GPT Review — Phase 2F-C MCP Production Patch Design + Opus/Codex Reviews — 2026-05-14

Verdict

Claude Code xhigh design is accepted as a strong production-design baseline, but implementation is not approved yet. The design must be revised once more to incorporate Opus and Codex review changes, then saved as Phase 2F-C Rev2.

Key decisions

1. New process is accepted: Claude Code xhigh investigates and writes design first; GPT/Opus/Codex review; then implementation is approved separately.
2. D1 list_documents optimization remains highest priority and should be implemented first, separately from transport.
3. D2a transport negotiation is required for production compliance but should not be claimed as proven root cause of ChatGPT patch hang.
4. D3 remove move_document from MCP tool schema/allowlists is approved in principle, but backward compatibility for internal /mcp must be explicitly decided in Rev2.
5. D4 vector sync double-work is important but must remain a separate phase until evidence proves listener behavior, loop safety, vector_status contract, and lag monitoring.
6. /mcp-gpt remains public default. /mcp-gpt-full must not be exposed publicly yet.

Required Rev2 corrections

• D1 must include LIKE escaping for %, _, and backslash, ESCAPE '\', collation/index strategy, legacy rows missing document_id, safe revision casting, pagination cap and keyset/offset policy.
• D2a must define exact Accept negotiation, SSE frame format, notification handling, GET /mcp behavior, headers (Cache-Control, X-Accel-Buffering), and test matrix. It must say D2a may not fix ChatGPT patch UI hang by itself.
• D3 must choose policy for move_document across /mcp internal, /mcp-gpt, /mcp-gpt-full, MCP_TOOLS, and dispatcher.
• D4 must require canary proof of double embedding/Qdrant writes before implementation.

Implementation gate

Implement only after Phase 2F-C Rev2 is written to KB and reviewed.