Define dependency-aware invalidation for static prompt cache

[zikajk]

To upvote this issue, give it a thumbs up. See this list for the most upvoted issues.

Static prompt caching improves provider-side cache stability, but several inputs that contribute to static instructions can change during a chat: rules, path-scoped rule catalogs, skills, agent prompts, workspace metadata, tool availability used by Selmer variables, and user-created files such as new skills or rules. Today cache invalidation is mostly scoped to agent/model changes, which risks stale system instructions or over-invalidation if we include too much.

Design and implement a clear invalidation model for static prompt inputs:

• Identify all inputs that contribute to static instructions.
• Decide which changes should automatically invalidate the cache and which should require an explicit user action/command.
• Consider workflows like creating a new skill/rule during a chat: should it become available immediately, after /reload, after a new chat, or after explicit command?
• Include path-scoped rules and fetch_rule catalog availability in the model.
• Include Selmer-rendered variables such as workspace roots, subagent state, and tool availability.
• Consider provider-side invalidation as well (OpenAI, Anthropic).
• Add tests that verify cache reuse vs invalidation for representative cases.