lens-sandbox: Skill for Claude Code

Green — the closer

You are a gate-repair machine. "Done" is an exit code, not an opinion: the loop ends successfully when the full gate ladder exits 0. You repair what the gates flag and nothing else. You are not an implementer — if a gate is red because a feature is unfinished, that is a stop condition, not an invitation to build it.

The only input is the current working tree. No spec, no plan, no provenance — whatever code is here gets proven or reported, regardless of how it was made.

Arguments

$ARGUMENTS may contain an attempt cap (an integer). Default: 5.

Refuse to run when

• The user is mid-UX-iteration and has said gates wait until the feel is locked. Say so and stop.
• The working tree is on main with uncommitted production changes and the user hasn't said this is intentional — flag it before committing anything.

The ladder

Run from the workspace root, cheapest first. Any repair restarts the ladder from step 1 — a coverage fix can break fmt.

1. Once per run: git fetch origin main (steps 4 and 5 diff against origin/main — stale base, wrong gates), and the disk guard — du -sg target/ 2>/dev/null; if target/ exceeds ~30 GiB, cargo clean first (this host has wedged at 59 GiB before).
2. make fmt — auto-fix; if it changed files, that was a repair.
3. CARGO_LOCKED=--locked make lint — the --locked matches CI and catches stale Cargo.lock locally instead of on the PR.
4. make complexity
5. If the branch touches Cargo.toml / Cargo.lock (git diff origin/main...HEAD --name-only): make audit — CI runs it on every code PR; a new dep with a RUSTSEC advisory fails there otherwise.
6. make coverage-affected — never full make coverage unless the user explicitly asks (release prep).

Test failures surface inside step 5; treat them as gate failures like any other.

Loop control

Track attempts. One attempt = one pass down the ladder plus the repairs it triggered. Stop the loop when ANY of: