fix(stream): skip [DONE] terminator for Claude SSE clients

[NomenAK]

Summary

Anthropic SSE streams terminate naturally on the event: message_stop frame — they do not emit an OpenAI-style data: [DONE]\n\n terminator. Emitting one to Claude clients (Capy, claude-cli, Anthropic SDK) causes parser errors and observable UI artifacts (assistant text rendered in the wrong panel, followups looping with retries).

Gate the [DONE] emission so it only fires for OpenAI clients (Chat Completions / Responses). Symmetric to the existing clientExpectsResponsesStream gate that already silences [DONE] for the Responses API.

Related Issues

• Closes #
• Related to #

Motivation

Captured artifact for a /v1/messages BYOK call to Claude Opus through OmniRoute :

event: message_stop
data: {"type":"message_stop"}

: x-omniroute-cache-hit=false
: x-omniroute-latency-ms=498
: x-omniroute-tokens-in=28328
: x-omniroute-model=claude-opus-4-7

data: [DONE]   ← does not belong on Anthropic SSE

data: [DONE] is not part of the Anthropic Messages SSE spec (docs). When the client encounters it after message_stop, the strict parsers in the Anthropic SDKs raise Unexpected event "[DONE]" and the client either drops the response or misclassifies the trailing text as a stream error, surfacing as : "the assistant's reply is rendered in the wrong UI section / retries the request".

The OpenAI Responses API has the same constraint — it terminates on response.completed, not [DONE] — and the codebase already has a clientExpectsResponsesStream gate for that case. This PR is the symmetric gate for Claude clients.

Changes

open-sse/utils/stream.ts :

• Introduce clientExpectsClaudeStream alongside the existing clientExpectsResponsesStream flag.
• Both [DONE] emission sites (post-stream end and post-error) now check !clientExpectsResponsesStream && !clientExpectsClaudeStream before writing the terminator.

Validation

• Manual : capture artifact pre/post — confirmed [DONE] no longer appears after message_stop on /v1/messages calls.
• OpenAI Chat Completions endpoint still receives [DONE] (regression check).
• OpenAI Responses endpoint still skips [DONE] (existing behavior preserved).

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request modifies the SSE stream utility to prevent emitting the [DONE] terminator for Claude-formatted streams, which avoids parser errors in specific SDKs. The reviewer suggests consolidating the logic for emitting the terminator into a single variable, shouldEmitDoneTerminator, to improve code readability and simplify the conditional checks in multiple locations.

[gemini-code-assist]

Use the consolidated shouldEmitDoneTerminator variable here to simplify the condition.

Suggested change

	if (!clientExpectsResponsesStream && !clientExpectsClaudeStream) {
	if (shouldEmitDoneTerminator) {

[gemini-code-assist]

Use the consolidated shouldEmitDoneTerminator variable here to simplify the condition.

Suggested change

	if (!clientExpectsResponsesStream && !clientExpectsClaudeStream) {
	if (shouldEmitDoneTerminator) {

[NomenAK]

Addressed @gemini-code-assist's review — refactored to consolidate the gate into a single shouldEmitDoneTerminator variable used at both [DONE] emission sites. Commit history kept single via amend. CI rerunning.

[gemini-code-assist]

Thanks for the update, NomenAK. Consolidating the gate into shouldEmitDoneTerminator is a much cleaner approach and significantly improves maintainability by ensuring consistent behavior across both emission sites. This looks good.

[diegosouzapw]

Thanks @NomenAK. I synced this branch with release/v3.8.0, added stream coverage proving Claude SSE clients do not receive a [DONE] terminator, and validated it locally with:\n\nnode --import tsx/esm --test tests/unit/stream-utils.test.ts\n\nI'll merge this through GitHub so the contribution is credited correctly.