3ca425416elangchain 1.4.2 ships `langchain.mcp`, whose MCPAdapter wraps fastmcp.Client and ClientGroup directly (resolve_tool, list_tools cache_mode, call_tool, Client.new, elicitation callbacks, MCPConfig, legacy and auto modes). The workflow now runs langchain's own MCP tests against the built wheels, and a new smoke drives MCPAdapter in-process, over stdio, MCPConfig, HTTP, SSE, a proxy, and a ClientGroup of a legacy HTTP server and a modern stdio server. Every smoke also runs through a FastMCP proxy in front of a stdio server and covers shapes this release cycle got wrong: `|` and non-ASCII template literals, comma-joined and exploded list query params, a float schema count, and a timed-out call followed by another on the same connection. Gaps the proxy already had in 4.0.5 are reported as known gaps that fail the run once they start passing. Server stderr goes to a log file, and the SDK's SSE reader is quieted where a check cancels a call on purpose. langchain-mcp-adapters' smoke is renamed smoke_langchain_mcp_adapters.py. Claude-Session: https://claude.ai/code/session_01KfHgVhbYEhBCC5eSeqGiuG Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| _harness.py | ||
| README.md | ||
| server.py | ||
| smoke_langchain_mcp.py | ||
| smoke_langchain_mcp_adapters.py | ||
| smoke_pydantic_ai.py | ||
downstream smoke
Checks that FastMCP still works the way its consumers use it. The workflow builds FastMCP's wheels once, installs them where each consumer runs, and drives the consumer against servers started from the checkout.
| job | what it runs | FastMCP install |
|---|---|---|
| pydantic-ai test suite | pydantic-ai's own MCP tests at a pinned tag | wheels overlaid on pydantic-ai's locked environment, as their FastMCP 4 job does |
| langchain test suite | langchain.mcp's own unit and integration tests at a pinned tag |
wheels overlaid on langchain's locked test environment |
| pydantic-ai smoke | smoke_pydantic_ai.py |
client-only fastmcp-slim[client], as pydantic-ai-slim[mcp] installs it |
| langchain.mcp smoke | smoke_langchain_mcp.py, including a ClientGroup of a legacy HTTP server and a modern stdio server |
fastmcp, as langchain[mcp] installs it |
| langchain-mcp-adapters smoke | smoke_langchain_mcp_adapters.py |
none: the adapters pin mcp<2, so they reach FastMCP over the wire |
Each smoke runs its checks over stdio, streamable HTTP, SSE, and a FastMCP proxy
in front of a stdio server, all with bearer auth where the transport allows it.
The checks cover shapes past releases got wrong: template literals such as |
and non-ASCII, comma-joined and exploded list query params, elicitation in both
protocol eras, schema counts written as floats, and a timed-out call followed by
another on the same connection.
Each smoke prints one line per check, writes a table to the job summary, and
reports warnings raised in the consumer's process. A check listed in
_harness.py's PROXY_GAPS or LEGACY_PROXY_GAPS is a gap the proxy already
had in 4.0.5: it shows as ⚠ without failing the run, and fails the run if it
starts passing so the entry gets removed.
Pushes and PRs test the versions pinned in the workflow. The nightly run, or a
manual run with latest, tests each consumer's latest release.
running locally
uv sync
uv build --wheel --out-dir dist . fastmcp_slim
slim="fastmcp-slim @ file://$PWD/$(ls dist/fastmcp_slim-*.whl)"
full="fastmcp @ file://$PWD/$(ls dist/fastmcp-4*.whl)"
uv run --isolated --no-project --no-config \
--with 'pydantic-ai-slim[mcp]' --with "${slim/fastmcp-slim/fastmcp-slim[client]}" \
python tests/downstream/smoke_pydantic_ai.py
uv run --isolated --no-project --no-config \
--with 'langchain[mcp]' --with "$full" --with "$slim" \
python tests/downstream/smoke_langchain_mcp.py
uv run --isolated --no-project --no-config \
--with langchain-mcp-adapters --with langchain \
python tests/downstream/smoke_langchain_mcp_adapters.py