Every AI coding session ends with the same awkward moment. The agent announces the task is complete, the tests pass, and the code is ready. Maybe it is. But the agent saying so is the same agent that wrote every line, made a hundred small judgment calls along the way, and quietly convinced itself of each one. Asking it whether its own work is good is a bit like letting a student grade their own exam.
On October 28, 2026, MiniMax created a public repository called MCode Stack, and added install entries for Claude Code and Codex the next day. What is inside is not a new model and not a prompt collection. It is the actual working habits of the team that builds MiniMax Code—their skills, checklists, and daily workflows, published as files any compatible agent can load. The most interesting piece answers the grading-your-own-exam problem directly.
The clean-context reviewer
The skill is called independent-review, and its rule is radical in its simplicity: the reviewer sees exactly three things. The diff. The requirement the change was supposed to satisfy. And the real test output, including failures and skips.
That is the whole list. The reviewer does not get the author's conversation, its plan, its intermediate reasoning, or its self-assessment. The skill's own explanation of why is worth quoting in spirit: a reviewer that inherits the author's conversation tends to follow the same reasoning and miss the same things. So the reviewer's job is deliberately reframed. It is not asked "does the author's logic hold up?" It is asked "does this change meet the requirement?" Those are different questions, and the second one can only be answered honestly by someone who was not there for the writing.
The skill goes further than the setup. The reviewer is pointed at specific traps: assumptions the code relies on but never checks, requirement items the diff silently skips, tests that pass without actually exercising the requirement—tests written by the same agent share its blind spots—and behavior changes outside the stated scope. Findings come back in a fixed structure with a location, a reason, and a severity label of blocking, should-fix, or nit; anything without a location and a reason is treated as an opinion, not a defect.
And then, the part that saves you from over-trusting the reviewer: back in the main session, you do not apply its findings wholesale. Each one gets verified against the real code first, ideally with a failing test, and rejected findings are recorded so the next review does not raise them again.
What else is in the stack
The repository is organized as personal stacks—one per team member. The main one, ronny, from the MiniMax Code R&D lead, holds about fifteen skills grouped by moment in the working day. There is a checklist for deciding whether a task deserves a subagent at all—the companion article to independent-review is bluntly titled "More subagents isn't better." There are skills for splitting work into small reviewable pull requests, for auditing what your tools and injected context cost on every request, for keeping long sessions from going dull, for deciding how long a shell command should be allowed to run, and for handling hotfixes without panic-merging. A second stack adds a skill for spotting the loops where you keep serving as a manual step for your own agents.
Everything is in the common SKILL.md format, so most of it loads into MiniMax Code, Claude Code, Codex, or any other agent that supports skills. Installation is a plugin command or two—claude plugin install ronny --marketplace MiniMax-AI/mcode-stack—or, for a single skill, copying one folder into your agents' skills directory.
What this is not
Three clarifications before anyone over-reads the news. This is not MiniMax Code itself being open-sourced; the tool lives in a separate repository, and this one contains only workflows and skills. It is not a new model, and nothing here is free compute—whatever model your agent runs on still bills you the usual way. And while the repository is MIT-licensed overall, the engineering review skills are adapted from Google's Engineering Practices documentation and remain under CC-BY 3.0 with attribution, which matters if you plan to republish them.
One more honesty note: we have not installed or run these skills ourselves. A published checklist is not evidence that it catches more bugs—only that a serious team works this way every day and thought it worth sharing.
Why it is worth your attention anyway
The AI coding conversation has been obsessed with generation: bigger models, longer context, more parallel agents. MCode Stack is a bet that the next gains come from the layer around the agent—when to make it stop and check its own work, who reviews, and what the reviewer is allowed to see. The independent-review skill in particular names a failure mode most of us have felt but rarely formalized: the second opinion was never independent, because it read the first opinion's diary.
You can argue about whether an AI reviewing an AI counts as real review. But a reviewer with clean eyes, a fixed checklist, and a requirement to point at beats the alternative that most teams are actually running today, which is the author grading its own exam and everyone calling it done.