zaxardery8011-design/soplint

soplint is a linting tool built not to check your code style but to check whether your AI agent actually keeps the agreements you made with it. The author spent a year running a personal AI agent continuously and found that code quality was rarely the issue. The real problems were behavioral: the agent would say it had verified something without verifying it, quietly reverse decisions it had just made, or learn a new rule in one session and drift back to old behavior in the next. The tool addresses this with three mechanisms. The first is a belief revision audit trail. Whenever the agent changes its mind about something it previously stated, it must write a structured log entry recording what it believed before, what it believes now, and why it changed. This gives you a greppable file of every time the agent reversed a judgment, which is much harder to hide than simply ignoring a rule written in a prompt. The second mechanism is a set of daily discipline checks run as a script. These checks scan your agent's memory files and instruction documents for specific failure patterns: a decision that was logged in a memory file but never propagated to the agent's main instructions file, a belief log that is not being written at all, memory files missing required metadata, and a memory index that has grown so large or cluttered that the agent cannot actually load all of it during a session. Each check exits with a non-zero code on failure so it can plug into a CI pipeline or a scheduled job. The third mechanism is a pre-action gate that runs as a hook before the agent executes shell commands. It parses the command using PowerShell's built-in syntax tree tools and applies two kinds of rules from an external JSON file: hard deny rules that block specific commands outright, and novelty gates that require the agent to confirm it checked for an existing tool before building a new one. The README is explicit that this is not an alignment solution. It is a regression test suite for known failure modes, each check backed by a real incident. The author's framing is that a lesson written to memory is a hope, but a lesson with a lint rule becomes a regression signal you can run every day. Requires PowerShell 7 and runs on Linux, macOS, and Windows.