CrowdControl.Backend.Credentials (crowd_control v0.2.0)

Credential rewriting shared by every remote backend that fronts an egress proxy.

This exists as one module, used by every backend, for the same reason CrowdControl.Backend.Shell does. Backend.Docker passes the rewritten env through the exec API's Env array; Backend.Kubernetes renders it into an env file shipped over the exec stdin channel. The transport differs, the rewrite must not: the whole documented purpose of apply_credentials/2 is that the real provider key is removed rather than overridden, and a second, subtly different copy is precisely how one of them ends up merely overriding it.

Summary

Functions

apply_credentials(env, config)

Rewrite the CLI's credential env for egress-proxy mode.

Functions

apply_credentials(env, config)

@spec apply_credentials(
  map(),
  keyword()
) :: map()

Rewrite the CLI's credential env for egress-proxy mode.

When :proxy_url is set, the sandbox is pointed at the proxy and given a per-session token instead of the real key:

ANTHROPIC_BASE_URL = proxy_url
ANTHROPIC_API_KEY  = session_token

The real :api_key is removed, not merely overridden. Leaving it in place would hand the sandbox a working provider credential even though every request is nominally routed through the proxy — which defeats the entire point of running the sandbox on an isolated network. That silent double-injection is the failure mode this function exists to prevent, and docker_test.exs asserts it directly.

With no :proxy_url, the env is returned untouched.