Security boundary

Keep credential redemption outside model context

s-gw reduces accidental credential exposure by separating an agent request from local credential redemption. It is a preview control layer, not a replacement for endpoint or provider security.

Protected path

A request crosses the boundary. The raw value does not.

  1. 01
    Agent request

    The coding agent supplies a typed handle, command, environment binding, working directory, destination, and reason.

  2. 02
    Local decision

    A policy or user approval decides whether that bounded request may execute.

  3. 03
    Local execution

    The credential is resolved inside the local boundary and injected only into the approved child process.

  4. 04
    Sanitized return

    Known raw values are replaced before output returns to the caller, and the event is recorded locally.

Intended guarantees

What the current boundary is designed to enforce

  • Agent-facing interfaces use handles and metadata instead of raw values.
  • Execution must match the credential's allowed command or SSH policy.
  • Reusable approval is bounded by stored request attributes and expiry or login session.
  • Approved environment commands start with a cleared child environment plus allowlisted values.
  • Approval, execution, denial, failure, and policy activity are recorded locally.

Residual risk

Approval is not containment

A malicious approved executable can still misuse a credential. s-gw cannot defend a compromised operating-system account, kernel, credential store, or process, and output scanning cannot detect every transformed derivative of a secret.

Keep allowed commands narrow, review destinations and arguments, prefer low-privilege credentials, and keep independent provider-side controls enabled.

Required reading

Review the full threat model

Do this before relying on s-gw for sensitive workflows.

Read the threat model