Irrlicht: Skill for Claude Code

Review the Working Diff

/ir:code-review [effort] [base]

Argument	Values	Default
effort	low · medium · high · xhigh · max	medium
base	any git diff spec	origin/main...HEAD

Both are positional and both optional: /ir:code-review, /ir:code-review high, and /ir:code-review low origin/main...HEAD are all valid. A bare effort word is an effort, never a PR number — this skill reviews a local diff and never takes a PR argument.

Why this skill exists

The built-in /code-review is marked disable-model-invocation: the Skill tool refuses it and it cannot be reached from Bash either. It is a human-triggered path. That left ir:exec's review gate unrunnable (#1205), so this is the repo's own reviewer — same job, agent-invocable, and aware of conventions in AGENTS.md that a general-purpose reviewer wouldn't check.

It is not a replacement for the real thing. For a large or risky diff, a human running /code-review <tier> origin/main...HEAD — or /code-review ultra for the cloud multi-agent pass — is still the stronger review, and recommending that is a legitimate outcome of this skill.

1 · Resolve the diff

git fetch origin main                      # refresh the base before diffing
git diff --stat <base>                     # scope
git diff <base>                            # the review surface

This skill already defaults to origin/main...HEAD. Keep it — inside a worktree, both of the defaults you'd otherwise reach for by hand mislead: the local main ref is stale (never updated when other PRs merge, so main...HEAD drags in already-merged hunks), and once the branch has been pushed with git push -u, @{upstream}...HEAD is ~empty (upstream now points at the pushed branch, so it reviews nothing). origin/main...HEAD after a fetch is the honest surface.

Review only what the diff changed, plus whatever the change puts at risk. An unrelated pre-existing wart is not a finding.