performance-analysis

Build with debug symbols using cargo build --release, then record execution with a system profiler like perf record followed by perf report. This shows which functions consume the most CPU time during a task run.

Track prompt tokens, completion tokens, and tool calls per task to find waste. Common fixes include injecting fewer files into context, trimming the skill catalog with progressive disclosure, and avoiding repeated reads of the same file across turns.

Track prompt tokens per turn for context bloat, completion tokens per turn for over-generation, tool calls per task for wasted reads, and total cost per task. Aixlarity tracks session cost automatically.

Frequent auto-compact signals context pressure from accumulated conversation history or oversized injected files. It indicates the prompt assembly is carrying more context than needed and should be audited before adding new sources.

Never optimize based on intuition alone. Measure first with profiling data, since the actual bottleneck is often different from what developers assume, and premature optimization wastes effort on the wrong components.