Code Review

Code Review

Find bugs that affect results, reproducibility, data safety, or cluster use. Skip style nits.

Findings must be

• New in this change, unless asked for a broad audit.
• Impactful: name the input/scenario.
• Actionable: give the fix.

Severity

• P0: blocks commit/run/merge. Wrong result, lost data, secret leak, cannot rerun, serious cluster misuse.
• P1: fix before commit/merge. Realistic failure.
• P2: fix later.
• P3: nit. Do not report unless asked.

Verdict:

• LGTM: no P0/P1.
• FAIL: P0/P1 found.

Always check

• Hardcoded personal paths: /Users/..., C:\Users\..., /home/netid/..., /gpfs/scratch60/netid/....
• Raw data, secrets, env dirs, or large outputs staged.
• Missing seeds in simulation, bootstrap, train/test split, UMAP, random forest, stochastic optimizer.
• Silent data loss: dropna(), na.omit(), failed joins, duplicate drops, broad filters without counts.
• Lockfile drift: deps changed but uv.lock / renv.lock not updated.
• Raw data mutation: code writes into data/raw/ or overwrites inputs.
• HPC resource mismatch: requested 1 CPU but uses all cores; requested GPU but code does not use it.
• GPFS tiny-file storms.
• Package installs inside arrays.
• Scraping/API code without rate limit, cache, retry/backoff.

Process

1. Read changed files fully.
2. Check git status --short and staged files.
3. Run existing checks if practical.
4. Run a small realistic smoke test if practical.
5. For Slurm changes: check resources, thread env vars, paths, logs, resumability.

Output

## Findings

### [P1] Missing seed makes bootstrap estimates non-reproducible
**File**: scripts/bootstrap.py:42
**Issue**: `np.random.default_rng()` is called without a fixed seed.
**Fix**: create `rng = np.random.default_rng(42)` at the entry point and pass it in.

## Verdict
FAIL

If clean:

## Findings
None.

## Verdict
LGTM

Checklist

• Changed files read
• Staged/untracked files checked
• Seeds checked
• Lockfiles checked
• Smoke test/check run if practical
• Slurm/resource behavior checked if relevant

Further reading

• Google code review guide

Source: yale-som-hpc/claude-code-marketplace — distributed by TomeVault.