fix(community): build the remaining search clients off the event loop - #7933

Heads-up: open PR #7932 fixes this exact site with a stronger construction pattern — task-owned asyncio.create_task(asyncio.to_thread(...)) + asyncio.shield, which closes the client if the caller is cancelled during setup — plus its own backend/tests/blocking_io/test_jsonify_client_ssl.py and an overlapping AGENTS.md edit. Before this change, construction was synchronous on the loop, so there was no cancellation window here; plain await asyncio.to_thread(...) introduces one: if the awaiting task is cancelled while the worker thread is still building the client, the client gets constructed and its aclose() never runs (no sockets are open yet, so the impact is small, but it is exactly the window #7932's task ownership closes — the same gap shared by the InfoQuest/You.com/GroundRoute/Webz/Crawl4AI sites). The two PRs are functionally compatible, but they conflict on this line and in AGENTS.md, and both add a jsonify TLS test under backend/tests/blocking_io/. Worth coordinating with @Doy369 — ideally adopting the cancellation-safe _open_client shape here so the sweep lands the stronger pattern for all three providers, or letting #7932 rebase onto this.

Thanks for flagging the overlap. Would it work to keep the Jsonify fix in #7932 and scope #7933 to SearXNG and Jina? That would let both PRs move forward independently while keeping Jsonify's cancellation cleanup and regression tests. If you'd prefer #7933 to land first, I can rebase #7932 on top of it.

@BlueX888, following up on the overlap now that #7932 is updated to 5cdf0d6, rebased onto current main, and includes the cancellation-diagnostic fix from the latest review. Could you confirm this split?

• fix(jsonify): keep HTTPX TLS setup off the event loop #7932: Jsonify construction, cancellation cleanup and diagnostics, and its regression tests.
• fix(community): build the remaining search clients off the event loop #7933: SearXNG and both Jina client paths, with their regression tests.

Suggested order: land #7932 first, then rebase #7933. That would mean dropping the Jsonify tools.py hunk and the jsonify case/branches in test_remaining_search_clients_build_tls_off_loop. The AGENTS.md update can then build on the merged paragraph, retaining the Browserless/Jsonify cancellation guarantee while adding SearXNG and Jina to the off-loop list.

The shared-client helper will be a separate follow-up after these fixes land, initially covering Browserless and Jsonify. Please let me know if a different merge order works better for you and the maintainers.

Confirmed — #7932 keeps the Jsonify construction, cancellation cleanup and its regression tests; #7933 drops the Jsonify hunk and the jsonify case from test_remaining_search_clients_build_tls_off_loop, keeping SearXNG plus both Jina client paths. Landing #7932 first and then rebasing #7933 works for me. I will rebuild the backend/AGENTS.md paragraph on top of #7932's version so the Browserless/Jsonify task-owned guarantee stays intact and SearXNG/Jina are added to the off-loop list. Thanks for the clean split.

This enumeration is accurate at the head — I verified to_thread construction for RAGFlow, LightRAG, You.com, GroundRoute, Webz, Crawl4AI, Browserless, InfoQuest, jsonify, SearXNG and Jina, and Crawl4AI's presence resolves the #7906 follow-up; keeping the leading 'several community providers, including' preserves non-exhaustiveness. For completeness if this PR closes out the series: one on-loop httpx.AsyncClient construction remains in the community package outside this sentence's scope — aio_sandbox/backend.py:128 (wait_for_sandbox_ready_async) builds its client directly on the loop from async runtime paths. Sandbox infra rather than a search provider (and usually plain-HTTP local URLs), so fine to leave out of this change, but it would be the natural last follow-up.