chore(deps): upgrade pinecone 7 + aiohttp 3.14 to close transitive HTTP CVEs

[stuartc]

Short Description

Bumps langchain-pinecone (0.2.2 → 0.2.13) and the pinecone SDK (5.4.2 → 7.3.0) so we can finally move aiohttp off 3.10. That clears a big batch of transitive aiohttp HTTP CVEs.

We don't have Dependabot PRs configured, so I'm linking the alerts directly — this closes the 21 aiohttp ones (table at the bottom).

Implementation Details

The actual blocker was langchain-pinecone 0.2.2, which hard-pinned aiohttp<3.11 — so none of the patched aiohttp releases could even resolve. Bumping langchain-pinecone drops that cap (it moved to pinecone[asyncio]>=6), which in turn pulls the pinecone SDK up to 7.x and lets aiohttp go to 3.14.1, which is the real security target here.

No source changes were needed. The pinecone APIs we actually use (Pinecone(api_key=...), .Index(), list_indexes(), ServerlessSpec) are stable across v5→v7, and both create_index() calls already use keyword args (that's the one signature change in v7).

I kept langchain-core on 0.3.x on purpose — the 1.x migration is a separate PR further up this stack. The lock also sheds a pile of unused test-only transitives (pytest-asyncio, pytest-socket, vcrpy, langchain-tests, etc.) that were only present because of the old pinecone pin, which is why the lock diff looks bigger than the change really is.

Tested: unit suite is 9/9, and all six affected service modules import cleanly against pinecone 7.3.0. I couldn't run the live Pinecone/OpenAI smokes locally (no API keys in my checkout), so worth a quick sanity check in CI/dev before merge.

This is PR 1 of 3 in a stack — base is main. The other two build on top: #513 (pytest 9 + urllib3 + dependabot config), then #514 (LangChain 1.x + openai SDK 2.x). Merge order: this one first, then #513, then #514.

Alert	Severity	Summary
#49	High	HTTP parser auto_decompress zip-bomb DoS
#34	Low	Request/response smuggling via chunked trailer parsing
#50	Low	Unicode header-value parsing discrepancies
#51	Low	Unicode match groups in ASCII protocol-element regexes
#52	Low	Brute-force leak of internal static file path components
#53	Medium	DoS when bypassing asserts
#54	Medium	DoS via large payloads
#55	Medium	DoS via chunked messages
#56	Low	Cookie parser warning storm
#78	Medium	Unlimited trailer headers → uncapped memory
#79	Low	DoS via unbounded DNS cache in TCPConnector
#80	Low	CRLF injection via multipart content-type header
#81	Medium	UNC SSRF / NTLMv2 cred theft / local file read (Windows static handler)
#82	Medium	Multipart header size bypass
#83	Low	Late size enforcement for non-file multipart fields → memory DoS
#84	Low	Cookie & Proxy-Authorization leak on cross-origin redirect
#85	Low	Response splitting via \r in reason phrase
#86	Low	C parser accepts null bytes/control chars in header values
#87	Medium	Accepts duplicate Host headers
#102	Medium	Deserialization of untrusted data
#103	Medium	Cross-origin redirect with per-request cookies

AI Usage

Please disclose how you've used AI in this work (it's cool, we just want to know!):

• Code generation (copilot but not intellisense)
• Learning or fact checking
• Strategy / design
• Optimisation / refactoring
• Translation / spellchecking / doc gen
• Other
• I have not used AI

Put together with Claude Code — agent workflows did the dependency research, the bumps, and the verification; I reviewed and drove the grouping.

You can read more details in our Responsible AI Policy

[josephjclark]

Done in release 1.3.0