Summary


This PR adds DeepSeek prompt cache awareness to the TUI and introduces a new /cache inspect command for diagnosing cache-related prompt structure.

The main goal is to make DeepSeek context caching easier to reason about by separating stable reusable prompt prefixes from session history and dynamic request content.


Changes



DeepSeek cache-aware prompt structure



• Introduced a cache-oriented prompt layering model.
• Separates prompt content into clearer categories:
• static base prefix
• session history
• dynamic request content
• Keeps reusable prompt components at the front of the request where they are most likely to benefit from DeepSeek context caching.
  

Cache usage metrics



• Parses DeepSeek cache usage fields when available:
• prompt_cache_hit_tokens
• prompt_cache_miss_tokens
• Computes and displays cache hit information for recent requests.
• Handles providers or responses without cache fields gracefully.
  

/cache inspect



• Adds a /cache inspect command to inspect the rendered prompt structure without printing the full prompt text.
• Displays SHA-256 hashes for rendered prompt layers.
• Adds:
• Base static prefix hash
• Full request prefix hash
• static prefix stability status
• first divergence from the previous request
• Classifies rendered layers as:
• static
• history
• dynamic
• Helps distinguish expected request-history changes from actual static-prefix instability.
  

Safety and privacy



• /cache inspect does not print full prompt contents by default.
• Hashes are used for diagnostics instead of exposing full rendered prompt text.
  

Motivation


DeepSeek context caching can significantly reduce input cost when requests share a stable prefix. However, before this change it was difficult to tell whether cache misses were caused by changes in the reusable static prefix or by normal conversation history growth.

This PR adds both the cache-aware prompt design and the inspection tooling needed to debug cache behavior in real sessions.

In particular, /cache inspect makes it possible to verify that the static base prefix remains stable across turns while allowing the full rendered request to change as history, tool results, and user inputs evolve.


Example behavior


Across multiple turns in the same session, /cache inspect can now show:


• the same Base static prefix hash
• different Full request prefix hash
• Static base prefix stability: OK
• the first changed layer in the session history or dynamic request area
  
  This makes it easier to confirm that the reusable DeepSeek cache prefix is stable even when the full request changes.
  

Testing


Manual verification performed:


• Asked multiple questions in the same session.
• Ran /cache inspect after each request.
• Confirmed Base static prefix hash remains stable across turns.
• Confirmed Full request prefix hash changes as history grows.
• Confirmed cache hit / miss metrics are displayed when DeepSeek returns cache usage fields.
• Confirmed full prompt text is not printed by /cache inspect.