Auto-Compaction (Anthropic)

examples/compaction-anthropic.rs demonstrates enable_auto_compaction end to end on the direct Anthropic API: a fetch_archive tool returns ~16K tokens per call, the model fetches four volumes one at a time, and the API compacts the conversation into a compaction block mid-session. The example prints each compaction as it happens, then verifies the blocks were retained in session history and reports compaction-pass token usage.

Requires a compaction-capable model (Claude Sonnet 4.6+, Opus 4.6+, or the Claude 5 family). A full run consumes roughly 200K input tokens.

Run

export ANTHROPIC_API_KEY="sk-ant-..."
cargo run --example compaction-anthropic

What this example actually configures

• LlmProvider::Anthropic with model claude-sonnet-4-6
• enable_auto_compaction(50_000) — the Anthropic minimum trigger
• One local tool: fetch_archive (~16K tokens per volume)
• An on_compaction stream hook that prints the provider and summary preview
• A post-run report over session.messages and session.usage

Key builder setup

let agent = AgentBuilder::new("compaction-demo-anthropic")
    .provider(LlmProvider::Anthropic)
    .model("claude-sonnet-4-6")
    .system_prompt(
        "You are an expedition archivist. Fetch archive volumes with the \
         fetch_archive tool exactly as instructed, one volume per tool call.",
    )
    .enable_auto_compaction(50_000)
    .with_tool(Arc::new(fetch_archive()))
    .max_tokens(4096)
    .build()?;

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