cost-router

name	cost-router
description	Choose the best safe MCP lane across Codex, Z.ai, MiniMax, official MCPs, and local tools using intent, sensitivity, verification value, and then cost.

Cost Router

Core Rule

Despite the historical name, this skill is now intent-first. Choose the safest useful lane for the task intent; cost is secondary to safety, fit, and Codex-local verification.

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

Sensitivity

GREEN covers public docs, public repos, and generic technical prompts. YELLOW covers sanitized project-specific logs, diffs, or specs. RED covers secrets, API keys, credentials, private keys, wallet material, customer data, or sensitive proprietary code.

If content is RED, do not call external MCPs. Use Codex main and local tools only. Do not store that content in repo or vault artifacts.

The requested sensitivity is not trusted blindly. Route decisions must scan the provided context for API keys, JWTs, private keys, seed phrases, wallet material, OAuth tokens, database URLs, .env references, and customer-sensitive markers. Detected RED content overrides a GREEN or YELLOW request.

Intent Lanes

Intent	Default lane	Use
Trivial local work	local	No external call.
Logs, diffs, summaries, PR summaries	minimax-fast	Fast compression and first-pass synthesis.
Test ideas	minimax-fast	Breadth and speed.
Lightweight implementation support	zai-fast	Small agentic reasoning or command-following.
Debugging, architecture, auth, migrations, rollout risk	zai-deep	Deep analysis and hypothesis generation.
Claim adjudication / reviewer disagreement	zai-deep	Verdict with evidence, blocker classification, confidence.
Spec vs implementation review	zai-deep	Compare implementation to spec/runbook.
Current web research	zai-search or MiniMax search	Prefer Z.ai for depth; MiniMax for fast lookup.
Specific URL reading	zai-reader	Public/safe URLs only.
Public GitHub repo research	zai-repo	Public repos only unless sanitized and explicitly approved.
Screenshot, diagram, or chart understanding	zai-vision or minimax-vision	Analysis only, never generation.
RED/sensitive content	local	No external MCP.

For complexity 7 and above, Codex main owns the work with gates. External output can still be advisory when content is GREEN or sanitized YELLOW and the expected verification is clear.

External MCP Brief

Before sending context to Z.ai or MiniMax for non-trivial work, shape the request as:

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

Routing Decision Protocol

For substantive non-trivial work, record this decision before doing the work or before external delegation:

ROUTE_DECISION
sensitivity=GREEN|YELLOW|RED
intent=<logs|diff|summary|test-ideas|debugging|architecture|spec-review|claim-adjudication|research|repo-reading|url-reading|vision|implementation-support>
complexity=1-10
task_type=<legacy-compatible task type>
route=<local|minimax-fast|zai-fast|zai-deep|zai-search|zai-reader|zai-repo|zai-vision|minimax-vision|codex-subagent|fallback-local>
tool=<optional MCP tool>
reason=<short reason>
verification=<local verification expected>
fallback=<none or reason>

Skip the marker only for trivial tasks, RED-local work, explicit user opt-out from external models, unavailable MCPs, or context that cannot be safely sanitized. Missing markers are a warn-only runtime signal during rollout.

Recordkeeping

When routing externally, record sensitivity, intent, complexity, tool used, and whether Codex verified the result.

Use scripts/cost/route-task.py before external delegation, passing --text when any context would leave Codex. Use scripts/cost/redact-for-external.py before sending sanitized context; it exits non-zero and reports allowed_external=false for RED content. Use scripts/cost/estimate-context.py for rough context sizing, and scripts/cost/ledger.py for JSONL route records that never include raw prompt text or secrets.