mohdel 0.114.1 → 0.116.1

package/js/session/adapters/openai.js

@@ -77,6 +77,7 @@ export async function * openai (envelope, deps = {}) {
   let outputTokens = 0
   let thinkingTokens = 0
   let cachedInputTokens = 0
+  let cacheWriteTokens = 0
   let status = STATUS_COMPLETED
   /** @type {string | undefined} */
   let warning
@@ -132,6 +133,7 @@ export async function * openai (envelope, deps = {}) {
             outputTokens = event.response.usage.output_tokens ?? 0
             thinkingTokens = event.response.usage.output_tokens_details?.reasoning_tokens ?? 0
             cachedInputTokens = event.response.usage.input_tokens_details?.cached_tokens ?? 0
+            cacheWriteTokens = event.response.usage.input_tokens_details?.cache_write_tokens ?? 0
           }
           if (toolItems.size > 0) {
             status = STATUS_TOOL_USE
@@ -148,6 +150,7 @@ export async function * openai (envelope, deps = {}) {
             outputTokens = event.response.usage.output_tokens ?? 0
             thinkingTokens = event.response.usage.output_tokens_details?.reasoning_tokens ?? 0
             cachedInputTokens = event.response.usage.input_tokens_details?.cached_tokens ?? 0
+            cacheWriteTokens = event.response.usage.input_tokens_details?.cache_write_tokens ?? 0
           }
           break
 
@@ -177,11 +180,12 @@ export async function * openai (envelope, deps = {}) {
   // one from the other for the message-only count.
   const messageOutputTokens = Math.max(0, outputTokens - thinkingTokens)
 
-  // OpenAI counts cached_tokens as a SUBSET of input_tokens. Convert to
-  // mohdel's additive convention (cacheReadInputTokens is separate from
-  // inputTokens) by subtracting the cached portion before pricing. Both
-  // adapters and computeCost stay simpler with the additive shape.
-  const regularInputTokens = Math.max(0, inputTokens - cachedInputTokens)
+  // OpenAI counts cached_tokens and cache_write_tokens as SUBSETS of
+  // input_tokens. Convert to mohdel's additive convention (cacheRead/
+  // cacheWriteInputTokens are separate from inputTokens) by subtracting
+  // both portions before pricing. Both adapters and computeCost stay
+  // simpler with the additive shape.
+  const regularInputTokens = Math.max(0, inputTokens - cachedInputTokens - cacheWriteTokens)
 
   /** @type {import('#core/events.js').DoneEvent} */
   const done = {
@@ -192,6 +196,7 @@ export async function * openai (envelope, deps = {}) {
       inputTokens: regularInputTokens,
       outputTokens: messageOutputTokens,
       thinkingTokens,
+      ...(cacheWriteTokens > 0 && { cacheWriteInputTokens: cacheWriteTokens }),
       ...(cachedInputTokens > 0 && { cacheReadInputTokens: cachedInputTokens }),
       cost: costFor(
         catalogKey(envelope.model),
@@ -199,6 +204,7 @@ export async function * openai (envelope, deps = {}) {
           inputTokens: regularInputTokens,
           outputTokens: messageOutputTokens,
           thinkingTokens,
+          cacheWriteInputTokens: cacheWriteTokens,
           cacheReadInputTokens: cachedInputTokens
         }
       ),
@@ -273,6 +279,7 @@ function buildRequest (envelope, input, instructions) {
   if (envelope.identifier) {
     if (provider === 'openai') {
       request.safety_identifier = envelope.identifier
+      request.prompt_cache_key = envelope.identifier
     } else {
       request.user = envelope.identifier
     }

package/js/session/driver.js

@@ -12,8 +12,6 @@
  * @module session/driver
  */
 
-import readline from 'node:readline'
-
 import { run } from './run.js'
 import { runImage } from './run_image.js'
 import { runTranscription } from './run_transcription.js'
@@ -29,8 +27,6 @@ const PRECANCEL_CAP = 128
  * @returns {Promise<void>}
  */
 export async function drive (stdin, stdout) {
-  const rl = readline.createInterface({ input: stdin, crlfDelay: Infinity })
-
   /** @type {{callId: string, controller: AbortController} | null} */
   let currentCall = null
   /** @type {import('#core/envelope.js').CallEnvelope[]} */
@@ -38,6 +34,7 @@ export async function drive (stdin, stdout) {
   /** @type {(() => void) | null} */
   let queueNotify = null
   let stdinClosed = false
+  let framingError = null
   /** Cancel messages received before their envelope was dequeued.
    *  JS Sets are insertion-ordered, so `values().next()` is the
    *  oldest entry — cheap FIFO eviction at cap. */
@@ -51,7 +48,8 @@ export async function drive (stdin, stdout) {
     precancelled.add(callId)
   }
 
-  rl.on('line', (line) => {
+  const onLine = (line) => {
+    if (framingError) return
     const trimmed = line.trim()
     if (!trimmed) return
 
@@ -59,7 +57,17 @@ export async function drive (stdin, stdout) {
     try {
       obj = JSON.parse(trimmed)
     } catch (e) {
-      process.stderr.write(`session: failed to parse stdin line: ${e.message}\n`)
+      process.stderr.write(`session: malformed stdin line, exiting: ${e.message}\n`)
+      const error = {
+        message: 'SESSION_STDIN_MALFORMED',
+        detail: `stdin line is not valid JSON: ${e.message}`,
+        severity: 'error',
+        retryable: false,
+        type: 'SESSION_STDIN_MALFORMED'
+      }
+      stdout.write(JSON.stringify({ type: 'error', error }) + '\n')
+      framingError = Object.assign(new Error(error.message), { detail: error.detail })
+      if (queueNotify) { queueNotify(); queueNotify = null }
       return
     }
 
@@ -113,22 +121,35 @@ export async function drive (stdin, stdout) {
 
     envelopeQueue.push(obj)
     if (queueNotify) { queueNotify(); queueNotify = null }
-  })
+  }
 
-  rl.on('close', () => {
+  let stdinBuf = ''
+  stdin.setEncoding('utf8')
+  stdin.on('data', (chunk) => {
+    stdinBuf += chunk
+    let nl
+    while ((nl = stdinBuf.indexOf('\n')) !== -1) {
+      const line = stdinBuf.slice(0, nl)
+      stdinBuf = stdinBuf.slice(nl + 1)
+      onLine(line)
+    }
+  })
+  stdin.on('end', () => {
+    if (stdinBuf) onLine(stdinBuf)
     stdinClosed = true
     if (queueNotify) { queueNotify(); queueNotify = null }
   })
 
   while (true) {
-    // `stdinClosed` flips asynchronously inside the `rl.on('close')`
+    // `stdinClosed` flips asynchronously inside the `stdin.on('end')`
     // callback above, which also signals `queueNotify`. The linter's
     // no-unmodified-loop-condition rule can't see callback mutation,
     // so it false-positives here.
     // eslint-disable-next-line no-unmodified-loop-condition
-    while (envelopeQueue.length === 0 && !stdinClosed) {
+    while (envelopeQueue.length === 0 && !stdinClosed && !framingError) {
       await new Promise(resolve => { queueNotify = resolve })
     }
+    if (framingError) throw framingError
     if (envelopeQueue.length === 0) return
 
     const envelope = envelopeQueue.shift()

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.114.1",
+  "version": "0.116.1",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -83,20 +83,20 @@
     }
   },
   "optionalDependencies": {
-    "@clack/prompts": "^1.5.1",
-    "@opentelemetry/exporter-trace-otlp-grpc": "^0.219.0",
-    "@opentelemetry/sdk-node": "^0.219.0",
+    "@clack/prompts": "^1.7.0",
+    "@opentelemetry/exporter-trace-otlp-grpc": "^0.220.0",
+    "@opentelemetry/sdk-node": "^0.220.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.114.1"
+    "mohdel-thin-gate-linux-x64-gnu": "0.116.1"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.105.0",
+    "@anthropic-ai/sdk": "^0.110.0",
     "@cerebras/cerebras_cloud_sdk": "^1.61.1",
-    "@google/genai": "^2.9.0",
+    "@google/genai": "^2.11.0",
     "@opentelemetry/api": "^1.9.1",
     "env-paths": "^4.0.0",
-    "groq-sdk": "^1.2.1",
-    "openai": "^6.44.0",
+    "groq-sdk": "^1.3.0",
+    "openai": "^6.46.0",
     "undici": "^7.24.5"
   },
   "lint-staged": {
@@ -104,9 +104,9 @@
   },
   "devDependencies": {
     "gpt-tokenizer": "^3.4.0",
-    "lint-staged": "^17.0.7",
-    "release-it": "^20.2.0",
+    "lint-staged": "^17.0.8",
+    "release-it": "^20.2.1",
     "standard": "^17.1.2",
-    "vitest": "^4.1.9"
+    "vitest": "^4.1.10"
   }
 }