fix: adapt UiPathChatAnthropic to the langchain-anthropic 1.7.0 raw-response contract

[andreibalas-uipath]

What changed and why

Follow-up to #132, which pinned langchain-anthropic<1.7.0 as an interim mitigation. This PR is the real fix, plus a downstream CI gate so this class of breakage is caught before publish in the future.

langchain-anthropic 1.7.0 contract fix

• UiPathChatAnthropic._create/._acreate now route through with_raw_response.create(...) (both the plain and betas branches). langchain-anthropic 1.7.0 changed these private methods to return a raw-response wrapper and calls .parse() on the result; the previous override bodies returned parsed objects, so every operation failed with 'Message'/'Stream'/'AsyncStream' object has no attribute 'parse'.
• langchain-anthropic bound on the anthropic and bedrock extras moves from >=1.4.1,<1.7.0 (the fix: bound langchain-anthropic<1.7.0 in uipath-langchain-client #132 pin) to >=1.7.0,<2.0.0.
• UiPathChatLiteLLM no longer lets ChatLiteLLM's before-validator mark streaming=False as explicitly set: langchain-core >= 1.4 treats that as a hard streaming opt-out, silently downgrading .stream()/.astream() to non-streaming invokes.
• New tests: tests/langchain/clients/anthropic/test_raw_response_contract.py and tests/langchain/clients/litellm/test_streaming_dispatch.py.

Downstream integration gate (new CI)

.github/workflows/test-downstream-langchain.yml + .github/scripts/inject_uv_overrides.py: on every PR that touches package source, build both wheels from the PR, inject tool.uv.override-dependencies into a checkout of uipath-langchain-python main (the UV_OVERRIDE env var used by uipath-python's cross-tests is silently ignored by uv project commands — verified), then run the chat-models and multimodal-invoke live integration testcases against alpha via the downstream's own run.sh/validate_output.sh — the two testcases that exercise the LLM client end to end.

A single downstream-gate check summarizes (passes when jobs were legitimately skipped), suitable for branch protection. skip:downstream label bypasses.

Prerequisite before the integration jobs can pass: repo secrets ALPHA_TEST_CLIENT_ID, ALPHA_TEST_CLIENT_SECRET, ALPHA_BASE_URL (same values as uipath-langchain-python's integration workflow). Until they're added, the two live jobs (and the gate) will fail on this PR. Do not mark downstream-gate required until secrets are in and a burn-in period has passed.

Affected packages

• uipath-langchain-client only (1.18.3 → 1.18.4). Core untouched (CI files are workflow-level).

Verification

• ruff check, ruff format --check, pyright clean; full suite green (2110 passed, cassette replay).
• Anthropic fix: with main's code + langchain-anthropic 1.7.0, the Anthropic suite fails 121 tests with exactly the errors from the uipath-langchain-python chat-models CI failure; with this branch it passes 133/133 including VCR replay of recorded gateway traffic for the failing CI config (anthropic.claude-haiku-4-5-20251001-v1:0, vendor_type="awsbedrock").
• Gate rehearsed locally against a clean clone of uipath-langchain-python main: wheels built, overrides injected (both [tool.uv] present/absent paths), the testcase resolution picks the PR wheels with wheel-metadata-driven transitive pins and survives the testcase's own uv add uipath-langchain[bedrock/vertex] steps, and the agent graph constructs all six chat-model classes against the injected packages. The only un-rehearsed step is the live testcase run (needs the alpha secrets).

Side-finding for a separate PR

Rehearsing the gate surfaced a pre-existing latent drift bug: the fireworks extra has no fireworks-ai bound, so fresh resolutions pull fireworks-ai 1.x, which breaks UiPathChatFireworks imports (fireworks.client.api_client removed; 0.16+ also collides with the google/vertexai gRPC protos). Hidden today by committed lockfiles holding 0.15.x. Fix (fireworks-ai>=0.15.0,<0.16.0) to land separately.

🤖 Generated with Claude Code