guardrail-designerlisted

# guardrail-designer Composes input/output **guards**: ordered validators run at a boundary, each with an explicit `on_fail` action. A guardrail is not an eval — it runs on live traffic and decides what happens *right now* when a check fails. ## When to use - Stripping or blocking PII (emails, secrets) before it is logged or returned. - Enforcing a response shape (valid JSON, max length) before downstream parsing. - Refusing or rewriting unsafe content instead of passing it through. - Validating untrusted *input* before it reaches the model. ## Run it ``` python -m harness.guardrails check --text "reach me at jane@example.com" --boundary output python .claude/skills/guardrail-designer/scripts/run.py check --text "x" --boundary input ``` ## The on_fail contract Every validator declares what to do when it fails — this is the heart of guardrail design: | `on_fail` | Effect | |-----------|--------| | `fix` | Deterministically repair (e.g. truncate to max length). | | `filter` | Remove the offending span (e.g. redact emails). | | `reask` | Signal the caller to regenerate (sets `reask=True`). | | `refrain` | Replace the whole output with a safe refusal. | | `raise` | Hard-stop with an error. | | `noop` | Record the failure but pass the text through. | ## Method 1. Decide the boundary: **input** (defend the model) or **output** (defend the user/logs). 2. Order validators cheapest-first; a `refrain` short-circuits the rest. 3. Choose `on_fail` per validator by b