perf
Structured performance investigation with baselines, profiling, and evidence-backed decisions - run by AI agents inside your editor.
Why
Performance work without structure leads to random micro-optimizations, unmeasured changes, and wasted time. /perf enforces a rigorous 10-phase methodology - derived from recorded real investigation sessions - that ensures every optimization is baselined, isolated, measured, and documented.
Use cases:
- Investigating a reported performance regression
- Establishing baselines before and after a refactor
- Finding breaking points under load
- Profiling CPU and memory hotspots with language-native tools
- Making evidence-backed continue/stop decisions on optimization efforts
Installation
agentsys install perf
Requires agentsys to be set up in your project.
Quick Start
/perf --phase setup --scenario "API response time regression" --command "npm run bench" --version v1.2.0
This initializes an investigation, records the scenario, and prepares for baselining. The orchestrator walks you through each phase sequentially, checkpointing progress after every step.
To resume an in-progress investigation:
/perf --resume
How It Works
Every investigation follows 10 phases in order. The orchestrator enforces strict rules: sequential benchmarks only, 60-second minimum run durations, one change per experiment, and checkpoint commits after each phase.
-
Setup - Confirm the scenario, success criteria, and benchmark command. The benchmark must emit metrics between PERF_METRICS_START / PERF_METRICS_END markers.
-
Baseline - Run the benchmark for 60+ seconds with a 10-second warmup. Results are stored as baselines/<version>.json. This is the reference point for all comparisons.
-
Breaking point - Binary search across a parameterized range to find the threshold where performance degrades. Uses 30-second runs (the only exception to the 60-second rule).
-
Constraints - Run the benchmark under CPU and memory limits (default: 1 CPU, 1GB RAM). Compare against the unconstrained baseline to identify resource sensitivity.
-
Hypotheses - Generate up to 5 hypotheses about the root cause, each backed by evidence from git history and code analysis. No guessing - every hypothesis needs a file path or commit reference.
-
Code paths - Use repo-map to identify entry points, hot files, and call chains relevant to the scenario. Narrows the search space before profiling.
-
Profiling - Run language-specific profilers: --cpu-prof for Node.js, JFR for Java, cProfile for Python, pprof for Go. Capture file:line hotspots and flame graphs.
-
Optimization - Apply one change at a time. Validate with 2+ benchmark runs. Revert between experiments to keep the baseline clean.
-
Decision - Based on measured improvement, decide whether to continue optimizing or stop. The verdict and rationale are recorded in the investigation log.
-
Consolidation - Write the final baseline, close the evidence log, and mark the investigation complete.
Usage
/perf --phase setup --scenario "startup time" --command "node bench.js" --version v2.0.0
/perf --resume # continue where you left off
/perf --resume --phase profiling # jump to a specific phase
/perf --resume --phase baseline --runs 5 --aggregate median
/perf --resume --phase optimization --change "replaced O(n^2) loop with hash lookup"
/perf --resume --phase decision --verdict continue --rationale "20% improvement"
Key Flags
| Flag | Description |
|---|
--resume | Continue the active investigation |
--phase <name> | Target a specific phase |
--command <cmd> | Benchmark command (must emit PERF_METRICS markers) |
--version <ver> | Baseline version label |
--duration <sec> | Override benchmark duration (default 60s) |
--runs <n> | Number of runs for multi-run benchmarks |
--aggregate <method> | median, mean, min, or max (default median) |
--change <summary> | Description of the optimization being tested |
--verdict <v> | continue or stop |
Investigation Artifacts
State is persisted under {state-dir}/perf/: investigation.json (active state), investigations/<id>.md (evidence log), and baselines/<version>.json (metrics per version).
Architecture
agent-sh