csv-quality-gate

csv-quality-gate is a command-line data quality gate that runs CSV preflight validation, failing fast before a pipeline ingests broken, incomplete, duplicated, or junk input.

It runs batch quality checks on a CSV and returns pass, warn, or fail (with matching exit codes) before expensive pipeline steps burn time on bad input. It checks for missing required columns, empty files, empty critical cells, duplicate rows, and — under the outreach profile — suspicious company-name patterns. Teams can declare their own columns, thresholds, and patterns in a small TOML/JSON config, every issue points at the affected line numbers (never cell values), and the same gate runs as a pre-commit hook or a GitHub Action. It is stdlib-only: no third-party runtime dependencies.

The problems it is built for:

• "We keep running expensive pipeline steps on broken CSVs."
• "A batch run fails 20 minutes in because the input CSV was junk."
• "We only discover missing required columns after the job already started."
• "Duplicate rows and empty contact fields keep polluting our batch runs."
• "The gate says 12% duplicates, but which rows?"
• "Our CSVs have email and order_id, not company."
• "I want CSV preflight validation, not a whole data platform."

Quickstart (60 seconds)

pip install csv-quality-gate
csv-quality-gate check leads.csv --profile outreach

Example output for a CSV with a missing column and a borderline duplicate rate:

csv-quality-gate: FAIL
file: leads.csv
profile: outreach
rows: 125
  ERROR: missing required column: person_name
  WARNING: duplicate rate 12% exceeds warning threshold 10%
    evidence: column=company affected=15 rows at line(s) 4, 9, 15, 22, 31 (+10 more)

The process exits 0 on pass, 1 on warnings only, and 2 on fail, so you can wire it directly into a shell script, a pre-commit hook, or a CI step.

Try it with the included fixtures

You can verify the gate before connecting it to your own pipeline. From a checkout, install the package and run the included clean and broken inputs:

git clone https://github.com/hermes-labs-ai/csv-quality-gate.git
cd csv-quality-gate
python -m pip install .
csv-quality-gate check examples/clean.csv
csv-quality-gate check examples/promptfoo-dataset/tests-broken.csv \
  --config examples/promptfoo-dataset/csv-quality-gate.toml \
  --profile promptfoo

The first command exits 0 with PASS; the second exits 2 with bounded line-number evidence for the broken rows. This gives a first-use check without creating a sample CSV or guessing which profile and config to choose.

Install

pip install csv-quality-gate

For development:

pip install -e ".[dev]"

As an agent skill (Claude Code, Codex CLI, Gemini CLI)

The repository root is a portable Agent Plugin (plugin.json) that ships one skill, skills/csv-quality-gate/SKILL.md. It teaches the agent to run this gate on the CSV files you name and report the status and line-number evidence without overclaiming.

Host	Install	Read back
Claude Code	claude plugin marketplace add hermes-labs-ai/csv-quality-gate claude plugin install csv-quality-gate@csv-quality-gate	claude plugin list
OpenAI Codex CLI	codex plugin marketplace add hermes-labs-ai/csv-quality-gate codex plugin add csv-quality-gate@csv-quality-gate	codex plugin list
Gemini CLI	gemini extensions install https://github.com/hermes-labs-ai/csv-quality-gate --ref main	gemini skills list
skills.sh	npx skills add https://github.com/hermes-labs-ai/csv-quality-gate --skill csv-quality-gate	npx skills list

What each host reads:

• Claude Code reads .claude-plugin/marketplace.json (source .) and .claude-plugin/plugin.json.
• Codex reads the repo marketplace .agents/plugins/marketplace.json (source ./, the root) and the portable plugin.json.
• Gemini CLI reads gemini-extension.json and discovers the skill under skills/. Keep --ref main: without a ref, Gemini CLI installs the latest GitHub release archive, and releases up to v0.3.0 predate gemini-extension.json.

Installing the skill does not install the Python package. The skill uses an installed csv-quality-gate or runs the pinned release with uvx csv-quality-gate==0.3.0.

Usage

csv-quality-gate check leads.csv
csv-quality-gate check leads.csv --profile outreach
csv-quality-gate check leads.csv --profile generic --json
csv-quality-gate check leads.csv --config csv-quality-gate.toml --profile leads
csv-quality-gate check leads.csv --max-examples 20
csv-quality-gate check data/*.csv

Options for check:

• --profile NAME — a built-in profile (generic, outreach) or one declared in --config.
• --config FILE — TOML or JSON file with project-specific profiles (see Custom profiles).
• --max-examples N — how many affected line numbers each issue lists (default 5; 0 keeps counts only).
• --json — machine-readable output.

Several paths can be checked in one call. Text mode prints one block per file, JSON mode emits an array (a single path keeps the plain object), and the exit code is the worst status across files.

Exit codes:

• 0 pass
• 1 warnings only
• 2 fail (or 2 when the file does not exist, is not UTF-8, the profile is unknown, or the config is invalid)

Profiles

Built-in profiles:

• generic
  • checks for a required company column, empty company cells, duplicate company values, and empty files
• outreach
  • requires company and person_name, with higher empty-rate tolerances and an added suspicious company-name heuristic for GTM/contact pipelines

The thresholds for each profile are defined in src/csv_quality_gate/profiles.py.

Custom profiles

Declare your own profiles in a TOML (Python 3.11+) or JSON file and pass it with --config. A copyable example lives in examples/csv-quality-gate.toml.

[profiles.leads]
extends = "outreach"                 # optional: inherit a built-in's columns, patterns, and rates
required_columns = ["email", "company", "person_name"]
critical_columns = ["email", "person_name"]
duplicate_column = "email"
empty_warning_rate = 0.05
empty_fail_rate = 0.20
duplicate_warning_rate = 0.02
duplicate_fail_rate = 0.10
suspicious_column = "email"
suspicious_patterns = ['^[^@]+$']    # case-insensitive regular expressions
suspicious_warning_rate = 0.20
suspicious_fail_rate = 0.50

csv-quality-gate check data/leads.csv --config csv-quality-gate.toml --profile leads

Rules:

• Every key is optional. Without extends, a profile starts with no columns, no patterns, and the generic thresholds.
• A rate of 0.0 means "any occurrence": a check with no affected rows never raises an issue, so a clean file still passes.
• Config is data only: column names, rates between 0 and 1, and regex strings. Unknown keys, out-of-range rates, a warning rate above its fail rate, and invalid regexes are rejected with a precise message, and the CLI returns a normal fail receipt (exit 2) instead of a traceback.
• Built-ins stay available next to your profiles. A custom profile with the same name as a built-in replaces it for that run only.
• Regexes come from your own repository; keep them simple, since the gate does not guard against pathological patterns.

Output

Text mode (default):

csv-quality-gate: FAIL
file: leads.csv
profile: outreach
rows: 125
  ERROR: missing required column: person_name
  WARNING: duplicate rate 12% exceeds warning threshold 10%
    evidence: column=company affected=15 rows at line(s) 4, 9, 15, 22, 31 (+10 more)

Every empty-rate, duplicate-rate, and suspicious-value issue carries bounded evidence: the affected column, the total number of affected rows, and up to --max-examples physical line numbers (the header is line 1; a quoted record that spans several lines reports its last line). Cell values are never printed, so receipts stay safe to attach to CI logs or tickets. Duplicate evidence points at the second and later occurrences, so those are the lines to remove.

JSON mode (--json) emits an object with path, profile, rows, status, issues[], and (when --config is used) config. Each issue has severity and message; issues backed by rows add an evidence object:

csv-quality-gate check leads.csv --json

{
  "path": "leads.csv",
  "profile": "outreach",
  "rows": 125,
  "status": "fail",
  "issues": [
    {"severity": "error", "message": "missing required column: person_name"},
    {
      "severity": "warning",
      "message": "duplicate rate 12% exceeds warning threshold 10%",
      "evidence": {"column": "company", "total": 15, "rows": [4, 9, 15, 22, 31]}
    }
  ]
}

Limitations / What it does not do

• Heuristics are intentionally simple: empty-rate, duplicate-rate, and regex-based name patterns. They do not learn from your data.
• It validates shape and obvious noise, not semantic correctness — it cannot tell whether company values are real, only whether they are present, unique, and not obviously junk.
• The outreach profile is opinionated. Its suspicious-name patterns and thresholds were chosen for GTM contact lists and should not be treated as universal truth.
• Duplicate and empty checks operate on the columns a profile names; it does not auto-detect which columns matter. Built-ins use company and person_name; use a config file for anything else.
• Evidence is line numbers and counts only. It never quotes cell values, so it cannot tell you what a bad value was, only where it is.
• It validates each CSV file independently and assumes UTF-8 (BOM-tolerant), comma-separated input.
• It is not a data quality platform: no lineage, no profiling reports, no schema inference, no row-level remediation.

When to use it

• Before enrichment, outreach, ETL, or batch scoring runs
• As a pre-commit hook or CI step for checked-in CSV inputs
• As a preflight gate before expensive pipeline work

When not to use it

• When you need semantic validation of the data itself
• When your input is not CSV
• When you need a full data quality framework with lineage and profiling

pre-commit

Run the same gate on staged CSV files before they are committed. Add this to your .pre-commit-config.yaml (full example in examples/pre-commit-config.yaml):

repos:
  - repo: https://github.com/hermes-labs-ai/csv-quality-gate
    rev: v0.3.0
    hooks:
      - id: csv-quality-gate
        args: [--profile, outreach]
        # or: args: [--config, csv-quality-gate.toml, --profile, leads]

The hook receives every staged .csv file, prints one report block per file, and blocks the commit on warn or fail, matching the CLI exit codes. Narrow it with pre-commit's files: pattern if only some CSVs should be gated, and tune thresholds through a config file rather than skipping the hook.

To try the hook from a local checkout before pinning a release, point a config at the checkout path and a commit:

repos:
  - repo: /path/to/csv-quality-gate
    rev: <commit sha>
    hooks:
      - id: csv-quality-gate
        args: [--profile, outreach]

pre-commit run --config that-file.yaml csv-quality-gate --files data/leads.csv

pre-commit try-repo . csv-quality-gate --files data/leads.csv also works for the default generic profile (try-repo cannot pass hook args).

CI / GitHub Actions

Use this repository directly as a composite Action. It installs the packaged CLI, runs the selected profile, and always writes a JSON receipt at $GITHUB_WORKSPACE/csv-quality-gate-receipt.json. The status and receipt outputs remain available even when the Action exits with a warning or failure.

- id: csv_gate
  uses: hermes-labs-ai/csv-quality-gate@v0.3.0
  with:
    csv-path: data/leads.csv
    profile: leads
    config: csv-quality-gate.toml   # optional; omit to use built-in profiles

- run: echo "${{ steps.csv_gate.outputs.status }}"

The only inputs are csv-path, profile, and the optional config file path; the Action deliberately accepts no free-form command or shell arguments. It returns the same exit codes as the CLI: 0 for pass, 1 for warn, and 2 for fail. The receipt contains the same bounded evidence as --json, so it is safe to upload as a workflow artifact. Use continue-on-error: true on a calling step if your workflow needs to inspect warning or failure outputs before deciding how to proceed.

Set summary: true to write a compact result to the GitHub Actions job summary. It contains only status, profile, row count, and issue counts; it never includes the CSV path or cell values. Evidence column labels are omitted unless you also set summary-columns: true; enable that only when those labels are suitable for the job summary. The additive tool-version and summary-written outputs identify the installed package version and whether this invocation rendered the summary; the JSON receipt is unchanged.

The receipt path is fixed per workspace, so do not run more than one instance in parallel in the same workspace. The Action validates the package's existing CSV heuristics only; it does not add schema inference, semantic verification, or arbitrary CLI options.

A ready-to-copy install-based workflow also lives in examples/github-action.yml.

GitHub Marketplace

action.yml carries the metadata a Marketplace listing needs (name, description, author, branding). Listing is a maintainer step taken on the GitHub release form, not something the repository does on its own; the steps are in Releasing. Whether you reach the Action through Marketplace or this repository, the uses: line above is the same: pin a release tag (@v0.3.0) or a commit SHA.

Recipes

Short, copyable setups that put the gate in front of a specific tool. Each uses a custom profile and ships with one passing and one failing fixture you can run from a checkout:

• dbt seeds — gate seeds/*.csv before dbt seed.
• promptfoo datasets — gate a CSV test set before promptfoo eval.

Development

pip install -e ".[dev]"
ruff check .
python3 -m pytest -q
pre-commit try-repo . csv-quality-gate --files tests/fixtures/clean.csv   # optional hook smoke test

Part of the Hermes Labs reliability stack

csv-quality-gate is part of the Hermes Labs reliability stack — open-source tools that catch silent failure modes in production AI and data pipelines. csv-quality-gate guards the data that goes into a pipeline; it is complementary to, not a replacement for, the agent- and prompt-level tools in the stack.

About Hermes Labs

Hermes Labs is an AI reliability engineering studio for product and engineering teams shipping production agents and LLM applications. We find the structural AI failures standard evals miss, then harden retrieval, memory, agents, and the language layers around production AI systems with runtime controls and defensible evidence.

Browse the open-source catalog or contact roli@hermes-labs.ai.

License

MIT — see LICENSE.

Citation

If you use this software, please cite it using the metadata in CITATION.cff.