feat(a2a): read agent credentials from a file or a command

[edenreich]

Summary

Two new credential sources in the auth block of agents.yaml, next to token_env and oidc, plus an optional scopes list on oidc, and three examples that put an agent behind Google, Microsoft Entra ID and Amazon Cognito.

auth:
  token_file: /var/run/secrets/kubernetes.io/serviceaccount/token
auth:
  token_command: [gcloud, auth, print-identity-token, --audiences=https://billing.example.com]

• token_file is re-read on every request, so a token rotated in place by the kubelet, spiffe-helper or Vault Agent is picked up. It is the shape a workload gets from its platform, with no provisioned secret. This is also how kagent carries its workload credential.
• token_command runs a cloud CLI or a broker without a shell and reuses its stdout until the token expires (the exp claim when it is a JWT, five minutes otherwise). It is honoured from ~/.infer/agents.yaml only: a project file arrives with a clone and the agent card fetch runs without approval. A failing command is reported by name and exit status, never by its output.
• oidc.scopes is requested on top of the card's scopes. Entra requires api://<app>/.default on client credentials and the ADK card declares none.

Why

The existing OIDC handling is spec-correct and ahead of the official A2A SDKs, which only attach a pre-acquired token. What it lacked is a way to use the token a platform or cloud already mints: Google issues ID tokens from the metadata server (no client-credentials grant), Azure managed identity and Entra Agent ID mint tokens by token exchange, Kubernetes and SPIRE hand workloads a token file. Both additions are a few lines in auth.go and add no dependency.

Examples

examples/a2a-auth-gcp, examples/a2a-auth-entraid and examples/a2a-auth-aws mirror the gateway repository's auth-gcp, auth-entra and auth-cognito examples: a mock agent trusts the provider as its issuer, get-token.sh fetches the provider's token, and a tokenless mock model drives the delegation. GCP and AWS use token_file, Entra runs the client-credentials grant from the oidc block.

Verification

• task precommit:run passes, go test ./... passes.
• examples/a2a-auth: all five scenarios pass with this build.
• examples/a2a-auth-gcp wiring against a static-token mock agent with a hand-written .token: the task completes, a wrong token is reported as an authentication failure.
• token_command from ~/.infer/agents.yaml authenticates infer agents status, the same block in a project .infer/agents.yaml is refused with token_command is only read from ~/.infer/agents.yaml.
• The cloud examples themselves need the provider's account and were not run against a real issuer.

Docs

docs/agents-configuration.md, docs/a2a-connections.md and docs/examples.md are updated here.

Docs ticket: inference-gateway/docs#992

Follow-up, not in this PR: native auth.spiffe via the SPIFFE Workload API, and switching examples/a2a-gateway-auth from a pasted token to auth.oidc.

[inference-gateway-releaser]

🎉 This PR is included in version 0.227.0 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀