Will SGLang 0.5.20 break your Responses API workflow?

SGLang’s v0.5.20 release notes make response persistence an upgrade decision for anyone serving /v1/responses. The endpoint no longer retains results in memory by default. The server must start with --enable-response-store for the storage-dependent paths.

The compatibility question is what your client does after it creates a response:

Client operation With the default, store-off configuration
Retrieve a response later Returns HTTP 400
Chain a turn with previous_response_id Returns HTTP 400
Send a background request Returns HTTP 400

The release notes also call out background=true with store=false as rejected. These are documented v0.5.20 behaviors, not results from a test of a particular client.

Deployment topology changes the answer. A non-PD service can opt into the response store with --enable-response-store if those workflows are required. The release notes say a PD deployment cannot enable that store and rejects the stateful features at admission. For a client that depends on retrieval, chaining, or background requests, enabling a flag is therefore not an available PD migration plan in this release.

Before the upgrade, trace one real client journey and record whether it reads an earlier response by ID, passes previous_response_id, or requests background execution. Test each operation the client actually uses against a v0.5.20 service with the intended startup flags and topology. Check the response status and the client’s error path, including any fallback that assumes a stored response exists. If one of these operations is essential, choose a topology that can enable the store or redesign that client workflow before moving it to PD.

This checklist is a migration recommendation based on the release note. It does not establish how unrelated endpoints behave or claim a successful upgrade for any deployment.