model-router

name	model-router
description	Route sanitized work to intent-appropriate Z.ai, MiniMax, ralph_coding_models, official MCPs, and local Codex tools while keeping Codex main as final owner.

Model Router

Core Rule

Codex main owns decisions, file edits, synthesis, and verification. Z.ai and MiniMax are MCP-backed advisors or workers, never direct Codex model_provider backends.

Before any MCP-backed route, scan the prompt/context for RED-sensitive material. API keys, JWTs, private keys, seed phrases, wallet material, OAuth tokens, database URLs, .env references, and customer-sensitive markers block externalization even if the caller requested GREEN or YELLOW.

Tool Layers

Ralph Coding Model MCP

Use ralph_coding_models for sanitized coding support. It provides validate_coding_models, route_coding_task, zai_coding_deep for GLM-5.1, zai_coding_fast for GLM-5-Turbo, minimax_agentic_fast for MiniMax-M2.7-highspeed, and ensemble_counterpart.

Z.ai Official MCPs

Use Z.ai official MCPs or configured aliases for external context. zai_web_search and web-search-prime cover current web search. zai_web_reader and web-reader read a specific URL. zai_zread and zread handle public GitHub repository research. zai_vision handles image, screenshot, diagram, chart, or video understanding.

MiniMax Official MCP

Use minimax_coding_tools.web_search for fast external search and minimax_coding_tools.understand_image for quick image checks.

Intent Routing

Use intent, sensitivity, and expected local verification value before considering cost:

• minimax-fast: logs, diffs, summaries, PR summaries, and test ideas.
• zai-fast: lightweight implementation support and small command-following/agentic reasoning.
• zai-deep: debugging, architecture, auth, migrations, rollout risk, claim adjudication, spec review, and failure analysis.
• zai-search: current web research.
• zai-reader: specific public/safe URL reading.
• zai-repo: public GitHub repository research.
• zai-vision or minimax-vision: screenshot, diagram, chart, and UI understanding only.
• local: trivial work, RED content, unavailable MCPs, or context that cannot be safely minimized.

Use scripts/cost/route-task.py for deterministic intent lane selection before external delegation. It keeps task_type, route, and protocol_route compatibility while adding intent, lane, verification, route_decision, and external_mcp_brief.

For substantive non-trivial work, expose the selected route as a ROUTE_DECISION block or append it through scripts/cost/ledger.py. RED always maps to local work.

External MCP Brief

Before sending context to Z.ai or MiniMax for non-trivial work, use the brief shape emitted by the router:

EXTERNAL_MCP_BRIEF
tool=<Z.ai|MiniMax>
role=<debug analyst|spec reviewer|claim adjudicator|log summarizer|researcher|vision analyst|implementation advisor>
sensitivity=<GREEN|YELLOW-sanitized>
context_minimized=yes
task=<specific question>
constraints=<what not to change, what assumptions matter>
required_output=
- findings or verdict
- evidence
- confidence
- risks
- recommended next action
codex_final_owner=yes

Forbidden

Never use Z.ai or MiniMax for image generation, video generation, music generation, voice generation, voice cloning, TTS, secrets, private keys, wallet material, credentials, customer data, database credentials, .env content, OAuth/JWT material, or RED-sensitive content.

Image generation uses GPT Images 2 only.

Failure Mode

If a preferred MCP is unavailable, continue locally when safe, record the degraded route, and do not fake validation.