/qa-regress

A change is innocent until the code it touches says otherwise. This skill weighs exactly the distance between the regression base and the present, runs the repo's own tests over that distance, and pins whatever changed without a test to hold it. It runs in a fresh session and re-derives everything — a verdict that leaned on the Dev session's memory is not a verdict.

Two commitments hold absolutely. Source code is never edited under any circumstance — not to fix a caught regression, not to make a written test pass, not "just to tidy"; the moment QA improves the code it verifies, it has become a second implementer and its verdict is worthless. And the only things this skill writes into the repo are tests in their own commit and its report.

Independence first

Before anything runs, re-derive the context a fresh session owes itself:

• Read the target repo's config under docs/agentic-skills/. stack.md above all: its pinned test command is the only runner this skill uses, and its recorded test layout says where tests live. No step assumes a language, framework, or runner the repo did not pin. Absent config means derive both from the repo's own manifests and test files — say which evidence stood in.
• Read docs/agentic-skills/tracker.md for how this repo's tracker takes a comment, in case the run was invoked against a ticket. Absent means no ticket comment happens.
• Read GLOSSARY.md when present and name findings in its vocabulary. Absent means proceed silently.
• Read docs/adr/ for every record plausibly touching the changed area — a regression is sometimes a deliberate, documented decision, and that distinction changes a finding's grade, not its truth.

Establish the regression base

The regression base is the point in history this run compares against.