Changelog

Unreleased

• The composite GitHub Action now accepts newline-separated csv-paths for a batch check. csv-path remains supported for one file; exactly one path input is required. Batch receipts are JSON arrays and retain the CLI's worst-status exit behavior.
• Opt-in Action summaries now aggregate batch receipts without exposing paths or issue messages.

0.3.1 - 2026-09-12

• The repository root is now a portable Agent Plugin (plugin.json, Agent Plugins 1.0.0) with Claude Code, Codex CLI, and Gemini CLI manifests that all resolve one skill, skills/csv-quality-gate/SKILL.md. Runtime code is unchanged.

• Added an opt-in composite Action job summary with status, profile, row count, issue counts, and an opt-in bounded evidence-column list. The JSON receipt and CLI exit codes are unchanged; the Action also exposes an additive tool-version output.

• The composite Action declares author and Marketplace branding in action.yml; inputs, outputs, and behavior are unchanged.

• Added ecosystem recipes with one passing and one failing fixture each: examples/dbt-seed.md (gate seeds/*.csv before dbt seed) and examples/promptfoo-dataset.md (gate a CSV test set before promptfoo eval).

• Documented the release and Marketplace-listing steps in CONTRIBUTING.md.

0.3.0 - 2026-09-05

• Added project-specific profiles: check --config FILE --profile NAME loads TOML (Python 3.11+) or JSON profiles declared under [profiles]. Profiles are data only (column names, thresholds, regex strings) and may extends a built-in. Invalid files are rejected with a precise message and a normal fail receipt. Built-in profiles and their behavior are unchanged.
• Added bounded evidence to issues: each empty-rate, duplicate-rate, and suspicious-value issue now carries the affected column, the count of affected rows, and up to --max-examples (default 5) physical line numbers. Cell values are never emitted. JSON issues gain an optional evidence object; issues without evidence keep the 0.2.0 shape.
• The check command accepts several CSV paths; text output prints one block per file, JSON output becomes an array, and the exit code is the worst status.
• Added a pre-commit hook (.pre-commit-hooks.yaml, id csv-quality-gate) so the same gate runs on staged CSV files before commit.
• The composite Action accepts an optional config input; csv-path and profile behave exactly as before.
• Exposed suspicious_warning_rate and suspicious_fail_rate on Profile (defaults keep the previous fixed 10% / 30%). Public API additions: Profile, Evidence, ConfigError, load_config, and keyword-only profile, max_examples, and config parameters on validate_csv.
• CI now tests Python 3.10 and 3.11 and exercises the pre-commit hook.
• Fixed: a file that is not valid UTF-8 now yields a normal fail receipt instead of a traceback (which the Action previously reported as warn).
• Fixed: a header with surrounding whitespace (for example company ) is now matched and read as that column instead of being counted as fully empty.
• Fixed: a header-only file no longer adds a false missing required column error next to csv has no data rows; the status is still fail.
• A warning or fail rate of 0.0 now means "any occurrence"; a check with no affected rows never raises an issue. Built-in thresholds are unaffected.
• Options may appear between CSV paths (check a.csv --json b.csv).

0.2.0 - 2026-09-05

• Added a root composite GitHub Action with typed CSV path and profile inputs.
• The Action writes a JSON receipt and exposes its status and receipt path while preserving the CLI's pass/warn/fail exit codes.

0.1.2 - 2026-08-04

• Clarified the README, limitations, and outreach-profile heuristic without changing validation behavior.
• Added canonical package links and modernized license metadata.
• Added a tag-bound OIDC Trusted Publishing workflow for PyPI.

0.1.1 - 2026-05-30

• Added citation and Zenodo metadata.
• Aligned the public repository and Hermes Labs identity surfaces.

0.1.0

• Initial public release.
• Added generic and outreach validation profiles.
• Added text and JSON output modes.
• Added fixture-based tests for pass, warning, and fail cases.