[WaylandYang]

Maya-Kid's #768, rebased onto dev after #764 turned the streaming buffer into bytes. The conflict was mechanical: keep the Vec<u8> buffer, add their ended and got.

What it does. chat_tools_stream_with and chat_stream_raw get the completion check that #757 gave only to chat_at_streaming: a frame carrying finish_reason, or data: [DONE], marks the turn finished, and reaching EOF without either yields Interrupted { got } instead of a Turn or a silent end. Interrupted is already retryable in transient().

Reverting their source and keeping their tests shows what the check is for:

---- tests::a_tool_stream_cut_before_its_finish_is_an_error stdout ----
called `Result::unwrap_err()` on an `Ok` value: [Turn(AssistantTurn { content: None,
  tool_calls: [ToolCall { id: "call_1", name: "lookup", arguments: "{\"name\":" }] })]

Without it the agent receives a tool call whose arguments are the fragment {"name": and has to work out for itself that the JSON never finished.

One consequence worth stating, since the description did not. api/chat.rs streams chat_stream_raw and on an error yields an error event and returns before conversations::append_message. So a cut stream now shows the reader the partial text followed by an error, and the assistant turn is not saved at all; before, a short turn was saved with a clean done. Losing half an answer is the right call — half a thing is not a thing anywhere else in this ledger — but it is a decision, not an inheritance.

Their branch could not be rebased from here, so this carries their commit with authorship intact and #768 closes with it.

30 utopia-llm tests pass, clippy across utopia-llm and utopia-server is clean, fmt clean.

🤖 Generated with Claude Code