{"name":"dev-jahn-autoresearch","owner":{"name":"ClaudePluginHub"},"plugins":[{"name":"dev-jahn-autoresearch","source":{"source":"github","repo":"dev-jahn/autoresearch-plugin"},"description":"Karpathy-style autonomous LLM research loop adapted for generic ML codebases. Ships two skills — setup and run — plus an `ar` helper CLI that owns launch orchestration, metric extraction, atomic checkpoint commit, and chain-mode transitions so the main Claude session stays context-light across hundreds or thousands of iterations. v0.3.0 drops the wandb/accelerate monoculture: metric backend is pluggable (wandb / tensorboard / log-scan / custom with auto-detection), distributed-framework resolution is auto-detected (accelerate/deepspeed/fsdp/ddp/lightning/none), a new `hydra` entry pattern renders a Hydra-override-style train wrapper, and `--checkpoint-glob` gives priority-0 control over checkpoint discovery for Lightning / plain-torch / custom layouts.","version":"0.3.0","strict":true,"keywords":["research","autonomous","training","ml","llm","autoresearch","karpathy","wandb","tensorboard","hydra","lightning","experiment-tracking"],"category":"data"}]}