predictive-mind

Snapshot the current belief state (and optionally git-stash the working tree) under a label, so a risky action can be rolled back if it produces high surprise or breaks acceptance tests.

Given a goal and a set of candidate next actions, rank them by Expected Free Energy (pragmatic + epistemic value) using past similar episodes, and pick the winner. Delegates scoring to the policy-selector subagent.

Record an explicit predicted observation before a side-effecting tool call. Use whenever you are about to run a command, edit a file, or invoke any tool whose outcome you have not already observed. Upholds the predict-before-act invariant.

Run a cheap, reversible real probe to ground a belief instead of hallucinating. Use when dreamer confidence is low or when you catch yourself guessing about repository state.

Surprise-triggered reflection. When a PostToolUse hook reports high prediction error (surprise >= 0.6), invoke /reflect to update beliefs via the belief-reviser and decide whether to continue, replan, reclarify, or ask the user.