Skip to content

AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist agents (FastAPI + Next.js).

fix(deps): move the mcp gateway's dependency cutoff past fastembed 0.8.0 (#288)

Rufus8098ed552e5b
diff --git a/docker/Dockerfile b/docker/Dockerfile
index 2eb0aedf..5acac5fa 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -27,7 +27,7 @@ RUN pip install uv
# if this step succeeded: `|| true` keeps the build green if GitHub is briefly
# unreachable, and such an image fetches the commit — and needs network — on
# its first gateway start. `uv` itself (installed above) is not pinned.
-RUN uvx --exclude-newer 2026-09-22T00:00:00Z \
+RUN uvx --exclude-newer 2026-09-23T00:00:00Z \
--from git+https://github.com/SenteLabsAI/extensible-mcp@ac2001a09646a8044210042e12e62974f4c9687c extensible-mcp --help \
|| true
diff --git a/packages/core/openexecutive/orchestrator/mcp_gateway.py b/packages/core/openexecutive/orchestrator/mcp_gateway.py
index 8339c2c1..c2917490 100644
--- a/packages/core/openexecutive/orchestrator/mcp_gateway.py
+++ b/packages/core/openexecutive/orchestrator/mcp_gateway.py
@@ -40,8 +40,12 @@ _UVX_CMD = "uvx"
# cutoff together, deliberately; tests/unit/
# test_extensible_mcp_pin.py fails if docker/Dockerfile's pre-warm drifts from
# _EXTENSIBLE_MCP_LAUNCH_ARGS.
+# The cutoff must admit fastembed 0.8.1 (published 2026-09-22T20:01Z): 0.8.0
+# pads MiniLM batches to a fixed 128 tokens but truncates at 256, so any batch
+# mixing shorter and 129-256-token texts fails to stack, and the Google
+# Workspace tool index is such a batch — the gateway died building it.
_EXTENSIBLE_MCP_REV = "ac2001a09646a8044210042e12e62974f4c9687c"
-_EXTENSIBLE_MCP_EXCLUDE_NEWER = "2026-09-22T00:00:00Z"
+_EXTENSIBLE_MCP_EXCLUDE_NEWER = "2026-09-23T00:00:00Z"
_EXTENSIBLE_MCP_GIT = f"git+https://github.com/SenteLabsAI/extensible-mcp@{_EXTENSIBLE_MCP_REV}"
_EXTENSIBLE_MCP_CMD = "extensible-mcp"
# Everything after `uvx` up to the command, shared with docker/Dockerfile's pre-warm.
diff --git a/packages/core/tests/unit/test_extensible_mcp_pin.py b/packages/core/tests/unit/test_extensible_mcp_pin.py
index 9e8c9763..555b45a6 100644
--- a/packages/core/tests/unit/test_extensible_mcp_pin.py
+++ b/packages/core/tests/unit/test_extensible_mcp_pin.py
@@ -57,6 +57,22 @@ def test_gateway_freezes_dependency_resolution_with_a_cutoff() -> None:
assert args[-3:] == ("--from", mcp_gateway._EXTENSIBLE_MCP_GIT, mcp_gateway._EXTENSIBLE_MCP_CMD)
+# fastembed 0.8.1's PyPI upload time. Any earlier cutoff resolves 0.8.0, which
+# pads to a fixed 128 tokens but truncates at 256, so embedding the Google
+# Workspace tool descriptions raises and the gateway never starts.
+_FASTEMBED_0_8_1_PUBLISHED = "2026-09-22T20:01:22Z"
+
+
+def test_cutoff_admits_the_fastembed_padding_fix() -> None:
+ args = mcp_gateway._EXTENSIBLE_MCP_LAUNCH_ARGS
+ cutoff = args[args.index("--exclude-newer") + 1]
+ # Same fixed-width UTC format (checked above), so string order is time order.
+ assert cutoff > _FASTEMBED_0_8_1_PUBLISHED, (
+ f"cutoff {cutoff} resolves fastembed 0.8.0, whose batch padding breaks the gateway's "
+ "tool index; keep it after fastembed 0.8.1's release"
+ )
+
+
def test_dockerfile_prewarm_matches_the_gateway_launch() -> None:
prewarms = _dockerfile_prewarm_args()
assert prewarms, "docker/Dockerfile no longer pre-warms extensible-mcp with uvx"