encrypted-fields

Encrypted Fields

Wick has a built-in encrypted-fields layer that lets credentials flow between an LLM and a connector without ever appearing as plaintext in the LLM's context window or in the audit log. Treat this as the default for any sensitive value.

When this skill activates

• Adding a new connector that takes a credential, API key, or token (Configs field).
• Adding an operation whose response includes a credential, OTP, session token, JWT, refresh token, or any value the LLM should pass forward but not learn.
• Reviewing an existing connector for plaintext-leak risk.
• The user mentions wick_enc_, "encrypt this value", "hide this from the LLM", or asks why a credential is showing up unmasked.
• Adding a sensitive Input field that the LLM may receive in a previous response and replay in a later call.

If the work has nothing to do with credentials or sensitive output, skip this skill.

How the layer works (in 30 seconds)

1. Tag the field secret in the Configs / Input struct.
2. connectors.Service.Execute automatically:
   • decrypts any wick_enc_<...> token in the input map (and the stored configs map) before calling ExecuteFunc, so your code only ever sees plaintext.
   • masks every occurrence of those plaintext values in the marshaled response with wick_enc_<token>, so the LLM never sees plaintext.
3. The audit log stores wick_enc_ (pre-decrypt) in request_json and wick_enc_ (post-encrypt) in response_json. Retry from history works because Execute re-decrypts under the retrier's key.
4. Keys are per-user — HKDF(master_key, salt=user_uuid, info="wick-enc"). A token issued for user A cannot be decrypted under user B's session.

The token format is wick_enc_<base64url(nonce ‖ AES-256-GCM(plaintext, nonce, derived_key))>. AES-256-GCM, 12-byte random nonce per encrypt, 32-byte derived key.

Recipes

Recipe 1: a Configs credential (most common)

type Creds struct {
    Endpoint string `wick:"required"`
    APIKey   string `wick:"required;secret"` // ← auto-masked in every response
}