admin/Hermes-ui
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Account for full context tokens in compression (#908)

Browse Source 
* Account for full context tokens in compression

* Fix group chat final context updates

---------

Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
ekko
2026-05-21 19:40:52 +08:00
committed by GitHub
parent b2ec321990
commit 39ead94352
16 changed files with 730 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/server/context-engine.test.ts
+108
View File
@@ -203,6 +203,114 @@ describe('ContextEngine.buildContext', () => {
expect(mockSummarize).not.toHaveBeenCalled()
})
it('records full context token estimates without compressing when under threshold', async () => {
const messages = makeMessages(3)
mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
const contextTokenEstimator = vi.fn().mockResolvedValue(19_379)
const result = await engine.buildContext({
roomId: 'room-1',
agentId: 'agent-1',
agentName: 'Claude',
agentDescription: 'Helper',
agentSocketId: 'agent-socket',
roomName: 'general',
memberNames: ['Alice'],
members: [{ userId: 'u1', name: 'Alice', description: '' }],
upstream: 'http://localhost:8642',
apiKey: null,
currentMessage: messages[messages.length - 1],
contextTokenEstimator,
})
expect(result.meta.compressed).toBe(false)
expect(result.meta.contextTokenEstimate).toBe(19_379)
expect(result.meta.messageTokenEstimate).toBeGreaterThan(0)
expect(contextTokenEstimator).toHaveBeenCalledWith(
expect.arrayContaining([{ role: 'assistant', content: expect.stringContaining('[Claude]') }]),
expect.stringContaining('"Claude"'),
)
expect(mockSummarize).not.toHaveBeenCalled()
})
it('uses full context token estimates to trigger group compression', async () => {
const messages = makeMessages(20)
mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
const result = await engine.buildContext({
roomId: 'room-1',
agentId: 'agent-1',
agentName: 'Claude',
agentDescription: 'Helper',
agentSocketId: 'agent-socket',
roomName: 'general',
memberNames: [],
members: [],
upstream: 'http://localhost:8642',
apiKey: null,
currentMessage: messages[messages.length - 1],
contextTokenEstimator: vi.fn().mockResolvedValue(120_000),
})
expect(result.meta.compressed).toBe(true)
expect(result.meta.contextTokenEstimate).toBe(120_000)
expect(mockSummarize).toHaveBeenCalledTimes(1)
expect(mockFetcher.saveContextSnapshot).toHaveBeenCalledTimes(1)
})
it('throws when group prompt and tools exceed threshold with too little history to compress', async () => {
const messages = makeMessages(4)
mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
await expect(engine.buildContext({
roomId: 'room-1',
agentId: 'agent-1',
agentName: 'Claude',
agentDescription: 'Helper',
agentSocketId: 'agent-socket',
roomName: 'general',
memberNames: [],
members: [],
upstream: 'http://localhost:8642',
apiKey: null,
currentMessage: messages[messages.length - 1],
contextTokenEstimator: vi.fn().mockResolvedValue(120_000),
})).rejects.toThrow('Context window is too small')
expect(mockSummarize).not.toHaveBeenCalled()
expect(mockFetcher.saveContextSnapshot).not.toHaveBeenCalled()
})
it('throws on snapshot path when overhead plus new messages exceed threshold without compressible history', async () => {
const messages = makeMessages(12)
mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
mockFetcher.getContextSnapshot = vi.fn().mockReturnValue({
roomId: 'room-1',
summary: 'Existing summary',
lastMessageId: 'msg-9',
lastMessageTimestamp: messages[9].timestamp,
updatedAt: Date.now(),
})
await expect(engine.buildContext({
roomId: 'room-1',
agentId: 'agent-1',
agentName: 'Claude',
agentDescription: 'Helper',
agentSocketId: 'agent-socket',
roomName: 'general',
memberNames: [],
members: [],
upstream: 'http://localhost:8642',
apiKey: null,
currentMessage: messages[messages.length - 1],
contextTokenEstimator: vi.fn().mockResolvedValue(120_000),
})).rejects.toThrow('Context window is too small')
expect(mockSummarize).not.toHaveBeenCalled()
expect(mockFetcher.saveContextSnapshot).not.toHaveBeenCalled()
})
it('splits into head/tail and compresses middle when over threshold', async () => {
const messages = makeMessages(20)
mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
tests/server/group-chat-member-sync.test.ts
+34
View File
@@ -27,6 +27,7 @@ vi.mock('../../packages/server/src/services/auth', () => ({
}))
import { AgentClients } from '../../packages/server/src/services/hermes/group-chat/agent-clients'
import { GroupChatServer } from '../../packages/server/src/services/hermes/group-chat'
import { groupChatRoutes, setGroupChatServer } from '../../packages/server/src/routes/hermes/group-chat'
function routeHandler(path: string, method: string) {
@@ -222,4 +223,37 @@ describe('Group Chat member/agent identity sync', () => {
members: [{ id: 'member-1', userId: 'human-1', name: 'Han', description: '', joinedAt: 1 }],
})
})
it('routes @mentions only from user messages, not agent replies', () => {
const server = Object.create(GroupChatServer.prototype) as any
const emit = vi.fn()
server.rooms = new Map([
['room-1', {
hasOnlineMember: vi.fn(() => true),
getOnlineMemberBySocketId: vi.fn((socketId: string) => socketId === 'agent-socket'
? { userId: 'agent-1', name: '丫鬟' }
: { userId: 'human-1', name: 'Human' }),
}],
])
server.socketUserMap = new Map([
['human-socket', 'human-1'],
['agent-socket', 'agent-1'],
])
server.userInfoMap = new Map([
['human-1', { name: 'Human', description: '' }],
['agent-1', { name: '丫鬟', description: '' }],
])
server.agentClients = { processMentions: vi.fn(async () => undefined) }
server.storage = {
saveMessageAndRefreshRoom: vi.fn((msg: any) => ({ message: msg, totalTokens: 123 })),
}
server.nsp = { to: vi.fn(() => ({ emit })) }
server.handleMessage({ id: 'human-socket' }, { roomId: 'room-1', content: '@all hi', role: 'user' }, vi.fn())
expect(server.agentClients.processMentions).toHaveBeenCalledTimes(1)
server.agentClients.processMentions.mockClear()
server.handleMessage({ id: 'agent-socket' }, { roomId: 'room-1', content: '@all agent says hi', role: 'assistant', mentionDepth: 1 }, vi.fn())
expect(server.agentClients.processMentions).not.toHaveBeenCalled()
})
})
tests/server/run-chat-compression.test.ts
+138
View File
@@ -162,6 +162,144 @@ describe('run chat compression trigger', () => {
)
})
it('uses full context estimates for compression threshold decisions', async () => {
const messages = Array.from({ length: 10 }, (_, index) => ({
id: index + 1,
session_id: 'session-1',
role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
content: `message ${index}`,
timestamp: index + 1,
tool_call_id: null,
tool_calls: null,
tool_name: null,
finish_reason: null,
reasoning_content: null,
}))
getSessionDetailMock.mockReturnValue({ messages })
calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 1_000, outputTokens: 0 })
compressorCompressMock.mockResolvedValue({
messages: [{ role: 'user', content: 'compressed by full context estimate' }],
meta: {
compressed: true,
llmCompressed: true,
totalMessages: 9,
summaryTokenEstimate: 1,
verbatimCount: 0,
compressedStartIndex: 0,
},
})
const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
const history = await buildCompressedHistory(
'session-1',
'default',
'http://upstream',
undefined,
vi.fn(),
new Map(),
{},
vi.fn(async () => 120_000),
)
expect(history).toEqual([{ role: 'user', content: 'compressed by full context estimate' }])
expect(compressorCompressMock).toHaveBeenCalledTimes(1)
})
it('emits full context token usage when the full estimate is under threshold', async () => {
const messages = Array.from({ length: 10 }, (_, index) => ({
id: index + 1,
session_id: 'session-1',
role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
content: `message ${index}`,
timestamp: index + 1,
tool_call_id: null,
tool_calls: null,
tool_name: null,
finish_reason: null,
reasoning_content: null,
}))
getSessionDetailMock.mockReturnValue({ messages })
calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 1_000, outputTokens: 900 })
const emit = vi.fn()
const contextTokenEstimator = vi.fn(async () => 19_379)
const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
const history = await buildCompressedHistory(
'session-1',
'default',
'http://upstream',
undefined,
emit,
new Map(),
{},
contextTokenEstimator,
)
expect(history).toHaveLength(9)
expect(contextTokenEstimator).toHaveBeenCalledWith(expect.arrayContaining([{ role: 'user', content: 'message 0' }]))
expect(emit).toHaveBeenCalledWith('usage.updated', expect.objectContaining({
event: 'usage.updated',
session_id: 'session-1',
inputTokens: 1_000,
outputTokens: 900,
contextTokens: 19_379,
}))
expect(compressorCompressMock).not.toHaveBeenCalled()
})
it('throws when fixed prompt and tool schemas exceed threshold before any history exists', async () => {
getSessionDetailMock.mockReturnValue({ messages: [] })
const emit = vi.fn()
const { buildCompressedHistory, ContextWindowTooSmallError } = await import('../../packages/server/src/services/hermes/run-chat/compression')
await expect(buildCompressedHistory(
'session-1',
'default',
'http://upstream',
undefined,
emit,
new Map(),
{},
vi.fn(async () => 120_000),
)).rejects.toBeInstanceOf(ContextWindowTooSmallError)
expect(emit).not.toHaveBeenCalledWith('usage.updated', expect.anything())
expect(compressorCompressMock).not.toHaveBeenCalled()
})
it('throws instead of compressing when full context is over threshold but history is too short', async () => {
const messages = Array.from({ length: 5 }, (_, index) => ({
id: index + 1,
session_id: 'session-1',
role: index === 4 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
content: `message ${index}`,
timestamp: index + 1,
tool_call_id: null,
tool_calls: null,
tool_name: null,
finish_reason: null,
reasoning_content: null,
}))
getSessionDetailMock.mockReturnValue({ messages })
calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 1_000, outputTokens: 0 })
const { buildCompressedHistory, ContextWindowTooSmallError } = await import('../../packages/server/src/services/hermes/run-chat/compression')
await expect(buildCompressedHistory(
'session-1',
'default',
'http://upstream',
undefined,
vi.fn(),
new Map(),
{},
vi.fn(async () => 120_000),
)).rejects.toBeInstanceOf(ContextWindowTooSmallError)
expect(compressorCompressMock).not.toHaveBeenCalled()
})
it('merges partial compression config with defaults', async () => {
const messages = Array.from({ length: 10 }, (_, index) => ({
id: index + 1,