Usage-only stream memory and shutdown

Problem

An OpenAI-compatible endpoint can keep a streaming request open while sending only usage metadata. This is unusual but valid enough for the SDK parser to continue yielding chunks.

Two layers retained every converted response for the lifetime of the stream:

• the OpenAI conversion pipeline kept all responses to find the pending finish chunk;
• the logging wrapper kept all responses to build one consolidated response at the end.

With a sustained metadata-only stream, both arrays grew without bound even though each new chunk replaced the same logical usage value. VP mode changes rendering but does not participate in either retention path.

The /quit action also only requested that background memory work stop. It did not abort the active model request, so the stream could continue allocating during the exit window.

Change

The OpenAI pipeline now keeps only its existing pending finish response. No history is needed to merge a later usage chunk.

The logging wrapper keeps responses that contain candidate parts or a finish reason, plus the final metadata response. This preserves the consolidated output and latest usage while making metadata-only retention constant-size.

The quit action calls the existing request cancellation path before requesting client shutdown and running exit cleanup. Literal /quit and /exit input also bypass the message queue while a response is active.

Scope

This does not add a new stream timeout, change VP behavior, or cap real model output. Streams that produce user-visible content still preserve that content for telemetry and optional OpenAI interaction logging.

Verification

• unit coverage for bounded metadata-only logging retention;
• unit coverage that /quit cancels an active request;
• existing finish/usage and duplicate tool-call stream tests;
• interactive reproduction against a local SSE server that returns a 429, an empty successful response, and then an unbounded usage-only stream;
• memory sampling and quit-latency evidence with VP enabled and disabled.

Local result

With a 512 MiB heap limit, the old build reached 515 MiB and crashed with VP enabled. The fixed build stayed between 101 and 131 MiB for 177 seconds while processing 13.57 million usage-only chunks. With VP disabled, it stayed between 101 and 130 MiB for 116 seconds while processing 9.62 million chunks. In both fixed runs, entering /quit closed the active server-side stream and exited the CLI normally.