mohdel 0.113.0 → 0.114.1

package/js/core/events.js

@@ -87,6 +87,10 @@
  *   Input tokens written to a fresh prompt cache breakpoint, billed at
  *   `cacheWritePrice`. Absent when the provider has no separate
  *   cache-write counter.
+ * @property {number} [cacheWrite1hInputTokens]
+ *   The 1h-TTL subset of `cacheWriteInputTokens` (Anthropic), billed at
+ *   `cacheWrite1hPrice`; the 5m portion is the remainder. A request may mix
+ *   TTLs. Absent when the turn has no 1h-TTL cache writes.
  * @property {number} [cacheReadInputTokens]
  *   Input tokens served from prompt cache, billed at `cacheReadPrice`.
  *   Absent when the provider has no prompt caching.

package/js/session/adapters/_pricing.js

@@ -27,8 +27,15 @@ import { getSpec, setCatalog } from './_catalog.js'
  * Optional fields fall back to other prices when absent:
  *   - `thinkingPrice` → `outputPrice`
  *   - `cacheWritePrice` → `inputPrice` (graceful for non-caching providers)
+ *   - `cacheWrite1hPrice` → `cacheWritePrice` (graceful when only one write tier is priced)
  *   - `cacheReadPrice` → `inputPrice`
  *
+ * `cacheWrite1hInputTokens` is the 1h-TTL subset of `cacheWriteInputTokens`
+ * (Anthropic bills 1h writes at a higher rate than 5m). The 5m portion is
+ * `cacheWriteInputTokens - cacheWrite1hInputTokens`. When the 1h field is
+ * absent or the catalog has no `cacheWrite1hPrice`, the whole write bills at
+ * `cacheWritePrice` — identical to the single-tier behaviour.
+ *
  * Token-counting convention: this function is purely additive across
  * `inputTokens`, `cacheWriteInputTokens`, `cacheReadInputTokens`,
  * `outputTokens`, and `thinkingTokens`. Adapters normalize provider-specific
@@ -36,7 +43,8 @@ import { getSpec, setCatalog } from './_catalog.js'
  *
  * @param {any} spec  Catalog entry, or `undefined`.
  * @param {{inputTokens?: number, outputTokens?: number, thinkingTokens?: number,
- *          cacheWriteInputTokens?: number, cacheReadInputTokens?: number}} usage
+ *          cacheWriteInputTokens?: number, cacheWrite1hInputTokens?: number,
+ *          cacheReadInputTokens?: number}} usage
  * @returns {number}
  */
 export function computeCost (spec, usage) {
@@ -45,6 +53,8 @@ export function computeCost (spec, usage) {
   const o = usage.outputTokens ?? 0
   const t = usage.thinkingTokens ?? 0
   const cw = usage.cacheWriteInputTokens ?? 0
+  const cw1h = Math.min(cw, usage.cacheWrite1hInputTokens ?? 0)
+  const cw5m = Math.max(0, cw - cw1h)
   const cr = usage.cacheReadInputTokens ?? 0
   const ip = resolveTier(spec.inputPrice, i)
   const op = resolveTier(spec.outputPrice, i)
@@ -53,9 +63,11 @@ export function computeCost (spec, usage) {
   const tpFinal = typeof tp === 'number' ? tp : op
   const cwp = resolveTier(spec.cacheWritePrice, i)
   const cwpFinal = typeof cwp === 'number' ? cwp : ip
+  const cw1hp = resolveTier(spec.cacheWrite1hPrice, i)
+  const cw1hpFinal = typeof cw1hp === 'number' ? cw1hp : cwpFinal
   const crp = resolveTier(spec.cacheReadPrice, i)
   const crpFinal = typeof crp === 'number' ? crp : ip
-  const total = (i * ip + cw * cwpFinal + cr * crpFinal + o * op + t * tpFinal) / 1_000_000
+  const total = (i * ip + cw5m * cwpFinal + cw1h * cw1hpFinal + cr * crpFinal + o * op + t * tpFinal) / 1_000_000
   return round(total)
 }
 

package/js/session/adapters/anthropic.js

@@ -114,6 +114,7 @@ export async function * anthropic (envelope, deps = {}) {
   let inputTokens = 0
   let outputTokens = 0
   let cacheWriteTokens = 0
+  let cacheWrite1hTokens = 0
   let cacheReadTokens = 0
   let thinkingChars = 0
   let status = STATUS_COMPLETED
@@ -140,6 +141,12 @@ export async function * anthropic (envelope, deps = {}) {
           if (event.message?.usage?.cache_creation_input_tokens) {
             cacheWriteTokens = event.message.usage.cache_creation_input_tokens
           }
+          // cache_creation breaks the write total down by TTL (5m + 1h).
+          // Surface the 1h subset so the cost layer can price it separately;
+          // a request may mix TTLs, so both tiers can be non-zero.
+          if (event.message?.usage?.cache_creation?.ephemeral_1h_input_tokens) {
+            cacheWrite1hTokens = event.message.usage.cache_creation.ephemeral_1h_input_tokens
+          }
           if (event.message?.usage?.cache_read_input_tokens) {
             cacheReadTokens = event.message.usage.cache_read_input_tokens
           }
@@ -249,6 +256,7 @@ export async function * anthropic (envelope, deps = {}) {
       outputTokens: messageOutputTokens,
       thinkingTokens: estimatedThinkingTokens,
       ...(cacheWriteTokens > 0 && { cacheWriteInputTokens: cacheWriteTokens }),
+      ...(cacheWrite1hTokens > 0 && { cacheWrite1hInputTokens: cacheWrite1hTokens }),
       ...(cacheReadTokens > 0 && { cacheReadInputTokens: cacheReadTokens }),
       cost: costFor(
         catalogKey(envelope.model),
@@ -257,6 +265,7 @@ export async function * anthropic (envelope, deps = {}) {
           outputTokens: messageOutputTokens,
           thinkingTokens: estimatedThinkingTokens,
           cacheWriteInputTokens: cacheWriteTokens,
+          cacheWrite1hInputTokens: cacheWrite1hTokens,
           cacheReadInputTokens: cacheReadTokens
         }
       ),

package/js/session/adapters/gemini.js

@@ -104,6 +104,7 @@ export async function * gemini (envelope, deps = {}) {
   let inputTokens = 0
   let outputTokens = 0
   let thinkingTokens = 0
+  let cacheReadTokens = 0
   let status = STATUS_COMPLETED
   /** @type {string | undefined} */
   let warning
@@ -164,6 +165,9 @@ export async function * gemini (envelope, deps = {}) {
         if (typeof chunk.usageMetadata.thoughtsTokenCount === 'number') {
           thinkingTokens = chunk.usageMetadata.thoughtsTokenCount
         }
+        if (typeof chunk.usageMetadata.cachedContentTokenCount === 'number') {
+          cacheReadTokens = chunk.usageMetadata.cachedContentTokenCount
+        }
       }
     }
   } catch (e) {
@@ -186,18 +190,23 @@ export async function * gemini (envelope, deps = {}) {
   }
 
   const end = String(process.hrtime.bigint())
+  // Implicit caching reports cachedContentTokenCount as a SUBSET of
+  // promptTokenCount. Convert to mohdel's additive convention by subtracting
+  // the cached portion before pricing, matching the OpenAI adapter.
+  const regularInputTokens = Math.max(0, inputTokens - cacheReadTokens)
   /** @type {import('#core/events.js').DoneEvent} */
   const done = {
     type: 'done',
     result: {
       status,
       output: currentOutput() || null,
-      inputTokens,
+      inputTokens: regularInputTokens,
       outputTokens,
       thinkingTokens,
+      ...(cacheReadTokens > 0 && { cacheReadInputTokens: cacheReadTokens }),
       cost: costFor(
         catalogKey(envelope.model),
-        { inputTokens, outputTokens, thinkingTokens }
+        { inputTokens: regularInputTokens, outputTokens, thinkingTokens, cacheReadInputTokens: cacheReadTokens }
       ),
       timestamps: { start, first: first ?? end, end }
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.113.0",
+  "version": "0.114.1",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -87,16 +87,16 @@
     "@opentelemetry/exporter-trace-otlp-grpc": "^0.219.0",
     "@opentelemetry/sdk-node": "^0.219.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.113.0"
+    "mohdel-thin-gate-linux-x64-gnu": "0.114.1"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.104.1",
+    "@anthropic-ai/sdk": "^0.105.0",
     "@cerebras/cerebras_cloud_sdk": "^1.61.1",
-    "@google/genai": "^2.8.0",
+    "@google/genai": "^2.9.0",
     "@opentelemetry/api": "^1.9.1",
     "env-paths": "^4.0.0",
     "groq-sdk": "^1.2.1",
-    "openai": "^6.42.0",
+    "openai": "^6.44.0",
     "undici": "^7.24.5"
   },
   "lint-staged": {
@@ -107,6 +107,6 @@
     "lint-staged": "^17.0.7",
     "release-it": "^20.2.0",
     "standard": "^17.1.2",
-    "vitest": "^4.1.8"
+    "vitest": "^4.1.9"
   }
 }