What it does
Performance Analyser is built for the moment a slowness complaint lands. You name the page, class or function and hand it the raw signals — database query count, hot-loop iterations, duplicated-code ratio, cold-start latency, and a baseline benchmark sample — and it turns them into a scored diagnosis, a before/after benchmark and a draft PR suggestion. Findings come back with the threshold that was crossed, not an adjective. A report says "dbQueryCount 64, warning at 20, critical at 60" rather than "this looks slow", which is the difference between a claim you can act on and one you have to re-verify yourself. The benchmark is derived rather than guessed. The "after" sample is computed from your baseline by applying each abnormal finding's threshold overshoot against a fixed fix-impact weight, so the same findings always produce the same numbers. It reports p50, p95, throughput and error rate for both sides with an improvement percentage. `benchmark` is callable on its own when you already know what is wrong and just want the projection. The PR suggestion groups the work by finding category and lists the files to touch from a fixed category-to-file mapping. It does not invent paths beyond that mapping, which is the usual way these tools lose credibility. Two caveats the agent states rather than buries: benchmark numbers are noisy unless the environment is isolated, so there is a quick versus full scope switch and an isolated-environment flag; and the findings and PR are advisory, requiring human review before merge. Everything numeric is deterministic. The language model only narrates the synthesis — naming the likely root cause, such as an N+1 query behind a high query count — and a template fallback keeps the agent useful with no API key at all. No connectors and no credentials: it analyses only the data you provide. Built for backend developers and tech leads who want performance work measured rather than argued about.
Example prompts
- Here are the metrics for OrderService.getOrderSummary — why is it slow?
- Score these infra signals and tell me what crosses a threshold
- Compute the before/after benchmark from this baseline
