[sbblack-qualus]

Add functionality to auto discover the highest version of psse on the machine that has an appropriate psspy package matching the python version from which the MCP server was run. I would suggest doing some homework and selecting an MCP run command setup in your AI harness that runs the MCP server purposely in a version of python that matches a version of psse you have on your machine (PSSE >= 35).

There was also some strange behavior where the injection of report output from api calls like psspy.diff to stdout interfered with FastMCP and crashed the MCP server. I have added wrappers to attempt to keep PSSE output away from STDIO. It seems to be more stable now. I suspect that proper use of the redirect module might be a more eloquent solution, but this seems to work for now.

[qian-harvard]

Thanks for working on this. PSS/E auto-discovery and keeping PSS/E's report output off the MCP stream are both useful. Unfortunately the branch can't be merged as it stands. In the merge commit (2c9e052), the conflict in PSSE/psse_mcp.py was resolved by keeping your fork's copy of the file whole. That copy dates from March, so the PR reverts everything that has landed in the PSS/E server since then:

1. The server no longer starts with the pinned SDK. It imports from mcp.server.fastmcp import FastMCP, but pyproject.toml pins mcp>=2,<3, where that module doesn't exist. Main uses from mcp.server.mcpserver import MCPServer as FastMCP. With main merged in, pytest aborts at collection: tests/test_psse_server.py and tests/test_sandbox.py raise ModuleNotFoundError.
2. The security hardening from #57 is gone. That covers:
   • the _PROHIBITED_PSSPY_COMMANDS blocklist, so launch_program, runrspnsfile, addmodellibrary and others become callable through run_psspy_command;
   • the checked_path sandbox on case and file arguments;
   • the identifier check in lookup_psspy_command, so lookup_psspy_command("../../pyproject") reads files outside the spec folder again.
3. PSS/E initialises at import time. Discovery and psseinit run on import and raise on any machine without PSS/E. Main made this lazy (_ensure_psse), and it reads the powermcp config paths (psse.python_lib / psse.bin), which this version drops.
4. The open_case fixes from #94 are reverted. The ierr from psspy.case is no longer checked, so a failed load reports success with the previous case's counts.
5. The list_dynamic_output_channels and read_dynamic_output tools from #97/#98 are deleted.

On the two features themselves:

• Auto-detect.
  • The candidate majors are hard-coded to [36, 35], so PSS/E 37 is never found.
  • PSSE_VERSION=36.2 raises ValueError.
  • The manual override is silently ignored unless both env vars are set.
  • The override defaults to psse36, so a PSS/E 35 user gets No module named 'psse36'.
• stdout. mcp 2.x already redirects fd 1 to stderr when the stdio transport starts. A tool that print()s or os.write(1, ...)s doesn't corrupt the JSON-RPC stream on main. The crash you saw looks like mcp 1.x behaviour. The part worth keeping is returning the captured text to the client (your _console_output), since otherwise psspy.diff output only reaches stderr.

Could you open a fresh branch from current main and port just the new behaviour onto main's version of the file?

• Move discovery into _resolve_psse_paths() so it stays lazy. Check powermcp config first, and don't cap the major version.
• Keep the _console_output capture inside run_psspy_command.
• Add tests to tests/test_psse_server.py using its fake psspy. CI doesn't have PSS/E, so those tests are what will cover this.