align compression token estimates (#749)

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

@@ -845,6 +845,7 @@ export const useChatStore = defineStore('chat', () => {
     // Capture session ID at send time — all callbacks use this, not activeSessionId
     const sid = activeSessionId.value!
     const isBridgeSlashCommand = activeSession.value?.source === 'cli' && content.trim().startsWith('/')
+    const isBridgeCompressCommand = isBridgeSlashCommand && /^\/compress(?:\s|$)/i.test(content.trim())
     const wasLiveBeforeSend = isSessionLive(sid)
     const shouldQueue = wasLiveBeforeSend && !isBridgeSlashCommand
 
@@ -1348,7 +1349,7 @@ export const useChatStore = defineStore('chat', () => {
         undefined,
       )
 
-      if (!isBridgeSlashCommand || !wasLiveBeforeSend) {
+      if (!isBridgeSlashCommand || isBridgeCompressCommand) {
         streamStates.value.set(sid, ctrl)
       }
     } catch (err: any) {

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

@@ -7,11 +7,11 @@ import {
   getSessionDetail,
 } from '../../../db/hermes/session-store'
 import { getCompressionSnapshot } from '../../../db/hermes/compression-snapshot'
-import { ChatContextCompressor, countTokens, SUMMARY_PREFIX } from '../../../lib/context-compressor'
+import { ChatContextCompressor, SUMMARY_PREFIX } from '../../../lib/context-compressor'
 import { getModelContextLength } from '../model-context'
 import { logger } from '../../logger'
 import { bridgeLogger } from '../../logger'
-import { calcAndUpdateUsage } from './usage'
+import { calcAndUpdateUsage, estimateUsageTokensFromMessages } from './usage'
 import type { ChatMessage } from '../../../lib/context-compressor'
 import type { SessionState, BridgeCompressionResult } from './types'
 
@@ -210,7 +210,8 @@ export async function forceCompressBridgeHistory(
 
   const upstream = getUpstream(profile).replace(/\/$/, '')
   const apiKey = getApiKey(profile) || undefined
-  const totalTokens = countTokens(JSON.stringify(history))
+  const beforeUsage = estimateUsageTokensFromMessages(history)
+  const totalTokens = beforeUsage.inputTokens + beforeUsage.outputTokens
   bridgeLogger.info({
     sessionId,
     profile,
@@ -234,7 +235,8 @@ export async function forceCompressBridgeHistory(
     if (m.name) msg.name = m.name
     return msg
   })
-  const afterTokens = countTokens(JSON.stringify(compressedMessages))
+  const afterUsage = estimateUsageTokensFromMessages(compressedMessages)
+  const afterTokens = afterUsage.inputTokens + afterUsage.outputTokens
   bridgeLogger.info({
     sessionId,
     profile,

packages/server/src/services/hermes/run-chat/handle-api-run.ts

@@ -19,7 +19,7 @@ import { readSseFrames } from './sse-utils'
 import { extractResponseText } from './response-utils'
 import { applyResponseStreamEvent, flushResponseRunToDb } from './response-stream'
 import { buildCompressedHistory, getOrCreateSession } from './compression'
-import { calcAndUpdateUsage } from './usage'
+import { calcAndUpdateUsage, estimateUsageTokensFromMessages } from './usage'
 import { handleMessage } from './message-format'
 import { countTokens, SUMMARY_PREFIX } from '../../../lib/context-compressor'
 import { getCompressionSnapshot } from '../../../db/hermes/compression-snapshot'
@@ -47,16 +47,14 @@ export async function loadSessionStateFromDb(sid: string, _sessionMap: Map<strin
     const snapshot = getCompressionSnapshot(sid)
     if (snapshot) {
       const newMessages = messages.slice(snapshot.lastMessageIndex + 1)
+      const newUsage = estimateUsageTokensFromMessages(newMessages)
       inputTokens = countTokens(SUMMARY_PREFIX + snapshot.summary) +
-        newMessages.filter(m => m.role === 'user').reduce((sum, m) => sum + countTokens(m.content || ''), 0)
-      outputTokens = newMessages
-        .filter(m => m.role === 'assistant' || m.role === 'tool')
-        .reduce((sum, m) => sum + countTokens(m.content || '') + countTokens(m.tool_calls + '' || ''), 0)
+        newUsage.inputTokens
+      outputTokens = newUsage.outputTokens
     } else {
-      inputTokens = messages.filter(m => m.role === 'user').reduce((sum, m) => sum + countTokens(m.content || ''), 0)
-      outputTokens = messages
-        .filter(m => m.role === 'assistant' || m.role === 'tool')
-        .reduce((sum, m) => sum + countTokens(m.content || '') + countTokens(m.tool_calls + '' || ''), 0)
+      const usage = estimateUsageTokensFromMessages(messages)
+      inputTokens = usage.inputTokens
+      outputTokens = usage.outputTokens
     }
 
     logger.info('[chat-run-socket] loaded session %s from DB (%d messages)', sid, messages.length)

packages/server/src/services/hermes/run-chat/handle-bridge-run.ts

@@ -6,13 +6,12 @@
 import type { Server, Socket } from 'socket.io'
 import { getSession, createSession, addMessage, updateSessionStats } from '../../../db/hermes/session-store'
 import { updateUsage } from '../../../db/hermes/usage-store'
-import { countTokens } from '../../../lib/context-compressor'
 import { logger, bridgeLogger } from '../../logger'
 import { AgentBridgeClient, type AgentBridgeMessage, type AgentBridgeOutput } from '../agent-bridge'
 import { contentBlocksToString, extractTextForPreview } from './content-blocks'
 import { buildCompressedHistory } from './compression'
 import { pushState, replaceState } from './compression'
-import { calcAndUpdateUsage } from './usage'
+import { calcAndUpdateUsage, estimateUsageTokensFromMessages } from './usage'
 import {
   flushBridgePendingToDb,
   ensureOpenBridgeAssistantMessage,
@@ -267,8 +266,9 @@ async function applyBridgeChunkAsync(
       emit('approval.resolved', payload)
     } else if (evType === 'bridge.compression.requested') {
       const bridgeHistory = await buildDbHistory(sessionId, { excludeLastUser: true })
+      const bridgeUsage = estimateUsageTokensFromMessages(bridgeHistory)
       const tokenCount = bridgeHistory.length > 0
-        ? countTokens(JSON.stringify(bridgeHistory))
+        ? bridgeUsage.inputTokens + bridgeUsage.outputTokens
         : ev.approx_tokens
       const payload = {
         event: 'compression.started',

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

@@ -5,8 +5,7 @@ import type { AgentBridgeClient } from '../agent-bridge'
 import { flushBridgePendingToDb } from './bridge-message'
 import { buildDbHistory, forceCompressBridgeHistory, getOrCreateSession, replaceState } from './compression'
 import { handleAbort } from './abort'
-import { calcAndUpdateUsage } from './usage'
-import { countTokens } from '../../../lib/context-compressor'
+import { calcAndUpdateUsage, estimateUsageTokensFromMessages } from './usage'
 import type { ContentBlock, QueuedRun, SessionState } from './types'
 
 type CommandName =
@@ -233,7 +232,8 @@ 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 tokenEstimate = history.length > 0 ? countTokens(JSON.stringify(history)) : 0
+        const usageEstimate = estimateUsageTokensFromMessages(history)
+        const tokenEstimate = usageEstimate.inputTokens + usageEstimate.outputTokens
         emit('compression.started', {
           event: 'compression.started',
           message_count: history.length,

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

@@ -11,6 +11,35 @@ import { countTokens, SUMMARY_PREFIX } from '../../../lib/context-compressor'
 import { logger } from '../../logger'
 import type { SessionState } from './types'
 
+type UsageTokenMessage = {
+  role?: string
+  content?: unknown
+  tool_calls?: unknown
+}
+
+function contentToUsageText(content: unknown): string {
+  if (typeof content === 'string') return content
+  if (!content) return ''
+  if (Array.isArray(content)) {
+    return content.map((block: any) => {
+      if (typeof block?.text === 'string') return block.text
+      if (typeof block?.type === 'string') return `[${block.type}]`
+      return String(block || '')
+    }).join('\n')
+  }
+  return String(content)
+}
+
+export function estimateUsageTokensFromMessages(messages: UsageTokenMessage[]): { inputTokens: number; outputTokens: number } {
+  const inputTokens = messages
+    .filter(m => m.role === 'user')
+    .reduce((sum, m) => sum + countTokens(contentToUsageText(m.content)), 0)
+  const outputTokens = messages
+    .filter(m => m.role === 'assistant' || m.role === 'tool')
+    .reduce((sum, m) => sum + countTokens(contentToUsageText(m.content)) + countTokens(String(m.tool_calls || '')), 0)
+  return { inputTokens, outputTokens }
+}
+
 export async function calcAndUpdateUsage(
   sid: string,
   state: SessionState,
@@ -26,16 +55,14 @@ export async function calcAndUpdateUsage(
     let outputTokens: number
     if (snapshot && msgs.length) {
       const newMessages = msgs.slice(snapshot.lastMessageIndex + 1)
+      const newUsage = estimateUsageTokensFromMessages(newMessages)
       inputTokens = countTokens(SUMMARY_PREFIX + snapshot.summary) +
-        newMessages.filter(m => m.role === 'user').reduce((sum, m) => sum + countTokens(m.content || ''), 0)
-      outputTokens = newMessages
-        .filter(m => m.role === 'assistant' || m.role === 'tool')
-        .reduce((sum, m) => sum + countTokens(m.content || '') + countTokens(m.tool_calls + '' || ''), 0)
+        newUsage.inputTokens
+      outputTokens = newUsage.outputTokens
     } else {
-      inputTokens = msgs.filter(m => m.role === 'user').reduce((sum, m) => sum + countTokens(m.content || ''), 0)
-      outputTokens = msgs
-        .filter(m => m.role === 'assistant' || m.role === 'tool')
-        .reduce((sum, m) => sum + countTokens(m.content || '') + countTokens(m.tool_calls + '' || ''), 0)
+      const usage = estimateUsageTokensFromMessages(msgs)
+      inputTokens = usage.inputTokens
+      outputTokens = usage.outputTokens
     }
     state.inputTokens = inputTokens
     state.outputTokens = outputTokens

tests/server/run-chat-usage.test.ts

@@ -0,0 +1,33 @@
+import { describe, expect, it } from 'vitest'
+import { countTokens } from '../../packages/server/src/lib/context-compressor'
+import { estimateUsageTokensFromMessages } from '../../packages/server/src/services/hermes/run-chat/usage'
+
+describe('run-chat usage token estimates', () => {
+  it('counts message content instead of serialized message payloads', () => {
+    const messages = [
+      { role: 'user', content: 'hello from user' },
+      { role: 'assistant', content: 'hello from assistant' },
+    ]
+
+    const usage = estimateUsageTokensFromMessages(messages)
+
+    expect(usage.inputTokens).toBe(countTokens('hello from user'))
+    expect(usage.outputTokens).toBe(countTokens('hello from assistant'))
+    expect(usage.inputTokens + usage.outputTokens).toBeLessThan(countTokens(JSON.stringify(messages)))
+  })
+
+  it('keeps assistant tool call tokens on the output side', () => {
+    const messages = [
+      {
+        role: 'assistant',
+        content: 'calling tool',
+        tool_calls: [{ id: 'call_1', type: 'function', function: { name: 'lookup', arguments: '{"q":"x"}' } }],
+      },
+    ]
+
+    const usage = estimateUsageTokensFromMessages(messages)
+
+    expect(usage.inputTokens).toBe(0)
+    expect(usage.outputTokens).toBe(countTokens('calling tool') + countTokens(String(messages[0].tool_calls || '')))
+  })
+})