@foldrun/core 2.0.0 → 2.1.0

package/src/translator.ts

@@ -22,7 +22,9 @@
 // and their results, streaming — and drops what a Chat-Completions endpoint
 // has no word for (cache_control, top_k, server-side tools, thinking on
 // endpoints that lack reasoning_effort). Anything dropped is said once in
-// the run trace. The translation is pure functions over JSON so the tests
+// the run trace. Those endpoints cache prompt prefixes on their own side;
+// the cached share they report comes back as cache_read_input_tokens, so a
+// step's record says how much of its context was re-read. The translation is pure functions over JSON so the tests
 // can drive every mapping without a network; the server around them is a
 // few dozen lines.
 
@@ -37,6 +39,34 @@ import { captureRefusalHeaders, isRefusalStatus, refusalLine } from "./refusal.t
 
 type Json = Record<string, unknown>;
 
+/**
+ * An OpenAI-shaped usage object as Anthropic's: fresh input, cached input,
+ * output. Both OpenAI shapes count the cached part INSIDE the prompt total
+ * (Chat's prompt_tokens_details.cached_tokens, Responses'
+ * input_tokens_details.cached_tokens; DeepSeek says prompt_cache_hit_tokens),
+ * so it is taken out of input here or the step would count it twice.
+ */
+export function anthropicUsage(u: Json | undefined, wire: TranslatorWire): { input: number; cached: number; output: number } {
+  if (!u) return { input: 0, cached: 0, output: 0 };
+  const total = Number((wire === "chat" ? u.prompt_tokens : u.input_tokens) ?? 0);
+  const details = (wire === "chat" ? u.prompt_tokens_details : u.input_tokens_details) as Json | undefined;
+  const raw = Number(details?.cached_tokens ?? u.prompt_cache_hit_tokens ?? 0);
+  const cached = Number.isFinite(raw) ? Math.max(0, Math.min(raw, total)) : 0;
+  return {
+    input: total - cached,
+    cached,
+    output: Number((wire === "chat" ? u.completion_tokens : u.output_tokens) ?? 0),
+  };
+}
+
+function usageBlock(u: { input: number; cached: number; output: number } | null): Json {
+  return {
+    input_tokens: u?.input ?? 0,
+    output_tokens: u?.output ?? 0,
+    ...(u && u.cached > 0 ? { cache_read_input_tokens: u.cached } : {}),
+  };
+}
+
 export type TranslatorWire = "chat" | "responses";
 
 export interface TranslatorSpec {
@@ -493,10 +523,7 @@ export function fromChatCompletion(res: Json, requestedModel: string, stopSequen
     content,
     stop_reason: stop.reason,
     stop_sequence: stop.sequence,
-    usage: {
-      input_tokens: Number(usage.prompt_tokens ?? 0),
-      output_tokens: Number(usage.completion_tokens ?? 0),
-    },
+    usage: usageBlock(anthropicUsage(usage, "chat")),
   };
 }
 
@@ -515,7 +542,7 @@ export class StreamTranslator {
   private tools = new Map<number, { block: number; argsSeen: boolean }>();
   /** The last choice's stop fields: finish_reason and any matched stop. */
   private finish: Json = {};
-  private usage: { input: number; output: number } | null = null;
+  private usage: { input: number; cached: number; output: number } | null = null;
   private id = `msg_${crypto.randomBytes(8).toString("hex")}`;
   private readonly model: string;
   private readonly stops: string[];
@@ -561,7 +588,7 @@ export class StreamTranslator {
     const out = this.start(chunk);
     const u = chunk.usage as Json | undefined;
     if (u && (u.prompt_tokens !== undefined || u.completion_tokens !== undefined)) {
-      this.usage = { input: Number(u.prompt_tokens ?? 0), output: Number(u.completion_tokens ?? 0) };
+      this.usage = anthropicUsage(u, "chat");
     }
     const choice = ((chunk.choices as Json[] | undefined) ?? [])[0];
     if (!choice) return out;
@@ -650,7 +677,7 @@ export class StreamTranslator {
       this.event("message_delta", {
         type: "message_delta",
         delta: { stop_reason: stop.reason, stop_sequence: stop.sequence },
-        usage: { input_tokens: this.usage?.input ?? 0, output_tokens: this.usage?.output ?? 0 },
+        usage: usageBlock(this.usage),
       }),
     );
     out.push(this.event("message_stop", { type: "message_stop" }));
@@ -691,10 +718,7 @@ export function fromResponses(res: Json, requestedModel: string): Json {
     content,
     stop_reason: stop,
     stop_sequence: null,
-    usage: {
-      input_tokens: Number(usage.input_tokens ?? 0),
-      output_tokens: Number(usage.output_tokens ?? 0),
-    },
+    usage: usageBlock(anthropicUsage(usage, "responses")),
   };
 }
 
@@ -710,7 +734,7 @@ export class ResponsesStreamTranslator {
   private nextIndex = 0;
   /** Responses output_index → Anthropic block index and kind. */
   private blocks = new Map<number, { block: number; kind: "text" | "tool"; argsSeen: boolean }>();
-  private usage: { input: number; output: number } | null = null;
+  private usage: { input: number; cached: number; output: number } | null = null;
   private incomplete: string | null = null;
   private sawTool = false;
   private sawRefusal = false;
@@ -812,7 +836,7 @@ export class ResponsesStreamTranslator {
     if (type === "response.completed" || type === "response.incomplete" || type === "response.failed") {
       const r = (ev.response as Json | undefined) ?? {};
       const u = r.usage as Json | undefined;
-      if (u) this.usage = { input: Number(u.input_tokens ?? 0), output: Number(u.output_tokens ?? 0) };
+      if (u) this.usage = anthropicUsage(u, "responses");
       const reason = (r.incomplete_details as Json | undefined)?.reason;
       if (typeof reason === "string") this.incomplete = reason;
       return out;
@@ -834,7 +858,7 @@ export class ResponsesStreamTranslator {
       this.event("message_delta", {
         type: "message_delta",
         delta: { stop_reason: reason, stop_sequence: null },
-        usage: { input_tokens: this.usage?.input ?? 0, output_tokens: this.usage?.output ?? 0 },
+        usage: usageBlock(this.usage),
       }),
     );
     out.push(this.event("message_stop", { type: "message_stop" }));