Workspace, Lark/IM and desktop users, plus operators needing the recent Todo and host-recovery fixes; stable CLI-only deployments can review the affected optional surfaces before upgrading.

What this release solves: More work state becomes visible and actionable, completed delivery is easier to verify, and desktop maintenance has a signed, recoverable App/runtime path.

Breaking changes: No. The 1.0 label marks a Workspace milestone, not blanket promotion of staged authority backends. Older desktop shells need one manual bootstrap replacement; optional delivery routes and host capabilities keep their explicit activation boundaries.

How to verify: Expect package version 1.0.0, inspect doctor output, then read back your Goal and the App/runtime identity for the installed source.

Contributors: @Duang777, @now-ing, @steven-kid, @yuefengw, @songoow. Milestone foundations also include @maxliux5; see the contribution details below.

loopx --version
loopx doctor
loopx machine-config inspect

Workspace milestone

1.0 marks the Personal Workspace milestone: a durable control surface for work that continues across sessions. The Workspace already existed before v0.5.4; this release brings its state and operation paths together.

  • Open the Workspace before slow Goal histories finish: a lightweight directory appears first, Goals load independently, and transient failures recover automatically, and stopped histories load on selection (#4007, #4009).
  • See all work-agent lanes, act from the workspace, and reliably read completed tasks (#3888, #3894, #3961).
  • Configure Goal capabilities and typed machine policy; inspect verified reports with bounded indexing (#3860, #3861, #3865, #3951).
  • Explore reusable, complex project stories through the real frontend and state APIs (#3990).

State kernel and integrations

  • Native Todo create/update, request-bound idempotency and attached-completion recovery improve real delivery; promoted claim retries now expose their exact identity, with cross-provider lease-race qualification (#3973, #3974, #3977, #3980, #3981, #3987, #3986).
  • Stage2C authority work remains staged. Kernel routing and legacy writer fences are foundations, not a declaration that every existing tenant has been promoted (#3882, #3895).
  • Reconnect existing Lark Goal Topics, reuse session history, and apply scoped outbound guidance before authorized messages (#3983, #3868, #3968, #3998, #3999, #3992).
  • Codex provider routing improves Fast/dualAuto and quota recovery; host recovery and heartbeat compatibility are hardened (#3984, #3892, #3880, #3913, #3890).
  • Read benchmark-native study views and bounded terminal insights without changing experiment authority or scoring (#3896, #3898, #3878, #3881).

Desktop and qualification

Signed App updates pair the shell with its runtime. Stable is independent of unrelated plugin releases; service startup failures leave recovery usable; channel selection survives polling (#3994, #3996). Native maintenance now matches the actual IPC origin; unavailable feeds and local status failures have distinct recovery messages. The behavior gate removes extra answer hints and adds adversarial diagnostic contrasts and mutation checks (#3997). Release checks now use bounded scheduler clocks, typed followthrough obligations and stable historical version anchors; CLI command owners and help/manpage coverage are aligned without raising budgets (#4002, #4003). No benchmark uplift or long-horizon outcome improvement is claimed.

Community Contributors

Special milestone thanks to @maxliux5 for the Workspace, tasks-first navigation, Manager/Lark integration and session/streaming foundations (#3149, #3167, #3274, #3385, #3292, #3285, #3073). These contributions predate the v0.5.4 comparison range; they are credited as foundations of 1.0.

Optional Capability Activation & Use

Set the shell variables in the commands below to your own existing Goal, Agent, reviewed configuration and input files. Configuration writes require the named explicit execution step.

Workspace and machine configuration

Activation: Run loopx dashboard; configure a Goal through its Capability settings. For typed machine policy, inspect namespaces, preview a JSON envelope, then apply that exact returned plan revision with --execute.

Validation: Use machine-config inspect to read effective policy and its revision; use configure-goal --goal-id "$GOAL_ID" for Goal readback.

Disable / rollback: Close the local dashboard with Ctrl-C. Preview machine-config remove --namespace "$NAMESPACE", then repeat with its --expected-plan-revision and --execute; rollback similarly uses a recorded --transaction-id, a fresh preview revision and explicit execution.

Authority boundary: Local machine configuration and Goal capability selection do not grant external writes, override user gates or promote staged authority-store tenants. Remote/SSH read models remain read-only.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/docs/guides/personal-workspace-user-guide.md

loopx machine-config describe
loopx machine-config inspect
loopx configure-goal --goal-id "$GOAL_ID"

Periodic reports

Activation: Ask the active agent to generate this week’s project report for a one-session Markdown/HTML report. Recurrence requires an explicit custom profile and host Automation; machine/Goal subscriptions require periodic_report.enabled: true and an explicit route_ref.

Validation: Inspect the weekly preset: active and generation_allowed should both be true. Read the generated artifacts and bounded source receipts.

Disable / rollback: One-session generation creates no recurring job. Pause its Automation or set a custom profile to enabled: false; disable the machine/Goal subscription to stop subscribed delivery.

Authority boundary: The weekly preset has no sink and does not send. An enabled subscription with an explicit route is standing delivery authority, still subject to provider, identity and readback gates.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/loopx/capabilities/periodic_report/README.md

loopx periodic-report inspect-profile --preset weekly --format json

Lark Goal Channels

Activation: In notification settings → Lark → Connections, choose the Goal, registered target Agent, chat, capture scope and ingress mode, then save. Reconnect an existing topic without creating a duplicate.

Validation: Read back the connection/session binding. Send a new @ message yourself; for Async inbox, drain the exact Goal/Agent to inspect the event.

Disable / rollback: Choose Disconnect on that Goal connection. This removes its topic route and preserves Goal data, sessions, history and other connections.

Authority boundary: Steering targets an exact live turn; Queuing targets the same exact session; Async inbox waits for drain. Capture scope does not enlarge agent authority or authorize cross-topic replies.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/docs/guides/personal-workspace-user-guide.md

loopx lark-inbox drain --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"

Workspace stories

Activation: From a source checkout with Python 3.11+ and Node 22.6+, run the module below and open its printed loopback URL.

Validation: Inspect three projects, each with four work roles, 18 delivery tasks, two owner decisions and two scheduled watches; inspect their calculated budgets and sensitivity tables.

Disable / rollback: Stop with Ctrl-C. For a persistent replay, use prepare --root /tmp/workspace-stories, then serve --root /tmp/workspace-stories; choose a new empty directory to start over.

Authority boundary: These are authored scenario replays through real LoopX APIs, not customer outcomes or live-agent completion claims. State is isolated; no scheduler, agent, external message, purchase or deployment is started.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/demo/workspace/README.md

python -m demo.workspace serve

Outbound guidance recall

Activation: In the existing agent-scoped Reward Memory experiment, add outbound_message.before_send with scoped_feedback, the exact peer_ref: agent:…, and automation.automatic_recall: true; apply the reviewed config below.

Validation: Read experiment-status. Use lark-inbox send with --provider-preflight and without --execute to validate a configured route without sending.

Disable / rollback: Set automation.automatic_recall: false or remove this surface and reapply the experiment configuration.

Authority boundary: Advisory preferences do not authorize sending. This integration covers Goal/Agent-bound lark-inbox send and reply, not arbitrary messaging tools or Goal Topic auto-replies. No raw outgoing message enters the recall query.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/loopx/capabilities/reward_memory/OUTBOUND.md

loopx configure-goal --goal-id "$GOAL_ID" --reward-memory-config "$REWARD_CONFIG" --reward-memory-agent "$AGENT_ID" --execute
loopx reward-memory experiment-status --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"

Codex provider routing

Activation: From the source checkout and the same activated Python environment, install this optional package and its managed manifest.

Validation: Run extension doctor and the packaged content-free request. This validates routing/host observations, not live account ownership.

Disable / rollback: Run loopx extension disable loopx-codex-provider-routing --execute. A separately deployed CPA operator is stopped/unloaded independently; disabling this read-only extension does not stop that service.

Authority boundary: The extension is a read-only contract compiler/qualifier, not a proxy or credential authority. CPA and the operator retain their own routing, installation and credential responsibilities.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/packages/loopx-codex-provider-routing/README.md

python3 -m pip install packages/loopx-codex-provider-routing
loopx extension install --manifest packages/loopx-codex-provider-routing/extension.toml --execute
loopx extension doctor loopx-codex-provider-routing --execute
loopx extension run loopx-codex-provider-routing --input-json packages/loopx-codex-provider-routing/examples/request.json --execute

Benchmark study readback

Activation: Pass an existing compact study manifest and public-safe local record store to study-dashboard; the command itself is the opt-in.

Validation: Inspect campaign/arm/case/run projections, declared denominators and provisional coverage. Add --four-arm-contract-json only for a qualified matching four-arm design.

Disable / rollback: Stop invoking the read-only projection. Disable any separately activated upload provider through its own extension lifecycle.

Authority boundary: This read model does not launch experiments, change scoring, create Todo authority or authorize uploading raw tasks, trajectories, logs or hidden evaluator data.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/loopx/capabilities/benchmark_toolkit/README.md

loopx benchmark study-dashboard --manifest-json "$STUDY_MANIFEST" --store "$STUDY_STORE" --format json

Signed desktop updates

Activation: Install the new macOS App once to bootstrap older shells. In Recovery & Updates, explicitly check stable or main, install, then restart to pair the App with its bundled runtime.

Validation: Read App version and the Workspace runtime identity; use loopx --version and loopx doctor. Stable reads the dedicated desktop-stable feed; main reads complete signed main builds.

Disable / rollback: Use Repair for the current App’s runtime or Restore previous version when a verified backup exists, then restart. Keep an external installation backup; runtime rollback does not promise reversal of incompatible future Goal schemas.

Authority boundary: Only explicit native actions install. Feeds cannot supply arbitrary browser commands or runtime URLs. macOS uses ad-hoc code signing plus updater signatures, not notarization; Python 3.11+ is required. Windows artifacts use manual updates.

Docs: https://github.com/huangruiteng/loopx/blob/v1.0.0/apps/desktop/loopx-control-plane/README.md

loopx --version
loopx doctor

Install / Update

Python 3.11+ and Node 22.6+ are required for the packaged control plane. Existing installs preserve their pip, pipx or archive owner through loopx update; inspect its plan first. PyPI users can also pin the milestone directly:

python3 -m pip install --upgrade loopx==1.0.0
loopx workflow-skills --install
loopx slash-commands --install
loopx doctor

Keep your previous installation and private Goal data backup. App rollback restores an installation, not arbitrary state-schema migrations. No blanket persisted-state migration is requested by this milestone.

中文摘要

让长程 Agent 的工作有一个持续可用的控制面：正在做什么、卡在哪里、等谁决策、实际交付了什么，都能在同一个 Workspace 里检查和操作。

本次聚焦完整工作状态、Capability 配置、报告、飞书协作及 App/运行时配套更新；它是工作区的 1.0 里程碑，而非首次出现前端。升级与最小验证命令见上方。

升级决策

谁需要升级： 工作区、飞书/IM 和桌面用户，以及需要近期 Todo、宿主恢复修复的 operator；稳定 CLI-only 部署可先核对受影响能力。

解决了什么： 更多工作状态可见可操作，完成结果可验证，桌面获得签名、可恢复的 App 与运行时配套维护路径。

是否有破坏性变更： 无。1.0 是工作区里程碑，不代表分阶段 authority 后端全部提升；旧桌面壳需一次手动替换，可选投递与宿主能力仍按明确授权启用。

如何验证： package version 应为 1.0.0，检查 doctor，再读回 Goal 配置与 App/runtime 的源码身份。

贡献者： @Duang777, @now-ing, @steven-kid, @yuefengw, @songoow；另感谢工作区基础贡献者 @maxliux5，范围区分见下文。

loopx --version
loopx doctor
loopx machine-config inspect

工作区里程碑

1.0 是 Personal Workspace 的里程碑：让跨会话持续推进的工作有可检查、可操作的长程控制面。工作区在 v0.5.4 之前已经存在，本次将更丰富的状态与操作路径汇合起来。

  • 先进入工作区，再逐个补齐 Goal 状态：轻量目录先展示，慢 Goal 不阻塞其他项目，短暂失败自动重试，已停止 Goal 的详情按需加载（#4007, #4009）。
  • 汇总各工作 Agent、即时操作并可靠读回已完成任务（#3888, #3894, #3961）。
  • 配置 Goal capability 与机器策略，检查经过验证的报告和有界索引（#3860, #3861, #3865, #3951）。
  • 用真实前端与状态 API 重放复杂项目，场景能力沉淀进仓库（#3990）。

状态内核与集成

  • 原生 Todo 创建/更新、请求级幂等与完成恢复改进真实交付；已提升路径的 claim retry 提供精确身份，并补充跨 provider 的 lease 竞争验证（#3973, #3974, #3977, #3980, #3981, #3987, #3986）。
  • Stage2C 仍分阶段推进，内核路由和旧 writer fence 不代表全部既有 tenant 已提升（#3882, #3895）。
  • 重连已有飞书 Goal Topic、复用会话历史，在已授权发送前召回范围内偏好（#3983, #3868, #3968, #3998, #3999, #3992）。
  • Codex 路由补齐 Fast/dualAuto、quota 恢复及宿主/心跳兼容（#3984, #3892, #3880, #3913, #3890）。
  • 只读研究视图与终态 insight 保留 benchmark 原生指标、权限与评分口径（#3896, #3898, #3878, #3881）。

桌面与资格验证

签名更新将 App 与运行时配套升级；stable 不受插件 Latest 干扰，服务启动失败后仍能恢复，通道选择不会被轮询覆盖（#3994, #3996）。原生更新接口修正 IPC origin 匹配，更新源不可用与本机状态读取失败分别给出恢复提示。行为测试移除额外答案提示，增加诊断诱导反例及判分器 mutation checks（#3997）。发布检查改用有界时钟、typed followthrough 义务与稳定历史版本锚点；CLI 模块职责、帮助和手册同步且不提高预算（#4002, #4003）。不据此宣称 benchmark 或长程结果提升。

社区贡献者

特别感谢 @maxliux5 为工作区、任务优先导航、Manager/飞书集成及会话/流式体验打下基础（#3149, #3167, #3274, #3385, #3292, #3285, #3073）。这些工作早于 v0.5.4，本次作为 1.0 基础贡献致谢，不计作本次 tag range 新增贡献。

可选能力启用与使用

下方变量应替换为你已有的 Goal、Agent、已审阅配置和输入文件；配置写入需显式执行。

Workspace and machine configuration

启用： 运行 loopx dashboard，在 Goal 的 Capability 设置中配置能力。机器策略先发现 namespace、预览 JSON，再用返回的精确 plan revision 和 --execute 应用。

验证： 用 machine-config inspect 读回生效策略与 revision；用 configure-goal --goal-id "$GOAL_ID" 读回 Goal 配置。

停用 / 回退： Ctrl-C 关闭本地 dashboard。先 machine-config remove --namespace "$NAMESPACE" 预览，再带返回的 --expected-plan-revision 与 --execute 删除；rollback 用已记录的 transaction ID、最新预览 revision 与显式执行。

权限边界： 配置不授予外发权限、不跳过审批，也不自动提升处于分阶段迁移中的 authority-store tenant；远端/SSH 读模型保持只读。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/docs/guides/personal-workspace-user-guide.md

loopx machine-config describe
loopx machine-config inspect
loopx configure-goal --goal-id "$GOAL_ID"

Periodic reports

启用： 向当前 Agent 明确请求“生成本周项目报告”，即可在当前会话生成 Markdown/HTML。定时报告需自定义 profile 与 Automation；机器/Goal 订阅需显式 enabled: true 和 route_ref。

验证： 检查 weekly preset 的 active、generation_allowed 均为 true，并读回报告产物与有界来源凭据。

停用 / 回退： 单次生成不会留下定时任务；暂停 Automation 或设置 profile 的 enabled: false 可停止定时路径；关闭机器/Goal 订阅可停止订阅投递。

权限边界： weekly preset 不含 sink、不发送；启用且指定 route 的订阅构成持续投递授权，仍受 provider、身份和读回检查约束。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/loopx/capabilities/periodic_report/README.md

loopx periodic-report inspect-profile --preset weekly --format json

Lark Goal Channels

启用： 在通知设置 → 飞书 → Connections 选择 Goal、已注册 Agent、群聊、捕获范围及 ingress 模式并保存；可以重连既有话题而不重复建话题。

验证： 读回连接与 Session 绑定；自己发一条新的 @ 消息，Async inbox 可用精确 Goal/Agent 的 drain 查看事件。

停用 / 回退： 在该 Goal 连接中选择 Disconnect，只删除话题路由，保留 Goal、会话、历史和其他连接。

权限边界： Steering 指向精确活跃 turn，Queuing 指向同一精确 Session，Async inbox 等待 drain；捕获范围不扩大 Agent 权限或跨话题回复授权。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/docs/guides/personal-workspace-user-guide.md

loopx lark-inbox drain --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"

Workspace stories

启用： 在 Python 3.11+、Node 22.6+ 的源码 checkout 中运行下面命令，打开打印的 loopback URL。

验证： 检查三个项目：每个有四个工作角色、18 项交付任务、两个 owner 决策及两个定时观察；可查看计算生成的预算和敏感性表。

停用 / 回退： Ctrl-C 停止。固定重放用 prepare --root /tmp/workspace-stories 后再 serve --root /tmp/workspace-stories；重开进度选择新的空目录。

权限边界： 场景通过真实 API 重放，来源为编写的场景，不冒充客户结果或实时 Agent 完成记录；状态隔离，不启动调度、Agent、外发、购买或部署。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/demo/workspace/README.md

python -m demo.workspace serve

Outbound guidance recall

启用： 在已有 Agent 范围的 Reward Memory 实验中配置 outbound_message.before_send、scoped_feedback、精确 peer_ref: agent:… 和 automation.automatic_recall: true，再应用审阅后的配置。

验证： 读回 experiment-status；对已配置路由使用 lark-inbox send --provider-preflight 且不带 --execute，可以检查但不发送。

停用 / 回退： 设置 automation.automatic_recall: false 或移除此 surface，再应用配置。

权限边界： 偏好建议不授予发送权限；仅覆盖 Goal/Agent 绑定的 send/reply，不拦截任意消息工具或 Goal Topic 自动回复；召回查询不包含原始待发消息。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/loopx/capabilities/reward_memory/OUTBOUND.md

loopx configure-goal --goal-id "$GOAL_ID" --reward-memory-config "$REWARD_CONFIG" --reward-memory-agent "$AGENT_ID" --execute
loopx reward-memory experiment-status --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"

Codex provider routing

启用： 在源码 checkout 的同一 Python 环境安装可选 package 及 managed manifest。

验证： 运行 extension doctor 及随包的无内容请求，验证路由/宿主观察合同，不代表验证实际账号归属。

停用 / 回退： 执行 loopx extension disable loopx-codex-provider-routing --execute；独立部署的 CPA operator 需单独停止/卸载服务，禁用此只读扩展不会停止它。

权限边界： 扩展是只读合同编译/资格检查器，不是代理或凭证 authority；CPA 与 operator 分别拥有自己的路由、安装和凭证职责。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/packages/loopx-codex-provider-routing/README.md

python3 -m pip install packages/loopx-codex-provider-routing
loopx extension install --manifest packages/loopx-codex-provider-routing/extension.toml --execute
loopx extension doctor loopx-codex-provider-routing --execute
loopx extension run loopx-codex-provider-routing --input-json packages/loopx-codex-provider-routing/examples/request.json --execute

Benchmark study readback

启用： 向 study-dashboard 提供已有的精简 manifest 和公开安全的本地记录 store；按次调用即为 opt-in。

验证： 查看 campaign/arm/case/run 投影、分母及 provisional 覆盖；只有匹配且合格的四臂设计才提供 four-arm contract。

停用 / 回退： 停止调用只读投影即可；若另行启用了上传 provider，通过该扩展自己的生命周期停用。

权限边界： 读模型不启动实验、不改评分、不创建 Todo authority，也不授权上传原始任务、轨迹、日志或隐藏评测数据。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/loopx/capabilities/benchmark_toolkit/README.md

loopx benchmark study-dashboard --manifest-json "$STUDY_MANIFEST" --store "$STUDY_STORE" --format json

Signed desktop updates

启用： 旧壳先手动安装新版 macOS App。在恢复与更新中显式检查 stable/main、安装并重启，使 App 与内置运行时配套。

验证： 读回 App 版本和工作区 runtime identity，并运行 version/doctor。stable 使用独立 desktop-stable feed，main 使用完整签名构建。

停用 / 回退： Repair 重装当前 App 的运行时；存在已验证备份时可恢复上版并重启。保留外部安装备份；运行时回滚不承诺逆转未来不兼容的 Goal schema。

权限边界： 仅显式原生动作执行安装，不接受浏览器任意命令/运行时 URL。macOS 为 ad-hoc code signing 加 updater 签名、尚未 notarize，需 Python 3.11+；Windows 手动更新。

文档： https://github.com/huangruiteng/loopx/blob/v1.0.0/apps/desktop/loopx-control-plane/README.md

loopx --version
loopx doctor

发布验证

Validation and provenance (English). Qualification is source-specific. Doubao evolving completed 21 scenarios, 42 actor attempts and 6 contrasts (66 live API calls), with no failed or skipped cases, on candidate 749f9004. That candidate also passed the TypeScript suite (625 passed, one optional skip), real isolated PostgreSQL integration (26 passed) and isolated native tests (25 passed). These are qualification results, not claims of benchmark uplift.

The initial full pytest run had 6,151 passes, two stale CLI-test assertions and 25 skips. After repairing the assertions for the extracted CLI owners, both hosted pytest shards and the aggregate gate passed on 7e94febe. The owner requested no further full-kernel rerun for the final frontend/desktop fixes. Final coverage combines this kernel baseline with scoped delta validation; it is not a new full-suite receipt for the tag.

On 40aaddb6, active-first loading and recovery passed Chromium/WebKit checks, the Lark binding-isolation regression passed, and the installer recovery smoke passed. On final candidate 14c5f877, native maintenance tests passed (24 passed, two environment-dependent cases not rerun in that unit invocation), release Clippy and dashboard TypeScript/build passed, five desktop packaging/feed tests passed, and the packaged browser smoke passed desktop/mobile update, IPC failure/readback recovery, feed failure, confirmation, redaction and startup recovery checks. The exact signed App was installed locally; runtime identity matched the final source and doctor passed required checks. Actual archive signature verification passed and tampering was rejected. Native GUI clickthrough was unavailable; browser and native contract checks are reported separately.

The earlier full-public fleet was not fully green: 512 checks ran with four failures involving the optional native Codex benchmark profile, nested canary installation, disk exhaustion during help/manpage generation and a promotion-concurrency installer timeout. The focused installer retry later passed. Remaining fleet gaps are retained as limitations; this release does not claim every optional benchmark/host profile passed. Doctor also reports non-blocking local skill-route and projection warnings, separate from its passing required checks.

验证与源码对应（中文）。 Doubao evolving 在候选 749f9004 上完成 21 个场景、42 次 actor 尝试和 6 组对照，共 66 次真实 API 调用，无失败或跳过；同一基线还通过 TypeScript（625 通过、1 个可选跳过）、真实隔离 PostgreSQL（26 通过）及隔离原生测试（25 通过）。这些是资格验证结果，不用于宣称 benchmark 提升。

最初全量 pytest 为 6,151 通过、2 个旧 CLI 断言失败、25 跳过。修正模块职责抽取后的旧断言后，7e94febe 的 GitHub 两个 pytest 分片及汇总门禁均通过。遵照维护者要求，最后的前端/桌面修复不再重跑完整内核；发布证据明确区分已有内核基线与最终增量验证。

40aaddb6 通过 active-first 渐进加载、Chromium/WebKit 恢复验证、飞书绑定隔离回归及安装恢复 smoke。最终候选 14c5f877 通过原生维护测试（24 通过；该次单测未重跑两个依赖隔离环境的用例）、release Clippy、前端类型检查和构建、5 项桌面打包/更新源测试，以及覆盖桌面/窄屏、IPC 状态读取恢复、更新源故障、安装确认、脱敏和启动恢复的浏览器回归。本机已安装这份签名 App，运行时身份匹配最终源码，doctor 必需检查通过；真实签名校验通过且能拒绝篡改。原生 GUI 点击验收不可用，未将浏览器验证冒称为原生 GUI 验收。

此前 full-public 的 512 项检查有 4 项失败，分别涉及可选原生 Codex benchmark profile、嵌套 canary 安装、帮助/手册检查时磁盘耗尽及 promotion 并发安装超时；后续安装恢复专项重试已通过。其余 fleet 缺口仍明确保留，不宣称所有可选 benchmark/宿主 profile 均通过。doctor 的本机 skill 路由及 projection 非阻断提示也与必需检查结果分开记录。

Compare: v0.5.4...v1.0.0

Contributors

maxliux5, songoow, and 4 other contributors
Assets 13
Loading