Maintainer Notes

Notes lane for rough ideas, investigation breadcrumbs, and hand-written reminders.

Open questions

• Prompt-cache minimum is a per-model fact, not a per-provider one. It lives on ProviderCapabilities.promptCacheMinTokens today, which is right for Anthropic (4096 on Opus) but will need a per-model table if a provider ever varies it across its own models.
• estimateTokens is chars/4. Good enough for the pre-flight spend estimate and the cache-minimum guard, but it is an estimate. An adapter that implements countTokens() should be preferred where precision matters.
• Response-cache coalescing is per-process. Fine for a single home server; a multi-process deployment would want the shared store F4 adds.