Auto-Compaction (OpenAI)

examples/compaction-openai.rs demonstrates enable_auto_compaction end to end on the OpenAI Responses API: a fetch_catalog tool returns ~4K tokens per call, the model fetches five sections one at a time, and the API compacts the conversation into encrypted compaction items mid-session. Appam replays the items automatically and prunes pre-compaction input on later turns. Works with appam's default stateless mode (store: false).

Run

export OPENAI_API_KEY="sk-..."
cargo run --example compaction-openai

What this example actually configures

• LlmProvider::OpenAI with model gpt-5-mini
• enable_auto_compaction(16_000) — comfortably above the compacted-window size so the server does not re-compact on every reasoning step
• One local tool: fetch_catalog (~4K tokens per section)
• An on_compaction stream hook (OpenAI summaries are opaque, so no summary text)
• A post-run report over session.messages and session.usage

Key builder setup

let agent = AgentBuilder::new("compaction-demo-openai")
    .provider(LlmProvider::OpenAI)
    .model("gpt-5-mini")
    .system_prompt(
        "You are a star catalog librarian. Fetch catalog sections with the \
         fetch_catalog tool exactly as instructed, one section per tool call.",
    )
    .enable_auto_compaction(16_000)
    .with_tool(Arc::new(fetch_catalog()))
    .max_tokens(4096)
    .build()?;

See the Context Compaction guide for how compaction works across providers.