test(e2e): decouple mcp upstream from the shared litellm service

The mcp-upstream service was a service_healthy dependency of the litellm
service, so every e2e suite's stack waited on it and would fail to start if a
future main-latest image broke the FastMCP import. Only the mcp suite needs
it, and it boots well before the proxy is live, so drop the depends_on edge.
Also remove the unused MCP_TRANSPORT env var; the upstream script hardcodes
the streamable-http transport.
This commit is contained in:
Yassin Kortam 2026-07-17 09:10:17 -07:00
parent ee1afeb444
commit b817b689b2

View file

@ -95,8 +95,6 @@ services:
condition: service_healthy
jaeger:
condition: service_healthy
mcp-upstream:
condition: service_healthy
env_file: .env
environment:
LITELLM_MASTER_KEY: sk-1234
@ -137,12 +135,13 @@ services:
# deterministic self-hosted upstream MCP server (FastMCP add/multiply over
# streamable-http), reachable by the litellm container at mcp-upstream:8090/mcp.
# The suite registers it through /v1/mcp/server at test time.
# Not a depends_on of litellm on purpose: only the mcp suite needs it, and it
# boots long before the proxy is live, so it must not gate the other suites'
# stack. The suite registers it through /v1/mcp/server at test time.
mcp-upstream:
image: ghcr.io/berriai/litellm:main-latest
entrypoint: ["python3", "/app/mcp_upstream_server.py"]
environment:
MCP_TRANSPORT: http
MCP_HOST: 0.0.0.0
MCP_PORT: "8090"
configs: