Skip to content

A few experiments with a Cotabby fork #829

Description

@mc-hamster

I’ve been working on an experimental fork of Cotabby, with plenty of help from GPT-6 Astra along the way. I’ve made some fairly substantial changes around prompt-cache reuse, finishing partially typed words, ghost-text stability, refreshed local screen context, and how dismissed suggestions are handled.

I also put together a 1,337-case phrase-prediction benchmark. On that benchmark, the changes brought exact next-word accuracy from roughly the upper 20s/low 30s to around 37%. It’s a controlled synthetic test—not a claim about every model or writing situation—but it’s been useful for checking whether an idea actually helps.

I'm using the 1,337-case benchmark as part of an agentic hypothesis-driven development loop: come up with an idea, test an idea, measure it against the baseline, keep what works, and iterate again.

I’ve also made signed and notarized release builds. If you’re curious, you’re welcome to try one:
https://github.com/mc-hamster/cotabby/releases

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    • Status
      Backlog

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions