Troubleshooting

MCP Container: no running task found: task mcp-xxx not found

Symptom

When a bot tries to use container tools (e.g., execute commands), the server logs show:

level=WARN msg="exec failed" provider=container_tool bot_id=xxx command=date error="no running task found: task mcp-xxx not found"

The containerd container logs may also show:

level=error msg="failed to delete task" error="rpc error: code = NotFound desc = container not created: not found"

Cause

The [mcp] data_root in config.toml is set to a host machine path (e.g., /Users/you/Code/Memoh/data), but the server and containerd containers use a Docker named volume mounted at /opt/memoh/data.

When the server creates an MCP container inside containerd, it uses data_root as the mount source. Since this host path does not exist inside the containerd container, runc fails with:

failed to fulfil mount request: open /Users/you/Code/Memoh/data/bots/xxx: no such file or directory

Solution

1. Set data_root to the in-container path in your config:

[mcp]
data_root = "/opt/memoh/data"

2. Clean up the stale containerd container (if it exists):

docker exec memoh-containerd ctr -n default containers rm mcp-<bot-id>

3. Restart the server:

docker compose restart server

Note: If you also run the server locally (outside Docker), keep the Docker config (conf/app.docker.toml) separate from your local config.toml, and update docker-compose.yml to mount the Docker-specific config instead.

MCP Container: Image update not taking effect after rebuild

Symptom

After updating Dockerfile.containerd (e.g., adding Node.js/Python to the MCP image), rebuilding and restarting the containerd container, MCP tools still fail with errors like:

exec: "npx": executable file not found in $PATH

Cause

The containerd entrypoint script (containerd-entrypoint.sh) skips image import if the image already exists in containerd's image store:

if ! ctr -n default images check "name==${MCP_IMAGE}" ...; then
  # import
fi

Since containerd_data is a persistent Docker volume, the old MCP image survives across container restarts. The new image embedded in the rebuilt Docker image is never imported.

Solution

1. Remove the old MCP image from containerd:

docker exec memoh-containerd ctr -n default images rm docker.io/library/memoh-mcp:latest

2. Restart the containerd container to trigger re-import:

docker compose restart containerd

3. Verify the new image was imported (size should be significantly larger if Node.js/Python were added):

docker exec memoh-containerd ctr -n default images ls

4. Delete the bot's MCP container and recreate it from the bot detail page so it uses the new image.