Thank you!!
I had pi write me a ralph loop plugin. You plan with it, it breaks it down into tasks. Task success means clearing the context - if it runs out of context, it clears and starts on the same task again (code changes from previous run are left intact).
Complete with little docker-like names for Ralph sessions and the ability to resume sessions should you need to.
Point is, I would generally do this because it costs less context I think, and better tracks progress.
Ping me if you are interested and want it. Otherwise, plenty around :)
https://github.com/jwdevantier/ralphish
:)
Remember, I didn't write a word of it - so if anything is not to your liking, it's no harder than pointing a model at it to fix it. One of the many powers of Pi is the way plugin- and API documentation is bundled in.
Most likely context length parameters aren’t set quite right for your model/provider and you genuinely run out of context during compaction