claude-model-relay

Model Relay for Claude Code

A three-stage model relay that routes non-trivial coding tasks through the right model for each phase of the work:

Fable 5 (relay-planner)  →  plan + complexity classification
        ↓
   HIGH → Opus 4.8   ┐
                     ├→ relay-implementer  →  diff + verification
   STANDARD → Sonnet ┘
        ↓
Fable 5 (relay-reviewer) →  APPROVE / REVISE

Why? Frontier models are best at planning and adversarial review; mid-tier models are cheaper and plenty capable at executing a well-specified plan. The relay spends your most expensive tokens where judgment matters (planning, reviewing) and routes the implementation to Opus 4.8 or Sonnet 5 based on a complexity rubric — with explicit fallback chains when a model hits its usage limit, so nothing degrades silently.

What's included

Install

npx github:darasoba/claude-model-relay

That copies the skill to ~/.claude/skills/model-relay/ and the three subagents to ~/.claude/agents/. Restart Claude Code to pick them up.

As a Claude Code plugin:

/plugin marketplace add darasoba/claude-model-relay
/plugin install model-relay@model-relay

Manual copy:

git clone https://github.com/darasoba/claude-model-relay.git
mkdir -p ~/.claude/skills/model-relay
cp claude-model-relay/skills/model-relay/SKILL.md ~/.claude/skills/model-relay/
cp claude-model-relay/agents/*.md ~/.claude/agents/

Usage

Ask Claude Code to use the relay for a task:

Use the model-relay skill to add rate limiting to the API routes.

Or make it automatic by adding this to your ~/.claude/CLAUDE.md:

## Model relay

For any non-trivial code change — a feature, a refactor, a multi-file change, or
a bug whose root cause isn't yet known — invoke the `model-relay` skill before
writing code. It runs: Fable 5 plans and classifies complexity → Opus 4.8
(HIGH) or Sonnet 5 (STANDARD) implements → Fable 5 reviews.

Skip the relay, and just do the work, for trivial mechanical edits, questions,
explanations, and read-only exploration. Say that you're skipping it and why.

Always tell me which model did each stage, and tell me whenever a stage falls
back to a different model because of a usage limit. Never degrade silently.

How it works

1. Plan — relay-planner (Fable 5, read-only) investigates the actual code, produces a step-by-step plan with exact files and symbols, and classifies the task HIGH (architectural, concurrency, security, unknown root cause, >3 files) or STANDARD (contained, well-specified). If the plan surfaces open questions that would change the implementation, the relay stops and asks you first.
2. Implement — relay-implementer executes the plan on Opus 4.8 (HIGH) or Sonnet 5 (STANDARD). It runs the plan's verification command and reports any PLAN DEVIATION rather than forcing a change through a false premise.
3. Review — relay-reviewer (Fable 5, read-only) re-runs the verification itself, reads the real diff, and returns APPROVE or REVISE with concrete failure scenarios. REVISE loops back to the implementer — capped at two rounds, then the disagreement is escalated to you.

Fallback chains

Claude Code's built-in fallbackModel only covers overload/unavailability — not usage-limit exhaustion. The relay handles that itself:

Guardrails: Fable never implements, Haiku never implements HIGH work, the reviewer never runs on the model that implemented, and every fallback is reported to you in one line.

Requirements

• Claude Code with access to the models you want in the chains (works best on plans with Fable 5 access; edit the agent frontmatter model: fields to fit your plan)

License

MIT