القائمة
Run cipherlib benchmarks and regenerate BENCHMARK.md plus its README embed safely, with the reference-machine guard. Use when measuring a performance change or when benchmark tables need refreshing.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Run the cipherlib release ritual — local pre-flight mirroring release-test.yml, CHANGELOG entry, single version-bump commit, matching vX.Y.Z tag, tag push, and CI watch through pub.dev publish. Use ONLY when the maintainer explicitly asks to release, publish, or bump the version.
Scaffold a new cipher algorithm, AES mode, or AEAD variant end-to-end in cipherlib house style — implementation, barrel wiring, canonical test file with KAT placeholders, integration parity, benchmark, README row, and CHANGELOG entry. Use whenever adding any new algorithm to this package.
The cipherlib verification gate — run after completing any change and before declaring it done. Ordered, fail-fast, CI-parity checks with a pass/fail report. Also covers when to add node-platform and integration runs.
استنادا إلى تصنيف SOC المهني
| name | benchmark |
| description | Run cipherlib benchmarks and regenerate BENCHMARK.md plus its README embed safely, with the reference-machine guard. Use when measuring a performance change or when benchmark tables need refreshing. |
BENCHMARK.md is generated output, and its numbers are only meaningful from
the reference machine (36GB Apple M3 Pro, compiled exe). The README embeds
BENCHMARK.md verbatim. This skill keeps all three honest.
sysctl -n machdep.cpu.brand_string
Apple M3 Pro → full run allowed, results may be committed.BENCHMARK.md or the README benchmark section, and you must say so in your
summary. Committed numbers from other machines mislead package users.A. Single algorithm (fast, for perf-change feedback):
sh scripts/compiled.sh benchmark/<algo>.dart
This runs the algorithm's matching test file first (KAT safety), compiles the
benchmark to build/, and runs it over 1MB/1KB/32B inputs. To measure a
change: run on the pre-change commit (git stash the work), record, restore,
run again, report the ratio per input size. Anything beyond ±5% on 1MB inputs
is signal; 32B numbers are noisy — say so when reporting them.
B. Full regeneration (only on the reference machine):
sh scripts/benchmark.sh
This runs the full VM test suite, compiles benchmark/benchmark.dart,
rewrites BENCHMARK.md, and appends the machine/SDK footer. It takes several
minutes (each benchmark warms up 100ms and exercises 2s, per size, per
library). Close other heavy processes first; a loaded machine skews results.
Diff BENCHMARK.md against git:
sh scripts/compiled.sh benchmark/<algo>.dart) before accepting it.**bold** winner marking and `N.NNx slow` annotations are
generated — never hand-edit them.dart --version.README.md contains BENCHMARK.md verbatim between a pair of identical
markers:
<!-- file: BENCHMARK.md -->
...embedded copy...
<!-- file: BENCHMARK.md -->
Replace everything between the two markers with the new BENCHMARK.md
content (keep the markers). The same mechanism embeds
example/cipherlib_example.dart between its own marker pair — leave that one
alone unless the example changed.
git diff should show only BENCHMARK.md + the README benchmark section
(plus whatever perf change prompted this).dart test -p vm) and state the measured delta per input size in your
summary. AGENTS.md §5 "Performance change" is the bar.Update benchmarks or
Update BENCHMARK.md after <change>.