Reviews run for a minimum of one week. The outcome of the review is decided on this date. This is the last day to make comments or ask questions about this review.
Eclipse Atesor
Atesor is LangGraph based multi-agent orchestration system that autonomously port x86/ARM based Linux packages to RISC-V. It also tests the builds on RISC-V based container images and native machines. It has been trained on 200+ manually ported package recipes from Go,Cmake build systems with more build systems support coming in future releases.
Eclipse Atesor provides a build system and package neutral agentic system which is fully autonomous and can self research, generate recipe, patch and test packages for RISC-V.
In-scope:
- RISC-V software porting
- LangGraph based agent building
- RISC-V container images (Docker on Alpine, Debian/Ubuntu)
- Native RISC-V machines for testing
- RISC-V packages management
- LLM fine-tuning and agentic design
Eclipse Atesor is an agentic, state-driven system that takes a Git repository URL (only required input) and returns a verified RISC-V build together with a complete porting recipe. The architecture is a hierarchical planner / supervisor / executor graph implemented with LangGraph, in which roughly 70% of routine analysis is handled by deterministic, zero-LLM scripted operations and the remaining steps are delegated to specialised agents that operate strictly inside a sandboxed RISC-V container. These LLMs are fine-tuned very precisely to handle build-neutral package porting.
The system is composed of the following coordinated agents:
PLANNER
Derives a high-level TaskPlan from the repository's structure, build system, and detected architecture-specific code.
SCOUT
Produces a concrete BuildPlan (configure, compile, install, verify) tailored to the detected build system (CMake, Make/Autotools, Meson, Ninja, Go modules, Cargo).
BUILDER
Executes the plan inside the RISC-V sandbox, captures structured output, and records every command, exit code, and duration.
FIXER
When the Builder reports an error, classifies it (dependency, compilation, linking, architecture, configuration, missing tools, …), retrieves relevant few-shot examples, and proposes either shell remediations or source patches that are re-applied through a validated patch pipeline.
SUPERVISOR
Verifies agent outputs against known anti-patterns (hallucinated commands, unsafe operations), routes the next step, and escalates when the system detects a loop, repeated error category, or budget exhaustion.
SUMMURIZER
On success, writes a Markdown porting recipe and persists the build to a recipe cache so that future requests for the same package short-circuit the LLM pipeline entirely.
The project further includes a lightweight few-shot memory system. Curated examples for each agent role live in JSON files in the repository; successful porting runs can append novel patterns back into this store under file-lock-protected I/O. This delivers measurable accuracy improvements without requiring a vector database or any external infrastructure.
A separate recipe cache records every package that has been successfully ported, keyed by package and sandbox profile, so that a project ported once can be re-built deterministically by anyone without re-invoking any LLM.
The Atesor trademark belongs to 10xEngineers. No problem foreseen passing it on to the Eclipse Foundation. The current codebase is licensed under MIT but we are willing to re-license it under the EPL v2 license. We own the complete codebase and are able to re-license it.
The final project's name might change before the project creation.
The OpenHW Group is the natural home for this project. The OpenHW Group's mission is to grow the open source RISC-V hardware and software ecosystem. Eclipse Atesor directly serves that mission by removing one of the most persistent friction points slowing RISC-V software adoption which is the manual, expert-driven process of porting third-party packages.
The project aims to reduce the RISC-V software porting efforts to minimal but automating the process using AI. We are able to successfully release v1.0 which is able to port Cmake/Go based packages to RISC-V autonomously with minimal human intervention.
We further aim to:
- Increase the build system support ( C#, Rust and more)
- Improve the accuracy of build recipes
- Centralized platform where RISC-V community can easily find all packages
- Port large packages/libs and applications
We propose a bi-annual major release cycle.
Atesor initial version has been released already with C/C++, Go based 150+ packages for Ubuntu and Alpine based systems.
The complete codebase in written by 10xEngineeers and contains no third-party licensed code.
- Log in to post comments
Eclipse Atesor is an agentic, state-driven system that takes a Git repository URL (only required input) and returns a verified RISC-V build together with a complete porting recipe. The architecture is a hierarchical planner / supervisor / executor graph implemented with LangGraph, in which roughly 70% of routine analysis is handled by deterministic, zero-LLM scripted operations and the remaining steps are delegated to specialised agents that operate strictly inside a sandboxed RISC-V container. These LLMs are fine-tuned very precisely to handle build-neutral package porting.
The system is composed of the following coordinated agents:
PLANNER
Derives a high-level TaskPlan from the repository's structure, build system, and detected architecture-specific code.
SCOUT
Produces a concrete BuildPlan (configure, compile, install, verify) tailored to the detected build system (CMake, Make/Autotools, Meson, Ninja, Go modules, Cargo).
BUILDER
Executes the plan inside the RISC-V sandbox, captures structured output, and records every command, exit code, and duration.
FIXER
When the Builder reports an error, classifies it (dependency, compilation, linking, architecture, configuration, missing tools, …), retrieves relevant few-shot examples, and proposes either shell remediations or source patches that are re-applied through a validated patch pipeline.
SUPERVISOR
Verifies agent outputs against known anti-patterns (hallucinated commands, unsafe operations), routes the next step, and escalates when the system detects a loop, repeated error category, or budget exhaustion.
SUMMURIZER
On success, writes a Markdown porting recipe and persists the build to a recipe cache so that future requests for the same package short-circuit the LLM pipeline entirely.
The project further includes a lightweight few-shot memory system. Curated examples for each agent role live in JSON files in the repository; successful porting runs can append novel patterns back into this store under file-lock-protected I/O. This delivers measurable accuracy improvements without requiring a vector database or any external infrastructure.
A separate recipe cache records every package that has been successfully ported, keyed by package and sandbox profile, so that a project ported once can be re-built deterministically by anyone without re-invoking any LLM.
- Log in to post comments
added the initial proposal.
Submitted by Akif Ejaz on Tue, 06/09/2026 - 03:26
added the initial proposal.