sg-process-check

Installation
SKILL.md

/sg-process-check — Diff-Driven Process Behavior Simulation

The backend counterpart of /sg-visual-run. Where the visual lane drives the browser to confirm a change in the UI, sg-process-check simulates the running process to show how a change affects its behavior — no browser.

The default is to run the code "in its head." It reads the diff and the code paths it touches, builds a model of the process (pipeline stages, data flow), and traces representative inputs through the OLD and NEW code by reasoning — predicting how the observable behavior differs. It can optionally really execute the cheap parts to anchor that reasoning, but it never needs to boot the stack. That is what makes it work on a 5-container app: the floor mode requires no infra.

It is scoped to the diff of the module you're working on (not the whole repo). Its oracle is before/after: the previous version is the reference. The question is not "is this correct in the absolute" — it's "did the observable behavior change, and was that intended?" The human decides. This is the behavior-level twin of sg-visual-fix's before/after screenshots.

⚠️ Reasoned ≠ measured — the honesty rule (authoritative)

A simulated trace is a prediction, not a measurement. Never present predicted behavior as observed. Every observation carries an evidence tag — reasoned (with confidence + assumptions) or measured (really executed). If a measured result contradicts the reasoned one, the measurement wins and the surprise is flagged. Anything you could not trace with confidence goes to uncovered — never a confident guess. The whole value is the human judging a clearly-labelled delta — not a confident-sounding guess dressed up as fact.

Hard rule — observe, never fix. Reports what the process does; never edits source. Remediation stays with sg-code-audit / sg-visual-fix. Even a one-line obvious fix — report it, do not edit. Even if the user will clearly want it fixed. Even if the fix is faster than describing it. Report, never edit.

Model guidance: use the strongest reasoning model available for --mode=reason (the reasoning IS the product); a fast general-purpose model suffices for hybrid/execute, where most of the work is mechanical orchestration.

Modes — a fidelity spectrum

Installs
1
GitHub Stars
7
First Seen
2 days ago
sg-process-check — bacoco/shipguard