scip-query: Skill for Codex

SCIP Hyper Optimization

This project-local skill restores the last standalone hyper-optimization method from commit bb02f016 and adapts its retired command names to the current scip-query interface.

Use it to make a command, indexer, watcher, workflow, or service faster without changing its observable result. A hyper-optimization campaign is a bounded engineering program that improves runtime, peak memory, or computational work against repeatable measurements. What distinguishes it from general refactoring is that measurements and output identity decide whether a change survives.

Choose the scale

• QUICK — one target, one already-plausible hot operation, and one before/after measurement can settle the decision. Record the run in docs/benchmarks/runs/YYYY-MM-DD-<target>.jsonl; a separate ledger is optional.
• CAMPAIGN — several targets, an unclear bottleneck, cold and warm paths that diverge, memory/runtime tradeoffs, or competing designs. Create the baseline, run history, ledger, profiles, and alternative-design track below.

If naming the operation to change requires investigation, use CAMPAIGN.

Essential terms

A measurement harness is a repeatable experimental setup: commands, fixtures or corpora, cache state, environment facts, and result records. Its defining trait is that another run can reproduce the state being compared.

A run history is a machine-readable measurement record with one row per command, subprocess, or profiled stage. Its defining trait is durability across code changes, so an improvement can be compared to the baseline that motivated it.

A profile span is one named and timed operation inside the target process, such as fingerprinting, compiler startup, conversion, validation, publishing, or cache loading. Its defining trait is that it assigns elapsed work and relevant cardinality to a particular part of the pipeline.

Hierarchical profiling is a bottleneck-localization method that times the whole pipeline, then subdivides only its dominant span. Its defining trait is progressive narrowing until the expensive operation is concrete enough to change.