marin: Skill for Claude Code

Simplify a branch or diff

Apply this workflow when cleanup of a branch or diff is explicitly requested. Use review-pr for independent findings without code edits. manage-pr owns publication and feedback after the cleanup is ready.

Load The Policies

Read these before reviewing:

• root AGENTS.md
• root TESTING.md
• docs/dev-guide/writing-style/ai-writing-donts.md
• the nearest module AGENTS.md and testing guide for each changed test

Read references/examples.md when the diff changes tests or prose.

Establish The Review Surface

Review the merge-base diff. Do not substitute commit history or the latest commit.

UPSTREAM=$(git rev-parse --verify origin/main 2>/dev/null || git rev-parse --verify main)
BASE=$(git merge-base HEAD "$UPSTREAM")
git status --short
git diff --stat "$BASE"
git diff --check "$BASE"
git diff "$BASE"
uv run python .agents/skills/simplify-code/scripts/scan_diff.py "$BASE"

List every changed test, comment, docstring, Markdown file, commit message, and draft PR title/body. Read the surrounding production code and existing tests. The scanner includes untracked files and returns candidates; inspect each one in context.

Gate Tests By Behavioral Value

For every changed test, finish this sentence:

This test fails when ___ user-visible or system behavior is wrong.

Delete or rewrite the test when the blank names an implementation detail, wording choice, supplied fixture value, or language guarantee.

Reject these patterns aggressively:

• config in, same config out: a constructed or lowered field equals the fixture
• constructor assignment, enum value, type, attribute, registration, or import exists
• exact log, help, error, command, or status prose that no machine consumes
• private state, helper call count, or assert_called_once_with
• production logic copied into the expected value
• is not None, does not raise, empty smoke tests, or permanent skips
• a large matrix whose rows exercise the same branch
• golden output produced by the implementation under test