Fix empty provider list from lazy langchain_community exports

[enichols76work]

Note: #72 also resolves this via a broader dependency modernization — see comment for details on the overlap/tradeoffs.

User description

Summary

• get_langchain_chat_models_info() read chat_models_module.__dict__.get(name) directly, which bypasses langchain_community.chat_models's lazy __getattr__-based exports (PEP 562) — every class always resolved to None, so provider discovery silently returned {}.
• This broke --list-providers, the interactive TUI's Target/Attack LLM provider menus, and instantiating any real (non-test-mocked) LangChain backend via ClientLangChain.
• Regression from 17f2502 ("Fix critical and high severity Dependabot vulnerabilities (Fix critical and high severity Dependabot vulnerabilities #70)"), which correctly bumped langchain-community to fix CVE-2025-68664 and moved the import from langchain.chat_models to langchain_community.chat_models, but didn't account for the newer package's lazy-loading mechanism. Uncaught for ~5 months because the existing test suite mocks chat_models_info entirely and never exercised the real discovery function.
• Fix: use getattr(chat_models_module, model_cls_name, None) so the lazy loader actually fires, wrapped in try/except so one broken lazy import can't crash the whole CLI at startup, with a logger.warning if that ever happens.

Testing

• Added tests/test_langchain_integration.py: a real-library regression test (now discovers 59 providers, was 0) and a synthetic test reproducing the exact lazy-__getattr__ mechanism, decoupled from langchain_community's actual internals.
• Added a test to tests/test_chat_clients.py guarding the real (unmocked) chat_models_info module attribute, since every existing test there patches it with a fake dict.
• Full suite passes: 96 passed.

Test plan

• pytest — full suite green
• python -m ps_fuzz.cli --list-providers — prints real provider names, not just the header
• python3 -c "from ps_fuzz.langchain_integration import get_langchain_chat_models_info; print(len(get_langchain_chat_models_info()))" — returns 59, was 0

---

Generated description

Below is a concise technical summary of the changes proposed in this PR:
Fix provider discovery in ps_fuzz.langchain_integration so ClientLangChain and the CLI menus can enumerate real LangChain chat backends again. Add regression coverage for lazy langchain_community.chat_models exports and the chat_models_info wiring.

Topic	Details
Provider discovery	Resolve chat model classes with getattr() so lazy langchain_community.chat_models exports are discovered instead of returning an empty provider list. Modified files (1) ps_fuzz/langchain_integration.py Latest Contributors(0) User Commit Date
Regression tests	Add regression tests that exercise real model discovery, lazy __getattr__ exports, and the unmocked chat_models_info cache used by ClientLangChain. Modified files (2) tests/test_chat_clients.py tests/test_langchain_integration.py Latest Contributors(0) User Commit Date

_{Review this PR on Baz | Customize your next review}

[baz-reviewer]

Merger

Waiting for CI and review to complete.

_{Commit 6543fd2 · Updated 2026-07-23 07:32 UTC}

---

_{Review this PR on Baz | Customize your next review}

[enichols76work]

Flagging for reviewers: #72 also fixes this same bug (empty provider list from langchain_community.chat_models's lazy exports), but via a different and much broader approach worth weighing against this PR:

• This PR: a minimal, one-line fix (chat_models_module.__dict__.get() → getattr()) that preserves all 59 dynamically-discovered langchain_community chat model backends, with no other behavior change.
• Update Python dependencies and LangChain integrations #72: replaces the whole discovery mechanism with a hardcoded CHAT_MODEL_REGISTRY of just 2 backends (open_ai, ollama), imported directly from langchain-openai/langchain-ollama instead of langchain_community. It's bundled with a much larger dependency modernization — raises the Python floor from >=3.9 to >=3.11 (dropping 3.9/3.10 support), bumps most dependencies to new major versions, and drops chromadb to an optional extra. It also fixes the pkg_resources deprecation warning in base64_injection.py/harmful_behavior.py/custom_benchmark.py the same way this repo would eventually need to.

Both resolve the immediate bug; #72 is the bigger, more invasive change. Wanted to surface the overlap so we don't land both / duplicate review effort — happy to close this one in favor of #72 if that's the preferred direction, or keep this as the low-risk fix to land now with #72's broader changes considered separately.