evalslisted
Install: claude install-skill rterakedis/leanwheel-skills
# Evals Skill
**Goal:** Give the flywheel the *non-deterministic* half of verification the
Build & Test Gate can't cover, and a **persistent regression net** so a later
story can't silently revert an earlier story's behavior — the exact failure mode
that motivated the Build & Test Gate, now made cumulative.
**Token philosophy (Pro plan):** the default eval type is `command` — a shell
command with an expected result. **Zero model tokens.** It works identically for
`swift test`, `xcodebuild`, `pytest`, `npm test`, `playwright`, or `curl`. The
`judge` type (LM-as-judge) is **opt-in and token-flagged** — used only where
behavior is genuinely non-deterministic and no command can assert it.
---
## Eval set layout
```
docs/evals/
README.md ← format reference (scaffolded by /setup)
epic-{n}.md ← accumulated cases for epic n, one block per case
e2e-{area}.md ← retro-fit cases registered by /e2e-tests (backfill net)
```
Each case is a fenced block:
```
### EVAL {epic}.{story}-{seq} — {short title}
type: command # command | judge
origin: story {epic}.{story} AC{n} # provenance
enabled: true
run: swift test --filter CartTotalTests # command type only
expect: exit-0 # exit-0 | output-contains:"<s>" | output-matches:/<re>/
# judge type only:
# target: git diff -- Sources/Cart.swift # what the judge reads
# rubric: |
# - Empty cart returns 0, never nil
# - Discounts never produce a negative total
```
Case