@viceme-ai/cli 0.8.0 → 0.8.2

package/CHANGELOG.md

@@ -1,5 +1,28 @@
 # Changelog
 
+## [0.8.2] - 2026-07-25
+
+### Fixes
+
+- accept typed source specs (`167656c`)
+
+## [0.8.1] - 2026-07-24
+
+### Fixes
+
+- 修正 Skill 编辑与等待契约 (`e5b92d9`)
+- 修正 ViceMe Skill 发布命令契约 (`23f140b`)
+- 明确步骤确认摘要路径 (`adc70a3`)
+- 六轮复审——confirm_steps 指引消除 preview 死锁 (`3b33b91`)
+- T2 复审修复——自然语言走 stdin 结构化传输/超时保留 edit-run ID/resume 对齐确认端点 (`852c234`)
+
+### Other Changes
+
+- separate business and protocol results (`f1f9f9d`)
+- 内置 Agent Skill 对齐交互步骤确认门(产品 3427) (`1992801`)
+- 中文化 T2 复审交接 (`175e838`)
+- add T2 review remediation handoff (`e689dff`)
+
 ## [0.8.0] - 2026-07-22
 
 ### Features

package/README.md

@@ -10,7 +10,7 @@ The official command-line client and Agent Skill for publishing external Skills
 
 [Install](#installation--quick-start) · [AI Agent Skills](#agent-skills) · [Auth](#authentication) · [Regions & profiles](#regions--profiles) · [Commands](#command-overview) · [Output contract](#json-output-contract) · [Security](#security-and-risk-controls) · [Development](#development)
 
-> **Rollout status:** the Core publication transport and stable-link path are implemented, and the exact Candidate preview → test run → result-confirmation gate is enforced: after `--yes`, the publication parks at `awaiting_action` with a typed `confirm_publish` action, and `job resume --decision confirm` is accepted only after the exact candidate has a succeeded, owner-accepted preview test run (otherwise 409 `preview_run_required`). Test runs, acceptance, and natural-language candidate edits are driven from the confirmation page (`next_action.payload.preview_url`) or the `/v1/skill-agent-publications/:id/preview-runs` and `/edits` endpoints; the CLI ships no separate commands for them. `--yes` confirms the publication request; it is not proof that the user reviewed the final Candidate.
+> **Rollout status:** the Core publication transport and stable-link path are implemented, and the metadata review plus exact Candidate preview → test run → result-confirmation gates are enforced. After `--yes`, the publication first parks at `meta_review` for `job metadata`, then at `awaiting_action` for `job preview`, optional `job edit`, `job run`, `job accept`, and `job resume`. Confirm is accepted only after the exact Candidate has a succeeded, owner-accepted preview run (otherwise 409 `preview_run_required`). A confirm receipt authorizes release but does not contain the final share link; run another bounded `job wait` until `share_published`. `--yes` confirms only the publication request, not the metadata or final Candidate.
 
 ## Why ViceMe CLI?
 
@@ -18,7 +18,7 @@ The official command-line client and Agent Skill for publishing external Skills
 - **Deterministic boundary** — the CLI performs typed protocol actions; it does not start another conversational Agent loop.
 - **Server-side compilation** — source parsing, LLM compilation, BuildRun materialization, and Release publication stay on ViceMe infrastructure.
 - **Stable publishing** — later releases of the same logical Agent keep the same share URL.
-- **Multiple source types** — accepts GitHub Skills, pasted Xiaohongshu/RedSkill expressions, archives, and local Skill folders.
+- **Multiple source types** — accepts typed GitHub and Xiaohongshu/RedSkill sources, archives, and local Skill folders.
 - **Secure by default** — on macOS, device-login credentials use AES-256-GCM encrypted files with Keychain-backed key material; other platforms retain their native credential manager. Explicit local overrides require a private profile file, public mutations require confirmation, and downloaded binaries are checksum-verified.
 - **Human and Agent login modes** — `viceme auth login` guides a person in the terminal, while Agent split-flows use explicit JSON.
 
@@ -83,7 +83,7 @@ Use the complete bootstrap command from the bundled Skill. The explicit npm regi
 npx --yes --registry=https://registry.npmjs.org --@viceme-ai:registry=https://registry.npmjs.org --package=@viceme-ai/cli@latest -- viceme install
 ```
 
-Read `data.authenticated` and `data.next_step` from the result. If authentication is already valid, continue to Step 4. If login is required, do not execute the human-oriented `data.next_step` inside the Agent; use the JSON split-flow in Step 2.
+Read `authenticated` and `next_step` from the formatted business result. If authentication is already valid, continue to Step 4. If login is required, do not execute the human-oriented `next_step` inside the Agent; use the JSON split-flow in Step 2.
 
 **Step 2 — Start device login when required**
 
@@ -91,7 +91,7 @@ Read `data.authenticated` and `data.next_step` from the result. If authenticatio
 viceme auth login --no-wait --json
 ```
 
-Return the exact `data.verification_url`; the CLI normalizes it to the prefilled `verification_url_complete` browser link when available. Include `data.user_code` only as a fallback if the browser asks for it. Preserve `data.device_code` for the continuation command, then stop the current turn. Do not request, print, or place an access token in the conversation.
+Return the exact `verification_url`; the CLI normalizes it to the prefilled `verification_url_complete` browser link when available. Include `user_code` only as a fallback if the browser asks for it. Preserve `device_code` for the continuation command, then stop the current turn. Do not request, print, or place an access token in the conversation.
 
 **Step 3 — Continue the same login in a later turn**
 
@@ -120,7 +120,7 @@ Continue only when authentication is valid and `skills doctor` reports a healthy
 viceme skill inspect https://github.com/acme/poster-skill --skill-root .
 ```
 
-Inspection is read-only. Follow the bundled `viceme` Skill for source-specific handling, Target selection, confirmation, bounded job waiting, and result reporting. If a publication ends at `binding_required`, run `viceme job bind <publication-id>`, give the signed ViceMe URL to the user, and stop. Downloading or forking is only an informational alternative; the CLI never performs it automatically. After the user binds the exact GitHub/Xiaohongshu channel account, inspect again and create a fresh ordinary publication rather than resuming the terminal one. Otherwise, when the publication parks at `awaiting_action`, guide the user through candidate preview and a test run with accepted result before resolving with `job resume --decision confirm` (see the rollout status above).
+Inspection is read-only. Follow the bundled `viceme` Skill for source-specific handling, Target selection, confirmation, bounded job waiting, and result reporting. If a publication ends at `binding_required`, run `viceme job bind <publication-id>`, give the signed ViceMe URL to the user, and stop. Downloading or forking is only an informational alternative; the CLI never performs it automatically. After the user binds the exact GitHub/Xiaohongshu channel account, inspect again and create a fresh ordinary publication rather than resuming the terminal one. At `meta_review`, show and resolve the metadata using the exact action ID and payload digest, then wait again. At `awaiting_action`, show the frozen Candidate summary, run and accept one exact-Candidate test result, and obtain the user's decision before `job resume`. After confirm, wait again until `share_published` before returning the share link.
 
 ## Regions & Profiles
 
@@ -219,14 +219,19 @@ viceme skill publish --resolution-id <resolution-id> --yes
 
 For GitHub, `--skill-root` is required and names the exact repository-relative directory containing `SKILL.md`; use `.` only for a root-level Skill. The calling Agent determines this path from the user input or read-only repository tree. ViceMe does not scan the repository to guess a Skill.
 
-### Xiaohongshu or RedSkill copied expression
+### Xiaohongshu or RedSkill
 
 ```bash
-viceme skill inspect --expression-stdin
+viceme skill inspect --source-stdin
 viceme skill publish --resolution-id <resolution-id> --yes
 ```
 
-The copied expression is untrusted data. ViceMe extracts a locator and fetches the source through an approved connector; it never executes marketplace installation text.
+The AI Host interprets the user's source intent and passes one typed JSON
+`SourceSpec` through stdin, for example
+`{"kind":"redskill","value":"ai-desk-card"}`. CLI/Core do not classify copied
+natural language with keyword or regex rules. Explicit platform intent is
+authoritative; ambiguous source requests must be clarified instead of silently
+substituting a same-name source from another provider.
 
 ### Archive or local Skill folder
 
@@ -254,30 +259,43 @@ viceme skill publish --file ./poster-skill-v2.zip \
 | `viceme skill inspect` | Freeze and inspect a source candidate without publishing |
 | `viceme skill publish` | Create or update a stable Skill Agent publication |
 | `viceme skill target` | Resolve existing logical Agent Targets and versions |
-| `viceme job` | Read, wait for, show a signed channel-binding URL, resume, explicitly retry, or cancel a durable publication |
+| `viceme job` | Read or wait for a publication, review metadata, preview/edit/test/accept its Candidate, show a signed channel-binding URL, resume an action, explicitly retry, or cancel |
 | `viceme skills` | Read, install, and diagnose the bundled Agent Skill |
 | `viceme update` | Update the npm launcher, verified binary, and bundled Skill together |
 
 Use `viceme <command> --help` for the exact flags. The release-checked machine-readable surface is stored in [`skills/viceme/references/command-manifest.json`](skills/viceme/references/command-manifest.json).
 
-## JSON Output Contract
+## Output Contract
 
-Automation-oriented data commands emit a stable JSON envelope by default. Interactive `viceme auth login` is the deliberate human-facing exception; AI Agents and scripts must use `--no-wait --json`, then continue with `--device-code <code> --json` in a later turn.
+ViceMe selects the smallest stable representation for each command:
 
-Success is written to **stdout** with exit code `0`:
+- Local/bootstrap commands such as `version`, `install`, `update`, `auth status`, `profile *`, and `skills doctor` write their formatted business result directly to **stdout**. They do not add `ok`, `data`, or unrelated build metadata.
+- `skills read` writes the requested file byte-for-byte without a JSON wrapper.
+- Interactive `viceme auth login` writes human guidance. AI Agents use `--no-wait --json`, then continue with `--device-code <code> --json`; those two commands return a formatted bare business object.
+- Publication protocol commands under `skill` and `job` keep a stable envelope because action receipts, durable status, and bounded-wait metadata form one cross-command protocol.
+
+A successful publication protocol result is written to **stdout** with exit code `0`:
+
+```json
+{
+  "ok": true,
+  "data": {}
+}
+```
+
+Only a bounded wait that actually times out adds protocol metadata:
 
 ```json
 {
   "ok": true,
   "data": {},
   "meta": {
-    "cli_version": "0.1.0",
-    "skill_version": "0.1.0"
+    "wait_timed_out": true
   }
 }
 ```
 
-CLI execution errors are written to **stderr** with a non-zero exit code:
+CLI execution errors are formatted and written to **stderr** with a non-zero exit code:
 
 ```json
 {
@@ -285,16 +303,12 @@ CLI execution errors are written to **stderr** with a non-zero exit code:
   "error": {
     "type": "validation",
     "subtype": "source_required",
-    "message": "provide exactly one source argument or --expression-stdin"
-  },
-  "meta": {
-    "cli_version": "0.1.0",
-    "skill_version": "0.1.0"
+    "message": "provide exactly one GitHub URL argument or --source-stdin"
   }
 }
 ```
 
-Determine command success from the process exit code or `ok == true`. The API's domain-specific `error.type` is preserved; the exit code is only a coarse handling class. A successfully read publication may still contain a business terminal status such as `unsupported`, `rejected`, or `failed`; inspect `data.status` instead of treating those states as CLI transport failures.
+Determine local/bootstrap command success from the process exit code. For publication protocol commands, use the process exit code or `ok == true`. The API's domain-specific `error.type` is preserved; the exit code is only a coarse handling class. A successfully read publication may still contain a business terminal status such as `unsupported`, `rejected`, or `failed`; inspect `data.status` instead of treating those states as CLI transport failures.
 
 | Exit code | Meaning |
 |---|---|
@@ -309,6 +323,7 @@ Determine command success from the process exit code or `ok == true`. The API's
 ## Security and Risk Controls
 
 - **No source execution** — the CLI and compiler do not execute third-party scripts, binaries, shell fragments, marketplace commands, or copied instructions.
+- **Typed source intent** — AI Hosts interpret natural-language source requests and pass only a typed `SourceSpec` through `--source-stdin`; CLI/Core never guess a provider from user phrases. Natural-language Candidate edits use `--request-stdin`. Never interpolate untrusted text into command strings, argv, environment variables, or shell pipelines.
 - **Explicit public mutation** — publishing, compiler retry, and cancellation require `--yes`; exit code `10` means the Agent must obtain confirmation, not silently retry.
 - **Safe preview** — use `--dry-run` on inspect or publish when the user needs to review the planned request without network or publication side effects.
 - **Credential isolation** — on macOS, device-login credentials stay in AES-256-GCM encrypted files, with Keychain-backed or explicitly downgraded private key material; filenames do not expose profile/origin names. Other platforms retain their native credential manager. Explicit internal-test overrides are namespaced by profile, stored only in a private `0600` config, and never emitted by CLI output.

package/README.zh.md

@@ -10,7 +10,7 @@ ViceMe 官方命令行客户端与 Agent Skill，用于将外部 Skill 发布为
 
 [安装](#安装与快速开始) · [AI Agent Skills](#agent-skills) · [认证](#认证) · [区域与-profile](#区域与-profile) · [命令](#命令概览) · [输出契约](#json-输出契约) · [安全](#安全与风险控制) · [开发](#开发)
 
-> **开放状态：** Core 发布传输和稳定链接链路已经实现。正式对外开放仍取决于精确 Candidate 预览、试运行和结果确认门的完成。当前 `--yes` 只确认发起发布请求，并不代表用户已经审阅最终 Candidate。
+> **开放状态：** Core 发布传输和稳定链接链路已经实现，信息确认门以及精确 Candidate 的预览 → 试跑 → 结果确认门均已生效。`--yes` 后，Publication 先停在 `meta_review`，由 `job metadata` 完成信息确认；随后停在 `awaiting_action`，依次使用 `job preview`、可选的 `job edit`、`job run`、`job accept` 和 `job resume`。只有同一 Candidate 的试跑成功且结果已由发布者接受时才能确认，否则返回 409 `preview_run_required`。确认回执只授权发布，不包含最终分享链接；需要再次有界执行 `job wait` 直到 `share_published`。`--yes` 只确认发起发布请求，不代表用户已确认信息或最终 Candidate。
 
 ## 为什么选择 ViceMe CLI？
 
@@ -18,7 +18,7 @@ ViceMe 官方命令行客户端与 Agent Skill，用于将外部 Skill 发布为
 - **确定性边界** — CLI 执行类型化协议操作，不会再启动一层对话式 Agent Loop。
 - **服务端编译** — 来源解析、LLM 编译、BuildRun 固化和 Release 发布均在 ViceMe 基础设施中完成。
 - **稳定发布** — 同一个逻辑 Agent 后续发布新版本时继续使用同一个分享链接。
-- **支持多种来源** — 支持 GitHub Skill、小红书或 RedSkill 复制口令、压缩包和本地 Skill 目录。
+- **支持多种来源** — 支持结构化 GitHub 与小红书/RedSkill 来源、压缩包和本地 Skill 目录。
 - **默认安全** — 在 macOS 上，设备登录凭证保存在 AES-256-GCM 加密文件中，主密钥默认由系统 Keychain 保护；其他平台继续使用原生凭证管理器。显式本地覆盖要求私有 Profile 文件，公开变更需要确认，下载的二进制文件必须通过校验和验证。
 - **人类与 Agent 双登录模式** — `viceme auth login` 在终端中引导用户，Agent 跨回合流程则显式使用 JSON。
 
@@ -83,7 +83,7 @@ viceme skills doctor
 npx --yes --registry=https://registry.npmjs.org --@viceme-ai:registry=https://registry.npmjs.org --package=@viceme-ai/cli@latest -- viceme install
 ```
 
-读取结果中的 `data.authenticated` 和 `data.next_step`。如果认证已经有效，直接进入第 4 步。如果需要登录，Agent 不要执行面向人类的 `data.next_step`，而应使用第 2 步的 JSON 跨回合流程。
+读取格式化业务结果中的 `authenticated` 和 `next_step`。如果认证已经有效，直接进入第 4 步。如果需要登录，Agent 不要执行面向人类的 `next_step`，而应使用第 2 步的 JSON 跨回合流程。
 
 **第 2 步 — 需要时启动设备登录**
 
@@ -91,7 +91,7 @@ npx --yes --registry=https://registry.npmjs.org --@viceme-ai:registry=https://re
 viceme auth login --no-wait --json
 ```
 
-向用户返回准确的 `data.verification_url`；存在 `verification_url_complete` 时，CLI 会把这个已预填设备码的浏览器直达链接规范化为 `verification_url`。只有浏览器要求输入时，才把 `data.user_code` 作为备用信息提供。保留 `data.device_code` 供后续命令使用，然后结束当前回合。不要在对话中索取、打印或传递访问令牌。
+向用户返回准确的 `verification_url`；存在 `verification_url_complete` 时，CLI 会把这个已预填设备码的浏览器直达链接规范化为 `verification_url`。只有浏览器要求输入时，才把 `user_code` 作为备用信息提供。保留 `device_code` 供后续命令使用，然后结束当前回合。不要在对话中索取、打印或传递访问令牌。
 
 **第 3 步 — 在后续回合继续同一个登录流程**
 
@@ -120,7 +120,7 @@ viceme skills list
 viceme skill inspect https://github.com/acme/poster-skill --skill-root .
 ```
 
-inspect 是只读操作。后续应按照随包发布的 `viceme` Skill 处理不同来源、Target 选择、用户确认、有界任务等待和结果返回。若 Publication 终结为 `binding_required`，运行 `viceme job bind <publication-id>`，把服务端签名的 ViceMe 链接交给用户后停止；下载或 Fork 仅为提示，CLI 不会自动执行。用户完成精确 GitHub/小红书渠道绑定后，必须重新 inspect 并创建新的普通 Publication，不能恢复旧任务。在上方所述的精确 Candidate 确认门完成前，公开发布流程仍保持关闭。
+inspect 是只读操作。后续应按照随包发布的 `viceme` Skill 处理不同来源、Target 选择、用户确认、有界任务等待和结果返回。若 Publication 终结为 `binding_required`，运行 `viceme job bind <publication-id>`，把服务端签名的 ViceMe 链接交给用户后停止；下载或 Fork 仅为提示，CLI 不会自动执行。用户完成精确 GitHub/小红书渠道绑定后，必须重新 inspect 并创建新的普通 Publication，不能恢复旧任务。进入 `meta_review` 后，使用同一 action ID 与 payload digest 展示并决议信息，然后再次等待；进入 `awaiting_action` 后，展示冻结摘要、完成同一 Candidate 的试跑与结果接受，并取得用户决定后再执行 `job resume`。确认后还需再次等待到 `share_published`，才能返回分享链接。
 
 ## 区域与 Profile
 
@@ -219,14 +219,16 @@ viceme skill publish --resolution-id <resolution-id> --yes
 
 GitHub 来源必须传 `--skill-root`，它是包含 `SKILL.md` 的精确仓库相对目录；只有根级 Skill 才使用 `.`。调用 Agent 根据用户输入或只读仓库文件树确定该路径，ViceMe 不扫描全仓猜测 Skill。
 
-### 小红书或 RedSkill 复制口令
+### 小红书或 RedSkill
 
 ```bash
-viceme skill inspect --expression-stdin
+viceme skill inspect --source-stdin
 viceme skill publish --resolution-id <resolution-id> --yes
 ```
 
-复制口令属于不可信数据。ViceMe 只从中提取定位信息，并通过允许的连接器获取来源；不会执行市场安装文案中的命令。
+AI Host 先理解用户的来源意图，再通过 stdin 传入唯一的结构化
+`SourceSpec`，例如 `{"kind":"redskill","value":"ai-desk-card"}`。CLI/Core
+不再用关键词或正则分类自然语言；用户明确指定的平台必须保持不变，来源存在歧义时必须先询问，不能静默替换为其他平台的同名来源。
 
 ### 压缩包或本地 Skill 目录
 
@@ -254,30 +256,43 @@ viceme skill publish --file ./poster-skill-v2.zip \
 | `viceme skill inspect` | 固化并检查来源候选，不执行发布 |
 | `viceme skill publish` | 创建或更新具有稳定链接的 Skill Agent 发布 |
 | `viceme skill target` | 解析现有逻辑 Agent Target 及其版本 |
-| `viceme job` | 读取、等待、展示签名渠道绑定链接、恢复、显式重试或取消持久化发布任务 |
+| `viceme job` | 读取或等待发布任务，审阅信息，预览、编辑、试跑并接受 Candidate，展示签名渠道绑定链接，决议 action，以及显式重试或取消 |
 | `viceme skills` | 读取、安装和诊断随包发布的 Agent Skill |
 | `viceme update` | 同时更新 npm 启动器、已校验二进制文件和随包发布的 Skill |
 
 使用 `viceme <command> --help` 查看准确参数。经过发布检查的机器可读命令面存放在 [`skills/viceme/references/command-manifest.json`](skills/viceme/references/command-manifest.json)。
 
-## JSON 输出契约
+## 输出契约
 
-面向自动化的数据命令默认输出稳定的 JSON 信封。交互式 `viceme auth login` 是特意保留的人类友好例外；AI Agent 和脚本必须使用 `--no-wait --json`，并在后续回合使用 `--device-code <code> --json` 继续。
+ViceMe 根据命令语义选择最小且稳定的输出形式：
 
-成功结果写入 **stdout**，退出码为 `0`：
+- `version`、`install`、`update`、`auth status`、`profile *`、`skills doctor` 等本地/引导命令，将格式化后的业务结果直接写入 **stdout**，不附加 `ok`、`data` 或无关构建元数据。
+- `skills read` 按原始字节输出目标文件，不添加 JSON 包装。
+- 交互式 `viceme auth login` 输出面向人的引导；AI Agent 使用 `--no-wait --json`，并在后续回合用 `--device-code <code> --json` 继续，这两个命令返回格式化的裸业务对象。
+- `skill` 和 `job` 下的发布协议命令继续使用稳定 Envelope，因为 action receipt、持久状态与有界等待元数据共同构成跨命令协议。
+
+发布协议命令成功时写入 **stdout**，退出码为 `0`：
+
+```json
+{
+  "ok": true,
+  "data": {}
+}
+```
+
+只有有界等待真实超时时才附加协议元数据：
 
 ```json
 {
   "ok": true,
   "data": {},
   "meta": {
-    "cli_version": "0.1.0",
-    "skill_version": "0.1.0"
+    "wait_timed_out": true
   }
 }
 ```
 
-CLI 执行错误写入 **stderr**，退出码非零：
+CLI 执行错误以格式化形式写入 **stderr**，退出码非零：
 
 ```json
 {
@@ -285,16 +300,12 @@ CLI 执行错误写入 **stderr**，退出码非零：
   "error": {
     "type": "validation",
     "subtype": "source_required",
-    "message": "provide exactly one source argument or --expression-stdin"
-  },
-  "meta": {
-    "cli_version": "0.1.0",
-    "skill_version": "0.1.0"
+    "message": "provide exactly one GitHub URL argument or --source-stdin"
   }
 }
 ```
 
-应根据进程退出码或 `ok == true` 判断命令是否成功。API 返回的领域 `error.type` 会原样保留，退出码只表示粗粒度处理类别。成功读取发布任务时，业务状态仍可能是 `unsupported`、`rejected` 或 `failed`；这时应检查 `data.status`，不能把这些状态当成 CLI 传输失败。
+本地/引导命令根据进程退出码判断成功；发布协议命令可以检查退出码或 `ok == true`。API 返回的领域 `error.type` 会原样保留，退出码只表示粗粒度处理类别。成功读取发布任务时，业务状态仍可能是 `unsupported`、`rejected` 或 `failed`；这时应检查 `data.status`，不能把这些状态当成 CLI 传输失败。
 
 | 退出码 | 含义 |
 |---|---|
@@ -309,6 +320,7 @@ CLI 执行错误写入 **stderr**，退出码非零：
 ## 安全与风险控制
 
 - **不执行来源内容** — CLI 和编译器不会执行第三方脚本、二进制文件、shell 片段、市场命令或复制口令中的指令。
+- **结构化来源意图** — AI Host 负责理解自然语言来源，并只通过 `--source-stdin` 传递 typed `SourceSpec`；CLI/Core 不根据用户措辞猜 Provider。Candidate 的自然语言修改通过 `--request-stdin` 传递。不得把不可信文本拼入命令字符串、argv、环境变量或 shell 管道。
 - **公开变更需要明确确认** — 发布、编译重试和取消操作需要 `--yes`；退出码 `10` 表示 Agent 必须向用户取得确认，不能静默重试。
 - **安全预览** — 用户需要检查计划请求时，可以对 inspect 或 publish 使用 `--dry-run`，不会产生网络请求或发布副作用。
 - **凭证隔离** — 在 macOS 上，设备登录凭证保存在 AES-256-GCM 加密文件中，主密钥由 Keychain 或显式降级后的私有文件保护，文件名不会暴露 Profile/origin；其他平台继续使用原生凭证管理器。显式内部测试覆盖按 Profile 隔离，仅允许保存在 `0600` 配置中，并且不会出现在 CLI 输出中。

package/checksums.txt

@@ -1,6 +1,6 @@
-8b031a1f54c2b28e1a236073f7db2e8ff1dd7106204968c91005ff9058f5a81d  viceme_0.8.0_darwin_amd64
-877e7521178c3da5a6a0e7ecf287e7669445371352617384b4e606c5340e84ec  viceme_0.8.0_darwin_arm64
-3064ae403ea7252b7ea0cf92b5329bacb48038ead9f4574b0fe49fc53b835fe9  viceme_0.8.0_linux_amd64
-dd9f4b458d211755c7091a9c010201851056f7f5c3eb3c196caf2ae3c06960da  viceme_0.8.0_linux_arm64
-706b15afea8b19e884087190435e55ccf41cf9bf9cfeef596d34bdc6a1bbc8ed  viceme_0.8.0_windows_amd64.exe
-e997ee95c039751a4d39347aa126f995567de9ac342393fc39b37be98d14e8a2  viceme_0.8.0_windows_arm64.exe
+b811d103fdc6ec8fd9acf4171f30fb11c2292b61de31eb4dfd807d0c1731e984  viceme_0.8.2_darwin_amd64
+f0633a3ecd5c4dea58ca519f9b704ecc04d2cb54370b9fa35c7f15261c46b150  viceme_0.8.2_darwin_arm64
+7f9b6cc67b6c6052a9d3032293f57ffbb16084abd82703b2877b87bde8f768d3  viceme_0.8.2_linux_amd64
+3e94d4896bb91a141e2269e74cd7a83f810ad65e857c1018824b95f0c83d43ae  viceme_0.8.2_linux_arm64
+b91d2b2b00734f12fdb008a002d1cb679903bd348da9c993ce5b472e850b18d5  viceme_0.8.2_windows_amd64.exe
+6dcd67f67750836ea808dae7d450cbaabd1ceef424f771c8baebab1877f16d3b  viceme_0.8.2_windows_arm64.exe

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@viceme-ai/cli",
-  "version": "0.8.0",
+  "version": "0.8.2",
   "description": "Install and run the ViceMe CLI and ViceMe Agent Skill",
   "type": "module",
   "bin": {