claude-code

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

@rules/routing.md

Claude Code Skill

Source of truth: https://code.claude.com/docs/en/cli-reference.

真理来源：https://code.claude.com/docs/en/cli-reference

Defaults

默认设置

Parameter	Default
Model selection	Use Claude Code CLI default unless the user explicitly asks for `--model`
Effort	Use CLI default unless the user explicitly asks for `--effort`
Permission mode	`--permission-mode default`
Headless mode	`-p` / `--print`
CI / scripted run	Add `--bare` so the call ignores local hooks, plugins, MCP, and CLAUDE.md
Resume target	`claude --continue` ( `-c` ) for the latest session in the current directory; `claude --resume` ( `-r` ) for a session by ID or display name

Do NOT ask the user for model or effort unless explicitly requested.

参数	默认值
模型选择	除非用户明确要求使用 `--model` ，否则使用Claude Code CLI的默认设置
工作力度	除非用户明确要求使用 `--effort` ，否则使用CLI的默认设置
权限模式	`--permission-mode default`
无头模式	`-p` / `--print`
CI/脚本化运行	添加 `--bare` ，使调用忽略本地钩子、插件、MCP和CLAUDE.md
恢复目标	当前目录中最新会话使用 `claude --continue` （ `-c` ）；通过ID或显示名称恢复会话使用 `claude --resume` （ `-r` ）

除非用户明确要求，否则不要询问用户关于模型或工作力度的选择。

Routing

路由规则

Use this skill when the request actually needs the

claude

CLI or a separate Claude Code session.

Read rules/routing.md before building a command when the request might be out of scope.
Route away to direct editing or another skill when the user wants generic writing, documentation cleanup, or local edits without needing the
```
claude
```
CLI itself.

当请求确实需要使用

claude

CLI或独立的Claude Code会话时，使用此技能。

当请求可能超出范围时，在构建命令前阅读rules/routing.md。
当用户需要通用写作、文档清理或无需
```
claude
```
CLI的本地编辑时，将请求转至直接编辑或其他技能。

Examples

示例

Positive triggers:

"Use Claude Code to review this repository and summarize the risks."
"Run
```
claude
```
in print mode and analyze this architecture."
"Continue the last Claude Code session and ask it to finish the patch."
"Resume the
```
auth-refactor
```
claude session and apply the next fix."
"Use
```
claude --bare -p
```
so this CI step does not pick up local hooks."

Negative triggers:

"Rewrite this runbook for readability."
"Create a new skill for our repo."

Boundary trigger:

"Research Claude Code permissions and tell me what they do." Use this skill only if the user wants the
```
claude
```
CLI involved; otherwise route to research or direct documentation work.

正向触发场景：

"Use Claude Code to review this repository and summarize the risks."
"Run
```
claude
```
in print mode and analyze this architecture."
"Continue the last Claude Code session and ask it to finish the patch."
"Resume the
```
auth-refactor
```
claude session and apply the next fix."
"Use
```
claude --bare -p
```
so this CI step does not pick up local hooks."

负向触发场景：

"Rewrite this runbook for readability."
"Create a new skill for our repo."

边界触发场景：

"Research Claude Code permissions and tell me what they do." 仅当用户希望涉及
```
claude
```
CLI时使用此技能；否则转至研究或直接文档处理工作。

Critical: Print Mode

关键：打印模式

Always use

-p

--print

for non-interactive Claude Code runs. Positional prompts without

-p

start the interactive REPL instead, so a script that omits

-p

will hang waiting for a TTY.

bash

undefined

对于非交互式Claude Code运行，始终使用

-p

--print

。不带

-p

的位置提示符会启动交互式REPL，因此省略

-p

的脚本会挂起等待TTY输入。

bash

undefined

Non-interactive (headless / SDK)

非交互式（无头/SDK）

claude --permission-mode default -p "your prompt here"

Interactive REPL (initial prompt only — does NOT exit)

交互式REPL（仅初始提示符——不会退出）

claude "your prompt here"


`-p` is the canonical SDK/CI entrypoint. The CLI was previously called "headless mode"; the `-p` flag is unchanged.

claude "your prompt here"


`-p`是标准的SDK/CI入口点。该CLI之前被称为“无头模式”；`-p`标志保持不变。

Bare Mode for CI and Scripts

CI和脚本的裸模式

Add

--bare

for any scripted or CI invocation. It skips auto-discovery of hooks, skills, plugins, MCP servers, auto memory, and CLAUDE.md so the call returns the same result on every machine. Bare mode also skips the OAuth keychain, so it requires

ANTHROPIC_API_KEY

ANTHROPIC_AUTH_TOKEN

, or an

apiKeyHelper

passed through

--settings

bash

ANTHROPIC_API_KEY=$KEY claude --bare -p "Summarize the diff" --allowedTools "Read"

For pipelines that need a long-lived OAuth token instead of an API key, generate one with

claude setup-token

and export it as

CLAUDE_CODE_OAUTH_TOKEN

(note: bare mode does NOT read

CLAUDE_CODE_OAUTH_TOKEN

— use

ANTHROPIC_API_KEY

apiKeyHelper

when

--bare

is set).

对于任何脚本化或CI调用，添加

--bare

。它会跳过钩子、技能、插件、MCP服务器、自动内存和CLAUDE.md的自动发现，使调用在每台机器上都返回相同的结果。裸模式还会跳过OAuth密钥链，因此需要

ANTHROPIC_API_KEY

、

ANTHROPIC_AUTH_TOKEN

或通过

--settings

传递的

apiKeyHelper

。

bash

ANTHROPIC_API_KEY=$KEY claude --bare -p "Summarize the diff" --allowedTools "Read"

对于需要长期OAuth令牌而非API密钥的流水线，使用

claude setup-token

生成令牌并将其导出为

CLAUDE_CODE_OAUTH_TOKEN

（注意：裸模式不会读取

CLAUDE_CODE_OAUTH_TOKEN

——设置

--bare

时请使用

ANTHROPIC_API_KEY

或

apiKeyHelper

）。

Running a Task

执行任务

Read references/recipes.md for concrete command recipes before changing permission modes, resuming a session, restricting tools, or adding extra directories.

在更改权限模式、恢复会话、限制工具或添加额外目录之前，请阅读references/recipes.md获取具体的命令示例。

Permission Mode Selection

权限模式选择

Six modes are supported. Pick the loosest mode the task safely allows.

Flag	When to use
`--permission-mode default`	General Claude Code usage with normal approval prompts
`--permission-mode plan`	Read-only analysis or planning with no file changes or shell execution
`--permission-mode acceptEdits`	The user explicitly wants Claude Code to write files (also auto-approves common filesystem commands like `mkdir` , `mv` , `cp` )
`--permission-mode auto`	Long autonomous tasks where prompt fatigue matters; a server-side classifier blocks risky actions but does not replace review
`--permission-mode dontAsk`	Locked-down CI: only `permissions.allow` rules and the read-only command set may run; everything else is auto-denied
`--permission-mode bypassPermissions`	Containers / VMs only — equivalent to `--dangerously-skip-permissions` , requires explicit user approval, and offers no protection against prompt injection

Auto mode requires Claude Code v2.1.83+ and is gated by plan, admin policy, model, and provider; if the CLI reports it as unavailable, do not retry.

支持六种模式。选择任务安全允许的最宽松模式。

标志	使用场景
`--permission-mode default`	带有正常批准提示的常规Claude Code使用
`--permission-mode plan`	只读分析或规划，不修改文件或执行shell命令
`--permission-mode acceptEdits`	用户明确希望Claude Code写入文件（也会自动批准常见的文件系统命令，如 `mkdir` 、 `mv` 、 `cp` ）
`--permission-mode auto`	长期自主任务，避免提示疲劳；服务器端分类器会阻止危险操作，但不能替代人工审核
`--permission-mode dontAsk`	锁定的CI环境：仅允许 `permissions.allow` 规则和只读命令集运行；其他所有操作都会被自动拒绝
`--permission-mode bypassPermissions`	仅适用于容器/虚拟机——等同于 `--dangerously-skip-permissions` ，需要用户明确批准，且对提示注入无保护

自动模式需要Claude Code v2.1.83+，并受计划、管理员策略、模型和提供商限制；如果CLI报告该模式不可用，请不要重试。

Command Discipline

命令规范

Start from

claude --permission-mode default -p "your prompt here"

Add
```
--model <model>
```
or
```
--effort <level>
```
only when the user explicitly asks.
Use
```
--output-format
```
only when the user wants something other than
```
text
```
. Supported values:
```
text
```
(default),
```
json
```
,
```
stream-json
```
.

For schema-validated structured output, pair

--output-format json

with

--json-schema '<JSON Schema>'

Add
```
--max-turns <N>
```
and/or
```
--max-budget-usd <dollars>
```
for autonomous print-mode runs that must not loop forever.
Use
```
--fallback-model <name>
```
in print mode when a graceful degrade beats a hard failure on overload.
Use
```
--add-dir <path>
```
when the task needs files outside the launch directory. (
```
--add-dir
```
grants file access; it does NOT load
```
.claude/
```
configuration from the added directory, except for
```
.claude/skills/
```
.)
Restrict tools with
```
--allowedTools
```
,
```
--disallowedTools
```
, or
```
--tools
```
(use
```
--tools ""
```
to disable all built-ins,
```
--tools "default"
```
for all, or a comma list like
```
"Bash,Edit,Read"
```
).

Add to / replace the system prompt with

--append-system-prompt

--append-system-prompt-file

--system-prompt

, or

--system-prompt-file

Load MCP servers with
```
--mcp-config <file-or-json>
```
; add
```
--strict-mcp-config
```
to ignore every other MCP source.

Ask before using

--dangerously-skip-permissions

. Prefer

--permission-mode acceptEdits

for normal file edits.

从

claude --permission-mode default -p "your prompt here"

开始构建命令。

仅当用户明确要求时，添加
```
--model <model>
```
或
```
--effort <level>
```
。
仅当用户需要非
```
text
```
格式的输出时，使用
```
--output-format
```
。支持的值：
```
text
```
（默认）、
```
json
```
、
```
stream-json
```
。
对于模式验证的结构化输出，将
```
--output-format json
```
与
```
--json-schema '<JSON Schema>'
```
配合使用。
对于必须避免无限循环的自主打印模式运行，添加
```
--max-turns <N>
```
和/或
```
--max-budget-usd <dollars>
```
。
当在过载情况下优雅降级优于硬失败时，在打印模式下使用
```
--fallback-model <name>
```
。
当任务需要启动目录之外的文件时，使用
```
--add-dir <path>
```
。（
```
--add-dir
```
授予文件访问权限；它不会加载添加目录中的
```
.claude/
```
配置，
```
.claude/skills/
```
除外。）
使用
```
--allowedTools
```
、
```
--disallowedTools
```
或
```
--tools
```
限制工具（使用
```
--tools ""
```
禁用所有内置工具，
```
--tools "default"
```
启用所有工具，或使用逗号分隔的列表如
```
"Bash,Edit,Read"
```
）。

使用

--append-system-prompt

、

--append-system-prompt-file

、

--system-prompt

或

--system-prompt-file

添加或替换系统提示符。

使用
```
--mcp-config <file-or-json>
```
加载MCP服务器；添加
```
--strict-mcp-config
```
以忽略所有其他MCP源。
使用
```
--dangerously-skip-permissions
```
前请询问用户。对于常规文件编辑，优先使用
```
--permission-mode acceptEdits
```
。

Resuming a Session

恢复会话

bash

undefined

bash

undefined

Latest session in the current directory

当前目录中的最新会话

claude --continue -p "continue the previous task" # short: claude -c -p "..."

claude --continue -p "continue the previous task" # 简写：claude -c -p "..."

Specific session by ID or by display name

通过ID或显示名称恢复特定会话

claude --resume "auth-refactor" -p "continue with this follow-up" # short: -r claude --resume <session-id> -p "continue with this follow-up"

claude --resume "auth-refactor" -p "continue with this follow-up" # 简写：-r claude --resume <session-id> -p "continue with this follow-up"

Pin a UUID so a script always reuses the same session

固定UUID，使脚本始终重用同一个会话

claude --session-id 550e8400-e29b-41d4-a716-446655440000 -p "..." # must be a valid UUID

claude --session-id 550e8400-e29b-41d4-a716-446655440000 -p "..." # 必须是有效的UUID

Resume sessions linked to a pull request

恢复与拉取请求关联的会话

claude --from-pr 123 -p "address review comments"


Use `--continue` for the latest conversation in the current directory.
Use `--resume` for a specific session by **ID or by display name** (set with `--name` / `-n`, or by `/rename` mid-session).
Use `--session-id` only with a real UUID — the CLI rejects other strings.
When resuming, keep the existing session's behavior unless the user explicitly asks to change the model, effort, or permission mode.
Add `--fork-session` only when the user wants to branch from the existing session instead of reusing it.

claude --from-pr 123 -p "address review comments"


使用`--continue`恢复当前目录中的最新对话。
使用`--resume`通过**ID或显示名称**恢复特定会话（通过`--name`/`-n`设置，或在会话中使用`/rename`设置）。
仅当使用真实UUID时才使用`--session-id`——CLI会拒绝其他字符串。
恢复会话时，保持现有会话的行为，除非用户明确要求更改模型、工作力度或权限模式。
仅当用户希望从现有会话分支而非重用时，添加`--fork-session`。

After Completion

完成后操作

Summarize the result, including any warnings or partial output.

Tell the user they can resume with

claude --continue

(

-c

claude --resume <id-or-name>

(

-r

), or

claude --from-pr <pr>

Ask whether to continue, adjust the prompt, or switch back to direct work.

总结结果，包括任何警告或部分输出。

告知用户可以使用

claude --continue

（

-c

）、

claude --resume <id-or-name>

（

-r

）或

claude --from-pr <pr>

恢复会话。

询问用户是否继续、调整提示符或切换回直接工作模式。

Critical Evaluation

关键评估

Treat Claude Code as a colleague, not an authority.

Trust your own grounded knowledge when you are confident.
Verify disagreements with current docs or primary sources before accepting a claim.
Remember that a separate Claude Code session can still be wrong or stale.
Let the user decide when there is genuine ambiguity.

将Claude Code视为同事，而非权威。

当你有把握时，相信自己的扎实知识。
在接受某个说法之前，先通过当前文档或原始来源验证分歧。
记住，独立的Claude Code会话仍然可能出错或信息过时。
当存在真正的歧义时，让用户决定。

Authentication

身份验证

The CLI reads credentials in this precedence order: cloud provider env vars (

CLAUDE_CODE_USE_BEDROCK

_VERTEX

_FOUNDRY

) →

ANTHROPIC_AUTH_TOKEN

→

ANTHROPIC_API_KEY

→

apiKeyHelper

→

CLAUDE_CODE_OAUTH_TOKEN

→ subscription OAuth from

/login

Browser login: run
```
claude
```
once and follow the prompt, or
```
claude auth login
```
(use
```
--console
```
for Console billing,
```
--sso
```
to force SSO).
Inspect / sign out:
```
claude auth status
```
or
```
/status
```
from inside a session;
```
claude auth logout
```
or
```
/logout
```
to clear credentials.
Long-lived OAuth for CI:
```
claude setup-token
```
prints a token; export it as
```
CLAUDE_CODE_OAUTH_TOKEN
```
(not read by
```
--bare
```
).

Direct API: set

ANTHROPIC_API_KEY

for

X-Api-Key

ANTHROPIC_AUTH_TOKEN

for

Authorization: Bearer

through a gateway.

CLI按以下优先级读取凭据：云提供商环境变量（

CLAUDE_CODE_USE_BEDROCK

_VERTEX

_FOUNDRY

）→

ANTHROPIC_AUTH_TOKEN

→

ANTHROPIC_API_KEY

→

apiKeyHelper

→

CLAUDE_CODE_OAUTH_TOKEN

→ 来自

/login

的订阅OAuth。

浏览器登录：运行一次
```
claude
```
并按照提示操作，或运行
```
claude auth login
```
（使用
```
--console
```
进行控制台计费，
```
--sso
```
强制使用SSO）。
检查/登出：运行
```
claude auth status
```
或在会话内使用
```
/status
```
；运行
```
claude auth logout
```
或在会话内使用
```
/logout
```
清除凭据。
CI用长期OAuth令牌：
```
claude setup-token
```
会打印一个令牌；将其导出为
```
CLAUDE_CODE_OAUTH_TOKEN
```
（裸模式不会读取此令牌）。

直接API：设置

ANTHROPIC_API_KEY

用于

X-Api-Key

，设置

ANTHROPIC_AUTH_TOKEN

用于通过网关的

Authorization: Bearer

。

Error Handling

错误处理

```
command not found: claude
```
: tell the user Claude Code CLI is not installed; they can run
```
claude install stable
```
after setup.
Auth errors: confirm precedence above (e.g. an unset
```
unset ANTHROPIC_API_KEY
```
may be needed when a subscription is active), then re-run
```
claude auth login
```
or
```
claude auth status
```
to confirm.
Permission blocks: retry with an appropriate
```
--permission-mode
```
(
```
plan
```
for read-only,
```
acceptEdits
```
for file edits) or adjust
```
--allowedTools
```
/
```
--disallowedTools
```
. Do not escalate to
```
--dangerously-skip-permissions
```
without explicit user approval.
Session not found: run
```
claude --resume
```
without an argument to pick from a list, or switch to
```
claude --continue
```
for the current directory;
```
--session-id
```
requires a UUID.
Auto-mode unavailable: this is gated by plan, admin policy, model, and provider — it is not a transient outage; fall back to
```
default
```
or
```
acceptEdits
```
.
Invalid flag or model errors: check
```
claude --help
```
, then re-run with supported options.
```
claude --help
```
does not list every flag — consult the CLI reference for the full list.

```
command not found: claude
```
：告知用户Claude Code CLI未安装；他们可以在设置后运行
```
claude install stable
```
。
身份验证错误：确认上述优先级（例如，当订阅处于活动状态时，可能需要取消设置
```
unset ANTHROPIC_API_KEY
```
），然后重新运行
```
claude auth login
```
或
```
claude auth status
```
确认。
权限阻止：使用适当的
```
--permission-mode
```
（只读操作使用
```
plan
```
，文件编辑使用
```
acceptEdits
```
）重试，或调整
```
--allowedTools
```
/
```
--disallowedTools
```
。未经用户明确批准，不要升级到
```
--dangerously-skip-permissions
```
。
会话未找到：不带参数运行
```
claude --resume
```
从列表中选择，或切换到
```
claude --continue
```
恢复当前目录的会话；
```
--session-id
```
需要UUID。
自动模式不可用：这受计划、管理员策略、模型和提供商限制——不是临时故障； fallback到
```
default
```
或
```
acceptEdits
```
模式。
无效标志或模型错误：检查
```
claude --help
```
，然后使用支持的选项重新运行。
```
claude --help
```
并未列出所有标志——请查阅CLI参考获取完整列表。