KB-56DB

Phase 2F-E2 — ChatGPT MCP Patch Hang Diagnosis

2 min read Revision 1
phase-2fmcpchatgptpatch_documentdiagnosisgateway-bugbackup-channel

Phase 2F-E2 — ChatGPT MCP Patch Hang Diagnosis — 2026-05-14

Verdict

ChatGPT MCP App read flow works, but write flow is not reliable: after user grants access for patch_document, the request does not reach VPS/AgentData. Backend is not the bottleneck.

Evidence

  • Target document knowledge/test/chatgpt-mcp-final-verify.md stayed at revision 2 and still contained 2026-05-13; ChatGPT patch was not applied.
  • AgentData logs over 20 minutes contained list/get calls but zero patch_document, update_document, or upload_document calls from ChatGPT MCP App.
  • Internal MCP patch speed check succeeded in 386 ms with HTTP 200 and revision bump.
  • Nginx showed no /gpt-mcp/ timeout/error.
  • Therefore the patch request is held/dropped before reaching nginx/AgentData, likely in ChatGPT MCP App / OpenAI gateway / UI approval flow.

Operational conclusion

  • Server side D1/D2a/D3 is sufficient for the minimal backup MCP channel for read operations and backend write correctness.
  • ChatGPT MCP App cannot be relied on for write CRUD right now.
  • For GPT write operations, use the restored Custom Action/OpenAPI route or Claude Code/Codex until ChatGPT MCP App write behavior is fixed upstream.
  • Do not open D4 because of this UI/platform issue; D4 may improve backend write latency but cannot fix a request that never reaches VPS.

Next action

  • Treat MCP App as read/backup channel only for now.
  • Keep Custom Action connector as primary GPT Council read/write channel.
  • Return to priority: miếng thông tin / automation workflow.
  • Create separate TD: report ChatGPT MCP App write hang to OpenAI with evidence.