@zanii/blackbox 0.21.1 → 0.21.2

package/dist/cost/index.d.ts

@@ -17,6 +17,9 @@ export interface ModelPrice {
     /** Audio in and out (else `input` / `output`). */
     input_audio?: number;
     output_audio?: number;
+    /** Image tokens in and out (else `input` / `output`): OpenAI's and Gemini's image models. */
+    input_image?: number;
+    output_image?: number;
     /** µUSD per 1,000 web-search requests (Anthropic's server tool); absent: not priced. */
     web_search_per_1k?: number;
     /** Prices once the call's prompt is over `above` tokens (Gemini's >200k tier), highest first. */
@@ -56,6 +59,8 @@ export interface Tokens {
     cache_write_1h?: number;
     input_audio?: number;
     output_audio?: number;
+    input_image?: number;
+    output_image?: number;
 }
 export interface CallCost {
     seq: number;
@@ -74,6 +79,8 @@ export interface CallCost {
     /** spec/cost.md §1a: the uplift applied, when one matched. */
     uplift_percent?: number;
     source?: "reported";
+    /** The answer carried no usage (`usage_missing`): unpriced, never counted as free. */
+    no_usage?: true;
 }
 export interface CostReport {
     v: 1;

package/dist/cost/index.js

@@ -2,12 +2,20 @@
 // Mirrors sdks/python/src/zanii_blackbox/cost.py.
 import { createHash } from "node:crypto";
 const CLASSES = ["input", "output", "cache_read", "cache_write"];
-const EXTRA = ["cache_write_1h", "input_audio", "output_audio"];
+const EXTRA = [
+    "cache_write_1h",
+    "input_audio",
+    "output_audio",
+    "input_image",
+    "output_image",
+];
 const OPTIONAL_PRICES = [
     "context_window",
     "cache_write_1h",
     "input_audio",
     "output_audio",
+    "input_image",
+    "output_image",
     "web_search_per_1k",
 ];
 /** Parses and checks a price table; its sha256 is of the exact file bytes. */
@@ -99,12 +107,20 @@ export function tokensOf(usage) {
         const cached = n("cachedContentTokenCount");
         const audioIn = n("promptTokensDetails_AUDIO");
         const audioOut = n("candidatesTokensDetails_AUDIO");
+        const imageIn = n("promptTokensDetails_IMAGE");
+        const imageOut = n("candidatesTokensDetails_IMAGE");
         return extras({
-            input: Math.max(0, n("promptTokenCount") - cached - audioIn) + n("toolUsePromptTokenCount"),
-            output: Math.max(0, n("candidatesTokenCount") - audioOut) + n("thoughtsTokenCount"),
+            input: Math.max(0, n("promptTokenCount") - cached - audioIn - imageIn) +
+                n("toolUsePromptTokenCount"),
+            output: Math.max(0, n("candidatesTokenCount") - audioOut - imageOut) + n("thoughtsTokenCount"),
             cache_read: cached,
             cache_write: 0,
-        }, { input_audio: audioIn, output_audio: audioOut });
+        }, {
+            input_audio: audioIn,
+            output_audio: audioOut,
+            input_image: imageIn,
+            output_image: imageOut,
+        });
     }
     if ("inputTokens" in usage)
         // Bedrock Converse (spec/providers.md §5)
@@ -130,18 +146,28 @@ export function tokensOf(usage) {
     // Anthropic's cache names are its own: they win over the Responses shape, since Anthropic also
     // sends output_tokens_details (thinking_tokens) since Claude Sonnet 5
     const anthropicCache = "cache_read_input_tokens" in usage || "cache_creation_input_tokens" in usage;
+    // OpenAI's transcriptions write input_token_details (singular); its images, image_tokens
     if (!anthropicCache &&
-        Object.keys(usage).some((k) => k.startsWith("input_tokens_details_") || k.startsWith("output_tokens_details_"))) {
+        Object.keys(usage).some((k) => k.startsWith("input_tokens_details_") ||
+            k.startsWith("output_tokens_details_") ||
+            k.startsWith("input_token_details_"))) {
         const cached = n("input_tokens_details_cached_tokens");
         const written = n("input_tokens_details_cache_write_tokens");
-        const audioIn = n("input_tokens_details_audio_tokens");
+        const audioIn = n("input_tokens_details_audio_tokens") + n("input_token_details_audio_tokens");
         const audioOut = n("output_tokens_details_audio_tokens");
+        const imageIn = n("input_tokens_details_image_tokens");
+        const imageOut = n("output_tokens_details_image_tokens");
         return extras({
-            input: Math.max(0, n("input_tokens") - cached - written - audioIn),
-            output: Math.max(0, n("output_tokens") - audioOut),
+            input: Math.max(0, n("input_tokens") - cached - written - audioIn - imageIn),
+            output: Math.max(0, n("output_tokens") - audioOut - imageOut),
             cache_read: cached,
             cache_write: written,
-        }, { input_audio: audioIn, output_audio: audioOut });
+        }, {
+            input_audio: audioIn,
+            output_audio: audioOut,
+            input_image: imageIn,
+            output_image: imageOut,
+        });
     }
     // Anthropic: cache_creation_input_tokens is both lifetimes; the 1-hour part is priced apart
     const hour = n("cache_creation_ephemeral_1h_input_tokens");
@@ -167,7 +193,12 @@ export function requestsOf(usage) {
 }
 /** A call's prompt, in tokens: what a context tier is decided on. */
 export function promptTokens(t) {
-    return t.input + t.cache_read + t.cache_write + (t.cache_write_1h ?? 0) + (t.input_audio ?? 0);
+    return (t.input +
+        t.cache_read +
+        t.cache_write +
+        (t.cache_write_1h ?? 0) +
+        (t.input_audio ?? 0) +
+        (t.input_image ?? 0));
 }
 /** The prices a call pays: a context tier when its prompt is over one, then its service tier. */
 export function effectivePrice(price, tokens, serviceTier) {
@@ -203,6 +234,8 @@ export function callMicroUsd(tokens, price, requests) {
     sum += (tokens.cache_write_1h ?? 0) * (price.cache_write_1h ?? price.cache_write);
     sum += (tokens.input_audio ?? 0) * (price.input_audio ?? price.input);
     sum += (tokens.output_audio ?? 0) * (price.output_audio ?? price.output);
+    sum += (tokens.input_image ?? 0) * (price.input_image ?? price.input);
+    sum += (tokens.output_image ?? 0) * (price.output_image ?? price.output);
     sum += (requests?.web_search ?? 0) * (price.web_search_per_1k ?? 0) * 1000;
     if (!Number.isSafeInteger(sum))
         throw new RangeError("cost overflow: tokens × price beyond 2^53");
@@ -288,9 +321,19 @@ export function sessionCost(lines, prices) {
         if (replayed.has(e.meta.request_seq))
             continue;
         const usage = e.meta.usage;
-        if (typeof usage !== "object" || usage === null)
-            continue;
         const model = typeof e.meta.model === "string" ? e.meta.model : null;
+        if (typeof usage !== "object" || usage === null) {
+            // an answer that said nothing of what it cost is unpriced, never silently free
+            if (e.meta.usage_missing === true)
+                calls.push({
+                    seq: e.seq,
+                    model,
+                    tokens: { input: 0, output: 0, cache_read: 0, cache_write: 0 },
+                    micro_usd: null,
+                    no_usage: true,
+                });
+            continue;
+        }
         const priced = priceCall({
             model,
             usage: usage,

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "0.21.1";
+export declare const VERSION = "0.21.2";

package/dist/version.js

@@ -1,2 +1,2 @@
 // The SDK version, on its own so the CLI can print it without loading the whole SDK.
-export const VERSION = "0.21.1";
+export const VERSION = "0.21.2";

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@zanii/blackbox",
   "license": "Apache-2.0",
-  "version": "0.21.1",
+  "version": "0.21.2",
   "description": "The flight recorder for AI agents: sessions, a zero-loss spool, framework hooks, approvals, and offline verification of the gateway's hash-chained record.",
   "keywords": [
     "ai-agents",