Oct 23, 2026
Launcher for seeding Forgejo instances and orchestrating sandboxed LLM agent turns across interdependent repositories
A coordinator loop wakes one maintainer agent per repository on a fixed cadence, each turn executing branches, pull requests, reviews, and releases inside a Harbor sandbox while a separate evaluator scores downstream dependents the agents never observe. The resulting checkpoint-to-bundle pipeline supports controlled studies of emergent multi-agent software engineering but remains specialized tooling for research groups rather than a general developer platform.
Executable RL environment that lets agents diagnose and repair distributed PyTorch training failures inside sandboxed execution
The environment launches synthetic multi-process Gloo jobs, injects one of four fault families, captures trusted launcher records of batches and checkpoints, then re-executes every candidate patch inside a network-less gVisor sandbox so an external verifier can replay optimizer transitions and sample accounting. This creates a narrow but well-defined benchmark for LLM agents and RL methods that must produce repairs that are both executable and semantically correct rather than merely passing superficial checks.
