Idea: Muse Wonder as a second model PR-review Lens

[don-petry]

Add Meta's Muse Wonder as a second model reviewer, alongside the pr-review Lens that is the only model Lens in config/defaults.json today. A reviewer from a different vendor catches a different class of miss than a second pass from the same model.

How it fits the contract

• Start as an Advisory reviewer (FR-32), not a required Lens. Its registry entry names a structured surface (a check run and its annotations). It gates nothing and is never waited for, so a quota notice or an outage blocks no merge. A Finding gates only if pr-review adopts it.
• Read from its own App (ADR 0015). It needs its own GitHub App identity, separate from pr-review and from the author (FR-29).
• Earn a place in the required Council with data. Run it advisory for a few weeks and track how often its Findings are adopted, how often they're false positives, what it costs per PR, and how long it takes. Promote it only if its adopted Findings are ones pr-review missed.

Access and cost (as reported, Oct 2026: verify on Meta's pages)

• Muse Wonder is on the Meta Model API at $4.00 per million input tokens and $20.00 per million output.
• Muse Wonder Full is subscription-only, through Muse Code Power ($50/mo) or Maximum ($200/mo), with no API.
• A Muse Code subscription covers Muse Code use, not API keys an app calls. A CI Lens therefore either pays per token on the API, or runs the Muse Code CLI headless under the subscription. The second needs the terms checked before relying on it.

Data handling

Muse Code subscriptions default to Contributor mode, which lets Meta train on prompts, code and output, and opting out later does not cover what was already sent. Running it on this public repo is acceptable. Disable Contributor mode, or don't run the Lens, on any private repo.

Open questions

1. Do Muse Code's terms allow a headless CI run under a personal subscription, or is the API the only legitimate route?
2. Wonder vs Wonder Full: is the subscription-only model worth $50/mo over pay-per-token Wonder for review?
3. Rate limits and latency per PR at max_diff_lines (4000).
4. Meta's benchmarks are vendor-reported and not independently replicated, so the advisory trial should be the evidence.

Related

The Meta Global AI Developer Hackathon (applications close Oct 25, build on Model API or Muse Code, free API credits) could fund the advisory trial.