Impact check

Repo facts

This skill needs the facts below. Look each one up in the repo's CLAUDE.md (or AGENTS.md) and the files it imports or points to (of a repo-facts file, only this skill's section: sed -n '/^## impact-check$/,/^## /p' <file>), then in local.md next to this file. If neither has it, ask the user once, then suggest the one line to add to CLAUDE.md so the next run finds it.

Purpose

Find what a change could break outside its lines, and back the verdict with code that ran, not a persuasive writeup. Listing callers is not the deliverable. The value is in breakage a symbol search cannot see.

Evidence labels

Label every claim the verdict rests on.

• run: a script or test exercised the real code, with the library version the app ships. Its output is pasted.
• read: a cited file:line (or the library's source) and a traced path showing the failure cannot be reached.
• assumed: neither. Allowed, but labelled, and never presented as settled.

A verdict with an assumed premise says so in its first line.

Steps

1. Read the change. Take the diff: staged, a branch, or a PR (gh pr diff <n>, gh pr view <n>). Note what changes implicitly: defaults, ordering, timing, types, formats, error paths.
2. Name the premise. It is the one or two facts that make the change safe if they hold, for example "the removed cache key is never read after startup". Most risky-looking changes reduce to one. Spend the effort here, not on a long list.
3. Look past the symbol search. It misses:
   • serialized forms: API JSON, DB columns, cache keys, queue messages, files, URLs;
   • other services or languages that read the same data;
   • feature flags, environment and config;
   • lifecycle and timing: async ordering, teardown, retries, render semantics;
   • the library at its lockfile version and any local patch: read its source in node_modules or the equivalent;
   • generated code, locale, timezone, platform.
4. Prove the premise. Write a throwaway script at the Repo facts location, outside the repo. It imports the real module or the same library version and fails loudly if the premise is false. Run it, paste the command and output, then delete it. If that is costly, label the premise assumed and say what would prove it.
5. Weigh each risk. Give how it breaks, file:line, likelihood, cost, and the check that would catch it. List what was checked and found fine separately, one line of reason each. Cite only code seen in this run, report an empty search as such, never invent a caller or an API.
6. Second opinions, only on request and for a wide change. Run 2-3 read-only general-purpose subagents with model: "sonnet" and effort: "medium" on one question, merge and note where they disagree. Agents without subagents (Codex, Antigravity) skip this.

Output

Change

What it does, including the implicit part.

Premise

The fact or facts, each one's label, and the pasted proof.

Risks

Real ones only, with the step 5 fields.

Cleared

What was checked, and why it is fine.

Before merging

The smallest test or repro that fails on the real risk, with the proof script kept to become it.

Strip secrets and private data before posting anything outside the conversation.

If local.md exists next to this file, read it; where it disagrees with this file, it wins.