v0.53.0

Highlights

Ollama vision capability detection. OllamaBackend now probes the model manifest at load time and advertises supportsVision in BackendCapabilities when the loaded model exposes a vision encoder (#1892). UI layers can reliably gate image-attach controls on capabilities.supportsVision without hard-coding model-name patterns — no caller change required.

let caps = await backend.capabilities()
if caps.supportsVision {
    // safe to attach images
}

Documentation

• Resolved WWDC 2026 LanguageModelExecutor / CoreAI trait-stub descriptions (#1895).

Tests

• Added repeatable local real-model integration and perf sweep (#1888).
• Capability-based tool-call discovery for Ollama E2E with aligned setup docs (#1890).
• Live streaming cancellation E2E with deterministic post-conditions (#1891).
• Made KeepAlivePolicyTests robust against CI scheduling starvation (#1894).

Continuous Integration

• Added on-demand companion compat check against an arbitrary core ref (#1889).