Scitex Repro

Reproducibility helpers for scientific Python experiments — unified seeded random-state for `random`/`numpy`/`torch`/`tf`, experiment ID + timestamp generation, and deterministic hashing of numpy arrays. Public API (9 symbols) — random state (`RandomStateManager` — class that fixes seeds across `random`, `numpy.random`, `torch`, `tensorflow`, `os.environ[PYTHONHASHSEED]` in one call; `get(seed=42)` factory returning a configured instance; `reset(seed=None)` rewinds all RNGs), ID/timestamp generation (`gen_ID(n=8)`, `gen_id(n=8)` alias, `gen_timestamp()`, `timestamp()` alias — producing strings suitable for experiment/run directory names), array hashing (`hash_array(arr)` — deterministic hex digest of a numpy array's values for experiment fingerprinting), legacy shim (`fix_seeds(seed=...)` — deprecated wrapper kept for older scripts). No CLI, no MCP tools. Drop-in replacement for hand-rolled "seed all frameworks" functions (`random.seed` + `np.random.seed` + `torch.manual_seed` + `torch.cuda.manual_seed_all` + `tf.random.set_seed` boilerplate), ad-hoc timestamp strings via `datetime.now().strftime(...)` in run dirs, and bespoke `hashlib.sha256(arr.tobytes())` one-liners. Use whenever the user asks to "seed this experiment deterministically", "seed numpy and torch together", "generate an experiment ID for this run", "make an output folder name like 20261023_1530_abc12345", "hash a numpy array for reproducibility", "reset the RNG mid-experiment", or mentions `RandomStateManager`, `gen_ID`, `hash_array`, `scitex.repro`, seed-everything.

mkdir -p ~/.claude/skills/scitex-repro
curl -L https://claudskills.com/skills/scitex-repro/SKILL.md \
  -o ~/.claude/skills/scitex-repro/SKILL.md