fix(chat): remove upstream usage values and pre-send inputTokens overwrite

- Remove all evt.usage/parsed.usage references, only use local countTokens
- Remove pre-send inputTokens calculation that was overwriting resume value
  with compressed context, causing incorrect context drop (70k → 40k)
- run.completed now recalculates inputTokens with current snapshot + full
  messages including new ones from this run

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

packages/client/src/stores/hermes/chat.ts

@@ -783,13 +783,6 @@ export const useChatStore = defineStore('chat', () => {
               if (lastMsg?.isStreaming) {
                 updateMessage(sid, lastMsg.id, { isStreaming: false })
               }
-              if (evt.usage) {
-                const target = sessions.value.find(s => s.id === sid)
-                if (target) {
-                  target.inputTokens = evt.usage.input_tokens
-                  target.outputTokens = evt.usage.output_tokens
-                }
-              }
               // Server-computed usage (local countTokens, snapshot-aware)
               if ((evt as any).inputTokens != null) {
                 const target = sessions.value.find(s => s.id === sid)
@@ -1101,13 +1094,6 @@ export const useChatStore = defineStore('chat', () => {
           if (lastMsg?.isStreaming) {
             updateMessage(sid, lastMsg.id, { isStreaming: false })
           }
-          if (evt.usage) {
-            const target = sessions.value.find(s => s.id === sid)
-            if (target) {
-              target.inputTokens = evt.usage.input_tokens
-              target.outputTokens = evt.usage.output_tokens
-            }
-          }
           // Server-computed usage (local countTokens, snapshot-aware)
           if ((evt as any).inputTokens != null) {
             const target = sessions.value.find(s => s.id === sid)

packages/server/src/services/hermes/chat-run-socket.ts

@@ -436,15 +436,6 @@ export class ChatRunSocket {
       const headers: Record<string, string> = { 'Content-Type': 'application/json' }
       if (apiKey) headers['Authorization'] = `Bearer ${apiKey}`
 
-      // Save input token count after compression (the actual context sent to model)
-      if (session_id && body.conversation_history) {
-        const state = this.sessionMap.get(session_id)
-        if (state) {
-          state.inputTokens = (body.conversation_history as any[]).reduce(
-            (sum, m) => sum + countTokens(m.content || ''), 0)
-        }
-      }
-
       const res = await fetch(`${upstream}/v1/runs`, {
         method: 'POST',
         headers,
@@ -568,7 +559,7 @@ export class ChatRunSocket {
             }
           }
 
-          // Track usage — recalculate with current snapshot + full messages
+          // Track usage — self-calculate with countTokens + snapshot
           if (parsed.event === 'run.completed') {
             const sid = session_id
             if (sid) {