[naji247]

Summary

mcp-compatibility.yml (job gating):

   schedule:
-    - cron: '0 0 * * *'
+    - cron: '0 2 * * *'
 jobs:
+  changes:        # git diff --name-only base...head -> outputs.sdk ('true' on non-PR events)
   discover-mcp-versions:
+    needs: changes
+    if: github.event_name != 'pull_request' || needs.changes.outputs.sdk == 'true'
   discover-fastmcp-versions:      # same needs/if
   test-dependency-groups:         # same needs/if

test-fastmcp-compatibility and test-without-fastmcp skip with their discover job. send-failure-notification (if: failure()) skips. report-compatibility (if: always()) still runs and passes.

mcp-prerelease-compatibility.yml (path filter):

 on:
+  pull_request:
+    branches: [ main ]
+    paths: [src/**, tests/**, pyproject.toml, uv.lock, .github/scripts/**, .github/workflows/mcp-prerelease-compatibility.yml]
   schedule:
     - cron: '0 2 * * *'   # unchanged

SDK-code paths: src/**, tests/**, pyproject.toml, uv.lock, .github/scripts/**, plus the workflow's own file.

Evidence

• Before: compat: push main, every PR to main, cron 0 0 * * *, dispatch. prerelease: cron 0 2 * * *, dispatch.
  After: compat: push main, PR to main (compat jobs gated on SDK paths), cron 0 2 * * *, dispatch. prerelease: PR to main on SDK paths, cron 0 2 * * *, dispatch. yaml.safe_load passes for both files. actionlint was not available on the devbox, so it did not run.

Merge Danger

Door: two-way

Blast Radius: CI

report-compatibility is the only required check. It lives in mcp-compatibility.yml, so that workflow uses job gating and the check still runs on every PR. A docs-only PR now skips all compat matrices and the prerelease workflow.

Fix: test-without-fastmcp (1.9.4) red on main since 2026-11-02

anyio 4.16.0 (published 2026-11-01 11:20 UTC) now raises ClosedResourceError in a receive() that waits on a memory stream when the stream closes. mcp below 1.10 does not catch that in BaseSession._receive_loop. The first client disconnect kills the Streamable-HTTP session manager's task group, and every later request gets a 500. mcp 1.10.0 added the except anyio.ClosedResourceError guard. Only the 1.9.4 leg breaks, because tests/conftest.py runs e2e/official only on mcp 1.9.2 and later. The bug is upstream: a plain mcp 1.9.4 server without agentcat.track() shuts down the same way.

mcp-compatibility.yml, test-without-fastmcp install step:

         if [[ "${{ matrix.mcp-version }}" == 1.* ]]; then
           uv pip install "pydantic<2.12"
         fi
+
+        # anyio 4.16 ... mcp 1.10.0 added the guard. Keep those legs on anyio 4.15.
+        if [[ "${{ matrix.mcp-version }}" == 1.[0-9].* ]]; then
+          uv pip install "anyio<4.16"
+        fi
@@ Show installed MCP version
         uv pip show mcp | grep "Version:" || echo "MCP not found"
+        uv pip show anyio | grep "Version:" || echo "anyio not found"

Evidence

• Before: RuntimeError: Task group is not initialized. Make sure to use run(). from mcp/server/streamable_http_manager.py:137; 24 e2e HTTP tests fail (PR run, main #582). Main #579 and #580 ran the same commit (8bb1a65); the only package changes between them were anyio 4.15.1 → 4.16.0, pytest-asyncio 1.4.1 → 1.4.1.post1, python-dateutil 3.0.0 → 4.0.0. Locally, with CI's exact package set, only swapping anyio flips tests/e2e/official (24 failed ↔ 32 passed).
• After: locally, mcp 1.9.4 + anyio 4.15.1: 868 passed, 30 skipped. PR CI: see the checks on this PR.

Merge Danger

Door: two-way

Blast Radius: CI

The pin only applies to the mcp 1.0–1.9 legs of test-without-fastmcp; mcp 1.10+ and 2.x legs still test the newest anyio. The pin hides a real upstream problem: a Customer on mcp 1.9.x who installs fresh today gets anyio 4.16, and their Streamable-HTTP server stops serving after the first client disconnects. AgentCat does not cause or fix that. The package metadata and README stay unchanged; whether to tell Customers "use mcp>=1.10 or anyio<4.16 with Streamable HTTP" is a separate owner decision.