Skip to content

Latest commit

 

History

History
22 lines (18 loc) · 2.17 KB

scripts.md

File metadata and controls

22 lines (18 loc) · 2.17 KB

Scripts

Everything under scripts/ reproduces a number in a report or packages a release. Each script prints its usage with no arguments. They read the clone copies they are given and never execute code inside them. The Python helpers under scripts/lib/ hold the scoring rules so that every baseline is scored by the same code as our own engine.

script reproduces report
gold-score.sh call-edge precision and recall of our engine (or code-review-graph through the adapter) against fixtures/gold/<project>/calls.jsonl engine parity, Plan 6
compare-engines.sh the full head-to-head with code-review-graph: Jaccards by kind, timing, the gold score engine parity
h2h-codegraph.sh codegraph-mcp scored on the same gold call labels, next to our engine engine parity ("codegraph head-to-head")
map-score.sh map precision@5 and recall@5, the relate band and the absent check against fixtures/gold-map/<clone>/features.jsonl mapping gold
map-baselines.sh the feature-mapping baselines (plain full-text OR search, codegraph symbol search) scored by the same rules engine parity ("Feature-mapping baselines")
gold-map-pairs.sh regenerates fixtures/gold-map/pairs.jsonl from the family pack, in a fixed order mapping gold
gold-map-merge.sh merges two readers' labels into the committed gold, with agreement and Cohen's kappa per clone mapping gold
install.sh, install.ps1 the user-facing installers: fetch the release asset for the platform, check it against SHA256SUMS, install it (scripts/test/install-script.test.mjs runs the shell one against local assets; the PowerShell one is untested until a Windows run) README Install

The floors in fixtures/gold/FLOORS.json and fixtures/gold-map/FLOORS.json are measured minus 0.01 and are only ever raised; cargo test -p comparegraph checks them when COMPAREGRAPH_STUDY points at a directory holding the clone copies named in the fixtures' READMEs, and says so on stderr when it does not.

Per-language graph digests (one fixture for each language id, re-golded with UPDATE_LANGUAGE_DIGESTS=1) are described in language-fixtures.md.