redline
A Claude Code plugin for automatic code review, adversarial review, and rescue delegation — powered by Codex.
Works with your existing OpenAI subscription, or route through OpenRouter for access to any model (GPT, Claude, Gemini, DeepSeek, and more).
The model decides
Redline's key principle: Claude decides what help it needs. After each response, a lightweight Stop hook asks whether code changes were made. If so, Claude evaluates the context and picks the most helpful action:
/redline:review — standard code review/redline:adversarial — challenge design decisions, probe hidden assumptions, test failure modes/redline:rescue — delegate a task to Codex as a smart friend
No hardcoded triggers, no diff thresholds. The model is in the best position to decide.
How it works
Claude Code Stop hook (fires after each response)
→ reminds Claude to consider /redline:... commands
→ Claude decides based on what it just did:
run a review, challenge the design, delegate to Codex, or skip
→ suppressed when already responding to a hook (no loops)
A non-user-invocable skill description stays in Claude's context at all times, providing the decision-making guidance. The hook is just a minimal nudge.
Reviews happen automatically — no manual invocation needed. You can also run any command directly at any time.
Install
/plugin install redline@alexanderatallah/redline
Then run /redline:setup to configure your provider, model, and effort level.
Development
claude --plugin-dir ./plugins/redline
Commands
| Command | Description |
|---|
/redline:setup | Configure provider (OpenAI or OpenRouter), model, effort, and routing |
/redline:review | Run a standard code review on uncommitted changes |
/redline:adversarial | Challenge design decisions, probe assumptions, test failure modes |
/redline:rescue <task> | Delegate a task to Codex for help when stuck |
/redline:review
Standard code review. Runs codex exec review --uncommitted in the background, assesses findings, and presents issues to the user.
/redline:adversarial
Goes beyond bug-finding. Challenges design decisions, probes hidden assumptions (what is the code silently relying on?), identifies failure modes (race conditions, resource exhaustion, stale state), and questions trade-offs. Asks: is this the right approach?
/redline:rescue
When you're stuck — hand the problem to Codex. Describe what you're working on and what you need help with. Codex works on it in the background. Results are presented faithfully — Claude doesn't filter or second-guess them. You decide which suggestions to act on.
Configuration
During /redline:setup, configure:
- Provider — use your existing OpenAI subscription, or route through OpenRouter for model choice
- Model (OpenRouter only) —
openai/gpt-5.4 (default), openrouter/auto, or any OpenRouter model slug
- Effort (OpenRouter only) — reasoning effort: minimal, low, medium, high (default: medium)
- Provider variant (OpenRouter only) —
:nitro (fastest, default), :floor (cheapest), or standard routing
Authentication
Redline supports two authentication methods:
OpenAI subscription — if Codex is already authenticated (codex login), Redline can use it directly. No additional setup needed.
OpenRouter — route through OpenRouter for access to any model. Set your API key via:
# Environment variable
export OPENROUTER_API_KEY=sk-or-...
# Or run OAuth login during setup
/redline:setup
Customization
Every command is a plain markdown file in commands/. Edit them to fit your project:
- Focus the review — add "pay special attention to SQL injection and auth boundaries" to
review.md
- Change the adversarial persona — make it focus on performance, security, or accessibility instead of general design
- Adjust rescue behavior — tell Codex to always write tests, or to explain its reasoning step-by-step
No scripts to modify, no config flags to learn. Just edit the markdown and /reload-plugins.
Why Redline?
Compared to other ways of reviewing Claude's code:
| Redline | Other plugins |
|---|
| Models | OpenAI subscription or any model via OpenRouter | Typically locked to one provider |
| Automatic reviews | Stop hook triggers automatically, model decides when to review | Manual invocation only |
| Customizable | Edit plain markdown commands to change review behavior | Commands are often hardcoded or complex to modify |
| Simplicity | ~13 files, no build step | Often 30+ files across scripts, agents, and configs |
Requirements
alexanderatallah