A new coding agent, higher spending limits, and smarter model handling
· 2 min read
Meta's Muse Code joins as a supported coding agent, auto-reload limits rise to $10,000, and chat handles unsupported models and flaky AI responses more gracefully.
Meta's Muse Code joins the lineup of supported coding agents, auto-reload limits jump to $10,000, and chat gets sturdier around unsupported models and flaky AI responses. A batch of behind-the-scenes reliability work also means fewer stalled sessions and failed requests.
New: Meta's Muse Code agent
- Meta's Muse Code is now available as a native coding agent across terminal, chat, and iOS sessions, with full session and quota management support.
Billing
- The auto-reload monthly spending limit has been raised to $10,000 (up from $1,000), with clearer inline validation when a value is out of range.
- Fixed a bug where credit purchases could fail due to malformed checkout return URLs.
Chat and models
- Chat now warns and blocks image attachments on models that don't support them, greying out incompatible options instead of letting the request fail.
- Added support for moderation and classifier models, plus a new "legacy" status for older but still-supported chat models.
Voice assistant
- The voice assistant can now read open roadmap feedback items aloud.
- Fixed voice assistant turns that occasionally timed out or returned silently.
Reliability
- AI requests now automatically retry with an alternate model or fail over to a backup provider when blocked by security filters or rejected credentials, keeping work moving instead of failing outright.
- Sessions and paused tasks on legacy agents now recover and resume automatically instead of stalling, and occasional JSON formatting glitches in AI responses are corrected automatically.
Admin tools
- Admin cost reconciliation now flags cost-vs-usage inconsistencies for easier auditing.
Setup
- Docker Desktop setup on Windows now detects and installs WSL 2 prerequisites automatically.
Plus 2 smaller improvements.