@vincemakes/kiso-provider-anthropic 0.13.0 → 0.15.0

package/dist/index.d.ts

@@ -18,6 +18,8 @@ import type { Adapter } from "@vincemakes/kiso-core";
 export interface AnthropicProviderConfig {
     readonly apiKey?: string;
     readonly baseUrl?: string;
+    /** PH-1c.1: opt-in prompt caching (see AnthropicAdapterOptions). */
+    readonly promptCaching?: boolean;
 }
 /**
  * High-level factory (round 7): builds the adapter FROM CONFIG, owning the SDK
@@ -27,4 +29,14 @@ export interface AnthropicProviderConfig {
  * never through a hoisted root.
  */
 export declare function createAnthropicProvider(config?: AnthropicProviderConfig): Adapter;
-export declare function createAnthropicAdapter(client: Anthropic): Adapter;
+/** PH-1c.1 (finding PH-F13): opt-in prompt caching. OFF by default —
+ *  injecting cache_control changes the request bytes, and a cost
+ *  behavior does not flip silently (the default flips only after the
+ *  paired bench proves it at a release). ON places exactly two
+ *  ephemeral breakpoints: the system block, and a ROLLING one on the
+ *  last message's last content block — the next request's stable
+ *  prefix reaches it, which is the whole trick. */
+export interface AnthropicAdapterOptions {
+    readonly promptCaching?: boolean;
+}
+export declare function createAnthropicAdapter(client: Anthropic, adapterOpts?: AnthropicAdapterOptions): Adapter;

package/dist/index.js

@@ -26,9 +26,10 @@ export function createAnthropicProvider(config = {}) {
         ...(config.apiKey !== undefined ? { apiKey: config.apiKey } : {}),
         ...(config.baseUrl !== undefined ? { baseURL: config.baseUrl } : {}),
     });
-    return createAnthropicAdapter(client);
+    return createAnthropicAdapter(client, { ...(config.promptCaching !== undefined ? { promptCaching: config.promptCaching } : {}) });
 }
-export function createAnthropicAdapter(client) {
+export function createAnthropicAdapter(client, adapterOpts = {}) {
+    const caching = adapterOpts.promptCaching === true;
     return {
         async *stream(options) {
             // D4: the stream CREATION is inside the error normalization —
@@ -38,9 +39,15 @@ export function createAnthropicAdapter(client) {
             try {
                 stream = client.messages.stream({
                     model: options.model,
-                    ...(options.systemPrompt !== undefined ? { system: options.systemPrompt } : {}),
+                    ...(options.systemPrompt !== undefined
+                        ? {
+                            system: caching
+                                ? [{ type: "text", text: options.systemPrompt, cache_control: { type: "ephemeral" } }]
+                                : options.systemPrompt,
+                        }
+                        : {}),
                     max_tokens: options.maxTokens ?? 4096,
-                    messages: toAnthropicMessages(options.messages),
+                    messages: caching ? withCacheBreakpoint(toAnthropicMessages(options.messages)) : toAnthropicMessages(options.messages),
                     ...(options.tools?.length ? { tools: toAnthropicTools(options.tools) } : {}),
                     ...(options.temperature !== undefined ? { temperature: options.temperature } : {}),
                 }, 
@@ -262,6 +269,22 @@ function toAnthropicMessages(messages) {
     }
     return out;
 }
+/** PH-1c.1 — the rolling conversation breakpoint: the LAST message's
+ *  LAST content block gains cache_control, so the next request's
+ *  stable prefix (this whole conversation) is a cache hit. String
+ *  content normalizes to one text block; an empty message is left
+ *  alone. Pure — the input array is never mutated. */
+function withCacheBreakpoint(messages) {
+    const last = messages[messages.length - 1];
+    if (last === undefined)
+        return messages;
+    const blocks = typeof last.content === "string" ? [{ type: "text", text: last.content }] : [...last.content];
+    const tail = blocks[blocks.length - 1];
+    if (tail === undefined)
+        return messages;
+    blocks[blocks.length - 1] = { ...tail, cache_control: { type: "ephemeral" } };
+    return [...messages.slice(0, -1), { ...last, content: blocks }];
+}
 function toAnthropicBlock(block) {
     return block.type === "text"
         ? { type: "text", text: block.text }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vincemakes/kiso-provider-anthropic",
-  "version": "0.13.0",
+  "version": "0.15.0",
   "description": "kiso Anthropic adapter — official SDK wire events mapped to the kiso event union.",
   "type": "module",
   "license": "MIT",
@@ -22,7 +22,7 @@
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.115.0",
-    "@vincemakes/kiso-core": "0.13.0"
+    "@vincemakes/kiso-core": "0.15.0"
   },
   "devDependencies": {
     "@types/node": "^26.1.2",