STDIN support

[mjrusso]

(Note: I'm filing this ticket for posterity. I took a stab at implementing support for reading from STDIN, and in the process came to appreciate that the feature is way more complicated than it's worth.)

A bit of backstory. I wanted to quickly build a simple MCP server using the STDIO transport, via the official Python SDK, like so:

Mix.install([
  {:pythonx, "~> 0.3.0"}
])

Pythonx.uv_init("""
[project]
name = "mcp-test"
version = "0.0.0"
requires-python = "==3.13.*"
dependencies = [
  "mcp == 1.3.0"
]
""")

defmodule MCPServer do
  import Pythonx, only: :sigils

  def main do
    ~PY"""
    from mcp.server.fastmcp import FastMCP

    mcp = FastMCP("Demo")

    @mcp.tool()
    def add(a: int, b: int) -> int:
        "Add two numbers"
        return a + b

    mcp.run()
    """
  end
end

MCPServer.main()

This throws an exception, as pythonx doesn't support reading from STDIN.

I thought, "how hard could this be?" and tried my hand at implementation.

Specifically, I added an Elixir process that reads from STDIN, writing each line (via a new NIF) to a shared data structure that the Python evaluations can read from. (There is a lot of complexity here and I never quite got it working properly; the data structure exists outside of any given evaluation, which is its own can of worms. To do this properly you'd need bookkeeping to keep track of which data each evaluation has seen, etc., and there's a lot of other edge cases to work through.)

All that being said, practically speaking, I don't think STDIN is terribly useful for pythonx outside of a few niche use cases. I would have been better served spending the time implementing a STDIO MCP server in Elixir directly. (Sidenote: there is an existing MCP library —https://github.com/kEND/mcp_sse — but it is SSE transport only.)

In any event, I'm filing this ticket for documentation purposes, not as a feature request :)

[jonatanklosko]

@mjrusso thanks for the notes. We may support stdin in the future, but this has a low priority. I think the way it should work is that our Python stdin override sends a message to an Elixir process (similar to stdout), that process reads from Elixir IO and then passes the result back to Python. My initial thought was to do as in cocoa-xu/nif_call, but that is probably too much complexity for this feature. Another idea I just had would be to wait for the "response" in Python rather than C, with something like asyncio future, that the Elixir process would resolve. I may experiment with this at a later point.

[mjrusso]

Thanks @jonatanklosko, that all makes sense, and sounds like a good line of exploration.