Module redact 

Secret redaction for outgoing tool output.

winx pipes shell output, file contents, and saved memory straight into the model’s context - env, cat .env, CI logs, etc. This scrubs high-confidence credential patterns from any text leaving the server (and from persisted ContextSave memory) before it reaches the model or disk.

Only named, low-false-positive patterns are matched (provider key prefixes, JWTs, PEM private-key blocks, user:pass@ URLs). There is no entropy heuristic, so ordinary code is never mangled. Set WINX_NO_REDACT=1 to disable entirely (e.g. when you knowingly need the raw value).

A redacted span becomes [REDACTED:<rule>] - the rule name only, never any part of the secret.

Functions

redact

Scrub credential patterns from text. Returns Cow::Borrowed (no allocation) when nothing matched, so the common case is cheap. The regex crate is linear-time, so this is safe on adversarial/large input.

redact_json

Recursively redact every string inside a JSON value (for structuredContent).