Public in-process runtime for embedding Everruns.

The runtime crate exposes an in-memory execution surface that runs the same core atoms (input, reason, act) used elsewhere in the system, but without the durable engine, gRPC worker boundary, or control-plane server.

This is the intended public entrypoint for embedders who want to:

run sessions in their own process
provide their own platform definition (capabilities, drivers, harnesses)
seed harnesses, agents, sessions, and workspace files directly in code
replace the default in-memory stores with custom runtime backends
inspect the assembled turn context before or after executing a turn
reuse runtime-owned host phase execution from durable or server-backed hosts
map plan_next_host_turn(...) onto their own queue, retry, or in-memory host

For a runnable example, see:

cargo run -p everruns-runtime --example in_process_runtime
cargo run -p everruns-runtime --example inspect_context

Example

use everruns_core::{
    CapabilityRegistry, DriverRegistry, InputMessage, LlmProviderType, ModelWithProvider,
    PlatformDefinition,
};
use everruns_core::capabilities::TestMathCapability;
use everruns_runtime::InProcessRuntimeBuilder;

let mut capabilities = CapabilityRegistry::new();
capabilities.register(TestMathCapability);

let platform = PlatformDefinition::new(capabilities, DriverRegistry::new());

let runtime = InProcessRuntimeBuilder::new()
    .platform_definition(platform)
    .single_session(|s| {
        s.harness("math", "You are a calculator.")
            .harness_display_name("Math")
            .with_capability("test_math")
            .agent("math-agent", "Use tools when needed.")
            .agent_display_name("Math Agent")
            .agent_max_iterations(8)
            .session_title("Math Session")
    })
    .llm_sim(everruns_core::llmsim_driver::LlmSimConfig::fixed("4"))
    .default_model(ModelWithProvider {
        model: "llmsim-model".into(),
        provider_type: LlmProviderType::LlmSim,
        api_key: Some("fake-key".into()),
        base_url: None,
    })
    .build()
    .await?;

let session_id = runtime.default_session_id().expect("single_session id");
let result = runtime
    .run_turn(
        session_id,
        InputMessage::user("What is 2 + 2?"),
    )
    .await?;
assert!(result.success);
# Ok::<(), everruns_core::AgentLoopError>(())

Real-disk workspace

Embedders who want built-in capabilities (file_system, agent_instructions, skills, ...) to read and write a real directory on disk can configure [RealDiskSessionFileSystemFactory] on their [PlatformDefinition]. Every capability that goes through ToolContext.file_store or SystemPromptContext.file_store picks it up automatically.

See the runnable examples for the full wiring:

cargo run -p everruns-runtime --example real_disk_agent_instructions
cargo run -p everruns-runtime --example real_disk_file_system_tools

And specs/file-store.md for the trait contract.

everruns-runtime 0.8.33

Example

Real-disk workspace