OCR Enhanced Code Review

A Codex skill that combines the deterministic OpenCodeReview (OCR) CLI with Codex reasoning for high precision code reviews.

Repository layout

• skill/ — the installable skill, including its instructions, agent metadata, helper script, and validation dependencies.
• tests/ — tests for the skill metadata and OCR preparation helper.
• LICENSE — MIT License.

Usage

From a target repository, run the bundled preparation helper:

uv run --project skill skill/scripts/prepare_review.py --repo /path/to/repository --packets

The helper asks OCR for the reviewable file set and applicable rules without invoking an LLM. With --packets, it also emits schema-2, deterministic rule-group packets containing bounded diff segments and changed-line anchors. Packets are capped at 128 KiB by default and oversized groups or files are split deterministically; adjust this with --packet-max-bytes. Codex then reviews those packets and emits structured finding JSON; normalize_findings rejects out-of-scope or unchanged-line findings, deduplicates, and sorts the final output. Schema-1 manifests remain valid inputs.

The LLM is responsible for risk and semantic reasoning only. OCR/scripts own selection, rule resolution, changed-line mapping, packet construction, finding validation, deduplication, ordering, and rendering. A plan is requested only for packets above the configured churn/file thresholds (--plan-lines, --plan-files).

Development

Install validation dependencies and run the test suite:

uv sync --project skill --group validation
uv run --project skill pytest

The project requires Python 3.10 or newer.