Hi Mikkel,
Thanks again for your contribution of the new sampler to PyMC3.
As for the pCN algorithm, I think you can take a look at pymc3.smc, where it isolate the likelihood term alone: https://github.com/pymc-devs/pymc3/blob/ee57bed9aa5f73dc48081dd3c82267dd4f0f2834/pymc3/smc/smc.py#L127
My recommendation is to follow the pattern in SMC and program it into a specialized sampler, you can initialized a class method that contains the state of the last likelihood.