RuntimeError: Cannot patch execution_info before it has been set on LangGraph Cloud

[formertheorist]

When I run my deployed graph (chat_agent) on LangSmith / LangGraph Cloud, the run fails immediately with:

RuntimeError: Cannot patch execution_info before it has been set

In LangSmith, the failure shows up right at the start of the trace (e.g. around __START__). I suspect a platform API vs Python library interaction (not certain), because on traces that work, I see LANGSMITH_LANGGRAPH_API_VERSION 0.7.90. On failing traces, it shows 0.7.96.

My setup

langgraph.json uses "dependencies": ["."], and pyproject.toml has langgraph>=0.2.0. Graph built with Deep Agents (create_deep_agent), MemorySaver checkpointer, if that helps triage.

What I tried:

I dug through the LangSmith build logs for a good build. I see the base image, e.g.
gcr.io/langchain-prod/langgraph-executor-unlicensed:0.7.94-py3.11,
and lines like

• langgraph-api==0.7.94,
• langgraph-runtime-inmem==0.27.0,
• langgraph-cli==0.4.19

I don’t see a clear langgraph==… line for the core langgraph library, and searching for “already satisfied” / langgraph== didn’t surface it. my understanding is that the core package may already ship in the executor image unless my install step upgrades it.

Questions:

1. Is this execution_info / patch_execution_info error a known issue with a specific pairing of platform API / executor / langgraph versions?
2. What’s the recommended way to determine which langgraph (PyPI) version is actually executing my graph in Cloud (given the executor base image)?
3. What’s the recommended way to pin that version from a project that uses dependencies: ["."] — e.g. should langgraph==x.y.z in pyproject.toml reliably override what’s in the executor image?

happy to provide repo details, langgraph.json, and traceback if useful.

[formertheorist]

Update: resolved

raising the minimum langgraph version in pyproject.toml (and redeploying) fixed the RuntimeError: Cannot patch execution_info before it has been set failures in LangSmith.

I’m still curious why >= in pyproject.toml wasn’t enough to get the latest version.
My understanding after digging:
langgraph>=0.2.0 only means “pip is allowed to install something this new or newer.” On LangGraph Cloud, the executor base image already includes a langgraph build. If that build satisfies the constraint, it may not upgrade it to the newest release, so I end up on an older library. In my case, pinning a higher version for langgraph forced the resolver to pull a newer version.

However, I’d expect LangSmith’s platform API and the langgraph library to stay compatible, and I’m sure that’s the goal. so I wonder why wouldn’t the the executor base image come with a compatible langgraph.

Leaving this here in case someone else sees the same error after a platform version change.