OpenAI

RondoFlow can back an Assistant (internally an Agent) with OpenAI instead of the Claude Code CLI. An OpenAI Assistant talks directly to the OpenAI Responses API — streaming the reply, optionally letting the model search the web, and continuing the same conversation across turns.

OpenAI Assistants run on the same execution wiring as every other Assistant, so they slot into Workflows, Orchestration, and Discussions without any special handling. They differ only in what runs behind the scenes: an HTTP API call rather than a started (spawned) local process.

Supported models

The Assistant editor offers a curated list of chat models. The stored value is the raw model id sent to the Responses API, so the list can grow without code changes.

The default model for a new OpenAI Assistant is gpt-4.1-mini.

Deep-research models

When you turn on Deep research, RondoFlow ignores the chat model above and auto-switches the request to a dedicated deep-research model with web search forced on.

The fast model (o4-mini-deep-research-2025-06-26) is selected by default. Deep-research runs are agentic and can take many minutes, so RondoFlow uses a much wider request timeout for them (30 minutes vs. 10 minutes for a normal chat run).

Setup

OPENAI_API_KEY=sk-...

Tools and modes

Each OpenAI Assistant stores a small providerConfig: a model id plus two toggles.

Web search

Turn on Web search to add OpenAI’s web_search tool to the run. When the model searches, RondoFlow surfaces the activity in the run feed as tool-use / tool-result events, so you can see when it reaches out to the web. (These are mapped from OpenAI’s response.web_search_call.in_progress / …searching events to a tool_use and from …completed to a tool_result; no search query payload is captured.) See Monitoring.

Deep research

Turn on Deep research to hand the task to a deep-research model. This:

• Auto-switches the request to the configured deep-research model (the chat-model dropdown is disabled while it’s on).
• Forces web search on (the deep-research models require it) — the web-search toggle is locked.
• Streams the model’s progress through reasoning summaries (reasoning.summary: 'auto') as it works.

Deep research is slower and costlier than a normal chat run. Reach for it when you want a thorough, multi-source answer rather than a quick reply.

Multi-turn continuation

Within a single run, follow-up messages continue the same conversation. RondoFlow captures the response id returned by each completed response and passes it as previous_response_id on the next turn, so the model keeps full context across turns without you resending history.

What carries over from a Claude Assistant

An OpenAI Assistant shares the Assistant data model, but only the parts that map onto an API request are used at run time:

• Personality (Persona) and Skills — both reach the model. The request instructions is the persona plus the markdown of any enabled Skills (each Skill’s SKILL.md is appended), not the persona alone. The incoming message becomes the request input.
• Workspace and agent memory — folded into the same prompt the way they are for any Assistant.

Claude-CLI-only concerns are skipped for API providers: the allowed-tools list, Connections (MCP Servers), external folders, and permission modes do not apply to an OpenAI Assistant. Its only tools are the web_search / deep-research capabilities driven by providerConfig.

How providerConfig maps to the request

The stored config drives model and tool selection at run time:

interface ProviderConfig {
  model: string        // chat model id (ignored when deepResearch is on)
  webSearch: boolean   // adds the web_search tool
  deepResearch: boolean // auto-switches to a deep-research model
  deepResearchModel?: string // optional explicit deep-research override
}

• deepResearch: false → uses model; adds { type: 'web_search' } only when webSearch is true.
• deepResearch: true → uses deepResearchModel (or the fast default) and forces the web_search_preview tool, regardless of model / webSearch.

Cost estimates

The Responses API does not return a dollar figure, so RondoFlow computes a coarse estimate from a built-in per-model rate table (USD per 1M tokens) and reports it on the run’s usage event. Treat it as an approximation, not a billing source of truth.

Any model not in this table falls back to an estimate of $2 input / $8 output per 1M tokens.

Next steps