Test streaming and tool calls separately

Streaming is more than visible text

First confirm streaming support for the current model and protocol. Use a short, non-sensitive prompt to test event parsing. Network chunks do not necessarily contain complete events; buffer and assemble data according to the protocol rather than treating every read as a complete JSON object.

Record time to first content and total completion time. Test normal termination, user cancellation, disconnection, and service errors. If usage arrives in a separate event, verify receipt. Missing usage is unknown, not zero consumption.

Begin with a harmless tool

Tool calls require support from the model, interface, and client together. Start with a local function returning fixed test text. Give it no access to payments, email, file deletion, or real database writes. A model-generated call is a request to validate, not permission to execute arbitrary code.

• Validate the tool name, argument types, required fields, and allowed values.
• Preserve call identifiers and return results using the current protocol.
• Limit call counts, execution duration, and result size.
• Execute only complete, validated arguments; repeated events must not duplicate side effects.

Record explicit pass conditions

Test streaming text and non-streaming tool calls individually before combining them. Then inspect partial arguments, multiple calls, continuation after results, termination reasons, and final output. One successful example does not establish complete compatibility.

After interruption, preserve application task state and identify completed steps. Offer understandable results and a recovery path without silently repeating execution. Repeat affected checks after client upgrades, model changes, or protocol adjustments, recording passes, failures, and untested cases.