explain-modelmuxlisted

# How modelmux works Explain the mental model, adapting depth to the audience. The whole system is small and readable — point people at the actual files. ## The constraint Claude Code picks its upstream from **one global** environment variable, `ANTHROPIC_BASE_URL`. There's no per-subagent model setting on the client side. So if you want *some* agents on Claude and *others* on a different model, the decision has to happen **outside** the client — in a proxy that every request flows through. ## The proxy `src/server.ts` is a `Bun.serve` reverse proxy. For each request it: 1. Extracts routing **signals** (`src/signals.ts`). 2. Picks a target via a first-match **cascade** (`src/route.ts` + `routes.toml`). 3. Rewrites auth for the chosen upstream (`src/upstreams.ts`) and forwards it. SSE streams pass straight through. Two upstreams: **Anthropic** (`api.anthropic.com`, "passthrough" = keep the model Claude Code asked for) and **OpenRouter** (its Anthropic-compatible endpoint, so no request translation is needed). ## The signals (why it routes by header, not model name) From each request (`src/signals.ts`): - **`x-claude-code-agent-id`** — present *only* on subagent requests, so it marks `isSubagent`. This is the key signal, and routing on it sidesteps a Claude Code bug where a subagent's requested model can silently fall back to the parent's. - **`x-app`** — `cli` for foreground, `cli-bg` for background work. - **`<<route:alias>>`** — an explicit tag placed in an