fix command compression token estimate (#751)

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

@@ -66,6 +66,24 @@ export async function buildDbHistory(
   }).filter((m): m is ChatMessage => m !== null)
 }
 
+export function estimateSnapshotAwareHistoryUsage(
+  sessionId: string,
+  history: ChatMessage[],
+): { messageCount: number; tokenCount: number } {
+  const snapshot = getCompressionSnapshot(sessionId)
+  const messages = snapshot
+    ? [
+        { role: 'user', content: SUMMARY_PREFIX + snapshot.summary },
+        ...history.slice(snapshot.lastMessageIndex + 1),
+      ]
+    : history
+  const usage = estimateUsageTokensFromMessages(messages)
+  return {
+    messageCount: messages.length,
+    tokenCount: usage.inputTokens + usage.outputTokens,
+  }
+}
+
 export async function buildCompressedHistory(
   sessionId: string,
   profile: string,
@@ -210,12 +228,13 @@ export async function forceCompressBridgeHistory(
 
   const upstream = getUpstream(profile).replace(/\/$/, '')
   const apiKey = getApiKey(profile) || undefined
-  const beforeUsage = estimateUsageTokensFromMessages(history)
+  const beforeUsage = estimateSnapshotAwareHistoryUsage(sessionId, history)
-  const totalTokens = beforeUsage.inputTokens + beforeUsage.outputTokens
+  const totalTokens = beforeUsage.tokenCount
   bridgeLogger.info({
     sessionId,
     profile,
     historyMessages: history.length,
+    snapshotAwareMessages: beforeUsage.messageCount,
     bridgeProvidedMessages: Array.isArray(_messages) ? _messages.length : 0,
     tokenEstimate: totalTokens,
     snapshotAware: true,

packages/server/src/services/hermes/run-chat/session-command.ts

@@ -3,9 +3,9 @@ import { addMessage, clearSessionMessages, createSession, getSession, renameSess
 import { logger } from '../../logger'
 import type { AgentBridgeClient } from '../agent-bridge'
 import { flushBridgePendingToDb } from './bridge-message'
-import { buildDbHistory, forceCompressBridgeHistory, getOrCreateSession, replaceState } from './compression'
+import { buildDbHistory, estimateSnapshotAwareHistoryUsage, forceCompressBridgeHistory, getOrCreateSession, replaceState } from './compression'
 import { handleAbort } from './abort'
-import { calcAndUpdateUsage, estimateUsageTokensFromMessages } from './usage'
+import { calcAndUpdateUsage } from './usage'
 import type { ContentBlock, QueuedRun, SessionState } from './types'
 
 type CommandName =
@@ -232,12 +232,11 @@ export async function handleSessionCommand(
       const emit = (event: string, payload: any) => emitToSession(ctx.nsp, ctx.socket, sessionId, event, payload)
       try {
         const history = await buildDbHistory(sessionId, { excludeLastUser: true })
-        const usageEstimate = estimateUsageTokensFromMessages(history)
+        const usageEstimate = estimateSnapshotAwareHistoryUsage(sessionId, history)
-        const tokenEstimate = usageEstimate.inputTokens + usageEstimate.outputTokens
         emit('compression.started', {
           event: 'compression.started',
-          message_count: history.length,
+          message_count: usageEstimate.messageCount,
-          token_count: tokenEstimate,
+          token_count: usageEstimate.tokenCount,
           source: 'command',
         })
         const result = await forceCompressBridgeHistory(