Bug Description

In run_agent.py around line 9188, _think_text is assigned the full assistant_message.content, then reasoning XML tags are stripped with re.sub. When the model does not use <think>, <REASONING_SCRATCHPAD>, or <reasoning> tags in its response, _think_text ends up being the complete reply text — identical to what is streamed via message.delta events.

This causes the reasoning.available SSE event to fire with text equal to the full reply, making clients that display both events show duplicate content.

Steps to Reproduce

1. Start the Gateway server
2. Send a message that produces a plain-text response (no reasoning tags in the model output)
3. Subscribe to the SSE stream for that run
4. Observe that reasoning.available fires with text equal to the full reply, and message.delta events carry the same content

Current Code

# run_agent.py ~L9188
_think_text = assistant_message.content.strip()
_think_text = re.sub(
    r'</?(?:REASONING_SCRATCHPAD|think|reasoning)>', '', _think_text
).strip()

re.sub removes the tag delimiters but keeps all content between and outside them, so a plain reply like "Hello world" passes through unchanged.

Expected Behavior

reasoning.available should only fire when the model's response actually contains reasoning tags. The text field should contain only the content inside those tags.

Suggested fix:

_think_match = re.search(
    r'<(?:REASONING_SCRATCHPAD|think|reasoning)>(.*?)</(?:REASONING_SCRATCHPAD|think|reasoning)>',
    assistant_message.content,
    re.DOTALL,
)
_think_text = _think_match.group(1).strip() if _think_match else ""

Impact

Any client consuming the structured SSE stream will display reasoning and reply as identical content when the model responds without explicit reasoning tags.