Triage batch: model limits, streaming, and rendering fixes
Highlights
- Claude Opus 5 is now in the built-in Anthropic catalog and works without manual configuration
- Anthropic requests derive their max output tokens per model instead of a flat 32K cap
- OpenRouter streaming no longer drops content when responses arrive with CRLF or multi-line SSE events
Improvements
- Todo plans ask for user intent once at the plan level instead of per item
- Requests show a distinct 'sending request' phase instead of sitting on 'connecting'
- Thinking/reasoning display is off by default for new users
- The /login heading lists unconfigured providers as dim rows for a cleaner header
- /model requests the route-expanded catalog and rebuilds an already-open picker
Fixes
- Kimi K3 now resolves to its full 1M context window everywhere
- The context meter reflects catalog context limits for Anthropic models
- Failed LaTeX renders are cached so redraws stop respawning latex processes
- PDF extraction failures are recovered instead of crashing the session
- Swarm history truncation no longer panics on multi-byte characters
- Anthropic OAuth sessions no longer advertise built-in tools that are not available
- Desktop builds no longer break on Windows from Unix-only IPC
Full changelog: v0.56.0...v0.57.0
Platform availability
- Linux x86_64: available
- Linux aarch64: available
- macOS Apple Silicon: available
- macOS Intel: available
- Windows x86_64: available
- Windows ARM64: available
- FreeBSD x86_64: available