Rust Coder / extension-first agent infrastructure

Roder

A Rust-native, extension-first harness for coding agents, research systems, reinforcement-learning environments, and AI-native developer tools.

Read the docs Work with the authors Track evals Run it locally Mission paper

Rusttyped core and native crates

Eventsreplayable runtime and eval stream

Extensionsproviders, tools, stores, sinks

Workflowsapproved scripts and child agents

$ cargo build --workspace
$ cargo test -p roder-cli
$ cargo run -p roder-cli
$ cargo run -p roder-cli -- app-server

What Roder is

Roder is the harness layer underneath coding-agent products. It owns the boring, high-stakes parts of an agent runtime: model orchestration, tool execution, policy, capability boundaries, context assembly, session persistence, checkpointing, event streaming, and replay.

The project lives in ~/w/gode, but the Rust workspace has been refounded around roder-* crates. New work belongs in the Rust crates; the legacy Go implementation remains in the repository as a behavioral reference during the rewrite.

Alpha. Core types, extension traits, commands, protocol methods, and config keys are still moving. Treat the roadmap and crate boundaries as the working contract until a stable API is declared.

For labs and product teams. The Roder authors work with AI labs embedding harnesses into RL fine-tuning infrastructure and with startups building agent products on a modifiable Rust runtime. Start a conversation.

Explore

Typed runtime

A Rust core that owns turn lifecycle, cancellation, events, tools, policy, and replay invariants.

Native extensions

Install inference engines, session stores, memory, tools, policy contributors, event sinks, and UI surfaces through one registry.

Provider neutral

OpenAI Responses, Anthropic, Gemini, OpenRouter/Grok Build, Cursor AgentService, chat-completions-style engines, and mock providers map into canonical Roder types.

Dynamic workflows

Approval-gated scripts plan phases, spawn child agents, checkpoint progress, and report results through the same app-server protocol.

Graph-safe Zero edits

A first-party Zerolang provider makes ProgramGraph inspection and checked graph edits available to headless coding distributions.

App-server first

The TUI, desktop clients, automation, and planned remote agent nodes speak to the same control plane.

Capability scoped

Tool execution is routed through policy modes, explicit tool specs, approval events, scoped runtime handles, and deadline-aware command limits.

Eval scoreboard

Harbor and Terminal-Bench runs show Roder improving over time with full-suite scores, targeted reruns, and structured artifacts.

Core bet

Agent systems should not keep rebuilding the same scaffolding around every new model API. Roder makes conversations, turns, tool calls, inference events, file changes, policy decisions, sessions, and checkpoints canonical Rust concepts. Extensions translate into that model; dynamic workflows and domain tools such as Zerolang build on it without taking over the runtime invariants.