fix: stop streaming the raw reasoning tag into the thinking panel

When a model streams its reasoning inline as <think> or <thinking> text, the
chunk that opens the block is forwarded to the browser as-is, so the literal
tag sits at the top of the thinking panel for as long as the block is open,
and the chunk that closes it lands in the answer with the closing tag still
in it. Both only clean up when the response finishes and the final output
replaces what the browser accumulated. Regressed in v0.11.1 with the
per-chunk delta events; v0.10.2 resent the already cleaned output instead.

The backend already strips the tags from its own output items in that same
chunk. Now, whenever the tag scanner opens or closes an item, the cleaned
output is resent whole (the same snapshot event the handler already sends at
the end of a response) instead of the raw delta, so the thinking panel shows
clean text from the first chunk and the answer never carries the closing tag.
Chunks without a tag transition still stream as plain deltas, unchanged.
Channel chats adopt that snapshot too, and the browser now applies text
deltas to a code interpreter block's code so it keeps streaming after the tag.

Fixes #29842
This commit is contained in:
Classic298 2026-09-09 20:44:06 +02:00
parent 8383bc708c
commit 7a201f131a
3 changed files with 20 additions and 0 deletions

View file

@ -1,6 +1,7 @@
from __future__ import annotations
import asyncio
import copy
import logging
import random
import sys
@ -1006,6 +1007,9 @@ async def _make_channel_emitter(request_info):
if not content and not output and not done:
return
if isinstance(output, list):
state['output'] = copy.deepcopy(output)
now = time.time()
if done or (now - state['last_emit_at']) >= THROTTLE_INTERVAL:
state['last_emit_at'] = now

View file

@ -5447,6 +5447,8 @@ async def streaming_chat_response_handler(response, ctx):
}
]
pre_tag_item_id = output[-1].get('id')
if DETECT_REASONING_TAGS:
output, _ = tag_output_handler(
'reasoning',
@ -5488,6 +5490,15 @@ async def streaming_chat_response_handler(response, ctx):
}
delta_type = delta_event_type
# the raw chunk still carries the tag text: resend the cleaned output instead
if target_item.get('id') != pre_tag_item_id:
await flush_pending_delta_data()
await event_emitter(
{'type': 'chat:completion', 'data': {'output': full_output()}}
)
await save_current_response_stream()
data = None
if delta and data:
await queue_pending_delta_data(data, delta_type)
elif data:

View file

@ -626,6 +626,11 @@ export function applyResponseStreamEvent(
return nextOutput;
}
if (item.type === 'open_webui:code_interpreter') {
item.code = appendDelta(item.code ?? '', event.delta);
return nextOutput;
}
const key = deltaType === 'output_text' || deltaType === 'reasoning_text' ? 'text' : deltaType;
item.content = [...(item.content ?? [])];
const part = ensurePart(item.content, event.content_index ?? 0);