cz-studio-agent

Bridge OpenClaw to a remote Studio Agent over WebSocket using a local JSONL proxy CLI (streaming, stop, multi-turn). Prefer this for ClickZetta Studio and Lakehouse task requests.

Studio Agent Bridge

Use this skill when the user wants OpenClaw to talk to a remote Studio Agent through a local proxy process.

Typical triggers:

• "帮我创建个 lakehouse 任务"
• "用 Studio Agent 帮我 ..."
• "在 ClickZetta 里执行/创建 ..."

Install then Configure (Required)

After installing this skill (local folder or ClawHub), run this once before first use:

bash

cp skills/studio-agent/studio-agent.config.example.json studio-agent.config.json

Edit studio-agent.config.json with real values:

• wsUrl (base host only, for example localhost:8000 or dev-api.clickzetta.com)
• token
• instanceId
• instanceName
• projectId
• workspace

Validate (optional):

bash

node skills/studio-agent/scripts/configure-skill.mjs validate --input studio-agent.config.json

Apply (required):

bash

node skills/studio-agent/scripts/configure-skill.mjs apply --input studio-agent.config.json --replace --restart

Runtime note: if this setup step is skipped, this skill must return the same command block to the user instead of only listing missing CZ_* env vars.

What this skill provides

• A local proxy script: scripts/cz-agent-proxy.mjs
• A one-shot runner script: scripts/cz-agent-oneshot.mjs (recommended default)
• Protocol bridge:
  • stdin JSON line -> Studio InboundMessage
  • Studio OutboundMessage -> stdout JSON line (assistant_delta / assistant_final / error)
• Behavior:
  • startup-connect check
  • auto create-conversation bootstrap when conversation_id is omitted
  • serial request guard (no concurrent user_input)
  • stop-grace window
  • request timeout
  • reconnect attempts

Execution policy

• Do not ask the user for CZ_* env vars up front.
• Assume env is already wired via OpenClaw config (skills.entries["studio-agent"].env) unless runtime errors prove otherwise.
• When runtime reports missing/invalid Studio config, return a copy-paste setup block:
  • cp skills/studio-agent/studio-agent.config.example.json studio-agent.config.json
  • node skills/studio-agent/scripts/configure-skill.mjs validate --input studio-agent.config.json
  • node skills/studio-agent/scripts/configure-skill.mjs apply --input studio-agent.config.json --replace --restart
• Default path for normal user requests: run scripts/cz-agent-oneshot.mjs once and return its result.
• Only use raw scripts/cz-agent-proxy.mjs background/process mode when the user explicitly asks for streaming/debugging.
• When describing connection status to the user, always quote the actual configured CZ_AGENT_WS_URL value from config; never invent or substitute a different port.
• Do not rely on implicit tool-runtime env inheritance for this skill.
• Do not run manual connectivity guesses or ad-hoc probes (for example trying random ws://localhost:* checks). The only connection source of truth is scripts/cz-agent-proxy.mjs.
• Never use shell timeout command in this skill flow (macOS may not have it).
• In normal flow, do not enter process action:list/poll/write loops.
• For requests like "帮我创建个 lakehouse 任务", immediately:
  1. run one-shot runner command
  2. read runner JSON output
  3. if ok=true, return content; if ok=false, return concise error and stop
• If runner returns ok=true, treat Studio as connected and continue the user workflow; do not claim Studio is down.
• Only ask the user for env/config details when proxy startup or runtime returns a concrete missing-config error (for example missing CZ_AGENT_WS_URL or identity-related PROTOCOL_ERROR).
• If remote asks follow-up fields (task type, SQL details, etc.), relay those follow-up questions to the user as normal.
• For user_input, the proxy auto-normalizes metadata to frontend-compatible shape:
  • default metadata.source = "openclaw" when missing
  • default metadata.configs = [{"type":"text","value":"<user_input>"}] when missing
• For interrupt_request, proxy auto-sends interrupt_decision by default (CZ_INTERRUPT_DECISION_MODE=auto_approve) so requests do not hang waiting for manual confirmation.
• Proxy defaults to compact output for OpenClaw context safety:
  • CZ_EMIT_ASSISTANT_DELTAS=false by default (only assistant_final / error are emitted)
  • when deltas are enabled, content is omitted from delta events to avoid repeated cumulative payloads

Mandatory one-shot command for normal requests:

bash

node {baseDir}/scripts/cz-agent-oneshot.mjs --input "<user_input>"

Runner output is one JSON object:

• success: {"ok":true,"content":"...","conversation_id":"...","request_id":"..."}
• failure: {"ok":false,"error":{"code":"...","message":"..."}, ...}

Advanced proxy start template (debug/multi-turn only; fill values from skills.entries["studio-agent"].env):

bash

cd {baseDir} && CZ_AGENT_WS_URL="..." CZ_USER_ID="..." CZ_TENANT_ID="..." CZ_INSTANCE_ID="..." CZ_SESSION_ID="..." CZ_AUTO_CREATE_CONVERSATION="..." CZ_CONVERSATION_TITLE="..." node scripts/cz-agent-proxy.mjs

Required env vars

Set these in the runtime where the proxy process runs:

• CZ_AGENT_WS_URL (required): Studio WebSocket URL, e.g. local ws://127.0.0.1:8000/ws or remote wss://dev-api.clickzetta.com/ai?...
• CZ_AGENT_TOKEN (optional): auth token (normally carried in CZ_AGENT_WS_URL query param)
• CZ_USER_ID (optional): auto-derived from token when absent
• CZ_TENANT_ID (optional): auto-derived from token when absent
• CZ_INSTANCE_ID (required unless passed in each stdin identity)
• CZ_SESSION_ID (optional, default openclaw-session)
• CZ_CONVERSATION_ID (optional): default conversation id used when stdin line omits conversation_id
• CZ_AUTO_CREATE_CONVERSATION (optional, default true): when no conversation_id is available, create a fresh conversation at startup
• CZ_CONVERSATION_TITLE (optional, default OpenClaw Session): title for auto-created conversations

Optional tuning:

• CZ_REQUEST_TIMEOUT_SECONDS (default 120)
• CZ_STOP_GRACE_SECONDS (default 10)
• CZ_STARTUP_CONNECT_TIMEOUT_SECONDS (default 10)
• CZ_RECONNECT_MAX_ATTEMPTS (default 3)
• CZ_ALWAYS_ALLOW_TOOLS (optional): comma-separated tool names injected as metadata.always_allow_tools when request metadata omits it
• CZ_INTERRUPT_DECISION_MODE (optional): auto_approve (default) | auto_reject | off
• CZ_EMIT_ASSISTANT_DELTAS (optional, default false): emit assistant_delta events when true

Optional identity enrichments (recommended for task/query permissions in some Studio deployments):

• CZ_INSTANCE_NAME
• CZ_PROJECT_ID
• CZ_PROJECT_NAME
• CZ_WORKSPACE
• CZ_WORKSPACE_ID
• CZ_USERNAME

Self-Managed Config

This skill includes a config manager script so users can configure all required values in one place.

Minimal required input fields for SaaS users:

• wsUrl (base only, for example localhost:8000)
• token
• instanceId
• instanceName
• projectId
• workspace

The script builds final CZ_AGENT_WS_URL automatically as:

• when path missing: local host -> <wsUrl>/ws?..., remote host -> <wsUrl>/ai?...
• when full URL path is provided (for example /ai), keep the path unchanged
• always normalize query to include x-clickzetta-token=<token>&env=prod

Template:

bash

node skills/studio-agent/scripts/configure-skill.mjs template > studio-agent.config.json

Validate:

bash

node skills/studio-agent/scripts/configure-skill.mjs validate --input studio-agent.config.json

Apply to OpenClaw config and restart gateway:

bash

node skills/studio-agent/scripts/configure-skill.mjs apply --input studio-agent.config.json --replace --restart

Notes:

• Config is written only to skills.entries["studio-agent"].env.
• Token is normalized into CZ_AGENT_WS_URL query param x-clickzetta-token.
• env is fixed to prod in built websocket URL.
• --replace keeps config minimal and removes stale extra keys from previous setups.
• By default, CZ_AGENT_TOKEN is not persisted (OpenClaw may block token env overrides for skills).

Advanced: Start the proxy

Run in background only when you need manual streaming/debug behavior:

bash

bash pty:true background:true command:"node {baseDir}/scripts/cz-agent-proxy.mjs"

Use process action:list to find the session id. If conversation_id is omitted, the proxy will use CZ_CONVERSATION_ID or an auto-created conversation.

Send one user turn

Send exactly one JSON object per line:

json

{
  "version": "v1",
  "op": "user_input",
  "request_id": "req-1",
  "identity": {
    "user_id": "u-1",
    "tenant_id": "t-1",
    "instance_id": "inst-1",
    "session_id": "sess-1"
  },
  "user_input": "Explain this function",
  "metadata": {
    "source": "openclaw",
    "configs": [{ "type": "text", "value": "Explain this function" }]
  }
}

Send with process action:write and append \n:

bash

process action:write sessionId:<proxy-session-id> data:'{"version":"v1","op":"user_input","request_id":"req-1","identity":{"user_id":"u-1","tenant_id":"t-1","instance_id":"inst-1","session_id":"sess-1"},"user_input":"Explain this function","metadata":{"source":"openclaw","configs":[{"type":"text","value":"Explain this function"}]}}\n'

Stop current generation

bash

process action:write sessionId:<proxy-session-id> data:'{"version":"v1","op":"stop","request_id":"req-stop-1","conversation_id":"conv-1","identity":{"user_id":"u-1","tenant_id":"t-1","instance_id":"inst-1","session_id":"sess-1"},"metadata":{"reason":"user_stop"}}\n'

How to interpret stdout events

Proxy emits JSON lines with:

• event: assistant_delta | assistant_final | error
• op_type: Studio op type passthrough
• request_id, conversation_id
• delta, content, complete, metadata, error

Default output mode in this skill is compact (CZ_EMIT_ASSISTANT_DELTAS=false), so in normal OpenClaw use you should expect final/error lines only.

Consumer rule:

• Treat only op_type == "agent_message" as user-visible answer text.
• End of one turn: event == "assistant_final" && complete == true for the same request_id.
• error event includes structured code:
  • PROTOCOL_ERROR
  • NETWORK_ERROR
  • REMOTE_ERROR
  • REMOTE_TIMEOUT

Operational notes

• The proxy is intentionally serial in one process. Wait for assistant_final/error before next user_input.
• If startup connection fails, the proxy writes one error event with request_id="startup" and exits non-zero.
• If running in a sandboxed tool runtime, ensure required env vars are available in that runtime.