Cleanup the UA detection method

nicegui/client.py

@@ -7,6 +7,7 @@
 import uuid
 from collections import defaultdict
 from collections.abc import Callable, Iterable
+from functools import cached_property
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, ClassVar, cast
 
@@ -153,18 +154,7 @@ def __exit__(self, *_) -> None:
 
     def build_response(self, request: Request, status_code: int = 200) -> Response:
         """Build a FastAPI response for the client."""
-        accept = request.headers.get('accept', '')
-        user_agent = request.headers.get('user-agent', '')
-        # NOTE: This simple check doesn't handle quality values (q=) or wildcards (*/*).
-        # It works for the real use case: agents sending exactly `Accept: text/markdown`.
-        explicit_markdown = 'text/markdown' in accept and 'text/html' not in accept
-        # UA-based fallback: clients like claude.ai's WebFetch send `Accept: */*` despite
-        # being agents (vs. Claude Code which sends `text/markdown, text/html, */*`).
-        # When the page opts in and the request looks like an agent with no explicit
-        # text/* preference, prefer markdown. Explicit `text/html` always wins HTML.
-        wildcard_accept = not accept or accept.strip() == '*/*'
-        ua_fallback = wildcard_accept and 'text/html' not in accept and _looks_like_agent(user_agent)
-        if self.page.resolve_markdown() and (explicit_markdown or ua_fallback):
+        if self.page.resolve_markdown() and self._did_user_request_markdown(request):
             parts = []
             if title := self.resolve_title():
                 parts.append(f'# {title}')
@@ -472,28 +462,44 @@ def prune_instances(cls, *, client_age_threshold: float = 60.0) -> None:
         except Exception:
             log.exception('Error while pruning clients')
 
+    @cached_property
+    def _agent_ua_regex(self):
+        """
+        Known live-user / agent fetchers (verified April 2026).
+
+        Vendor docs change frequently; refresh this list when new live-user agents are announced.
+
+        Authoritative pages at the time of writing:
+        - support.claude.com (claudebot / claude-user / claude-searchbot)
+        - platform.openai.com/docs/bots (gptbot / oai-searchbot / chatgpt-user)
+        - docs.perplexity.ai (perplexitybot / perplexity-user)
+        - google's search central docs cover google-cloudvertexbot
+        - gemini-deep-research and google-agent (project mariner) are documented in their
+          respective product pages rather than on the common-crawlers page.
+
+        Substring match is intentional: vendors append version + embed-product suffixes
+        (e.g. `claude-user (claude-code/2.1.121; ...)`).
+        """
+        return re.compile(
+            r'claude-?(bot|user|searchbot)|gptbot|oai-searchbot|chatgpt-user|'
+            r'perplexity(bot|-user)|google-(cloudvertexbot|agent)|gemini-deep-research',
+            re.IGNORECASE,
+        )
+
+    def _did_user_request_markdown(self, request) -> bool:
+        accept = request.headers.get('accept', '')
+        if 'text/markdown' in accept and ('text/html' not in accept or accept.find('text/markdown') < accept.find('text/html')):
+            return True
+        if accept and '*/*' not in accept:
+            return False
+        # UA-based fallback: clients like claude.ai's WebFetch send `Accept: */*` despite
+        # being agents (vs. Claude Code which sends `text/markdown, text/html, */*`).
+        # When the page opts in and the request looks like an agent with no explicit
+        # text/* preference, prefer markdown. Explicit `text/html` always wins HTML.
+        user_agent = request.headers.get('user-agent', '')
+        return bool('text/html' not in accept and self._agent_ua_regex.search(user_agent))
+
 
 def _is_prefetch(request: Request) -> bool:
     purpose = (request.headers.get('Sec-Purpose') or request.headers.get('Purpose') or '').lower()
     return 'prefetch' in purpose and 'prerender' not in purpose
-
-
-# Known live-user / agent fetchers (verified April 2026). Vendor docs change frequently;
-# refresh this list when new live-user agents are announced. Authoritative pages at the time
-# of writing: support.claude.com (ClaudeBot / Claude-User / Claude-SearchBot),
-# platform.openai.com/docs/bots (GPTBot / OAI-SearchBot / ChatGPT-User),
-# docs.perplexity.ai (PerplexityBot / Perplexity-User), Google's Search Central docs cover
-# Google-CloudVertexBot; Gemini-Deep-Research and Google-Agent (Project Mariner) are
-# documented in their respective product pages rather than on the common-crawlers page.
-# Substring match is intentional: vendors append version + embed-product suffixes
-# (e.g. `Claude-User (claude-code/2.1.121; ...)`).
-_AGENT_UA_REGEX = re.compile(
-    r'claude-?(bot|user|searchbot)|gptbot|oai-searchbot|chatgpt-user|'
-    r'perplexity(bot|-user)|google-(cloudvertexbot|agent)|gemini-deep-research',
-    re.IGNORECASE,
-)
-
-
-def _looks_like_agent(user_agent: str) -> bool:
-    """Heuristic: does the User-Agent header belong to a known live-user agent fetcher?"""
-    return bool(_AGENT_UA_REGEX.search(user_agent))

tests/test_markdown_response.py

@@ -58,6 +58,19 @@ def page():
     assert 'text/markdown' in response.headers['content-type']
 
 
+async def test_ua_fallback_respects_markdown_coming_before_html(user: User):
+    """A known-agent UA asking for `text/markdown, text/html` gets markdown because it's requested first."""
+    @ui.page('/', markdown=True)
+    def page():
+        ui.label('Hello')
+
+    response = await user.http_client.get('/', headers={
+        'Accept': 'text/markdown, text/html',
+        'User-Agent': 'GPTBot/1.1',
+    })
+    assert 'text/markdown' in response.headers['content-type']
+
+
 async def test_ua_fallback_respects_explicit_html_preference(user: User):
     """A known-agent UA asking for `text/html` still gets HTML — explicit Accept always wins."""
     @ui.page('/', markdown=True)