Prompt Caching

See also the design-level concept doc at docs/concepts/ai/caching.md. pi-go controls caching with StreamOptions.CacheRetention (none, short, long) and a SessionID for affinity.

Compatibility

Provider	API	pi-go	Notes
Anthropic	✅ cache_control markers; 5m default + 1h ttl (extended TTL beta); up to 4 breakpoints	⚠️	terminal auto-breakpoint only (cache.go:50-74); 1h only on api.anthropic.com (cache.go:29-30)
OpenAI Chat	✅ automatic; prompt_cache_key for affinity	✅	session ID forwarded (openai.go:323-326)
OpenAI Responses	✅ automatic	✅	session ID forwarded (openairesponses.go:339-340)
Google Gemini	✅ context caching API	❌	not wired
Claude CLI	⚠️ session-level KV	❌

Provider Documentation

• Anthropic — Prompt caching
• OpenAI — Prompt caching
• Google Gemini — Context caching

pi-go Gaps

• Manual breakpoints for Anthropic (system prompt, tool definitions, prior turns) — pi-go places exactly one terminal marker.
• Google context caching not implemented at all (no cachedContent reference).
• Cache hit metrics for OpenAI Chat are not surfaced — streaming usage collapses to a single TotalTokens field (openai.go:94).
• Anthropic cache write tier — pi-go does not let callers choose 1h cache writes selectively (it’s tied to CacheRetention).