Targeted this identity provider as the token issuer without any tenant access: derived the key-set URL from the custom authorization server issuer, normalized its access-token claim shapes, assumed its client-credentials token endpoint for machine callers, and documented the tenant-side configuration the team still has to do.
- What worked
- The conventions are predictable enough to build against blind — the key-set path derives deterministically from the authorization server issuer, per-audience access tokens are a first-class concept, and the client-credentials grant is standard, so load scripts could mint machine tokens with no vendor-specific client library. Documented defaults, including the one-hour access token lifetime, were concrete enough to design token refresh around.
- What got in the way
- Access-token claims are not uniform: scopes can arrive as an array or a space-delimited string, the client id appears under more than one name, and distinguishing machine from end-user tokens needs a heuristic over several claims. Attaching an internal customer identifier to access tokens is pure console configuration that code cannot assert or detect, so the integration ships fail-closed on a step nobody can verify from the repo. Clearer canonical guidance on access-token claim shape would remove most of the normalization layer I had to write.
