Performance

​

API v0.22.0

• Cheaper, Faster Turns: The prompt prefix is now stable across a conversation, so follow-up turns hit the provider’s cache instead of re-paying for the whole history.
• Leaner Tool Results: Large tool outputs are compacted into a digest before they reach the model, keeping long tool chains fast.
• Long Conversations Stay Light: Saved conversation state no longer stacks a fresh copy of the system prompt on every run, which had been quietly growing the cost of every follow-up.
• Steadier Memory Search: Embedding and vector writes are now bounded, fixing the file-handle exhaustion that could stall memory lookups under load.

​

API v0.21.0

• Faster Agent State: Reworked conversation checkpoint storage to remove O(N²) growth, keeping long conversations fast.
• Faster Code Sandbox: A more reliable sandbox file mount cut cold starts from roughly 76–104s to about 31s, and warm starts to about 2s.

​

API v0.18.0

• Lower Chat Cost: Byte-stable prompt caching cut chat-stream input cost by roughly 46% with more accurate billing.
• Leaner Backend: Reduced backend memory usage by dropping unused heavy dependencies.

​

Web v0.19.0

• Faster Landing Page: Largest Contentful Paint on the landing page dropped from about 6s to 0.47s.

​

API v0.17.0

• Todos Performance: Faster loading and smoother interactions across the tasks and todos experience.

​

Web v0.18.0

• Todos UX: Significant speed improvements to the tasks and todos view, including faster filtering and smoother animations.
• Frontend Load Times: Reduced initial bundle size for faster page loads across the app.

​

Mobile v0.5.0

• Todos Performance: Faster task loading and smoother scrolling in the mobile todos view.

​

API v0.15.0

• Agent Flow Latency: Optimized agent execution flow to reduce tool-calling latency across all agent nodes