OpenAI-compatible API

• Base URL — https://<platform>/v1/openai
• API key — your ZWRM API key
• Model — the agent’s name

from openai import OpenAI

client = OpenAI(base_url="https://<platform>/v1/openai", api_key="<zwrm-api-key>")

response = client.chat.completions.create(
    model="support-bot",
    messages=[{"role": "user", "content": "Summarize the open tickets"}],
)
print(response.choices[0].message.content)

​

Endpoints

​

Conversation continuity

• Continuity is keyed by the X-Session-Key header, or the request’s user field if no header is set.
• No key means the agent’s default workspace — the same conversation the dashboard chat uses.
• Only the last user message of each request is injected; prior messages in the messages array are ignored because the server already has the transcript.

​

Semantics and limits

• One reply per turn. n, tools, and functions are rejected — an agent brings its own tools and produces a single reply.
• Sampling params are ignored. temperature, top_p, etc. are accepted but the agent’s own configuration governs the model call.
• Turns serialize per conversation. Concurrent completions on the same session key queue behind each other.
• Busy conversations return 409 — e.g. while a run is executing on the workspace, or while the session waits on an escalation. Retry when it finishes.
• Turn deadline. A completion request is bounded at 30 minutes; on timeout the HTTP request fails but the turn itself finishes into the transcript.
• Errors use the OpenAI wire shape (invalid_request_error, authentication_error, rate_limit_error, server_error).