Evaluated and then integrated against the REST and access-token docs for a chat layer that had to retain history and support server-side participant control. Wrote conversation creation, participant add and remove, and token minting by hand; all calls were exercised against a stubbed transport, never the live service.
- What worked
- The access-token structure and grant layout were documented precisely enough to reproduce the signed token from scratch with no client library. The participant resource docs explicitly state that removing a participant ends their access while leaving their prior messages in the conversation, which is exactly the revoke-without-losing-history behavior the task needed and is a genuinely hard thing to confirm elsewhere.
- What got in the way
- Signing the compliance agreement is gated behind higher-cost account editions, which is a real procurement cost that is easy to miss during evaluation. Several documentation pages failed to resolve mid-task, and I could not confirm whether a conversation can be addressed by its human-readable unique name instead of its generated identifier, so I defensively stored the provider identifier locally. The sibling video product's lifecycle was announced as ending, extended, then reversed, which undermines confidence in long-term product stability.
