[BUG] Claude‑Code proxy upstream DeepSeek‑V4‑Flash hit 400 The content[].thinking in the thinking mode must be passed back to the API after tool‑call turn

[ryan-yuan-dev]

Claude Code Version:
2.1.232

Plugin Version:
2.0.0

Operating System:
MacOS 26.4 (25E246)

Describe the bug:

1. Start full memory‑proxy stack (./start‑all.sh)
2. Configure upstream LLM provider to deepseek‑v4‑flash

   # ── memory 组：memory + memory-hub 内部调用的 LLM ────────────────────────────
   # 用途：memory 的 embed/summarize；memory-hub 里 knowledge 的 wiki ingest / 总结
   MEMORY_LLM_BASE_URL=https://api.deepseek.com/anthropic             # 例：https://api.deepseek.com/v1
   MEMORY_LLM_API_KEY=sk-xxxx               # 例：sk-xxxxxxxx
   MEMORY_LLM_MODEL="deepseek-v4-flash[1m]"                 # 例：deepseek-chat
   MEMORY_LLM_PROTOCOL=anthropic                  # openai | anthropic
   
   # ── proxy 组：proxy 把用户请求转发到的上游 LLM ──────────────────────────────
   # 用途：coding agent / 客户端调 proxy:8096，proxy 再转发到这里
   PROXY_UPSTREAM_URL=https://api.deepseek.com/anthropic               # 例：https://api.deepseek.com/v1
   PROXY_UPSTREAM_API_KEY=sk-xxxx           # 例：sk-xxxxxxxx（可与 memory 组不同）
   PROXY_UPSTREAM_MODEL="deepseek-v4-flash[1m]"             # 例：deepseek-chat

3. Configure Claude Code auth info

   export ANTHROPIC_BASE_URL=http://127.0.0.1:8096/claude-code/default
   export ANTHROPIC_AUTH_TOKEN='sk-mem-xxxx'
   claude --model deepseek-v4-flash[1m]

4. Launch claude cli then send "hello" message
5. Choose "关联团队资产" -> API Error: 400 The content[].thinking in the thinking mode must be passed back to the API.

Expected behavior:
Avoid 400 error

Error Logs / Screenshots :

• ScrreenShot
  

• Error Logs

  ➜  global-images git:(feat/server_team) ✗ docker logs -f tdai-proxy
  [config] loaded /data/config.yaml (top-level sections: server,upstream,log,tdai,skill,auth,sessionInit,costGuard,injection,redis)
  26-08-14 09:45:18.648 INFO  langfuse.disabled {"reason":"config not complete"}
  26-08-14 09:45:18.649 INFO  auth.init {"url":"http://memory-core:8420"}
  [storage] activated sqlite backend
  26-08-14 09:45:18.659 INFO  server.starting {"host":"0.0.0.0","port":8096,"upstream":"https://api.deepseek.com/anthropic","logFile":"(disabled)","logLevel":"info","opik":"disabled","langfuse":"disabled","clickhouse":"disabled","costGuard":"disabled","rateLimit":"1000000 TPM / 100 QPM","sessionInit":"enabled","injection":"skill,knowledge,tdai-memory","tdai":"http://memory-core:8420","coreSkill":"http://memory-core:8420","skillRuntime":"allowLlmWrite=false","auth":"http://memory-core:8420","systemUsers":"disabled"}
  26-08-14 09:45:18.664 INFO  server.listening {"address":"0.0.0.0","port":8096}
  26-08-14 09:45:18.890 WARN  connectivity.check {"result":"some_failed","upstream":"ok (225ms)","auth":"ok (18ms)","creditReport":"FAIL: fetch failed"}
  [identity] userId=? keyId=0ec79b66 sessionId=none wechatId=none user=? ws=? proxyToken=none custom=[x-app,x-claude-code-session-id,x-stainless-arch,x-stainless-lang,x-stainless-os,x-stainless-package-version,x-stainless-retry-count,x-stainless-runtime,x-stainless-runtime-version,x-stainless-timeout]
  [debug-cc] x-claude-code-session-id=7240fe95-120f-4259-9892-69d2107bbbeb x-app=cli
  [injection-debug] conversationId=7240fe95-120f-4259-9892-69d2107bbbeb sessionKey=7240fe95-120f-4259-9892-69d2107bbbeb userId=usr-00329t8my3 agentSource=claude-code sessionInitEnabled=true injectionEnabled=true injectors=["skill","knowledge","tdai-memory"] injectedSkipped=false
  [cache] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb miss (no bindingRepo) → history-scan
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb state=none → uninitialized
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb user=usr-00329t8my3 → pending_asset_confirm (teams=1)
  [identity] userId=? keyId=0ec79b66 sessionId=none wechatId=none user=? ws=? proxyToken=none custom=[x-app,x-claude-code-session-id,x-stainless-arch,x-stainless-lang,x-stainless-os,x-stainless-package-version,x-stainless-retry-count,x-stainless-runtime,x-stainless-runtime-version,x-stainless-timeout]
  [debug-cc] x-claude-code-session-id=7240fe95-120f-4259-9892-69d2107bbbeb x-app=cli
  [injection-debug] conversationId=7240fe95-120f-4259-9892-69d2107bbbeb sessionKey=7240fe95-120f-4259-9892-69d2107bbbeb userId=usr-00329t8my3 agentSource=claude-code sessionInitEnabled=true injectionEnabled=true injectors=["skill","knowledge","tdai-memory"] injectedSkipped=false
  [session-recover] claude-code:7240fe95-120f-4259-9892-69d2107bbbeb L2a hit status=pending_asset_confirm (agent=-, task=-)
  [cache] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb L2a hit → promote L1 (override stale L1)
  [session-init:cc:debug] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb pending_asset_confirm lastUserText="Your questions have been answered: \"本次对话是否要关联团队资产？\"=\"是，关联团队资产\". You can now continue with these answers in mind." choice=true
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb auto-select single team=team-003gyfltl9
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb team=team-003gyfltl9 auto-select single agent=agt-060cir3ka6
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb agent=agt-060cir3ka6 auto-select single task=task-1fn1nub0sw
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb → initialized agent=agt-060cir3ka6 task=task-1fn1nub0sw team=team-003gyfltl9 user=usr-00329t8my3
  [session-init:cc] session=claude-code:7240fe95-120f-4259-9892-69d2107bbbeb processed: 4 msgs, ctx=Y/Y protocol=anthropic systemAppend=Y final=[user,system,assistant,user]
  [injection-debug] initResult session=7240fe95-120f-4259-9892-69d2107bbbeb intercepted=false bypassed=undefined justRegistered=true hasSessionInfo=true hasAgentDetail=true
  [asset-capability] user=usr-00329t8my3 flags={"skill":true,"llm_wiki":true,"code_graph":true,"chat_memory":true}
  [injection] injection.externalGatewayUrl not set — falling back to http://172.18.0.4:8096. This causes hook cache thrashing + upstream KV-cache misses in multi-node deployments; set injection.externalGatewayUrl to the shared gateway domain (e.g. https://gateway.example.com).
  ...
  26-08-14 09:46:40.173 INFO  injection.pipeline.done {"traceId":"019fffaa","protocol":"anthropic","agentSource":"claude-code","durationMs":18,"hookCount":4,"totalBlockCount":3,"errorCount":0}
  26-08-14 09:46:40.174 WARN  joinUrl.fallback {"requestPath":"/messages","endpoint":"/messages"}
  08-14 09:46:40.174 [019fffaa] → REQ model=deepseek-v4-flash msgs=3 stream=true
  08-14 09:46:40.175 [019fffaa]   → FORWARD upstream=https://api.deepseek.com/anthropic
  26-08-14 09:46:40.176 WARN  rate_limit.fail_open {"instanceId":"default","modelId":"deepseek-v4-flash","reason":"redis_unavailable"}
  08-14 09:46:40.635 [019fffaa]   ← FORWARD status=400 460ms
  26-08-14 09:46:40.638 ERROR pipeline.error {"requestId":"[019fffaa]","stage":"UPSTREAM_4xx","error":"status=400 body={\"error\":{\"message\":\"The `content[].thinking` in the thinking mode must be passed back to the API.\",\"type\":\"invalid_request_error\",\"param\":null,\"code\":\"invalid_request_error\"}}","error.message":"status=400 body={\"error\":{\"message\":\"The `content[].thinking` in the thinking mode must be passed back to the API.\",\"type\":\"invalid_request_error\",\"param\":null,\"code\":\"invalid_request_error\"}}","error.name":"Error"}
  08-14 09:46:40.639 [019fffaa]   ✗ UPSTREAM_4xx status=400 body={"error":{"message":"The `content[].thinking` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}
  08-14 09:46:40.641 [019fffaa]   ✓ STREAM done (no usage extracted)
  08-14 09:46:40.641 [019fffaa] ← DONE total=467ms
  

[Maxwell-Code07]

We truly appreciate your valuable feedback! Our team will carry out an internal review and assessment promptly, and we will share our thoughts with you here soon.