Block: ZelaxyArena Block

Execute AI model completions using the ZelaxyArena streaming engine. Supports multi-turn conversations, system prompts, tool calls, and full token cost tracking — with the response streamed in real time. Reach for this block when you want to run a completion through the full ZelaxyArena agent (including its workspace tools and snapshot), rather than a bare provider call.

Overview

Property	Value
Type	`zelaxy-arena`
Category	`blocks`
Color	`#7C3AED`

When to Use

Run multi-turn AI conversations that need to carry forward a conversationId between workflow steps
Stream token-by-token responses to the Zelaxy UI during long generations
Invoke the full ZelaxyArena agent with workspace-scoped tools, rather than a standalone LLM provider
Pass a custom systemPrompt to shape model behavior without modifying the global agent configuration
Control sampling creativity with temperature and cap output length with maxTokens
Capture detailed token usage and cost breakdowns for billing or analytics workflows

Configuration

Model

ID: model | Type: short-input | Layout: half | Required

The model identifier to use for the completion, for example gpt-5.6-terra or claude-sonnet-5. This value is forwarded directly to the ZelaxyArena execution API, so use any model string that the engine supports.

Conversation ID

ID: conversationId | Type: short-input | Layout: half | Optional

Pass an existing conversation ID (format: conv_xxxxxxxx) to continue a prior exchange, reusing the conversation's stored context. Leave blank to let the block generate a new UUID and start a fresh conversation. The generated or provided ID is always echoed back in the conversationId output so you can thread it into later blocks.

System Prompt

ID: systemPrompt | Type: long-input | Layout: full | Optional

Optional system-level instructions sent to the model before the user messages. Use this to define the assistant's persona, constraints, or task-specific guidance. Supports {{blockName.field}} references to inject dynamic values from upstream blocks.

Messages

ID: messages | Type: long-input | Layout: full | Required

A JSON array of message objects, each with a role ("user", "assistant", or "system") and a content field. Example:

[{"role": "user", "content": "Summarize this document."}]

You can pass the output of an upstream block using {{blockName.messages}}, or construct the array inline. This is the primary conversation payload sent to the model.

Temperature

Sampling temperature controlling response randomness. 0 produces deterministic, focused output; 2 produces highly varied, creative output. Most tasks work well between 0.2 and 0.8.

Max Tokens

ID: maxTokens | Type: short-input | Layout: half | Optional

Maximum number of tokens to generate in the response. Leave blank to use the model's default limit. Set a value (e.g. 1024) to cap output length and control cost.

Enable Streaming

ID: stream | Type: switch | Layout: half | Optional

When enabled, the block streams the response token-by-token to the Zelaxy UI. The execution engine uses NDJSON framing internally; chunk, heartbeat, and final events are handled automatically. Enable this for long completions where you want live feedback in the canvas.

Inputs & Outputs

Inputs

model (string) — Model identifier for the completion
conversationId (string) — Conversation ID for multi-turn context; omit for a new conversation
systemPrompt (string) — System-level prompt text
messages (json) — Array of {role, content} message objects
temperature (number) — Sampling temperature between 0 and 2
maxTokens (number) — Maximum number of tokens to generate
stream (boolean) — Whether to stream the response token-by-token

Outputs

content (string) — Generated text content from the model
model (string) — Model that was used for the completion
conversationId (string) — Conversation ID to pass to subsequent turns for multi-turn continuity
tokens (json) — Token usage breakdown: { prompt, completion, total }
toolCalls (json) — Tool calls made during the completion, if any: { list: [{name, arguments, result, error, duration}], count }
cost (json) — Estimated cost breakdown: { input, output, total, currency }

Tools

The ZelaxyArena block has no entries in tools.access. It does not call the Zelaxy tool registry directly. Instead, it posts to the internal /api/zelaxy-arena/execute endpoint via the dedicated ZelaxyArenaBlockHandler. That endpoint runs the full ZelaxyArena agent — including any workspace-scoped tools attached to the agent — and streams results back over NDJSON. Tool calls made by the model during execution are surfaced in the toolCalls output.

YAML Example

zelaxy_arena_1:
  type: zelaxy-arena
  name: "ZelaxyArena Completion"
  inputs:
    model: "gpt-4o"
    systemPrompt: "You are a concise summarizer. Respond with a 3-sentence summary."
    messages: "{{starter.messages}}"
    temperature: 0.3
    maxTokens: 512
    stream: true
  connections:
    outgoing:
      - target: response_1

ZelaxyArena Block