Logging safely
Use redactForLog to keep keys and tokens out of your logs.
A secret that reaches a log is a secret that has leaked. redactForLog takes any value headed to a logger, and replaces anything shaped like a model provider key or a Baseera secret key with [redacted].
import { redactForLog } from "@baseera/core";
try {
await baseera.kb.upsert(input);
} catch (error) {
logger.error(redactForLog(error));
}What it does
It always returns a string.
| Input | Result |
|---|---|
An Error | "Name: message", redacted. The stack is not included. |
| A string | The string, redacted. |
| Anything else | JSON.stringify of the value, redacted. If it cannot be serialized, the string [redacted]. |
redactForLog("Request failed for sk-ant-api03-abcdefghijklmnop");
// "Request failed for [redacted]"
redactForLog({ key: "sk_" + "a".repeat(64), user: "u_1" });
// '{"key":"[redacted]","user":"u_1"}'redactSecrets(text) is the same for a plain string, and REDACTED is the replacement text.
What it matches
- Anthropic keys (
sk-ant-and at least 8 more characters). - OpenRouter keys (
sk-or-). - OpenAI style keys (
sk-, withproj-orsvcacct-allowed, and at least 16 characters). - Google keys (
AIzaand at least 20 characters). - Baseera secret keys (
sk_and at least 16 hex characters). - Bearer tokens (
Bearerand at least 16 token characters).
What it does not match
redactForLog is a safety net, not a guarantee. It matches shapes. It does not match a user token (a JWT), a webhook signing secret (whsec_), or a secret of your own. Do not log those at all.
Rules to follow
- Never log a key, a user token, a webhook secret or a request header that carries one.
- Never put a key in state, in a URL, in an error message or in a response body.
- Log the error
codeof aBaseeraError, not the whole request. - Do not log tool inputs or outputs. They can hold your users' data.
The SDK follows the same rules itself: it never logs keys, tokens, or tool inputs and outputs, and the user token is sent only as a header.