@thinkingai/ae-cli 6.1.27 → 6.1.29

package/README.md

@@ -17,7 +17,6 @@ The CLI is designed around:
 - Node.js 20 or later.
 - macOS, Linux, or native Windows (PowerShell / Command Prompt); WSL is optional.
 - Access to the npm registry for CLI installation.
-- Access to GitHub only when the bundled Skills source cannot be used.
 
 ## Installation
 
@@ -30,7 +29,7 @@ npx @thinkingai/ae-cli@latest install
 npx @thinkingai/ae-cli@latest install --format json
 ```
 
-The installer checks Node.js 20+, installs and verifies the CLI, then installs matching Skills globally in `~/.agents/skills`, writing the shared files once before automatically distributing copies to clients that need their own directory. No Agent selection is needed; Codex and Cursor can share these Skills even when installed later. `--dry-run`, `--format json`, and optional `--host <AE_HOST>` are supported. Unrelated project-only client warnings are hidden after successful verification; missing files and YAML errors remain failures. Being logged out is not an installation failure. See the [installation guide](./cli-installation-guide.md).
+The installer checks Node.js 20+, installs and verifies the CLI, then uses the bundled Skills tool to install matching global Skills for supported clients. First installation includes all bundled Skills when no business state has been confirmed. `--dry-run`, `--format json`, and optional `--host <AE_HOST>` are supported. Being logged out is not an installation failure. See the [installation guide](./cli-installation-guide.md).
 
 The Skills teach supported coding Agents, including Claude Code, Codex, Cursor, and similar tools, how to discover and call `ae-cli`.
 
@@ -51,6 +50,10 @@ ae-cli auth status
 
 Use the AE URL supplied by your AgenticEngine administrator. If no Host is configured, `ae-cli` guides existing customers to obtain that URL first. Only users without an AgenticEngine environment are directed to [request a trial](https://thinkingai.cn/request-demo).
 
+Business-based command filtering and Skill selection apply only to local CLI use by external Agents; built-in TE Agent sandboxes keep the complete command and Skill sets. After successful login, the CLI filters `--help` and capability discovery by the active environment’s enabled businesses and synchronizes managed global Skills. Analysis remains available by default. Run `ae-cli config refresh` after enabling a business to fetch its state and restore its Skills. Switching the active environment or account synchronizes again; a one-off `--host` query does not change global Skills.
+
+Only explicitly disabled businesses are removed. Older servers, missing states, and network failures do not trigger removal. Legacy client links to a verified, unmodified managed shared Skill are migrated automatically; other modified or unrecognized copies are preserved and reported. Start a new Agent session after synchronization; existing conversation context is retained.
+
 ## Environment-Bound Updates
 
 Each AE environment returns the exact `aeCliVersion` it expects. Use the unified update command to install that CLI version and its matching Skills:
@@ -62,11 +65,11 @@ ae-cli update --dry-run
 
 Updates use the same global Skills installation and verification as `install`. Older packages need no installer or generated manifest: their bundled Skills define the expected contents. If Skills fail after the CLI installs, the result reports partial completion and can be retried.
 
-Automatic synchronization runs during normal commands after login; host version checks are cached for up to 24 hours. `ae-cli update --host <AE_HOST>` checks immediately. After a downgrade, future updates run the older CLI's own updater, including its historical limitations. To use the new installer logic while following an older host, run `npx @thinkingai/ae-cli@latest update --host <AE_HOST>`; the target remains the host-required version.
+Automatic synchronization runs during normal commands after login; host version checks are cached for up to one hour. `ae-cli update --host <AE_HOST>` checks immediately. After a downgrade, future updates run the older CLI's own updater, including its historical limitations. To use the new installer logic while following an older host, run `npx @thinkingai/ae-cli@latest update --host <AE_HOST>`; the target remains the host-required version.
 
-`ae-cli update` installs the exact version required by the active host rather than npm `latest`. It synchronizes Skills from the installed npm package first and uses the matching GitHub tag only as a fallback.
+`ae-cli update` installs the exact version required by the active host rather than npm `latest`. Skills are synchronized from the verified target package; modified or unrecognized copies are preserved and reported.
 
-Starting with the `6.0.37` and `6.1.9` release lines, ordinary business commands can automatically upgrade or downgrade the CLI to the host-required version. After a successful synchronization, the old process exits with `AE_CLI_VERSION_SYNCED`; re-run the original command so it executes with the new CLI and Skills. Installation failures are reported without replacing the business command's JSON output.
+Starting with the `6.0.37` and `6.1.9` release lines, ordinary business commands can automatically upgrade or downgrade the CLI to the host-required version, including switching from `6.0.x` to `6.1.x`. Automatic synchronization requires stable `major.minor.patch` versions at both ends; development and other nonstandard versions are skipped. After successful synchronization, the new CLI automatically executes the original command once and returns its result and exit code. Arguments, piped input, working directory, and confirmation behavior are preserved. If the new CLI cannot start, `AE_CLI_VERSION_SYNCED` asks you to re-run manually; an executed command is never retried automatically. Installation failures are reported without replacing the business command's JSON output.
 
 Useful controls:
 
@@ -144,6 +147,7 @@ The table covers every current root command individually. Use `ae-cli --help` fo
 | Agent platform | `context` | Read the current product-page context authorized for the active Agent Run |
 | Agent platform | `memory` | AgentSpace memory, drafts, shared review, extraction jobs, revisions, and state transitions |
 | Agent platform | `team` | Agent Team and TeamRun execution, chat, results, and artifacts |
+| Agent platform | `evaluation` | Agent Evaluation datasets, LLM judge evaluators, runs, case results, exports, schedules, and model/agent configs |
 | Agent platform | `system` | Root/admin management of members, sandboxes, tools, models, usage, quotas, and channels |
 | General tools | `capability` | Capability discovery, schema inspection, validation, dry-run, and generic execution |
 | General tools | `auth` | Host-scoped login, status, and multi-account management |
@@ -157,6 +161,22 @@ The table covers every current root command individually. Use `ae-cli --help` fo
 | Agent platform | `automation` | Automation configuration, execution, history, and results |
 <!-- root-command-surface:end -->
 
+Review collected analysis questions with `ae-cli analysis scenario list --project-id 1 --days 7`. The `ae-analysis` Skill can deduplicate questions and assess badcases with concise, evidence-based reasons. See the [scenario review guide](skills/ae-analysis/references/scenario_list.md). Answers can include elapsed time; use `ae-cli analysis scenario get --help` to query linked command execution details.
+
+Evaluation commands follow the current AE Host in personal terminals and sandboxes; use `--host` to select another host. Evaluation requests use the HTTP or HTTPS scheme configured for that AE Host, consistent with other AE commands. Legacy `--endpoint` and `AE_EVAL_BASE_URL` values may only match the current AE Host; other values fail before credentials are sent. Remove these settings and use `--host` to select a host. Outside a sandbox, export `AE_EVAL_ACCESS_TOKEN` with a platform session token; sandboxes can obtain a session automatically. See [endpoint and authentication rules](skills/ae-evaluation/SKILL.md#endpoint-and-authentication).
+
+### Scheduled Feature experiments
+
+For a configured Feature experiment, save its start time and optional fixed deadline, check readiness, then submit it:
+
+```bash
+ae-cli experiment experiment save --project-id 1 --req '{"expId":"exp_123"}' --start-time '2026-11-01T16:00:00+08:00' --fixed-end-time '2026-11-08T16:00:00+08:00'
+ae-cli experiment experiment ready-check --project-id 1 --exp-id exp_123
+ae-cli experiment experiment manage --project-id 1 --exp-id exp_123 --target-status pending
+```
+
+Saving alone does not start an experiment. The server checks submitted experiments every half hour. Use `experiment experiment get` to verify status; see [Feature timing](skills/ae-experiment/references/save_experiment.md#feature-timing) for traffic checks, timing updates, and manual starts.
+
 ## Capability Gateway
 
 Capability Gateway is the preferred entry point for operations that do not need a dedicated curated command:
@@ -170,6 +190,10 @@ ae-cli capability dry-run analysis.report.list --input input.json
 ae-cli capability run analysis.report.list --input input.json
 ```
 
+Administrators can set each member's daily CLI call limit in External Access Management.
+
+For asynchronous exports, omit `--timeout-seconds` to use the server's default timeout. See the [export guide](skills/ae-analysis/references/analysis_data_export.md) for submitting, waiting for, and downloading results.
+
 `--input` accepts inline JSON, a JSON file path, `@<path>`, or `-` for stdin.
 
 Use `validate` while fixing complex nested input. Use `dry-run` when you need the final risk, output mode, cancellation support, or delete confirmation preview. Do not stack `validate` and `dry-run` for the same final payload by default; `dry-run` already validates it.
@@ -195,7 +219,7 @@ ae-cli auth status --host https://host-a.example.com
 # Import a CLI token copied from External Access Management (hidden prompt)
 ae-cli auth set-token --host https://host-a.example.com
 # Non-interactive input for local scripts or tests
-printf '%s' "$AE_CLI_TOKEN" | ae-cli auth set-token --host https://host-a.example.com --token-stdin
+printf '%s' "$IMPORTED_CLI_TOKEN" | ae-cli auth set-token --host https://host-a.example.com --token-stdin
 # For the uncommon case where one Host needs multiple accounts
 ae-cli auth login --host https://host-a.example.com --add
 ae-cli auth list --host https://host-a.example.com
@@ -217,9 +241,28 @@ A regular `auth login` keeps the simple one-Host/one-account behavior and replac
 
 `auth set-token` is for local use with a `cli_` token copied from External Access Management. It prompts without echo by default; `--token-stdin` supports non-interactive local testing. The CLI validates the token against the target Host before replacing the stored credential, encrypts it in the existing per-Host vault, and never includes the value in command output. Add `--add` to retain other accounts for the same Host.
 
+Inject credentials into the current CLI invocation through environment variables or a specified file without changing stored machine accounts. This supports automation scripts, CI/CD, shared Agents, and other workflows:
+
+```bash
+AE_CLI_TOKEN="$CALLER_CLI_TOKEN" AE_CLI_TOKEN_HOST=https://host-a.example.com ae-cli auth status
+AE_CLI_TOKEN_FILE=/secure/caller-token.json ae-cli auth status
+ae-cli --cli-token-file /secure/caller-token.json auth status
+```
+
+The token file is a read-only JSON object: `{"url":"https://host-a.example.com","token":"cli_REPLACE_WITH_USER_TOKEN"}`. The historical `cli-token` key is also accepted. Keep the file private to its owner (for example, mode `600`). Files must be regular JSON files no larger than 64 KiB. The file URL or `AE_CLI_TOKEN_HOST` selects the host for this invocation; an explicit `--host` must match it. External tokens are sent only to that origin; alternate service origins and HTTP redirects are rejected.
+
+Credential precedence is `--cli-token-file` > `AE_CLI_TOKEN` with required `AE_CLI_TOKEN_HOST` > `AE_CLI_TOKEN_FILE` > the existing stored-account/sandbox sources. An explicitly selected source that is empty, malformed, unreadable, or bound to a different host returns `EXTERNAL_CLI_TOKEN_INVALID`; it never falls back to another identity. External credentials also take precedence over sandbox identity headers for Agent calls. `auth status` reports `environment-cli-token` or `file-cli-token`, account and expiration when available, and never reveals the token. Authentication failures do not replace or delete stored credentials.
+
+External mode does not write the token or renewal state to the local credential vault. Renewal is best effort, uses the `cli-token` header, and is limited per host/token per process day; older servers without header-based renewal can still authenticate business calls but cannot automatically extend external tokens. `auth`, `auth login`, `auth set-token`, `auth use`, and `auth logout` return `EXTERNAL_CLI_TOKEN_ACTIVE` while external credentials are selected. Remove these process settings before managing machine accounts; `auth list` continues to list local accounts.
+
 The new CLI persists only CLI tokens, never access or refresh tokens. Multiple accounts live in an encrypted V1 vault while the active account is projected into the historical file shape for automatic downgrades. When the CLI upgrades again, it reconciles login, switch, or logout changes made by the old CLI.
 
 Login uses a cross-platform device-code flow. Use `--no-browser` when the environment cannot open a browser.
+
+On upgraded 6.1 deployments, authorization always uses the UBA device endpoints and the platform `/#/cli/device-activate` page, independently of Kubernetes or Agent state. Only an explicit missing authorization endpoint (HTTP 404) selects the supported legacy Agent flow; network failures, HTML responses and permission denials never change the issuer. Use `auth login --no-wait`, then resume with the returned `--device-code` and the same `--host` for non-interactive authorization. Existing host accounts, CLI tokens and command flags remain compatible.
+
+The CLI calls business endpoints directly; services enforce product License and user/project permissions. An unavailable component returns the actual API error. HTTP 404 can also mean a missing resource or an unsupported route, so it does not by itself establish that a component is not deployed.
+
 Trial guidance is emitted only when no Host is configured; normal commands and authentication flows for configured environments do not display it.
 
 ## Output and Safety
@@ -246,6 +289,9 @@ JSON flags generally accept inline JSON, `@file`, a file path, or `-` for stdin.
 
 ## Knowledge Bases
 
+Use `ae-cli kb candidate changes --name <name> --candidate <id>` to review changed sources, Wiki and Index before approving a company compilation. See [Publication previews](skills/ae-kb/references/publication.md) for entry details and bounded text diffs.
+For large approval previews, use `kb candidate changes --group wiki --search payment --offset 0 --limit 50`. The first successful compilation publishes automatically; later company KB updates require approval.
+
 Manage the server-side knowledge-base lifecycle:
 
 ```bash

package/README.zh.md

@@ -17,7 +17,6 @@ CLI 的核心设计包括：
 - Node.js 20 或更高版本。
 - 支持 macOS、Linux 和原生 Windows（PowerShell / CMD），无需安装 WSL。
 - 安装 CLI 时能够访问 npm 仓库。
-- 内置 Skills 来源不可用时，能够访问 GitHub。
 
 ## 安装
 
@@ -34,7 +33,7 @@ npx @thinkingai/ae-cli@latest install
 npx @thinkingai/ae-cli@latest install --format json
 ```
 
-安装器检查 Node.js 20+、安装并验证 CLI，再将同版本 Skills 全局安装到 `~/.agents/skills`，公共文件只写入一次，再自动向需要独立目录的客户端分发副本，无需选择 Agent；之后安装的 Codex、Cursor 也可共享这些 Skills。支持 `--dry-run`、`--format json` 和可选的 `--host <AE_HOST>`。验证成功后不显示无关的项目级客户端告警；文件缺失和 YAML 错误仍判为失败。未登录不算安装失败。详见[安装指南](./cli-installation-guide.md)。
+安装器检查 Node.js 20+、安装并验证 CLI，再通过内置的 Skills 工具向支持的客户端安装同版本全局 Skills。首次安装且尚无已确认业务状态时安装全量内置 Skills。支持 `--dry-run`、`--format json` 和可选的 `--host <AE_HOST>`。未登录不算安装失败。详见[安装指南](./cli-installation-guide.md)。
 
 Skills 会帮助 Claude Code、Codex、Cursor 等编码 Agent 理解、发现并调用 `ae-cli`。
 
@@ -55,6 +54,10 @@ ae-cli auth status
 
 请使用 AgenticEngine 管理员提供的 AE 地址。未配置 Host 时，`ae-cli` 会先引导已有客户向管理员获取地址；只有确认尚无 AgenticEngine 环境的用户，才会看到[申请试用](https://thinkingai.cn/request-demo)入口。
 
+按业务过滤命令和剪裁 Skills 仅适用于外部 Agent 使用的本机 CLI；TE Agent 内置沙箱始终保留完整命令与全部 Skills。登录成功后，CLI 按当前环境已开通的业务过滤 `--help` 和能力发现，并同步受管的全局 Skills；分析能力默认保留。后续开通业务时，执行 `ae-cli config refresh` 即可重新获取状态并补装对应 Skills。切换活动环境或账号会重新同步；单次 `--host` 查询不改变全局 Skills。
+
+仅明确关闭的业务会被移除；旧服务端、缺失状态或网络失败不会触发删除。指向已核验、未修改的受管共享 Skill 的客户端历史链接会自动迁移；其他用户修改和来源不明的副本保留并报告。同步后请新开 Agent 会话，已有会话的上下文不会被撤回。
+
 ## 环境版本同步
 
 每个 AE 环境都会返回其要求的精确 `aeCliVersion`。使用统一更新命令安装对应版本的 CLI 和 Skills：
@@ -66,11 +69,11 @@ ae-cli update --dry-run
 
 更新与 `install` 复用全局 Skills 安装和校验。旧包无需包含安装器或生成的清单，直接以目标包内置 Skills 为准。CLI 安装后 Skills 失败会返回部分成功信息，修复后可重复执行。
 
-自动同步在登录后的普通命令中触发，集群版本检查最长缓存 24 小时；`ae-cli update --host <AE_HOST>` 可立即查询。降级后，后续更新由旧版 CLI 自身执行，也受旧版实现限制。需要用新安装逻辑匹配旧集群时，可运行 `npx @thinkingai/ae-cli@latest update --host <AE_HOST>`，目标仍是集群要求的版本。
+自动同步在登录后的普通命令中触发，集群版本检查最长缓存 1 小时；`ae-cli update --host <AE_HOST>` 可立即查询。降级后，后续更新由旧版 CLI 自身执行，也受旧版实现限制。需要用新安装逻辑匹配旧集群时，可运行 `npx @thinkingai/ae-cli@latest update --host <AE_HOST>`，目标仍是集群要求的版本。
 
-`ae-cli update` 安装当前 Host 要求的精确版本，不依赖 npm `latest`。更新时优先从已安装的 npm 包同步 Skills，本地来源失败后才回退到对应的 GitHub tag。
+`ae-cli update` 安装当前 Host 要求的精确版本，不依赖 npm `latest`。Skills 从已验证的目标包同步；用户修改或来源不明的副本会保留并报告。
 
-从 `6.0.37` 和 `6.1.9` 维护线开始，普通业务命令可以自动升级或降级到环境要求的版本。同步成功后，旧进程会返回 `AE_CLI_VERSION_SYNCED`；重新执行原命令，即可使用新的 CLI 和 Skills。安装失败会给出提示，但不会用安装输出污染业务命令的 JSON 结果。
+从 `6.0.37` 和 `6.1.9` 维护线开始，普通业务命令可以自动升级或降级到环境要求的版本，包括从 `6.0.x` 切换到 `6.1.x`。自动同步要求本地和目标均为 `主版本.次版本.补丁号` 格式的正式版本，开发版等不规则版本会跳过。同步成功后，新版 CLI 会自动执行原命令一次，并返回该命令的结果和退出码；参数、管道输入、工作目录和确认行为保持不变。如果新版 CLI 无法启动，会返回 `AE_CLI_VERSION_SYNCED` 并提示手动重跑；已经执行的命令不会自动重试。安装失败会给出提示，但不会用安装输出污染业务命令的 JSON 结果。
 
 常用控制方式：
 
@@ -148,6 +151,7 @@ ae-cli capability list --domain analysis --jq '.data.capabilities[] | .id'
 | Agent 平台 | `context` | 读取当前 Agent Run 已授权的产品页面上下文 |
 | Agent 平台 | `memory` | AgentSpace 记忆、草稿、共享审核、提取任务、版本与状态治理 |
 | Agent 平台 | `team` | Agent Team 和 TeamRun 的执行、对话、结果及产物 |
+| Agent 平台 | `evaluation` | Agent Evaluation 评测数据集、LLM 裁判、评测任务、用例结果、导出、调度，以及模型/Agent 配置 |
 | Agent 平台 | `system` | root/admin 的成员、沙盒、工具、模型、用量、配额和渠道管理 |
 | 通用工具 | `capability` | 能力发现、Schema 查看、参数校验、dry-run 和通用执行 |
 | 通用工具 | `auth` | 按 Host 登录、查看状态及管理同一 Host 的多个账号 |
@@ -161,6 +165,22 @@ ae-cli capability list --domain analysis --jq '.data.capabilities[] | .id'
 | Agent 平台 | `automation` | 自动化配置、执行、历史和结果 |
 <!-- root-command-surface:end -->
 
+使用 `ae-cli analysis scenario list --project-id 1 --days 7` 查询已采集的分析问题，由 `ae-analysis` Skill 指导 Agent 去重、判断 badcase 并简要说明原因和证据。详见[分析问题复盘指南](skills/ae-analysis/references/scenario_list.md)。 回答可附带耗时；使用 `ae-cli analysis scenario get --help` 查看关联命令执行明细的查询方式。
+
+评测命令在普通终端和沙箱中均跟随当前 AE Host，可通过 `--host` 切换。评测请求沿用该 AE Host 配置的 HTTP 或 HTTPS 协议，与其他 AE 命令一致。旧 `--endpoint` 和 `AE_EVAL_BASE_URL` 只能与当前 AE Host 一致，其他值会在发送凭证前报错；请移除这些配置，统一使用 `--host` 切换。沙箱外需设置 `AE_EVAL_ACCESS_TOKEN` 平台 session token，沙箱内可自动获取会话。详见 [地址与鉴权规则](skills/ae-evaluation/SKILL.md#endpoint-and-authentication)。
+
+### Feature 实验定时开启
+
+对于已完成配置的 Feature 实验，保存开始时间和可选的固定截止时间，检查就绪状态后提交：
+
+```bash
+ae-cli experiment experiment save --project-id 1 --req '{"expId":"exp_123"}' --start-time '2026-11-01T16:00:00+08:00' --fixed-end-time '2026-11-08T16:00:00+08:00'
+ae-cli experiment experiment ready-check --project-id 1 --exp-id exp_123
+ae-cli experiment experiment manage --project-id 1 --exp-id exp_123 --target-status pending
+```
+
+仅保存不会开启实验，服务端每半小时检查已提交的实验。使用 `experiment experiment get` 确认状态；流量冲突检查、时间修改与手动开启规则见 [Feature 时间配置](skills/ae-experiment/references/save_experiment.md#feature-timing)。
+
 ## Capability Gateway
 
 不需要独立精选命令的能力，应优先通过 Capability Gateway 调用：
@@ -174,6 +194,10 @@ ae-cli capability dry-run analysis.report.list --input input.json
 ae-cli capability run analysis.report.list --input input.json
 ```
 
+管理员可在“外部访问管理”为成员配置每日 CLI 调用限额。
+
+异步导出时，省略 `--timeout-seconds` 即使用服务端默认超时。提交、等待和下载结果的方法见[导出使用指南](skills/ae-analysis/references/analysis_data_export.md)。
+
 `--input` 支持内联 JSON、JSON 文件路径、`@<path>`，或通过 `-` 从 stdin 读取。
 
 复杂嵌套参数尚在调整时使用 `validate`；需要确认最终风险等级、输出模式、取消能力或删除门禁时使用 `dry-run`。默认不要对同一份最终入参连续执行 `validate` 和 `dry-run`，因为 `dry-run` 已包含参数校验。
@@ -199,7 +223,7 @@ ae-cli auth status --host https://host-a.example.com
 # 导入从“外部访问管理”复制的 CLI Token（隐藏输入）
 ae-cli auth set-token --host https://host-a.example.com
 # 本地脚本或测试使用非交互输入
-printf '%s' "$AE_CLI_TOKEN" | ae-cli auth set-token --host https://host-a.example.com --token-stdin
+printf '%s' "$IMPORTED_CLI_TOKEN" | ae-cli auth set-token --host https://host-a.example.com --token-stdin
 # 少量需要同一 Host 多账号的场景
 ae-cli auth login --host https://host-a.example.com --add
 ae-cli auth list --host https://host-a.example.com
@@ -221,9 +245,28 @@ ae-cli config remove pre-production --yes
 
 `auth set-token` 用于在本地导入从“外部访问管理”复制的 `cli_` Token。默认使用不回显的隐藏输入，`--token-stdin` 支持本地测试等非交互场景。CLI 会先向目标 Host 校验 Token，再替换本地凭证，并写入现有的按 Host 加密凭据库；命令输出不会包含 Token。需要保留同一 Host 的其他账号时增加 `--add`。
 
+支持通过环境变量或指定文件，为当前 CLI 调用注入凭证，不改变机器上已保存的账号；适用于自动化脚本、CI/CD、公共Agent等场景。
+
+```bash
+AE_CLI_TOKEN="$CALLER_CLI_TOKEN" AE_CLI_TOKEN_HOST=https://host-a.example.com ae-cli auth status
+AE_CLI_TOKEN_FILE=/secure/caller-token.json ae-cli auth status
+ae-cli --cli-token-file /secure/caller-token.json auth status
+```
+
+Token 文件是只读 JSON 对象：`{"url":"https://host-a.example.com","token":"cli_REPLACE_WITH_USER_TOKEN"}`，也兼容历史字段 `cli-token`。文件应仅允许所有者读取，例如权限 `600`；必须是大小不超过 64 KiB 的普通 JSON 文件。文件中的 URL 或 `AE_CLI_TOKEN_HOST` 决定本次调用的 Host；显式 `--host` 必须与之匹配。外部 Token 仅发送到该 Origin，拒绝其他服务 Origin 和 HTTP 重定向。
+
+凭证优先级为 `--cli-token-file` > `AE_CLI_TOKEN`（必须同时提供 `AE_CLI_TOKEN_HOST`）> `AE_CLI_TOKEN_FILE` > 现有本地账号或沙箱来源。显式来源为空、格式错误、无法读取或 Host 不匹配时返回 `EXTERNAL_CLI_TOKEN_INVALID`，不会回退到其他身份。外部凭证也优先于 Agent 调用的沙箱身份头。`auth status` 的来源为 `environment-cli-token` 或 `file-cli-token`，并在后端支持时展示账号与到期时间，不输出 Token。认证失败不会替换或删除本地凭证。
+
+外部模式不将 Token 或续期状态写入本地凭据库。自动续期使用 `cli-token` 请求头，失败不阻断业务请求，每个进程按 Host、Token 和当天日期限频；旧后端不支持请求头续期时仍可认证业务请求，但不能自动延长外部 Token。选择外部凭证时，`auth`、`auth login`、`auth set-token`、`auth use` 和 `auth logout` 返回 `EXTERNAL_CLI_TOKEN_ACTIVE`。管理机器账号前先移除这些进程设置；`auth list` 仍只列出本地账号。
+
 新版 CLI 只持久化 CLI Token，不保存 access token 或 refresh token。多账号保存在加密的 V1 凭据文件中，同时维护旧文件格式的当前账号投影，保证 CLI 自动降级后仍可登录；再次升级时会合并旧 CLI 对投影的登录、切换或退出变更。
 
 登录使用跨平台设备码流程。当前环境无法打开浏览器时，可使用 `--no-browser`。
+
+已改造的 6.1 环境统一使用 UBA 设备授权接口与平台 `/#/cli/device-activate` 页面，不随 K8s 或 Agent 状态切换。只有授权端点明确不存在（HTTP 404）时才兼容旧 Agent 流程；网络失败、HTML 响应和权限拒绝均不会更换授权方。非交互环境使用 `auth login --no-wait`，再携带返回的 `--device-code` 和同一个 `--host` 继续授权。原有 Host 多账号、CLI Token 与命令参数保持兼容。
+
+CLI 直接请求对应业务接口，由服务端校验产品 License 和用户/项目权限。组件不可用时会返回实际接口错误；HTTP 404 也可能表示资源不存在或路由不支持，不能据此直接判断组件未部署。
+
 试用引导仅在尚未配置 Host 时出现；已配置环境的正常命令和认证流程不会展示该提示。
 
 ## 输出与安全
@@ -250,6 +293,9 @@ JSON 参数通常支持内联 JSON、`@file`、文件路径或通过 `-` 从 std
 
 ## 知识库
 
+公司知识库编译审批前，可用 `ae-cli kb candidate changes --name <name> --candidate <id>` 查看数据源、Wiki 和 Index 的变化。条目详情和文本差异用法见[审批预览](skills/ae-kb/references/publication.md)。
+审批变化较多时，可用 `kb candidate changes --group wiki --search payment --offset 0 --limit 50` 分页检索；首次成功编译自动发布，后续公司知识库更新进入审批。
+
 管理服务端知识库生命周期：
 
 ```bash

package/bin/ae-cli.js

@@ -1,7 +1,12 @@
 #!/usr/bin/env node
+import { readFileSync } from 'node:fs';
 if (Number(process.versions.node.split('.')[0]) < 20) {
   console.error('AE CLI requires Node.js 20 or later. Select a supported Node.js runtime and rerun.');
   process.exitCode = 1;
+} else if (process.argv.length === 3 && process.argv[2] === '--runtime-capabilities') {
+  // Keep discovery independent of application imports and credential initialization.
+  const capabilities = JSON.parse(readFileSync(new URL('./runtime-capabilities.json', import.meta.url), 'utf8'));
+  process.stdout.write(JSON.stringify(capabilities) + '\n');
 } else {
   await import('../dist/index.js');
 }

package/bin/runtime-capabilities.json

@@ -0,0 +1,5 @@
+{
+  "protocol": "ae-cli-runtime-capabilities",
+  "version": 1,
+  "features": ["kb_read_context_v1"]
+}

package/cli-installation-guide.md

@@ -20,13 +20,13 @@ Node.js 20 or later, npm, and npx must be available. The installer does not chan
 
 The running package determines the exact CLI version to install. The global executable is verified with `ae-cli --version`; PATH problems include the verified absolute path. Local npm links are protected by a small guard. Other installation conflicts use normal npm errors; the installer does not migrate existing installations.
 
-Skills are installed globally in `~/.agents/skills`. The shared files are installed once, then the Skills tool automatically distributes copies to detected clients that require their own directory. Shared-directory targets are skipped during distribution. Existing per-Skill links are replaced with copies when needed. Codex and Cursor support this shared directory, so installing Cursor later does not require a separate copy of the Skills. A client that only reads its own directory may need the same install command run again after that client is installed. User-global installation does not promise availability in remote/cloud clients or for every future client.
+The pinned Skills tool resolves supported global client directories, including the shared `~/.agents/skills` directory. AE does not maintain a separate client path table. A client that only reads its own directory may need the same install command run again after that client is installed. Project-only clients, unknown clients, remote/cloud clients, and manually imported Skills are outside automatic global synchronization. Overlapping source directories and unverified links are protected rather than replaced.
 
 The external tool is pinned to `skills@1.5.12` to preserve Node 20 support and skip copying a Skill onto its own source directory. The installer does not maintain its own Agent catalog, ask users to select clients, or select clients based on the invoking AI. Unrelated user Skills are retained.
 
 ## Verification and output
 
-The installed CLI package's `skills/` directory defines the expected Skills and contents at runtime, including for historical packages without an installer or generated manifest. A failed shared-directory installation can retry the exact release tag `ThinkingAIAgenticEngine/ae-cli#v<CLI_VERSION>`; it never falls back to the default branch. Missing or invalid bundled Skills are reported rather than guessing the expected set from another version. Client distribution failures do not trigger a source download retry. If the local source and fallback both fail, the top-level error retains the first cause, and `skills.attempts` records both attempts with their stage, source, cause, and failures.
+The verified target CLI package's complete `skills/` directory defines the expected contents, including for historical packages without an installer or generated manifest. The synchronizer uses this source for installation and restoration. Missing or invalid package sources, modified copies, ownership conflicts, and tool failures are reported; they do not trigger an unrestricted remote installation retry. `skills.attempts` retains the attempted source and failure details.
 
 Verification parses YAML/frontmatter and compares expected files in the shared global directory with the target package. Merely finding a directory, an `ae-*` name, or an exit code of zero is insufficient. Failures reported by the pinned tool for supported clients are also treated as failures, even when shared files exist.
 
@@ -37,6 +37,12 @@ Verification parses YAML/frontmatter and compares expected files in the shared g
 
 `--format json` emits one report on stdout. Progress goes to stderr; raw subprocess logs are not forwarded. Reports include `ok`, `dry_run`, `cli.package/version/path/verified`, `skills.directory/expected_count/installed_count/missing/source/version/failures/attempts`, `auth.host/authenticated`, and `next_steps`. Counts refer to unique Skills in the global directory. Use `--format table` for a human summary. Dry runs make no subprocess calls or installation/configuration writes.
 
+`install --dry-run --format json` returns that report directly, with a root-level `commands` array. `update --dry-run --format json` keeps the command envelope (`ok` and `data`); its `data` includes `action`, `dryRun`, `target`, optional `host`, `commands`, and `skillsSources`.
+
+Each `commands` array describes the current installation plan; `skillsSources` lists the current source choices. Array lengths and contents can change with the installer. Automation should run the complete `install` or `update` command, check its exit code and `ok`, and inspect the installation report. Do not depend on fixed array indexes or execute individual entries as a replacement for the installer.
+
+The managed synchronizer uses the verified package source. Replaying historical GitHub fallback or `skills add --copy` commands would bypass business selection and copy protection.
+
 ## Host and login
 
 `--host` is optional. It delegates status checks to the existing auth command; it does not independently save credentials or change the active host. Being logged out does not make a verified installation fail.
@@ -54,6 +60,21 @@ ae-cli auth status --host <AE_HOST>
 
 Never print, copy, or manually save any token.
 
+## Business selection after login
+
+First installation without confirmed business state installs all bundled Skills. After successful login, `config refresh`, an account switch, or an active-environment switch, config determines the optional business selection. Analysis and foundation Skills remain available. Only pristine managed copies of explicitly disabled businesses are removed; enabled businesses are restored from the package. Unknown states preserve existing copies. Installation, update, and automatic version synchronization reuse this selection, so an upgrade does not restore disabled Skills.
+
+```bash
+ae-cli config refresh
+ae-cli config show
+```
+
+The global selection follows one active local environment. An explicit `--host` refresh reads that environment's config without changing global Skills, even when its URL matches the active host. External-token and sandbox invocations do not manage local global Skills. Failed config or Skill synchronization does not revoke successful authentication. Partial synchronization reports protected or failed paths; resolve the reported conflict and retry. Existing copies are adopted once only when they exactly match a known package source. No regular backup is needed because the CLI package retains the complete resources.
+
+Older Skills installations may have created client links to the shared global directory. The synchronizer migrates an existing same-name link only when its destination is a managed shared Skill whose contents still match both its ownership record and the installed package. A confirmed disabled business removes the verified links before its shared copy; an enabled business replaces them with managed copies. Unknown business states leave links unchanged. Custom targets, modified shared contents, and links to package sources remain protected.
+
+Start a new Agent session after synchronization. Descriptions already loaded into a conversation cannot be withdrawn.
+
 ## Host-compatible upgrades and downgrades
 
 ```bash
@@ -73,7 +94,7 @@ To use the newer installer while still respecting the current host version, incl
 npx @thinkingai/ae-cli@latest update --host <AE_HOST>
 ```
 
-The newer command orchestrates the work; the installed target remains the host-required version. Required package versions and fallback tags must be published and accessible. Global CLI installation follows one target at a time; switching between differently versioned hosts can switch that installation again.
+The newer command orchestrates the work; the installed target remains the host-required version. Required package versions must be published and accessible. Global CLI installation follows one target at a time; switching between differently versioned hosts can switch that installation again.
 
 ## Advanced manual troubleshooting