[pcbeingused333]

Fixes the streaming proxy ending every OpenAI-format response with stop_reason: end_turn, even when the model asked for a tool call.

Cause

SSETranslator.processChunk handles finish_reason by emitting a message_delta with the mapped stop reason (tool_calls -> tool_use). When the upstream then sends [DONE], it emitted a second message_delta with a hard-coded end_turn. Anthropic-format clients take the last message_delta, so tool_use was overwritten and the tool call was treated as a finished turn.

Receipt (event sequence produced on main for an upstream stream ending in finish_reason: "tool_calls" followed by [DONE], as asserted by the new test):

event: message_delta  {"delta":{"stop_reason":"tool_use",...}}
event: message_delta  {"delta":{"stop_reason":"end_turn",...}}
event: message_stop

Fix

Track whether a stop reason was already emitted and, on [DONE], only send the fallback end_turn message_delta when none was. message_stop is still always sent. The case where [DONE] arrives with no finish_reason keeps the old fallback.

Tests

New test/sse-translator.local.mjs (added to npm test): tool_calls finish keeps a single tool_use; normal stop gives a single end_turn; [DONE] alone falls back to end_turn. Two of the three fail on main; all pass with the change.

---

This PR was prepared with the help of an AI assistant, and the author is responsible for the change.

[VickyXAI]

Thanks for the careful write-up and the tests. The bug you found in SSETranslator is real, but it turns out that class was never wired in: no file in the repo imports src/proxy/sse-translator.ts, in any commit of its history. The proxy pipes the gateway's stream through as-is and reads usage from its message_delta events itself, so the double message_delta could not reach a client.

Rather than fix code that never runs, we removed the file in 1c108dd, so I'm closing this PR.

One note in case you take on similar work: keeping the message_delta sent at finish_reason would also have dropped the trailing usage chunk that OpenAI-style streams send after it. Emitting the stop reason once, at [DONE], keeps both the stop reason and the final usage.