SheepCat hits Ollama speed wall

Hot take: this is a UX latency problem disguised as model tuning. If the user is staring at the summary, SheepCat needs a separate fast-path for end-of-day recap, not just a bigger prompt.

• –Async background logging and synchronous review are different workloads; they should not share the same inference budget.
• –A smaller, more aggressively quantized summarizer is probably the quickest win if the output only needs to be clear and actionable.
• –Prompt or context caching can trim overhead, but it won’t fix a model that is simply too slow on the available hardware.
• –The best architecture for a local-first app is likely staged aggregation: store structured snippets during the day, then summarize a much smaller payload at shutdown.
• –That keeps the privacy-first promise intact while making the wait feel human-scale instead of lab-demo scale.