subagentskills

.com agent skill

← all skills

determinism-harness

open cwc-science skl_determinism_harness
description (verbatim frontmatter)
Double-run determinism verification for any pipeline, generator, or seeded computation: runs a shell command TWICE in fresh temp working dirs, diffs the declared output files (byte-identical by default, or a normalized mode that first strips caller-declared volatile regex patterns like timestamps/run-ids), and returns a typed verdict {deterministic, first_divergent_file, diff_excerpt} plus a CLI with meaningful exit codes (0/1/2). Generalizes this repo's two real double-run precedents: the corpus pipeline's run-twice byte-identical reproducibility check and scripts/build-graph.py's byte-identical-on-unchanged-tree artifact discipline. Use when the user asks "is this pipeline deterministic", "verify reproducibility", "run it twice and diff", "check the generator doesn't churn", or before committing any generated artifact whose diff should mean something. Do NOT use for flaky-TEST hunting (run the test runner's own repeat tooling), for cross-machine reproducibility (this harness only certifies same-machine double-run identity — an honest, narrower claim), or on commands with side effects outside their working directory (each run gets a fresh temp cwd; external state is not isolated).
summary

Double-run determinism verification for any pipeline, generator, or seeded computation: runs a shell command TWICE in fresh temp working dirs, diffs the declared output files (byte-identical by default, or a normalized mode that first strips caller-declared volatile regex patterns like timestamps...

source

plugins/cwc-science/skills/determinism-harness/SKILL.md

provenance

created 2026-07-10 18:13:13 · JSON