Skills
Sandbox credential masking in Claude Code now decodes JWTs and re-signs AWS SigV4 — agents can call authed services without seeing the secret
Claude Code 2.1.224 (Aug 7) added structured masking to the sandbox: `extract` with `onExtractNoMatch` pulls a value out of a structured env var, `decode: "jwt"` with `maskClaims` redacts named claims inside a token rather than blanking the whole thing, and `awsPairs`/`sigv4` re-sign AWS requests on the way out. The SigV4 and JWT paths require `network.tlsTerminate`, since the proxy has to see the request to re-sign it. This closes the usual gap where you either hand an agent a live credential or lose the tool entirely — the agent gets a working request while the raw secret stays outside its context and transcript.
↳ Follow the thread