feat(protocol): 保留调用方 auto compact 窗口

CHANGELOG.md

@@ -4,6 +4,7 @@
 
 - 槽位种子页默认锁定：需确认解锁后才能改配置，播种前再次确认；时区/语言提供 en、cn、jp 基础选项。时区跟随当前 SOCKS5（空/非手动钉住）时展示出口时区，播种经 `timezone_follow_proxy` 写入；可改固定 IANA 时区或改回跟随。TZ/LANG/LC_ALL 仍不经 `extra_env` 覆盖契约键。
 - 槽位种子页：`settings.json` 的 env 行与 `theme` 可直接编辑并随播种写入；`autoUpdates`、`grove_enabled` 只读且播种时固定为 `false`；TZ/LANG/LC_ALL、遥测 kill-switch、Bedrock/Vertex、gateway hint 等同理只读（播种由契约重写）。缺 `CLAUDE_CODE_ATTRIBUTION_HEADER` 时默认 `0`。调试上传默认关，独立 query 拉最近 40 条上传，切换开关不重置未保存编辑；Go sidecar 写 `vms/<id>/run/telemetry-upload.log`，面板 `GET seed-settings?debug=1` 返回 `telemetry_log`。
+- 协议页可配置 `routing.compatibility.auto_compact_window`（100K–1M 或跟随模型）；入站在清理 settings 前提取调用方 `CLAUDE_CODE_AUTO_COMPACT_WINDOW`，经内部 `auto_compact_window` 字段进入当前 job，清理后的 body 不再携带该 env。
 - 槽位种子页按当前槽 `settings.json` 列出真实配置；`claude.setting` 可编辑播种；缺失文件返回 `null`、空文件返回空串，不用策略默认冒充槽内文件；遥测开关与 sidecar 进程状态分开展示，保存不因遥测进程刷新失败而 502。
 - fleet 遥测刷新在 kin-egress 二进制被替换（含 `(deleted)` 旧进程与脏 pidfile）时，会按 egress 配置扫停全部 helper、等待 `/proc/<pid>` 消失后再绑定新二进制；同一 SOCKS 出口的 stop/start 与 `ensureProxyEgress` 共用 per-proxy 锁，避免并发 telemetry 互相 SIGTERM。（#604）
 - 面板：侧栏「GitHub / 文档」两枚链接不再折行，字体偏宽时「文档」不会逐字竖排，空间不够先缩箭头再省略文字；VM 列表搜索框关闭浏览器自动填充，不再被回填成登录邮箱。

src/lib/admin/panel-api.mjs

@@ -475,6 +475,12 @@ export function validatePersonaRoutingPatch(body = {}) {
   if (Object.prototype.hasOwnProperty.call(compat, 'cch_enabled') && typeof compat.cch_enabled !== 'boolean') {
     problems.push('cch_enabled 必须是布尔')
   }
+  if (Object.prototype.hasOwnProperty.call(compat, 'auto_compact_window') && compat.auto_compact_window != null) {
+    const windowTokens = Number(compat.auto_compact_window)
+    if (!Number.isInteger(windowTokens) || windowTokens < 100_000 || windowTokens > 1_000_000) {
+      problems.push('auto_compact_window 必须是 100000–1000000 的整数')
+    }
+  }
   if (
     compat.thinking_display != null &&
     compat.thinking_display !== '' &&
@@ -1406,6 +1412,9 @@ function presentCompatibility(compat) {
     .trim()
     .toLowerCase()
   out.thinking_display = display === 'passthrough' ? 'passthrough' : 'summarized'
+  const windowTokens = Number(out.auto_compact_window)
+  out.auto_compact_window =
+    Number.isInteger(windowTokens) && windowTokens >= 100_000 && windowTokens <= 1_000_000 ? windowTokens : null
   return out
 }
 

src/lib/protocol/handle-protocol.mjs

@@ -35,7 +35,7 @@ import {
   createClaudeMessageAssembler,
   applyClaudeSSELineToMessage,
 } from './convert.mjs'
-import { sanitizeInboundBody, defaultSeedPolicy } from './seed-policy.mjs'
+import { sanitizeInboundBody, defaultSeedPolicy, callerAutoCompactWindow } from './seed-policy.mjs'
 import { fingerprintRequest } from './client-fingerprint.mjs'
 import { validateOfficialModel } from './models.mjs'
 import { handleCodexProtocol, handleCodexSearch } from './handle-codex.mjs'
@@ -619,11 +619,13 @@ export function createHandleProtocol(deps) {
       )
     }
 
+    const autoCompactWindow = callerAutoCompactWindow(inbound)
     let ctx = {
       path: pathName,
       protocol,
       body: sanitizeInboundBody(inbound, defaultSeedPolicy()),
       headers: { ...req.headers },
+      autoCompactWindow,
     }
     ctx = applyIntercept(cfg.intercept.rules, 'before_convert', ctx)
     const bodyCheck = validateRequestBody(protocol, ctx.body)
@@ -1291,6 +1293,7 @@ export function createHandleProtocol(deps) {
               preserveSystem: officialPassthrough,
               preserveThinkingBlocks: preserveCliHopThinkingBlocks,
               preserveCallerEnvironment,
+              autoCompactWindow: ctx.autoCompactWindow,
             })
             hopBody = await materializeRemoteImageSources(hopBody)
             if (panelRewrite) {

src/lib/protocol/outbound-attempt.mjs

@@ -190,11 +190,16 @@ export function prepareCliHopBody(
     preserveSystem = false,
     preserveThinkingBlocks = false,
     preserveCallerEnvironment = false,
+    autoCompactWindow = null,
   } = {},
 ) {
   if (requestContext?.purpose === 'auto_mode_classifier') return prepareClassifierBody(canonicalBody, { stream })
   if (preserveSystem) validateCacheBreakpoints(canonicalBody)
   let body = officialMessagesBody(canonicalBody, { stream })
+  const windowTokens = Number(autoCompactWindow)
+  if (Number.isInteger(windowTokens) && windowTokens >= 100_000 && windowTokens <= 1_000_000) {
+    body.auto_compact_window = windowTokens
+  }
   delete body.metadata
   if (!preserveSystem) {
     const leftover = stripCliOwnedSystem(body.system, { preserveCallerEnvironment })

src/lib/protocol/seed-policy.mjs

@@ -118,6 +118,16 @@ export function buildSlotSettingsEnv(pol = {}, { timezone, locale, extra } = {})
   return out
 }
 
+export function callerAutoCompactWindow(body = {}) {
+  const sources = [body?.settings?.env, body?.claude_settings?.env, body?.env]
+  for (const source of sources) {
+    if (!source || typeof source !== 'object' || Array.isArray(source)) continue
+    const value = Number(source.CLAUDE_CODE_AUTO_COMPACT_WINDOW)
+    if (Number.isInteger(value) && value >= 100_000 && value <= 1_000_000) return value
+  }
+  return null
+}
+
 export function buildSeedSettingsEnv(pol = {}, extra = {}) {
   return buildSlotSettingsEnv(pol, { extra })
 }

src/lib/transport/rust-kernel-supervisor.mjs

@@ -37,6 +37,11 @@ import {
 } from '../oauth/oauth-credentials.mjs'
 import { slotHost } from '../vm/slot-host.mjs'
 
+function normalizeAutoCompactWindow(value) {
+  const windowTokens = Number(value)
+  return Number.isInteger(windowTokens) && windowTokens >= 100_000 && windowTokens <= 1_000_000 ? windowTokens : null
+}
+
 const starts = new Map()
 const KVM_HEALTH_MS = { kvm: 120_000, tcg: 480_000 }
 
@@ -693,6 +698,13 @@ export function writeKernelConfig(
   const previousMax = Number(previous.max_request_bytes)
   const requestBytes = routing == null && previousMax > 0 ? previousMax : maxRequestBytes(routing)
   const cchEnabled = routing == null ? previous.cch_enabled !== false : routing?.compatibility?.cch_enabled !== false
+  const previousWindow = Number(previous.auto_compact_window)
+  const autoCompactWindow =
+    routing == null
+      ? Number.isInteger(previousWindow) && previousWindow >= 100_000 && previousWindow <= 1_000_000
+        ? previousWindow
+        : null
+      : normalizeAutoCompactWindow(routing?.compatibility?.auto_compact_window)
   const sessionIdleTtlSeconds =
     routing == null
       ? normalizeSessionIdleTtlSeconds(previous.session_idle_ttl_seconds)
@@ -726,6 +738,7 @@ export function writeKernelConfig(
     cli_version: OFFICIAL_CLI_VERSION,
     default_cache_ttl: defaultCacheTtl,
     cch_enabled: cchEnabled,
+    auto_compact_window: autoCompactWindow,
   }
 
   if (tz) config.timezone = tz

test/unit/go-worker-client.test.mjs

@@ -14,6 +14,8 @@ import {
   restoreUncommittedHop,
   semanticStatusForStreamError,
 } from '../../src/lib/transport/go-worker-client.mjs'
+import { prepareCliHopBody } from '../../src/lib/protocol/outbound-attempt.mjs'
+import { callerAutoCompactWindow, defaultSeedPolicy, sanitizeInboundBody } from '../../src/lib/protocol/seed-policy.mjs'
 import { extractOpenaiUsage } from '../../src/lib/protocol/openai-usage.mjs'
 import { callRustKernel, rustKernelHealth } from '../../src/lib/transport/rust-kernel-client.mjs'
 import { forwardApi, readApiJson } from '../../src/lib/transport/api-kernel-client.mjs'
@@ -164,6 +166,29 @@ test('cli-hop envelope keeps role=system turns that the VM betas do not declare'
   assert.deepEqual(out.body.system, body.system)
 })
 
+test('cli-hop keeps the caller compact window outside settings', () => {
+  const inbound = {
+    model: 'claude-opus-5-5',
+    messages: [{ role: 'user', content: 'hi' }],
+    settings: { theme: 'light', env: { CLAUDE_CODE_AUTO_COMPACT_WINDOW: '1000000', TZ: 'UTC' } },
+  }
+  const autoCompactWindow = callerAutoCompactWindow(inbound)
+  const sanitized = sanitizeInboundBody(inbound, defaultSeedPolicy())
+  assert.equal(callerAutoCompactWindow(sanitized), null)
+  const prepared = prepareCliHopBody(sanitized, { autoCompactWindow })
+  const out = finalizeWorkerPayload({
+    body: prepared,
+    reqHeaders: {},
+    exec: { homeDir: '', vm: { claude: { mode: 'setup-token', scope: 'user:inference' } } },
+    identity: null,
+    cliHop: true,
+  })
+  assert.equal(sanitized.settings, undefined)
+  assert.equal(out.body.settings, undefined)
+  assert.equal(out.body.env, undefined)
+  assert.equal(out.body.auto_compact_window, 1000000)
+})
+
 test('setup-token cli-hop summarizes hidden thinking by default and keeps updates when passthrough', () => {
   const exec = { homeDir: '', vm: { claude: { mode: 'setup-token', scope: 'user:inference' } } }
   const updates = finalizeWorkerPayload({

test/unit/kernel-router.test.mjs

@@ -631,6 +631,7 @@ test('writeKernelConfig separates container paths from host socket paths', () =>
   assert.equal(doc.https_proxy, undefined)
   assert.equal(doc.slots_per_worker, WRAP_SLOT_MAX)
   assert.equal(doc.default_cache_ttl, '1h')
+  assert.equal(doc.auto_compact_window, null)
   assert.match(written.socketPath.replace(/\\/g, '/'), /vms\/vm-09\/run\/kernel\.sock$/)
   assert.equal(fs.readFileSync(written.tokenPath, 'utf8').trim(), 'tok')
   const exec = { vmId: 'vm-09', homeDir: path.join(root, 'vms', 'vm-09', 'cli-home'), vm: { id: 'vm-09' } }
@@ -702,6 +703,7 @@ test('writeKernelConfig cli-hop writes local_cli without secrets', () => {
   assert.equal(doc.system_layout, 'identity')
   assert.equal(doc.cli_version, OFFICIAL_CLI_VERSION)
   assert.equal(doc.default_cache_ttl, '5m')
+  assert.equal(doc.auto_compact_window, null)
   assert.equal(doc.timezone, 'America/New_York')
   assert.equal(doc.proxy_url, '')
   assert.equal(doc.proxy_required, false)

web/src/features/settings/protocol-pane.tsx

@@ -3,7 +3,14 @@ import {
   SESSION_IDLE_TTL_SECONDS_MAX,
   SESSION_IDLE_TTL_SECONDS_MIN,
 } from '@/types/panel-routing'
-import { Brain, Check, Clock3, Database, Hash, ShieldCheck } from 'lucide-react'
+import {
+  Brain,
+  Check,
+  Clock3,
+  Database,
+  Scaling,
+  ShieldCheck,
+} from 'lucide-react'
 import {
   cacheTtlFromCompat,
   detectProxiedOfficialCcFromCompat,
@@ -32,6 +39,13 @@ const CLAUDE_CODE_FINGERPRINT: [string, string][] = [
   ['计费块', 'cc_version = UA 版本'],
 ]
 
+function autoCompactWindowValue(value: unknown) {
+  const tokens = Number(value)
+  if (tokens === 200_000 || tokens === 500_000 || tokens === 1_000_000)
+    return String(tokens)
+  return 'model'
+}
+
 export function ProtocolPane({
   compat,
   inference,
@@ -188,7 +202,34 @@ export function ProtocolPane({
         </SettingRow>
       </PaneSection>
       <PaneSection
-        icon={Hash}
+        icon={Scaling}
+        title='Auto compact 窗口'
+        desc='默认跟随请求模型。显式值替换模型窗口，请求里的 CLAUDE_CODE_AUTO_COMPACT_WINDOW 仍优先。'
+      >
+        <SettingRow
+          label='窗口'
+          desc='跟随模型时，Opus / Sonnet / Haiku 5.x 使用 1M，其余模型使用 200K。'
+        >
+          <ChipGroup
+            label='Auto compact 窗口'
+            value={autoCompactWindowValue(compat.auto_compact_window)}
+            options={[
+              ['model', '跟随模型'],
+              ['200000', '200K'],
+              ['500000', '500K'],
+              ['1000000', '1M'],
+            ]}
+            onChange={(value) =>
+              onCompatChange({
+                ...compat,
+                auto_compact_window: value === 'model' ? null : Number(value),
+              })
+            }
+          />
+        </SettingRow>
+      </PaneSection>
+      <PaneSection
+        icon={Clock3}
         title='CCH 统一附加'
         desc='出站 billing 的 cch。开启后缺失也补，不区分试用或订阅。'
       >