OpenAI Provider

Setup

export OPENAI_API_KEY="sk-..."

Quick Start

OpenAI is auto-detected from openai/..., gpt-..., o1-..., and o3-... model strings:

use appam::prelude::*;

#[tokio::main]
async fn main() -> Result<()> {
    let agent = Agent::quick(
        "openai/gpt-5.5",
        "You are a helpful assistant.",
        vec![],
    )?;

    agent
        .stream("Explain async/await in Rust.")
        .on_content(|text| print!("{}", text))
        .run()
        .await?;

    Ok(())
}

Explicit Builder Form

use appam::prelude::*;

let agent = AgentBuilder::new("openai-agent")
    .provider(LlmProvider::OpenAI)
    .model("gpt-5.5")
    .system_prompt("You are a helpful assistant.")
    .build()?;

Reasoning

The OpenAI config in this crate supports these reasoning effort levels:

• None
• Minimal
• Low
• Medium
• High
• XHigh

Use openai_reasoning(...) on the builder:

use appam::prelude::*;

let agent = AgentBuilder::new("reasoning-agent")
    .provider(LlmProvider::OpenAI)
    .model("gpt-daybreak-blue-latest")
    .system_prompt("You are a reasoning assistant.")
    .openai_reasoning(ReasoningConfig::high_effort())
    .build()?;

Convenience constructors include auto(), high_effort(), xhigh_effort(), low_latency(), minimal(), and no_reasoning().

The exact gpt-5.6-sol and gpt-daybreak-blue-latest models share configuration rules through the direct OpenAI and Azure OpenAI transports: their model-aware default remains High, and explicit XHigh is supported. Neither accepts None reasoning or sampling parameters. Requests preserve the selected model identifier, and unknown suffixed variants remain unsupported for XHigh.

GPT-5.5 Sampling Note

The current OpenAI config code only allows temperature, top_p, and top_logprobs for GPT-5.5 when reasoning is in None mode. Use ReasoningConfig::no_reasoning() if you need those sampling controls on GPT-5.5.

Model-Aware Helpers

The crate exposes model-aware helpers such as:

• default_reasoning_effort_for_model
• model_supports_xhigh_reasoning
• resolve_reasoning_effort_for_model

These are useful if your application chooses effort dynamically from a model string.

Example Binary

cargo run --example coding-agent-openai-responses