uv tool install version not verified — stale global binary causes opaque import errors

[wtthornton]

Problem

When tapps-mcp is installed globally via uv tool install or pip install --user, the .mcp.json config references just tapps-mcp (no path). If the user upgrades the source repo but forgets to reinstall the global tool, the MCP server runs the old version from ~/.local/bin/tapps-mcp.

This produces opaque errors like:

Error executing tool tapps_session_start: No module named 'tapps_core.experts.rag_embedder'

The error points to a module removed in v2.0.0, but the user believes they're running v2.3.0 because the repo is at v2.3.0.

Diagnosis Path (took ~30 minutes)

1. tapps_session_start → No module named 'tapps_core.experts.rag_embedder'
2. uv sync in repo → rebuilt venv, but error persists
3. Searched for rag_embedder in source → not found anywhere
4. Cleaned __pycache__ → no effect
5. which tapps-mcp → ~/.local/bin/tapps-mcp
6. tapps-mcp --version → v1.14.0 (expected v2.3.0)
7. uv tool install --force --editable packages/tapps-mcp → fixed

Suggested Fixes

1. tapps_session_start should report the binary path — include sys.executable or shutil.which('tapps-mcp') in the response so version mismatches are immediately visible
2. tapps_upgrade should check binary version — compare tapps-mcp --version against the repo version and warn if mismatched
3. tapps_doctor should detect this — check if the running server version matches the repo version
4. .mcp.json template could use full path — uv run --project /path/to/tapps-mcp tapps-mcp serve instead of bare tapps-mcp