anthropic prompt caching

1 files changed, 12 insertions, 0 deletions

diff --git a/libpanto/src/config.zig b/libpanto/src/config.zig
index 4aae81f..7e1e8e1 100644
--- a/libpanto/src/config.zig
+++ b/libpanto/src/config.zig
@@ -101,6 +101,18 @@ pub const AnthropicMessagesConfig = struct {
     /// between tool calls. Ignored when `thinking == .adaptive` (interleaving
     /// is automatic there) or `.disabled`.
     thinking_interleaved: bool = false,
+    /// Place one `cache_control` breakpoint on the last cacheable block of
+    /// each request, replicating Anthropic's "automatic caching" (a single
+    /// advancing breakpoint) via the broadly-supported per-block marker
+    /// rather than the top-level field.
+    ///
+    /// Defaults on: for a normal append-only multi-turn session it's a
+    /// near-pure win (reads 0.1x base input vs. a one-time 1.25x write).
+    /// Set false when the workload makes that write premium unrecoverable
+    /// — e.g. one-shot requests, or an embedder that aggressively rewrites
+    /// history so prefixes are never reused. There the 1.25x write is pure
+    /// overhead with no read to amortize it.
+    prompt_cache: bool = true,
 };
 
 /// Per-provider transport/auth/model configuration. Tagged by `APIStyle`.