MiniMax open-sourced MCode Stack, the agent skills and checklists its own coding team uses to build MiniMax Code
AspireAI community · MiniMax on GitHub
MiniMax published MCode Stack today, an MIT-licensed repository of the workflows and agent skills its MiniMax Code team uses internally. The README is explicit that it is not a prompt collection and not a list of best practices: it is the working material of people shipping code every day, covering how to keep AI-written code at a high bar, when a task should go to a subagent, how long to wait on a long command, how to stop long sessions from going dull, and what belongs in context on every turn.
The structure is a plugin marketplace rather than a flat file dump. The repository carries 130 files, 18 SKILL.md skill definitions across two contributor stacks (ronny and jiachen), a stack template for third parties, and skills covering code review, review comments, small PRs, handoff docs, trajectory review, tool-budget audits, context-tax audits and subagent-split checks. Installing one skill means copying a directory into ~/.agents/skills; installing the whole stack means pointing MiniMax Code's plugin marketplace at the repository. The repo was created on October 28 and already had 156 stars.
Two caveats worth knowing before you adopt it. The six engineering-* skills are adapted from Google Engineering Practices and stay under CC-BY 3.0 rather than the repo's MIT license, so the licensing is mixed by design. And this is a self-published artifact from one team's own workflow: it tells you what MiniMax's team does, not that those practices are the right ones for your codebase.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.