Summary
Split out from #222, where @stevemju noted in passing:
Agent Brain: 10.2.0 (also present on main) (btw >= 10.3.0 doesn't work for me because of a mcp dependency conflicts, but that's another issue)
This is that other issue. A user is pinned to 10.2.0 and cannot upgrade, so it blocks them from picking up every fix since — including the base_url fix in #222.
Investigation so far (2026-08-30)
Could not reproduce from a clean environment. Resolutions were run with uv pip compile --python-version 3.10 (the reporter is on Python 3.10) for:
| Package set |
Result |
agent-brain-rag==10.4.0 + agent-brain-cli==10.4.0 |
resolves clean |
the above + agent-brain-ag-mcp==10.4.0 |
resolves clean |
agent-brain-rag/cli/ag-mcp all ==10.3.2 |
resolves clean |
Published metadata was also checked and is correct — no path dependencies leaked into any published wheel (a real risk given agent-brain-cli/pyproject.toml carries agent-brain-ag-mcp = { path = "../agent-brain-mcp" } for local dev, which the release process swaps to a PyPI pin):
agent-brain-cli==10.4.0 → agent-brain-rag<11.0.0,>=10.4.0, agent-brain-uds<11.0.0,>=10.3.0
agent-brain-ag-mcp==10.4.0 → mcp<2.0.0,>=1.27.2, agent-brain-rag<11.0.0,>=10.4.0, agent-brain-uds<11.0.0,>=10.4.0
mcp itself declares requires_python >=3.10 at both 1.12.0 and 1.27.2, so there is no Python-version conflict.
Leading hypothesis
The conflict is environment-local, not a packaging defect: v10.4 raised the floor to mcp>=1.27.2 (Phase 67 needed mcp.server.auth). An environment that already holds an older mcp pinned by another MCP tool would fail to co-resolve, and pip's resolver reports that as a dependency conflict.
What's needed to close this
From the reporter:
- The exact install command (
pip install ... / uv pip install ..., and which packages).
- The full resolver error text, which names the conflicting pins.
pip freeze | grep -i mcp from the target environment before the upgrade.
- Whether installing into a fresh virtualenv succeeds.
Possible outcome
If confirmed as a pre-existing mcp pin, the fix is documentation (an install-troubleshooting note about the mcp>=1.27.2 floor from v10.4 and using a clean venv), not a metadata change.
Summary
Split out from #222, where @stevemju noted in passing:
This is that other issue. A user is pinned to 10.2.0 and cannot upgrade, so it blocks them from picking up every fix since — including the
base_urlfix in #222.Investigation so far (2026-08-30)
Could not reproduce from a clean environment. Resolutions were run with
uv pip compile --python-version 3.10(the reporter is on Python 3.10) for:agent-brain-rag==10.4.0+agent-brain-cli==10.4.0agent-brain-ag-mcp==10.4.0agent-brain-rag/cli/ag-mcpall==10.3.2Published metadata was also checked and is correct — no path dependencies leaked into any published wheel (a real risk given
agent-brain-cli/pyproject.tomlcarriesagent-brain-ag-mcp = { path = "../agent-brain-mcp" }for local dev, which the release process swaps to a PyPI pin):agent-brain-cli==10.4.0→agent-brain-rag<11.0.0,>=10.4.0,agent-brain-uds<11.0.0,>=10.3.0agent-brain-ag-mcp==10.4.0→mcp<2.0.0,>=1.27.2,agent-brain-rag<11.0.0,>=10.4.0,agent-brain-uds<11.0.0,>=10.4.0mcpitself declaresrequires_python >=3.10at both 1.12.0 and 1.27.2, so there is no Python-version conflict.Leading hypothesis
The conflict is environment-local, not a packaging defect: v10.4 raised the floor to
mcp>=1.27.2(Phase 67 neededmcp.server.auth). An environment that already holds an oldermcppinned by another MCP tool would fail to co-resolve, andpip's resolver reports that as a dependency conflict.What's needed to close this
From the reporter:
pip install .../uv pip install ..., and which packages).pip freeze | grep -i mcpfrom the target environment before the upgrade.Possible outcome
If confirmed as a pre-existing
mcppin, the fix is documentation (an install-troubleshooting note about themcp>=1.27.2floor from v10.4 and using a clean venv), not a metadata change.