Audit Python latency, CPU, memory, I/O, and throughput with reproducible profiles, controlled experiments, correctness checks, and regression gates.
A slow Python path rarely needs a random collection of faster-looking idioms. It needs a workload you can reproduce, a measurement that matches the user-visible problem, evidence that localizes the cost, and a change that preserves behavior.
This skill turns latency, CPU, memory, allocation, I/O, startup, and throughput problems into a controlled performance audit. It separates wall time from CPU and wait time, matches profilers to the symptom, records the environment and workload, and requires one-variable experiments. Every proposed optimization gets semantic-equivalence checks, a comparison rule, a stop condition, and a rollback.
Use it to diagnose a slow API, a memory-heavy batch, a concurrency collapse, or an optimization PR whose benchmark is too weak to support release. The output includes an evidence register, baseline matrix, causal findings, experiment ledger, correctness matrix, comparison results, and regression gates. It refuses percentage wins without the absolute baseline and keeps lab results separate from production claims.
Version 1.0.0 was reviewed on 2026-07-16 and evaluated with gpt-5.6-sol in 12 fresh read-only contexts. The candidate scored 100 across three cases, compared with 79.58 for the same agent without the skill and 79.58 for the exact free upstream. The cases covered API tail latency and unsafe caching, Python and native memory retention, and a multiprocessing change that broke deduplication semantics. This evaluation measures the supplied cases, not live Python performance or buyer outcomes. The reviewed package SHA-256 is 16cfa283181eafe6f40698ae7eb8f6df2dcea4657cffcf09102a970e648fb48b.
The package includes a read-only checker for an optional canonical JSON audit record. Passing it confirms structure, references, and verdict precedence. It does not run a profiler or prove that recorded measurements are correct. Maintenance and install support are handled by the platform-operated publisher for this listing.
Input
Review an endpoint that reports p95 latency between 420 and 780 ms. A one-run cProfile says JSON serialization is hot, but requests also wait on PostgreSQL and an HTTP fraud check. The proposed PR adds an unbounded cache and claims a 40% win from five local runs.
Output
A measure-first verdict; workload and environment gaps; wall, CPU, database, and external-wait separation; cache-key and tenant-isolation risks; comparable sample plan; one-variable experiments; semantic gates; production guardrails; and rollback triggers.
Input
Audit a Python 3.12 worker that processes 800 CSV files. RSS rises from 300 MB to 6.2 GB. tracemalloc shows retained dictionaries, a C extension is involved, and the proposed generator rewrite changes error ordering and retry boundaries.
Provide the Python code or repository context, the slow or memory-heavy path, the affected workload, current measurements or profiles, Python and dependency versions, environment constraints, correctness invariants, and permission to run or change anything. Ask for diagnosis, an experiment plan, implementation when authorized, or a release review. The skill keeps unrun commands and hypotheses separate from observed results.
The request, job, batch, data shape, concurrency, cache state, and user-visible latency, throughput, CPU, memory, allocation, I/O, startup, or cost problem.
Python implementation and version, operating system and architecture, dependency or lockfile revision, process model, CPU and memory limits, and relevant external services.
Available benchmark samples, profiles, traces, tests, expected values, exceptions, ordering, side effects, resource-lifecycle rules, and compatibility limits.
Whether the agent may only review supplied evidence, run local measurements, install profiling tools, prepare patches, or apply a contained change. Production attachment requires explicit authority.
SKILL.md; Python performance audit rubric; report and canonical JSON output contract; read-only JSON audit checker; Agent Skills interface metadata
No reviews yet.
Output
Separate RSS, Python allocation, native memory, and lifecycle findings; phase-bound snapshots; retention experiments; batch-size and cleanup hypotheses; error-order and retry equivalence tests; peak and steady-state gates; and an honest release verdict.
Input
Review a PR that replaces a stable Python deduplication loop with set conversion and multiprocessing. Its one-shot time.time benchmark drops from 2.1 to 0.8 seconds, but output order matters, duplicate diagnostics disappear, small inputs regress, and worker startup is not included.
Output
A do-not-release verdict; invalid benchmark findings; ordering and diagnostic regressions; size-distribution and startup-aware comparisons; algorithm-only and process-topology experiments kept separate; correctness gates; and a bounded alternative with regression budgets.
Creator
Vvareno
