Dependency support¶
The package declares Python 3.11 or later and pytest 9.0.3 or later. CI checks Python 3.11, 3.12, 3.13 and 3.14 with the newest stable pytest that pip resolves. A separate Python 3.11 lane checks exactly pytest 9.0.3. Resolved versions are printed in each CI run; the current lane moves as new stable versions ship.
The minimum core lane uses pytest-asyncio 1.3.0 and pytest-cov 5.0.0 as its
test harness, rather than installing the dev extra. pytest-asyncio 1.3.0 is
the first release that allows pytest 9, so it is also the [async] extra's
floor.
Floors without known vulnerabilities¶
No declared floor has a published advisory. CI's dependency audit fails if one appears, and also audits the current resolution with every extra and the hash-locked tool requirements, on every change and weekly. Floors raised for this reason:
- 0.7.0: pytest 9.0.3, the first release that fixes CVE-2025-71176 (other local users on a shared UNIX machine could interfere with pytest's temporary directories). This drops pytest 8.
- 0.6.1: pydantic 2.4.0, langchain-core 1.3.3 for the
langchainandlanggraphextras (the 0.3 line has unfixed advisories), and LangGraph 1.0.10.
Optional SDKs¶
| Extra | Exact tested floor | Current version tested on 2026-10-11 |
|---|---|---|
openai |
OpenAI 1.66.3 | 3.29.0 |
anthropic |
Anthropic 0.49.0 | 1.13.0 |
langchain |
langchain-core 1.3.3 | 1.7.7 |
judge |
DeepEval 4.2.8 | 4.2.8 |
mcp |
MCP SDK 2.2.0; Uvicorn 0.31.1 | MCP SDK 2.3.0 |
langgraph |
LangGraph 1.0.10 (with langgraph-prebuilt 1.0.8); langchain-core 1.3.3 | LangGraph 1.2.20; langchain-core 1.7.7 |
claude-agent |
Claude Agent SDK 0.2.140 | Claude Agent SDK 0.2.180 |
CI's real-SDK contract matrix installs these exact floors and a separate set of current stable releases on Python 3.12. OpenAI and Anthropic tests construct real clients and use in-process HTTP transports to check serialization, output, error handling, concurrent independent sessions and reset behavior. LangChain tests execute real Runnables. DeepEval contracts use both an offline model and real provider SDK transports, including the default Anthropic judge path. No provider API calls or credentials are needed for these contract tests.
The judge extra includes the Anthropic SDK required by its default judge. DeepEval 2.0 lacks the Anthropic and Gemini model wrappers used by the plugin; 4.2.8 is the verified floor for the current provider setup.
The OpenAI floor includes Responses support. The old Anthropic 0.30 floor
failed default client construction with current httpx because it passes the
removed proxies argument. Version 0.40 fixes construction but lacks the
thinking request argument used by the current DeepEval default judge. The
combined contract passes at 0.49.0. Current vendor SDKs use httpx2 internally,
while the plugin's HTTP adapter still uses httpx.
The current lane also executes the LangGraph state recipe with
langgraph==1.2.14; the observed checkpoint version is 4.2.0. LangGraph is not
part of the LangChain extra, and the minimum LangChain lane does not claim
compatibility with current LangGraph. See the state recipe.
The langgraph and claude-agent extras are tested in the MCP contract
lanes: the floors on Python 3.11 with MCP SDK 2.2.0, and current releases on
Python 3.12, 3.13 and 3.14. Claude Agent SDK 0.2.140 is the first release that allows
MCP SDK 2.x. CI tests LangGraph 1.0.10 with langgraph-prebuilt==1.0.8.
LangGraph 1.0.10 also allows prebuilt 1.0.13, which imports names LangGraph
1.0.x lacks; pin prebuilt if you stay on LangGraph 1.0. The OpenAI Agents SDK recipe
is tested with the current openai-agents release (0.24.0) and is not an
extra.
A dependency floor is a resolver constraint, not evidence that every release above it has been tested. Unit tests and real-SDK contracts check request construction, responses, errors, and conversation behavior without provider credentials. The async-to-sync bridge also exercises repeated real SDK requests against a local HTTP server, catching connection reuse across event-loop lifetimes. The manual live-API smoke workflow checks selected installed SDKs against real services and requires provider credentials; it is not the pull-request gate.
Before claiming a new SDK range, test the exact proposed minimum and a current stable version with real SDK clients and local transports, record resolved versions, and update this page and CI together. Test dependency resolution as well as adapter behavior. Add bounds only when a reproduced incompatibility justifies them. The judge extra has its own transitive requirements and should be tested separately from the core minimum lane.
Django¶
The model endpoint for pipeline tests needs no Django package or plugin. CI's
Django job runs an example project with python manage.py test and with
pytest-django on Python 3.12 and 3.14, using the current Django and
pytest-django releases (the job prints the resolved versions). It covers the
test settings calling install(), clients built at import time and in
AppConfig.ready(), and --parallel under the spawn and forkserver start
methods and the platform default, where each worker starts its own endpoint.
Under fork, the default on Linux before Python 3.14, a worker's first
model_endpoint block fails with an explanation instead of capturing requests
in the wrong process. Install tblib alongside Django's parallel runner: without
it, Django can't send any test failure back from a worker and the run crashes
with cannot pickle 'traceback' object (the explanation is still printed).
Packaging and tooling¶
The optional MCP harness has a separate Python 3.11 floor and
Python 3.12–3.14 current matrix. Its real SDK tests cover in-memory, stdio and
loopback Streamable HTTP transports with 2025-11-25 and 2026-07-28 protocol modes.
SDK 2.0.1 failed stdio construction; 2.2.0 is the verified floor. The <3 bound
keeps a future SDK major outside this tested integration. No MCP package is
required for the core agent contracts or existing chatbot suite.
CI builds and verifies both the sdist and wheel. A fresh environment installs the wheel and runs the packaged probes from an external consumer directory, covering plugin entry-point discovery and consumer fixture configuration. Both the chatbot suite and the packaged agent suite run on minimum/current pytest from installed wheels. The agent checks also verify versioned selection, paired baselines, missing configuration, required-coverage deselection and recorded violations. The reference Responses/MCP agent has offline SDK tests at the supported OpenAI/MCP floors and current versions; its installed example uses the real stdio transport without provider requests. The independent support application example uses only the OpenAI extra and direct local application APIs; its clean-wheel check explicitly verifies that MCP is absent. Both SDK floors and current versions exercise its offline model loop.
Ruff and Pyright are pinned to the same versions in the dev extra and pre-commit hooks. Update both files together and review any resulting lint or format changes. These tool pins are contributor tooling, not runtime limits.