Model Escalator — AI skill for Claude Code

model-escalator

model-escalator is a [Claude Code](https://code.claude.com) mod that runs subagents and in-process teammates on cheaper models and watches them. When one gets stuck (the same error again and again, failing tests that do not improve) it intervenes one step at a time: a nudge; a one-off consult with a senior model whose advice goes back to the agent; a hot swap that moves the same agent one model tier up and keeps its context; and finally a stop that hands control back to you. It also learns which model to start each kind of task on, and it meters what the work costs against the same work on your lead's model. The savings are estimates from API list prices, not guarantees: escalations and consults are part of the real cost, so a session can come out worse than the baseline.

Status: early release (see [CHANGELOG.md](CHANGELOG.md) for the current version). The detector's defaults are calibrated on real sessions, but the savings have not yet been measured with the A/B evals in `plugin/evals/`: treat the reported figures as estimates. Start with `observe` mode on a real project (see [Quick start](#quick-start)).

Requirements and installation

Requires Claude Code 2.1.287 or later.

From the marketplace in this repository, inside Claude Code:

/plugin marketplace add atozzi-dev/model-escalator
/plugin install model-escalator@model-escalator

For development, clone the repository and load the mod from its root (it reloads when the files change):

claude --plugin-dir ./plugin

Mods are not sandboxed: they run with your permissions. To see what the mod hooks and calls without running it, clone the repository and run `claude plugin validate plugin`; the output is listed in [What it touches](#what-it-touches).

Quick start

After installing, nothing needs configuring. The mod runs in `active` mode: it tracks every subagent and in-process teammate, leaves your lead model alone, and intervenes only on the agents it sees get stuck. Within