Skip to content

Model check: read the LangGraph (TypeScript) engine's Gemini client, and test the lock rule on its own lock - #1793

Merged
davidmckayv merged 1 commit into
feat/model-lock-langgraph-tsfrom
fix/lgts-probe-lock-tests
Nov 17, 2026
Merged

davidmckayv merged 1 commit into
feat/model-lock-langgraph-tsfrom
fix/lgts-probe-lock-tests

Conversation

@davidmckayv

Copy link
Copy Markdown
Contributor

Two fixes to #1778, found by its first real run against the LangGraph (TypeScript) engine.

What

  • The engine's Gemini client is read. The engine's package.json names @langchain/google, but the probe's NODE_CLIENTS listed only @langchain/google-genai, so a run recorded no Gemini client version for the engine. @langchain/google is added (@langchain/google-genai stays, for any image that ships it).
  • The lock rule is tested on a lock of its own. "Before its first run, nothing is offered" and "after a run, exactly the models that answered are offered" read the live MODEL_LOCK, so they could pass only while the engine's entry was empty. The first lock refresh after a real run would have failed them. Both now build on an unrun lock defined in the test.

Tests

  • The client-version test now installs @langchain/google too. Red before the change (the probe returned every client except it), green after.
  • With the lock a real run writes (64 models verified, 4 refused) in shared/model-lock.json, all 8 tests in scripts/model-compat-langgraph-ts.test.ts pass; before this change two failed. With the committed empty lock: 8 pass.

The run that found them

--tags source on the engine at feat/langgraph-ts, and again on the image the internal build published from the same commit, with the same result on every row: Anthropic 15 of 15, OpenAI 29 of 29, Google 20 of 24. The four Google refusals are the provider's: gemini-2.5-flash-lite is closed to new users, gemini-decision-1 is not a chat model, and gemini-omni-1.1-flash and gemini-omni-1.2-flash only take Google's Interactions API.

…and test the lock rule on its own lock

The engine ships @langchain/google, which the probe did not read, so a run recorded no Gemini client for it. Two tests read the live model lock and could only pass while it was empty, so the first lock refresh after a real run would have failed them; they now build on an unrun lock of their own.
@davidmckayv
davidmckayv merged commit 32d7855 into feat/model-lock-langgraph-ts Nov 17, 2026
14 of 15 checks passed
@davidmckayv
davidmckayv deleted the fix/lgts-probe-lock-tests branch November 17, 2026 23:28
davidmckayv added a commit that referenced this pull request Nov 17, 2026
…and test the lock rule on its own lock (#1793)

The engine ships @langchain/google, which the probe did not read, so a run recorded no Gemini client for it. Two tests read the live model lock and could only pass while it was empty, so the first lock refresh after a real run would have failed them; they now build on an unrun lock of their own.
davidmckayv added a commit that referenced this pull request Nov 18, 2026
…and test the lock rule on its own lock (#1793)

The engine ships @langchain/google, which the probe did not read, so a run recorded no Gemini client for it. Two tests read the live model lock and could only pass while it was empty, so the first lock refresh after a real run would have failed them; they now build on an unrun lock of their own.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant