@goodandready/dsh-subscriptions 0.6.11 → 0.6.12

package/README.md

@@ -272,6 +272,15 @@ Supports Anthropic's adaptive thinking (`thinking: { type: "adaptive" }`) and re
 
 ---
 
+## 🛡️ Token Usage Normalization & Session Projection Fix (Added in v0.6.12)
+
+- **Session Projection Fix (GitHub #4)**: Fixed an issue where ChatGPT Codex and other streaming provider responses caused DSH session recovery to fail on page reload with `received NaN, expected number on uncachedInputTokens and outputTokens`.
+- **Canonical `TokenUsage` Normalization**: Introduced `toTokenUsage()` across all streaming pipelines (`codexResponsesStream`, `openaiChatStream`, `anthropicStream`, and `googleStream`) to map raw vendor usage payloads into canonical DSH `TokenUsage` (`inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWriteTokens`, `reasoningTokens`, `totalTokens`).
+- **Strict Integer Guard**: All token counts are strictly guarded to non-negative finite integers (guaranteeing `0` fallback, never `NaN` or `undefined`). Disjoint uncached token calculation is performed when cached tokens are folded into prompt totals.
+- **Defense in Depth**: Added adapter-level stream validation in `SubscriptionAdapter` to prevent malformed or invalid token usage chunks from ever reaching DSH session persistence.
+
+---
+
 ## 🔒 OAuth Client Validation & Route Updates (Added in v0.6.11)
 
 - **OAuth Client ID Guard (GitHub #2)**: uildAuthorizeUrl and ntigravity.authorizeUrl now strictly require a non-empty clientId. If unconfigured, the endpoint returns an explicit HTTP 400 (missing_client_id) and the UI prompts the user to configure ntigravityClientId in plugin settings or use "From CLI" instead of directing the browser to a failing Google OAuth page.

package/README.ru.md

@@ -245,6 +245,15 @@ dsh-subscriptions:
 
 ---
 
+## 🛡️ Нормализация расхода токенов и защита проекций сессий (Добавлено в v0.6.12)
+
+- **Защита от повреждения сессий (GitHub #4)**: Устранена критическая ошибка, из-за которой сессии с моделями ChatGPT Codex повреждались после первого ответа и не открывались после перезагрузки с ошибкой `received NaN, expected number on uncachedInputTokens and outputTokens`.
+- **Каноническая нормализация `TokenUsage`**: Внедрена универсальная функция `toTokenUsage()` во всех потоковых адаптерах (`codexResponsesStream`, `openaiChatStream`, `anthropicStream`, `googleStream`) для приведения вендорских структур к стандартному интерфейсу DSH `TokenUsage` (`inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWriteTokens`, `reasoningTokens`, `totalTokens`).
+- **Защита от `NaN` и disjoint-расчет**: Все значения токенов гарантированно приводятся к целым неотрицательным числам (с безопасным fallback в `0`, без `NaN` и `undefined`). Для провайдеров, объединяющих кэш с промптом, автоматически вычисляется раздельный объем некэшированного ввода.
+- **Эшелонированная защита адаптера**: Добавлен перехват и валидация чанков расхода токенов на уровне `SubscriptionAdapter`, исключающий попадание некорректных данных в журнал и хранилище сессий DSH.
+
+---
+
 ## 🔒 Защита OAuth Client ID и обновление ссылок (Добавлено в v0.6.11)
 
 - **Защита OAuth Client ID (GitHub #2)**: uildAuthorizeUrl и ntigravity.authorizeUrl теперь строго требуют непустой clientId. Если параметр не задан, сервер возвращает явный код ошибки HTTP 400 (missing_client_id), а интерфейс выводит подсказку с предложением указать ntigravityClientId в настройках или нажать «📥 Из CLI», предотвращая отправку битых запросов в Google OAuth.

package/README.zh.md

@@ -167,6 +167,15 @@ dsh plugin --profile web add @goodandready/dsh-subscriptions
 
 ---
 
+## 🛡️ Token 用量规范化与会话投影修复 (v0.6.12 新增)
+
+- **会话投影崩溃修复 (GitHub #4)**：修复了使用 ChatGPT Codex 及其他流式模型时，首次回复后会话损坏、刷新页面报错 `received NaN, expected number on uncachedInputTokens and outputTokens` 且无法加载的问题。
+- **全格式通用 TokenUsage 转换**：在所有流式传输通道（`codexResponsesStream`、`openaiChatStream`、`anthropicStream`、`googleStream`）中引入统一的 `toTokenUsage()` 转换逻辑，将各服务商的原始用量数据标准化为 DSH 规范的 `TokenUsage` 接口（`inputTokens`、`outputTokens`、`cacheReadTokens`、`cacheWriteTokens`、`reasoningTokens`、`totalTokens`）。
+- **严格非负整数保底与独立缓存计算**：对所有 token 计数实施严格的非负整数约束（保底为 `0`，杜绝 `NaN` 与 `undefined`）。针对将缓存合并计入总 prompt 的供应商自动执行拆分，确保输入 token 数的准确性。
+- **适配器层纵深防御**：在 `SubscriptionAdapter` 中新增用量数据过滤与净化机制，从根源上阻止异常用量流入 DSH 会话持久化与状态恢复流程。
+
+---
+
 ## 🔒 OAuth Client ID 校验保护与控制台地址更新 (v0.6.11 新增)
 
 - **Google OAuth Client ID 校验保护 (GitHub #2)**：uildAuthorizeUrl 和 ntigravity.authorizeUrl 严格要求非空 clientId。未配置时服务端返回明确的 HTTP 400 (missing_client_id)，前端界面提示用户在插件设置中填写 ntigravityClientId 或使用“📥 从 CLI 导入”，避免浏览器跳转至 Google OAuth 400 报错页面。

package/lib/adapter.js

@@ -6,6 +6,7 @@ import { modelCatalog } from './messages.js'
 import { streamWithRotation } from './stream-rotate.js'
 import { pickFetch } from './proxy.js'
 import { quotaSnapshot, parseBody } from './ratelimit.js'
+import { toTokenUsage } from './wire.js'
 
 // #94: heuristic for test/preview/beta/legacy model ids.
 function isDeprecatedId(id) {
@@ -176,7 +177,17 @@ export class SubscriptionAdapter extends LlmAdapter {
             throw err
           }
         },
-      })) { yieldedAny = true; yield event }
+      })) {
+        if (event && event.type === 'usage') {
+          const safeUsage = toTokenUsage(event.usage)
+          if (!safeUsage) continue
+          yieldedAny = true
+          yield { ...event, usage: safeUsage }
+          continue
+        }
+        yieldedAny = true
+        yield event
+      }
       return
     } catch (err) {
       // #91: when the whole pool is exhausted and nothing has been streamed

package/lib/analyze-session.js

@@ -16,9 +16,9 @@ export function analyzeSessionEvents(events = []) {
     const usage = ev.usage || (ev.type === 'usage' ? ev.usage : null) || (ev.data && ev.data.usage)
     if (!usage) continue
 
-    const p = Number(usage.prompt_tokens || usage.input_tokens || usage.promptTokens || 0)
-    const c = Number(usage.prompt_tokens_details?.cached_tokens || usage.cache_read_input_tokens || usage.cached_tokens || usage.cachedTokens || 0)
-    const comp = Number(usage.completion_tokens || usage.output_tokens || usage.completionTokens || 0)
+    const p = Number(usage.inputTokens ?? usage.prompt_tokens ?? usage.input_tokens ?? usage.promptTokens ?? 0)
+    const c = Number(usage.cacheReadTokens ?? usage.prompt_tokens_details?.cached_tokens ?? usage.cache_read_input_tokens ?? usage.cached_tokens ?? usage.cachedTokens ?? 0)
+    const comp = Number(usage.outputTokens ?? usage.completion_tokens ?? usage.output_tokens ?? usage.completionTokens ?? 0)
 
     if (p <= 0 && comp <= 0) continue
 

package/lib/responses-stream.js

@@ -1,5 +1,5 @@
 import { iterateSse, jsonSse } from './sse.js'
-import { httpError } from './wire.js'
+import { httpError, toTokenUsage } from './wire.js'
 
 function responsesFailure(code, message) {
   const text = String(message || code || 'responses failed')
@@ -132,8 +132,8 @@ export class ResponsesStreamTranslator {
       }
       case 'response.completed': {
         this.terminated = true
-        const usage = event.response?.usage
-        if (usage !== undefined) chunks.push({ type: 'usage', usage })
+        const u = toTokenUsage(event.response?.usage)
+        if (u) chunks.push({ type: 'usage', usage: u })
         chunks.push({ type: 'finish', reason: { kind: this.sawToolCall ? 'tool-calls' : 'stop' } })
         return chunks
       }

package/lib/wire.js

@@ -57,6 +57,102 @@ export function tokenBlobFromOAuth(json, extra) {
   }
 }
 
+function toCount(v) {
+  if (typeof v === "number") {
+    return Number.isFinite(v) && v >= 0 ? Math.floor(v) : 0
+  }
+  if (typeof v === "string" && v.trim()) {
+    const n = Number(v)
+    return Number.isFinite(n) && n >= 0 ? Math.floor(n) : 0
+  }
+  return 0
+}
+
+/**
+ * Normalizes vendor usage metadata into canonical DSH TokenUsage format.
+ * Guarantees inputTokens and outputTokens are non-negative finite integer numbers (never NaN/undefined).
+ * Calculates disjoint uncached input tokens where providers fold cache hits into total prompt tokens.
+ * Returns null if raw is not an object or contains no token-related fields.
+ */
+export function toTokenUsage(raw) {
+  if (!raw || typeof raw !== "object") return null
+
+  const hasKnownField = (
+    "inputTokens" in raw || "input_tokens" in raw || "promptTokens" in raw || "prompt_tokens" in raw || "input" in raw ||
+    "outputTokens" in raw || "output_tokens" in raw || "completionTokens" in raw || "completion_tokens" in raw || "output" in raw ||
+    "totalTokens" in raw || "total_tokens" in raw || "totalTokenCount" in raw ||
+    "promptTokenCount" in raw || "candidatesTokenCount" in raw
+  )
+  if (!hasKnownField) return null
+
+  const rawCached = toCount(
+    raw.input_token_details?.cached_tokens ??
+    raw.prompt_tokens_details?.cached_tokens ??
+    raw.cache_read_input_tokens ??
+    raw.cached_tokens ??
+    raw.cacheReadTokens ??
+    raw.cachedContentTokenCount
+  )
+
+  let inputTokens = 0
+  if (raw.inputTokens !== undefined) {
+    inputTokens = toCount(raw.inputTokens)
+  } else if (raw.prompt_tokens !== undefined || raw.promptTokens !== undefined) {
+    const prompt = toCount(raw.prompt_tokens ?? raw.promptTokens)
+    inputTokens = rawCached > 0 && prompt >= rawCached ? prompt - rawCached : prompt
+  } else if (raw.input_token_details?.cached_tokens !== undefined) {
+    const prompt = toCount(raw.input_tokens ?? raw.input)
+    inputTokens = rawCached > 0 && prompt >= rawCached ? prompt - rawCached : prompt
+  } else if (raw.promptTokenCount !== undefined) {
+    const prompt = toCount(raw.promptTokenCount)
+    inputTokens = rawCached > 0 && prompt >= rawCached ? prompt - rawCached : prompt
+  } else {
+    inputTokens = toCount(raw.input_tokens ?? raw.input)
+  }
+
+  let outputTokens = 0
+  if (raw.outputTokens !== undefined) {
+    outputTokens = toCount(raw.outputTokens)
+  } else {
+    outputTokens = toCount(
+      raw.completion_tokens ??
+      raw.output_tokens ??
+      raw.completionTokens ??
+      raw.output ??
+      raw.candidatesTokenCount
+    )
+  }
+
+  const cacheWriteTokens = toCount(raw.cacheWriteTokens ?? raw.cache_creation_input_tokens)
+  const reasoningTokens = toCount(
+    raw.reasoningTokens ??
+    raw.output_token_details?.reasoning_tokens ??
+    raw.completion_tokens_details?.reasoning_tokens ??
+    raw.reasoning_tokens
+  )
+  const rawTotal = toCount(raw.totalTokens ?? raw.total_tokens ?? raw.totalTokenCount)
+
+  const out = {
+    inputTokens,
+    outputTokens,
+  }
+
+  if (rawCached > 0 || raw.cacheReadTokens !== undefined || raw.cache_read_input_tokens !== undefined || raw.input_token_details?.cached_tokens !== undefined || raw.prompt_tokens_details?.cached_tokens !== undefined || raw.cachedContentTokenCount !== undefined) {
+    out.cacheReadTokens = rawCached
+  }
+  if (cacheWriteTokens > 0 || raw.cacheWriteTokens !== undefined || raw.cache_creation_input_tokens !== undefined) {
+    out.cacheWriteTokens = cacheWriteTokens
+  }
+  if (reasoningTokens > 0 || raw.reasoningTokens !== undefined || raw.output_token_details?.reasoning_tokens !== undefined || raw.completion_tokens_details?.reasoning_tokens !== undefined) {
+    out.reasoningTokens = reasoningTokens
+  }
+  if (rawTotal > 0 || raw.totalTokens !== undefined || raw.total_tokens !== undefined || raw.totalTokenCount !== undefined) {
+    out.totalTokens = rawTotal > 0 ? rawTotal : (inputTokens + (out.cacheReadTokens || 0) + (out.cacheWriteTokens || 0) + outputTokens)
+  }
+
+  return out
+}
+
 function openBlock(kind, index) {
   return { kind, index, text: '', id: '', name: '' }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-subscriptions",
-  "version": "0.6.11",
+  "version": "0.6.12",
   "description": "Use ChatGPT Codex, Claude, Grok, Antigravity, Kimi, GLM, Cursor, Kiro, Copilot, Qwen, ERNIE, Spark, JetBrains, Perplexity, Replit, and Cody subscriptions as DeepSeek Harness LLM providers via OAuth.",
   "license": "MIT",
   "type": "module",