claude-muse launcher

# claude-muse

Claude Code's interface, running Meta's `muse-spark-1.3` instead of an Anthropic model.

You keep the client you already know - the keyboard, the permission model, the skills, the subagents, the worktrees - and swap only the model behind it. Useful for comparing the two models on identical work, and for spending a Muse Code subscription without learning a second agent's UI.

Your existing `claude` is untouched. `claude-muse` is a separate launcher with its own config profile, so nothing here can change how your normal sessions behave.

## Where this is

**It works, and it holds a long task.** Web search, subagents, file edits, streaming and the context readout all behave; the proxy repairs the shapes the endpoint rejects, learns the ones nobody anticipated, rides out rate limits, and answers `count_tokens` itself. Every response now logs why it stopped and what it contained, so a claim about the model's behavior is a number you can grep for rather than an impression.

**What still costs you something is the safety classifier.** It is a little over half of all requests, it carries no reasoning directive, and this endpoint has no setting for "do not reason" - its cheapest tier is `low`. So the classifier pays full reasoning on every mechanical yes/no question, and when it runs past Claude Code's patience the tool call is denied rather than judged. Allowing the tools you trust is the supported answer, and the profile ships with the edit tools already on that list.

Three things are deliberately not repaired: the classifier above, `count_tokens` (answered from a calibrated estimate, never a tokenizer fact), and web-search sources (reconstructed only for pages the model actually opened). Each is labeled where it appears rather than quietly smoothed over.

![One localhost hop between Claude Code and muse-spark: requests are repaired, retried and recorded, and responses return to Claude Code.](docs/architecture.svg)

That figure is the first of five: [the architecture gallery](docs/architecture.md) walks from system context down to the retry policy.

## Why a proxy is involved

The Meta Model API serves an **Anthropic-compatible surface** at `https://api.meta.ai/v1/messages` - same request shape, same `anthropic-version` header, same response envelope. Compatible, but a *subset*, and Claude Code routinely sends fields that subset rejects.

The failur