test: trust test modules under Haystack 3.0's deserialization allowlist by julian-risch · Pull Request #3537 · deepset-ai/haystack-core-integrations

julian-risch · 2026-07-02T12:43:38Z

Related Issues

Part of deepset-ai/haystack-private#446

Haystack 3.0 introduces a trusted-module allowlist for deserialization: from_dict/Pipeline.loads raise DeserializationError for classes and callables from modules outside haystack*, builtins, typing, and collections. Tests in ten integrations round-trip tools, splitting functions (hanlp), MCP transports, or OAuth token sources defined in the test modules themselves, which are not on that allowlist.

Proposed Changes:

For google_genai, hanlp, huggingface_api, langfuse, llama_cpp, mcp, nvidia, oauth, transformers, watsonx, plus (second commit) aimlapi, cometapi, meta_llama, mistral, openrouter, orcarouter, togetherai — whose test_serde_in_pipeline tests round-trip test-module tool functions through Pipeline.loads and hit the same guard once #3533/#3535 fix their other failures:

Add an autouse conftest fixture that sets HAYSTACK_DESERIALIZATION_ALLOWLIST="tests,test_*" — Haystack 3.0's documented opt-in mechanism, read per deserialization call, so monkeypatch.setenv is enough. Haystack 2.x ignores the variable entirely, so nothing changes there.
oauth: test_from_dict_unknown_source_type_raises now accepts DeserializationError in addition to ImportError — for an unknown/untrusted module, 3.0 refuses upfront with DeserializationError where 2.x fails at import time. (The guard working as intended for genuinely unknown types.)

How did you test it?

Haystack 2.x unit tests pass.
Haystack v3 branch (installed git+https://github.com/deepset-ai/haystack.git@v3 into the test envs): every trusted-modules DeserializationError failure is gone. google_genai, hanlp, langfuse, llama_cpp, mcp, and oauth unit suites are now fully green under v3. The remaining v3 failures in the other four are exactly the ones addressed by the sibling draft PRs: to_dict-format tests in huggingface_api/transformers/watsonx (test: make Tool/Agent serialization assertions version-agnostic for Haystack 2.x/3.x #3533) and lazy-client init tests in nvidia (fix: support Haystack 3.0's lazy OpenAI client initialization in OpenAI-inheriting generators #3536).
For the seven suites from the second commit I verified on a local merge of all four sibling branches (test: make Tool/Agent serialization assertions version-agnostic for Haystack 2.x/3.x #3533, test: guard ToolInvoker imports so chat-generator tests run under Haystack 3.0 #3535, fix: support Haystack 3.0's lazy OpenAI client initialization in OpenAI-inheriting generators #3536, this one): all seven unit suites are fully green under v3 on that combined state.

Notes for the reviewer

Checklist

I have read the contributors guidelines and the code of conduct
I have updated the related issue with new insights and changes
I added unit tests and updated the docstrings
I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test:.

🤖 Generated with Claude Code

Haystack 3.0 refuses to deserialize classes and callables from modules outside its trusted-module allowlist. Tests that round-trip tools, splitting functions, transports, or token sources defined in the test modules now fail with DeserializationError. Add an autouse fixture that allowlists the test modules via HAYSTACK_DESERIALIZATION_ALLOWLIST (ignored by 2.x), and let the oauth unknown-source test accept the new error type. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

github-actions · 2026-07-02T12:44:48Z