Search, caching, and execution
Understand Semble, prefix caching, and parallel tool calls.
Semble semantic search
Crux includes Semble to find implementations by meaning. The agent can explore with semantic_search, then read the relevant code. Focused snippets help keep unrelated content out of context. First use may require model resources and indexing, so cold-start and warm-search timings differ.
Find the code that releases the session lease after an interruption.
Use semantic search, then explain the key call relationships.Prefix caching
Crux organizes stable system prompts and project instructions into reusable prefixes, avoiding unnecessary changes. Providers with prompt caching can reuse matching prefixes. The interface displays the cache information returned by the provider.
Cached tokens still occupy context. Cache hit rate does not measure removed context or imply an equal discount on the total bill. Model changes, prefix changes, expiry, and provider policies affect reuse.
Project usage has reached 99% cache hit. This is not guaranteed for every request and is not a comparative benchmark conducted by this website.
Parallel calls in one round
Independent tools can execute in the same round, reducing model round trips and waiting. Dependent actions still run in order, such as reading configuration before making a change based on it.
Read the metrics
| Metric | Meaning |
|---|---|
| TTFT | Time from request to first token |
| tok/s | Output generation rate |
| Tokens | Reported input, output, or context usage |
| Cache hit | Cache usage reported by the provider |
| Cost | Estimate from known pricing and usage |
| Coding Plan usage | Supported provider's plan quota and remaining usage |
Compare fairly
Keep the model, task, context, machine, and network consistent. Record cold and warm runs separately. A high hit rate in one round does not replace measuring the time and cost of a complete task.