Lag length VAR/VECM; Causality testing procedures

[mono]

Dear pro-users,
I'm investigating daily data on commodity price series (5-day week, N=3,000, all I(1)). I use logarithmic levels and logarithmic first differences (viz. returns).

I could need some clarity with respect to the given approach. I really appreciate any commnent!

1. Use log level series in a VAR, irrespective of the order of integration, and AIC/SIC to determine the optimal lag length "p".
2. Run Johansen cointegration test with the above specified lag length p.
3. If cointegration is present, estimate a VECM of order (p-1).
4. Use residuals for further specification checking.

However, the point in case is that an optimal lag length for a VAR in levels with p=1 yields a lag length of p=0 in the first differenced series, hence also for a potential VEC model. Is that true indeed?
This does not make any sense and I wonder if I made any misstakes. Furthermore causality testing based on a stationary VAR is no issue, but how to do it within a VECM?
Use "Block exogenenity tests" instead? This also needs at least a lag length of one in the VAR specification.

Thank you for any comments upon this issue.