cx-ai-center

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

AI Center Skill

This is the tool for anything about AI/GenAI applications — both questions about their behavior (prompts/responses, quality, hallucinations, guardrails, security, cost/tokens, errors, latency — everything AI apps expose through their GenAI spans/tags) and actions to manage them (applications, evaluations/policies, policy↔app links, model pricing). If a request touches an AI application or its GenAI telemetry, use this skill.

Coralogix AI Center observes, evaluates, and guards GenAI/LLM applications. This skill answers questions about AI apps from two sources:

Configuration (this skill's
```
cx ai-center
```
commands): the AI application inventory, configured evaluations/policies, coverage, custom evaluations, and model pricing — none of which live in span telemetry.
Telemetry (GenAI spans): what users asked, how the model answered, cost, tokens, latency, errors, tool calls, and eval/guardrail verdicts — queried with cx spans '<DataPrime>'
. See references/ai-center-queries.md for the full, runnable query library, span schema, and playbooks.

Match the source to the question: "which apps lack guardrails" → config (

cx ai-center applications list

); "what are users asking my chatbot" → telemetry (

cx spans '…'

, reading the conversation from the GenAI spans). Some questions need both — e.g. "is my chatbot's PII policy actually catching PII?" joins config (is the policy enabled) with telemetry (the PII verdicts + the messages).

这是处理所有AI/GenAI应用相关事务的工具 —— 既可以回答关于AI应用行为的问题（提示词/响应、质量、幻觉、防护机制、安全性、成本/令牌、错误、延迟等所有AI应用通过GenAI spans/tags暴露的信息），也可以执行管理操作（应用程序、评估/策略、策略与应用关联、模型定价）。如果请求涉及AI应用或其GenAI遥测数据，请使用此Skill。

Coralogix AI Center用于观测、评估和防护GenAI/LLM应用。此Skill从两个来源获取AI应用相关问题的答案：

配置信息（通过此Skill的
```
cx ai-center
```
命令）：AI应用清单、已配置的评估/策略、覆盖范围、自定义评估以及模型定价 —— 这些信息均不存储在遥测链路数据中。
遥测数据（GenAI spans）：用户的提问内容、模型的回答内容、成本、令牌、延迟、错误、工具调用以及评估/防护判定结果 —— 通过**
```
cx spans '<DataPrime>'
```
**查询。完整的可运行查询库、链路数据 schema 和操作手册，请查看references/ai-center-queries.md。

根据问题类型匹配对应数据源：例如“哪些应用未启用防护机制”→ 配置信息（

cx ai-center applications list

）；“用户向我的聊天机器人提出了哪些问题”→ 遥测数据（

cx spans '…'

，从GenAI spans中读取对话内容）。部分问题需要同时使用两种数据源 —— 例如“我的聊天机器人的PII策略是否真的能识别出PII？”需要结合配置信息（策略是否启用）和遥测数据（PII判定结果+对话消息）。

Destructive Operation Safety

破坏性操作安全规范

All write operations (

create

update

delete

add

remove

set

) require interactive confirmation.

ai-center

is a risky command, so writes are also gated by

allow_risky_commands

~/.cx/config.toml

. To skip the prompt in scripts, pass

--yes

IMPORTANT: NEVER pass
--yes
without explicit user approval. Before executing any write:

Describe the exact operation to the user (what will be created/modified/deleted/linked).
Wait for the user to confirm.
Only then execute with
```
--yes
```
.

Read operations (

list

get

coverage

list-for-application

model-pricing get

) do not require confirmation and can be run freely.

所有写入操作（

create

、

update

、

delete

、

add

、

remove

、

set

）均需要交互式确认。

ai-center

属于高风险命令，因此写入操作还受

~/.cx/config.toml

中的

allow_risky_commands

配置限制。若要在脚本中跳过确认提示，请添加

--yes

参数。

**重要提示：未经用户明确批准，切勿使用

--yes

参数。**执行任何写入操作前：

向用户描述具体操作内容（将创建/修改/删除/关联的对象）。
等待用户确认。
获得确认后，再添加
```
--yes
```
参数执行操作。

读取操作（

list

、

get

、

coverage

、

list-for-application

、

model-pricing get

）无需确认，可直接执行。

Read-Only Mode

只读模式

Use

--read-only

(or

CX_READ_ONLY=1

) to block every write at the CLI level — safe for exploration.

使用

--read-only

（或

CX_READ_ONLY=1

）参数可在CLI层面阻止所有写入操作 —— 适合用于探索性操作。

Agent Mode

Agent模式

When running inside an AI agent (Claude Code, Cursor, Codex, …), cx detects it and — instead of showing a confirmation prompt that would hang forever (no human is there to type y/n) — stops immediately with an error telling you to get the user's approval, then re-run with

--yes

当在AI Agent（Claude Code、Cursor、Codex等）内部运行时，cx会自动检测环境，不会显示会导致程序挂起的确认提示（没有人工输入y/n），而是立即终止并抛出错误，提示你获取用户批准后，添加

--yes

参数重新执行。

No delete commands (by design)

无删除命令（设计如此）

The CLI intentionally exposes no delete for custom-evaluation policies, AI applications, or model pricing — even though the AI v3 API has those delete endpoints,

cx ai-center

does not surface them.

Custom-evaluation policy: can't be deleted; to take it off an app, detach with
```
custom-evaluations remove
```
(the policy object survives and can be re-attached).
Model pricing: no delete command. It's team-wide (not per-app), so to change or clear it, run
```
model-pricing set
```
with a new map (an empty map
```
{}
```
clears all overrides) —
```
set
```
replaces the whole set.

CLI刻意不提供自定义评估策略、AI应用或模型定价的删除命令 —— 尽管AI v3 API包含这些删除端点，但

cx ai-center

并未开放。

自定义评估策略：无法删除；若要从应用中移除，使用
```
custom-evaluations remove
```
命令解除关联（策略对象会保留，可重新关联）。
模型定价：无删除命令。它是团队级配置（而非应用级），若要修改或清空配置，使用
```
model-pricing set
```
命令传入新的映射表（空映射表
```
{}
```
会清除所有覆盖配置）——
```
set
```
命令会替换整个配置集合。

Golden rule

黄金准则

For content questions (quality, hallucination, sentiment, topics) read the actual conversation and cite the

traceID

— don't rely on verdict tags alone. The transcript lives in one of two conventions (

gen_ai.input.messages

output.messages

, or the older indexed

gen_ai.prompt.<n>

completion.<n>

tags); read it with the Reading conversations (content questions) queries in the library, which handle both and exclude the system prompt and tool traffic. Full guidance: references/ai-center-queries.md.

对于内容类问题（质量、幻觉、情感、主题），请读取实际对话内容并引用

traceID

—— 不要仅依赖判定标签。对话内容存储在两种约定格式中（

gen_ai.input.messages

output.messages

，或旧版索引格式

gen_ai.prompt.<n>

completion.<n>

标签）；可使用查询库中的「读取对话内容（内容类问题）」查询语句，该语句会兼容两种格式并排除系统提示词和工具流量。完整指南请查看：references/ai-center-queries.md。

CLI Commands

CLI命令

Show names to the user; use UUIDs only internally. When presenting results, refer to apps and evaluations by their human names (application/subsystem, evaluation name), not raw UUIDs. The UUID is only needed to call a by-id or write command — resolve it yourself from the matching

list

command (never guess or make the user paste a UUID).

**向用户展示名称；仅在内部使用UUID。**展示结果时，请使用应用和评估的人类可读名称（应用/子系统、评估名称），而非原始UUID。仅在调用按ID查询或写入命令时才需要UUID —— 请通过对应的

list

命令自行获取（切勿猜测或让用户粘贴UUID）。

Applications (inventory + guarded status)

应用程序（清单与防护状态）

Command	Purpose
`cx ai-center applications list`	List AI apps incl. `guardrailsIntegrated` (guarded) status
`cx ai-center applications list --evaluation-type <TYPE>`	Filter to apps using an eval type (repeatable)
`cx ai-center applications list --page-size <N> --page-offset <N>`	Paginate
`cx ai-center applications get <application-id>`	One application by UUID

命令	用途
`cx ai-center applications list`	列出所有AI应用，包含 `guardrailsIntegrated` （已启用防护）状态
`cx ai-center applications list --evaluation-type <TYPE>`	筛选使用指定评估类型的应用（可重复使用该参数）
`cx ai-center applications list --page-size <N> --page-offset <N>`	分页查询
`cx ai-center applications get <application-id>`	通过UUID查询单个应用

Evaluations (configured policies on apps)

评估（应用上配置的策略）

Command	Purpose
`cx ai-center evaluations list`	All configured evaluations
`cx ai-center evaluations list --application <app> --subsystem <sub>`	Scope to one app (the pair)
`cx ai-center evaluations list --evaluation-type <TYPE>`	Filter by type — `<TYPE>` is the API enum (e.g. `PII` , `TOXICITY` , `PROMPT_INJECTION` ; the keys from `coverage` ), not the lowercase form
`cx ai-center evaluations get <evaluation-id>`	One evaluation by UUID
`cx ai-center evaluations create --from-file eval.json`	Create/enable an evaluation (write)
`cx ai-center evaluations update <evaluation-id> --from-file patch.json`	Partial update (write)
`cx ai-center evaluations delete <evaluation-id>`	Remove an evaluation from its app (write)

命令	用途
`cx ai-center evaluations list`	查询所有已配置的评估
`cx ai-center evaluations list --application <app> --subsystem <sub>`	限定查询范围为单个应用（应用+子系统组合）
`cx ai-center evaluations list --evaluation-type <TYPE>`	按类型筛选 —— `<TYPE>` 为API枚举值（例如 `PII` 、 `TOXICITY` 、 `PROMPT_INJECTION` ；即 `coverage` 命令返回的键值），不支持小写形式
`cx ai-center evaluations get <evaluation-id>`	通过UUID查询单个评估
`cx ai-center evaluations create --from-file eval.json`	创建/启用评估（写入操作）
`cx ai-center evaluations update <evaluation-id> --from-file patch.json`	部分更新（写入操作）
`cx ai-center evaluations delete <evaluation-id>`	从应用中移除评估（写入操作）

Custom evaluations (policies) & application links

自定义评估（策略）与应用关联

Command	Purpose
`cx ai-center custom-evaluations list`	All custom evaluation policies
`cx ai-center custom-evaluations list-for-application <application-id>`	Policies linked to one app
`cx ai-center custom-evaluations create --from-file policy.json`	Create a custom policy (write)
`cx ai-center custom-evaluations update <id> --from-file patch.json`	Partial update (write)
`cx ai-center custom-evaluations add <evaluation-id> <application-id>`	Attach a policy to an app (write)
`cx ai-center custom-evaluations remove <evaluation-id> <application-id>`	Detach (reversible) (write)

By-id is prebuilt-only.
evaluations get <id>
fetches a prebuilt/configured evaluation. Custom policies have no get-by-id — find one via
custom-evaluations list
/
list-for-application
and match by
id
/name.

命令	用途
`cx ai-center custom-evaluations list`	查询所有自定义评估策略
`cx ai-center custom-evaluations list-for-application <application-id>`	查询与指定应用关联的策略
`cx ai-center custom-evaluations create --from-file policy.json`	创建自定义策略（写入操作）
`cx ai-center custom-evaluations update <id> --from-file patch.json`	部分更新（写入操作）
`cx ai-center custom-evaluations add <evaluation-id> <application-id>`	将策略关联到应用（写入操作）
`cx ai-center custom-evaluations remove <evaluation-id> <application-id>`	解除关联（可恢复）（写入操作）

仅预构建评估支持按ID查询。
evaluations get <id>
用于查询预构建/已配置的评估。自定义策略不支持按ID查询 —— 请通过
custom-evaluations list
/
list-for-application
命令查询，并通过
id
/名称匹配。

Coverage & model pricing

覆盖范围与模型定价

Command	Purpose
`cx ai-center coverage`	Map of each evaluation type → number of apps using it (coverage / gap analysis)
`cx ai-center model-pricing get`	Team's custom per-model pricing overrides
`cx ai-center model-pricing set --from-file prices.json`	Set team pricing (team-wide, new data only) (write)

The

--from-file

bodies for

evaluations

and

custom-evaluations

match the AI v3 API shape verbatim; use

to read JSON from stdin. For

evaluations create

target

is required and must be uppercase (

PROMPT

RESPONSE

); for

custom-evaluations create

name

instructions

, and

policyType

are required. Exception:

model-pricing set

takes just the raw

model→price

map — cx wraps it as

{"prices": …}

for you, so do not include the outer

prices

envelope. Each model maps to a price object; all four fields are optional doubles (USD per one million tokens), omit the ones that don't apply:

json

{
  "gpt-4o": {
    "inputPricePerMillionTokens": 2.5,
    "outputPricePerMillionTokens": 10,
    "cacheReadPricePerMillionTokens": 1.25,
    "cacheWritePricePerMillionTokens": 3.75
  }
}

An empty map

{}

clears all overrides (set replaces the whole set — it's team-wide, new data only).

model-pricing get

returns the wrapper

{ "pricing": { "id", "companyId", "prices": { … } } }

— the per-model overrides live under

prices

(empty when none are set).

命令	用途
`cx ai-center coverage`	生成各评估类型→使用该类型的应用数量的映射表（覆盖范围/缺口分析）
`cx ai-center model-pricing get`	查询团队的自定义模型定价覆盖配置
`cx ai-center model-pricing set --from-file prices.json`	设置团队定价（团队级配置，仅使用新数据）（写入操作）

evaluations

和

custom-evaluations

命令的

--from-file

参数对应的JSON结构与AI v3 API完全一致；可使用

从标准输入读取JSON。对于

evaluations create

命令，

target

为必填项且必须大写（

PROMPT

或

RESPONSE

）；对于

custom-evaluations create

命令，

name

、

instructions

和

policyType

为必填项。例外情况：

model-pricing set

仅接受原始的

模型→价格

映射表 —— cx会自动将其包装为

{"prices": …}

，因此请勿包含外层的

prices

包裹结构。每个模型对应一个价格对象，四个字段均为可选的浮点数（单位为每一百万令牌的美元价格），可忽略不适用的字段：

json

{
  "gpt-4o": {
    "inputPricePerMillionTokens": 2.5,
    "outputPricePerMillionTokens": 10,
    "cacheReadPricePerMillionTokens": 1.25,
    "cacheWritePricePerMillionTokens": 3.75
  }
}

空映射表

{}

会清除所有覆盖配置（

set

命令会替换整个配置集合 —— 这是团队级配置，仅使用新数据）。

model-pricing get

命令返回的结构为

{ "pricing": { "id", "companyId", "prices": { … } } }

—— 模型级覆盖配置位于

prices

字段下（无配置时为空）。

Common workflows

常见工作流

Inventory & guardrail gaps

清单与防护缺口

bash

undefined

bash

undefined

Which apps are NOT guarded?

哪些应用未启用防护？

cx ai-center applications list -o json | jq '[.[] | select(.guardrailsIntegrated==false)]'

undefined

cx ai-center applications list -o json | jq '[.[] | select(.guardrailsIntegrated==false)]'

undefined

Enable a policy on an app (write — confirm first!)

为应用启用策略（写入操作 —— 请先确认！）

bash

undefined

bash

undefined

1. Describe to the user; 2. get approval; 3. then:

1. 向用户描述操作内容；2. 获取批准；3. 执行以下命令：

cx ai-center evaluations create --from-file eval.json --yes

eval.json: { "application": "...", "subsystem": "...", "target": "PROMPT", "config": { "<type>": {...} }, "isEnabled": true }

eval.json示例: { "application": "...", "subsystem": "...", "target": "PROMPT", "config": { "<type>": {...} }, "isEnabled": true }

target

is REQUIRED and must be UPPERCASE — "PROMPT" or "RESPONSE" (the API rejects lowercase / a missing target).

target

为必填项且必须大写 —— "PROMPT"或"RESPONSE"（API会拒绝小写或缺失该字段的请求）。

undefined

undefined

Read the actual conversations (telemetry, not config)

读取实际对话内容（遥测数据，非配置信息）

Use

cx spans

with the query library in references/ai-center-queries.md — reading messages, cost, latency, errors, tool calls, and per-user analysis.

使用

cx spans

命令结合references/ai-center-queries.md中的查询库 —— 可读取消息内容、成本、延迟、错误、工具调用以及按用户维度的分析数据。

Key principles

核心原则

Config vs. telemetry: inventory / evaluations / policies / coverage / pricing →
```
cx ai-center
```
; content / cost / latency / errors / verdicts → GenAI spans via
```
cx spans
```
. Don't answer one from the other.
Confirm before writes. Describe the operation, get approval, then run with
```
--yes
```
.

配置信息 vs 遥测数据：清单/评估/策略/覆盖范围/定价 → 使用
```
cx ai-center
```
；内容/成本/延迟/错误/判定结果 → 通过
```
cx spans
```
查询GenAI链路数据。请勿混淆数据源。
写入操作前需确认。描述操作内容，获得用户批准后，再添加
```
--yes
```
参数执行。

cx-ai-center

Original

Translation

AI Center Skill

AI Center Skill

Destructive Operation Safety

破坏性操作安全规范

Read-Only Mode

只读模式

Agent Mode

Agent模式

No delete commands (by design)

无删除命令（设计如此）

Golden rule

黄金准则

CLI Commands

CLI命令

Applications (inventory + guarded status)

应用程序（清单与防护状态）

Evaluations (configured policies on apps)

评估（应用上配置的策略）

Custom evaluations (policies) & application links

自定义评估（策略）与应用关联

Coverage & model pricing

覆盖范围与模型定价

Common workflows

常见工作流

Inventory & guardrail gaps

清单与防护缺口

Which apps are NOT guarded?

哪些应用未启用防护？

Enable a policy on an app (write — confirm first!)

为应用启用策略（写入操作 —— 请先确认！）

1. Describe to the user; 2. get approval; 3. then:

1. 向用户描述操作内容；2. 获取批准；3. 执行以下命令：

eval.json: { "application": "...", "subsystem": "...", "target": "PROMPT", "config": { "<type>": {...} }, "isEnabled": true }

eval.json示例: { "application": "...", "subsystem": "...", "target": "PROMPT", "config": { "<type>": {...} }, "isEnabled": true }

target is REQUIRED and must be UPPERCASE — "PROMPT" or "RESPONSE" (the API rejects lowercase / a missing target).

target为必填项且必须大写 —— "PROMPT"或"RESPONSE"（API会拒绝小写或缺失该字段的请求）。

Read the actual conversations (telemetry, not config)

读取实际对话内容（遥测数据，非配置信息）

Key principles

核心原则

Related Skills

相关Skill

`target`
is REQUIRED and must be UPPERCASE — "PROMPT" or "RESPONSE" (the API rejects lowercase / a missing target).

`target`
为必填项且必须大写 —— "PROMPT"或"RESPONSE"（API会拒绝小写或缺失该字段的请求）。