runopsy-replay

Proposes a controlled re-run of part of a run. It never performs one.

Separating the proposal from the act is the whole point. A replay is the only thing in Runopsy that can change the world, so the risky decision — what gets executed again — is made by a person looking at a written plan, not by a scoring function that happened to rank a step highly.

Fail closed

The side-effect classifier only ever sees a tool's name. Arguments are stored as hashes by design, so it cannot know what a command actually did. A heuristic gate built on that has exactly one safe default: anything unrecognised needs human approval.

The asymmetry decides it. A false alarm costs one keystroke. A missed classification re-sends an email, re-charges a card, or deletes a branch a second time, and no quality of diagnosis afterwards undoes that.

What a plan will not promise

A plan states its own limits, because a replay whose caveats are invisible produces false confidence rather than evidence:

• No checkpoint at the chosen step — the session can fork but the working tree may not match, so a difference in outcome may come from the files rather than the change.
• Checkpoint and fork point at different steps — file state and message history would be restored to different moments.
• More than one variable changed — if the outcome differs, nothing says which change did it. Counterfactual validation needs one intervention at a time.
• A trace with gaps — the original conditions cannot be reconstructed from an incomplete recording.