Code Review Context
Guide contributors on how Codex builds bounded, cache-friendly review context for model inference—not a end-user code review checklist.
Overview
Code Review Context is an agent skill for the Ship phase that defines bounded, incrementally built model context rules for Codex code-review inference.
Install
npx skills add https://github.com/openai/codex --skill code-review-contextWhat is this skill?
- Six rules: no history rewrite, minimize cache-invalidating context churn, bounded injected items
- Hard cap: no context item larger than 10K tokens
- P0 flag for new individual items that can exceed ~1K tokens
- All fragments must live in `core/context` and implement ContextualUserFragment
- Incremental context build only—no unbounded injection
- 6 numbered context rules
- 10K token hard cap per injected item
- P0 manual review threshold for items crossing ~1K tokens
Adoption & trust: 1 installs on skills.sh; 89.7k GitHub stars; 3/3 security scanners passed (skills.sh audits); trending (+100% hot-view momentum).
What problem does it solve?
Review agents accumulate unbounded or reshuffled context that breaks caching, blows token limits, or hides oversized fragments from review.
Who is it for?
Developers working on Codex or building review agents who need cache-stable, capped context injection patterns.
Skip if: Solo builders seeking a generic PR review rubric, security audit steps, or language-specific style guides without touching agent context plumbing.
When should I use this skill?
Model-visible context for Codex code review—adding or changing injected fragments.
What do I get? / Deliverables
Contributors add only bounded ContextualUserFragment structs in `core/context` with explicit caps and P0 review for large new items.
- ContextualUserFragment implementations
- Bounded context injection spec compliant with listed rules
Recommended Skills
Journey fit
Material governs what enters the model during automated review in Codex—canonical shelf is Ship → review. Review subphase matches context fragments and history policy for inference-time code review, not security scanning or launch prep.
How it compares
Internal Codex context policy, not a standalone PR review skill package.
Common Questions / FAQ
Who is code-review-context for?
Codex contributors and advanced agent builders designing inference context for automated code review—not typical app-only solo shippers.
When should I use code-review-context?
During Ship review work when adding or changing injected review context in Codex, or when mirroring its bounded-fragment pattern in your own review agent.
Is code-review-context safe to install?
Use the Security Audits panel on this Prism page; the skill describes repository internals and has no install-side permissions of its own.
SKILL.md
READMESKILL.md - Code Review Context
Codex maintains a context (history of messages) that is sent to the model in inference requests. 1. No history rewrite - the context must be built up incrementally. 2. Avoid frequent changes to context that cause cache misses. 3. No unbounded items - everything injected in the model context must have a bounded size and a hard cap. 4. No items larger than 10K tokens. 5. Highlight new individual items that can cross >1k tokens as P0. These need an additional manual review. 6. All injected fragments must be defined as structs in `core/context` and implement ContextualUserFragment trait