volcengine-agent-identity（volcengine Agent 身份）

Agent Identity

Use the agent-identity plugin for UserPool OIDC login (入站授权), TIP token (工作负载访问令牌), credential hosting (出站授权 OAuth2, API key), and optional tool risk approval (权限管控 AuthZ).

Volcengine terminology: 用户池 (UserPool), 入站授权 (OIDC login), 出站授权 (credential fetch), 工作负载令牌 (TIP), 凭据托管 (credential hosting), 权限管控 (CheckPermission). Docs: Volcengine 智能体身份和权限管理.

Agent flow: When the user asks to log in, add credentials, check status, bind env, etc., call the corresponding tools directly. Do not suggest slash commands for those. Slash commands below are for user-initiated use (e.g. /identity approve <id> when the user must approve in chat; agent must never call identity_approve_tool).

Slash commands (user-initiated)

Tools Overview

Risk Detection and Approval

When authz.requireRiskApproval is on, the plugin classifies tool calls (e.g. exec, write, apply_patch) by risk. User-provided commands and file paths are evaluated:

• Rule-based: Destructive patterns (rm -rf, sudo, curl|bash), sensitive paths (/etc, ~/.ssh).
• LLM-based (optional): When rules return "medium", an LLM re-evaluates for context (authz.enableLlmRiskCheck).

High-risk calls require user approval. The approval message and block reason include the LLM risk explanation when available (e.g. "Pipe-to-shell: network fetch piped to shell execution").

Tool Parameters

identity_login

Starts OIDC login or refreshes TIP. Call when: "login", "登录", "sign in", "我需要先登录". Required before identity_fetch. No params.

identity_whoami

Brief identity check. Call when: "who am I", "查身份", "am I logged in", "当前登录状态"

Returns: sub, hasTip, loggedIn, sessionLoginAt, sessionExpiresAt, tipIssuedAt, tipExpiresAt, tipExpiresInSeconds, tipChain. No params.

identity_status

Full status including credentials and bindings. Call when: "status", "查看完整状态", "我的凭据和绑定", "show my credentials and bindings"

Returns: loggedIn, sub, hasTip, session (loginAt, expiresAt), tip (issuedAt, expiresAt, chain), credentialProviders, bindings. No params.

identity_list_credentials

Lists available credential providers and what the user has stored. Call this when the user wants to see what they can connect or what credentials they have.

User prompts: "有哪些服务可以连接", "what providers are available", "我添加了哪些凭据", "list my credentials", "show available providers"

{ "page": 2 }

Returns: providers, storedOnly, page, hasMore.

identity_fetch

Adds a credential for a provider (OAuth2 or API key). Call this when the user wants to add, get, or configure credentials.

User prompts that mean "call identity_fetch":

• English: "add/google my Google token", "get credentials for OpenAI", "connect my GitHub", "I need to use Google API", "set up API key for X", "authorize access to Y", "I want to use [provider] but have no key"
• 中文: "帮我添加/获取 Google 凭据", "配置 OpenAI 的 API key", "连接我的 GitHub", "我要用某某服务但没有密钥", "授权访问某平台", "添加某某的 token", "获取某某的凭证"

First ensure user is logged in (identity_whoami); if not, use identity_login. Then call identity_fetch with the provider. Use identity_list_credentials to discover available providers.

{ "provider": "google" }

{ "provider": "openai", "flow": "apikey", "returnValue": true }

Response:

• OAuth2-user: authUrl (user must open in browser). After authorization, success message sent to chat.
• OAuth2-m2m / apikey: success: true, message (completes immediately). If returnValue: true, also includes value (credential string) for same-turn use.

identity_set_binding

Binds a stored credential to an env var so tools can use it at runtime. Call this when the user wants tools/agent to have access to a credential.

User prompts: "让工具能用我的 Google 凭据", "bind/google my credential for tools", "把 Google token 注入给 agent", "inject my OpenAI key for API calls", "配置某某凭据给工具用"

Credential must exist first (identity_fetch). Common env vars: GOOGLE_ACCESS_TOKEN, OPENAI_API_KEY, GITHUB_TOKEN, etc.

{ "provider": "google", "envVar": "GOOGLE_ACCESS_TOKEN" }

If credential exists: binds it. Else: imports from process.env[envVar] as api_key (gateway must have that env set).

identity_unset_binding

{ "provider": "google" }

identity_approve_tool

Optional tool (not given to agent by default). For human approval, use /identity approve <id> or reply "approve" in chat. The agent must NOT call this tool to self-approve. The approval prompt includes the LLM risk reason when available.

{ "approval_id": "abc123" }

identity_risk_check

Evaluates risk of a command or tool call before execution. Call when: "这个命令安全吗", "is rm -rf dangerous", "check if this is risky", "帮我评估这个命令有没有风险"

*Provide either command or toolName. Returns risk, reason, source (rules or llm). Uses LLM when authz.enableLlmRiskCheck is true and rules return medium.

{ "command": "rm -rf /" }

{ "toolName": "write", "params": { "path": "/etc/hosts", "content": "..." } }

identity_list_risk_patterns

Returns built-in dangerous command patterns and sensitive paths. No params. Use to query what triggers high-risk approval.

{}

identity_config_suggest

Generates config snippets for the agent-identity plugin. Call when: user asks to configure identity, login, authz, risk approval, or "如何配置 identity 插件", "帮我配置登录", "怎么开启权限检查".

Returns: configPath, config (JSON to merge), instructions, nextSteps. When intent is identity or full, also returns identityDefaults (env vars, credential resolution order, config defaults, credential file format). User must manually add to openclaw.json and restart gateway.

{ "intent": "userpool", "lang": "zh" }

Workflow: Adding a Credential

1. Check login: identity_whoami (brief) or identity_status (full). If not logged in, use identity_login first (user opens auth URL).
2. Add credential: identity_fetch with provider. For OAuth2-user, tell user to open authUrl; success message sent when done.
3. Bind for tools (optional): identity_set_binding so the credential is injected as an env var when tools run.

Workflow: Checking Risk Before Running

1. Diagnose: identity_risk_check with command or toolName+params. Returns risk level and reason.
2. List patterns: identity_list_risk_patterns to see what triggers high-risk approval.

Configuration

Plugin config lives under plugins.entries.agent-identity.config:

• identity: Identity API (endpoint, credentials, workloadPoolName, workloadName, roleTrn). When roleTrn is set (AssumeRole), workload name is omitted; backend uses roleName. When workload not found (404), plugin auto-creates via CreateWorkloadIdentity then retries.
• userpool: OIDC (discoveryUrl, clientId, callbackUrl, or userPoolName+clientName)
• authz: Optional AuthZ (toolCheck, skillReadCheck, requireRiskApproval, enableLlmRiskCheck, llmRiskCheck, namespaceName, lowRiskBypass). When enableLlmRiskCheck is true, rules returning "medium" are re-evaluated via LLM; the risk reason is shown in approval prompts and block messages.

Notes

• Requires the agent-identity plugin to be enabled.
• /identity and tools require session context (channel + sender); use from an active chat.
• identity_risk_check and identity_list_risk_patterns do not require login.