code-review-c

Standalone adaptive Codex review skill. Version 0.1.0, MIT licensed.

Review quality comes first. Initial workload limits are provisional, not measured model capacities. The first release supplies onboarding tools and runs no paid calibration campaigns.

Use

Install the stable v0.1.0 tag into your Codex skills directory:

git clone --branch v0.1.0 --depth 1 https://github.com/dk-code-factory/code-review-c.git
skill_parent="${CODEX_HOME:-$HOME/.codex}/skills"
mkdir -p "$skill_parent"
test ! -e "$skill_parent/code-review-c" && cp -R code-review-c/codex/skills/code-review-c "$skill_parent/"

If that skill already exists, move the old directory aside before installing; keep it for rollback. Alternatively, add the tagged checkout as a local Codex plugin marketplace. The package contains exactly one skill, code-review-c. It needs no sibling checkout, external review plugin, or build step. Select it in Codex and request:

• “Review PR URL with code-review-c.”
• “Preview a strategy for base REV and head REV.”
• “Prepare model onboarding; do not run trials.”

Linux is the first-release support target. CI covers CPython 3.11–3.15; other platforms and Python versions are unverified. Git and native Codex agents support review/preview; authenticated gh supports GitHub PR pinning. Docker is needed only for isolated onboarding trials and container integration checks. Models and effort inherit the host selection. A missing or incompatible profile uses a provisional 58,000-byte total visible-input budget. No measured profiles are distributed. Binary, submodule and type changes remain explicit gaps requiring separate inspection; indivisible oversized hunks also remain gaps. Dirty worktrees are recorded, but evidence comes only from pinned commits.

The controller identifies semantic relationships and applicable aspects; deterministic helpers enforce exact hunk/aspect assignment, byte budgets, task dependencies and whole-PR essential gates. Tasks queue within native concurrency. Security and serious resource/performance checks remain essential after correctness findings. Optional improvements open only after complete, verified, clean essential coverage.

Read the skill, controller procedure, record contracts, and onboarding procedure, and operating guide.

Development

python3 scripts/check.py
python3 scripts/test.py

Runtime and tests use the Python standard library. Tests simulate reviewers and calibration receipts; they make no model calls. Onboarding preparation freezes fixtures and candidate strategies without inference. Model trials require explicit call/time limits and an immutable Docker image with Codex installed. Trial containers see only pinned source/evidence and temporary Codex authentication, keeping answer keys outside the reviewer filesystem. Three matched repetitions, held-out fixtures, critical-defect recall, precision and completion gates precede profile promotion.

On a POSIX host with Linux Docker, run python3 scripts/test_runtime.py for the same lifecycle checks run by Docker CI. The tests build an image with a fake Codex executable and synthetic authentication; they run no inference. They cover stdin, read-only source isolation, output capture, retained failures, trial timeout and interruption, and version-probe cleanup. The pinned Python base image may need to be downloaded. These checks do not measure review quality or model capacity.

Keep review artifacts, campaign data, local profiles, credentials, private rules, caches and benchmark results outside the package and reviewed repository. Provenance for adapted review guidance is pinned in PROVENANCE.json, with upstream license notices retained in NOTICE. This package does not change the source code-review-ad repository or its configured reviewers.