AI PR Review#
Manual-trigger AI security review for pull requests. Comment /review on any PR to get a focused security review.
Trigger Commands#
| Command | Model | Use When |
|---|---|---|
/review | Fast (default: gpt-5.4-mini) | Quick check, most PRs |
/review fast | Fast (default: gpt-5.4-mini) | Same as /review |
/review deep | Deep (default: gpt-5.4) | Complex changes, security-sensitive code |
What It Reviews#
The reviewer is tuned for Pipelock's security model. It flags:
- Weakened isolation or sandbox boundaries
- Implicit trust of model output
- Unsafe tool input/output handling
- Auth, policy, or permission bypass risk
- Race conditions in enforcement paths
- Missing validation where untrusted data crosses boundaries
- Logging or audit gaps
- Prompt injection escape vectors
It ignores style nits and generic suggestions. If nothing is wrong, it says so explicitly.
Setup#
Required GitHub Secrets#
Set these in Settings > Secrets and variables > Actions:
| Secret | Required | Description |
|---|---|---|
LITELLM_BASE_URL | If using LiteLLM | Your LiteLLM proxy URL (e.g., https://litellm.example.com/v1) |
LITELLM_API_KEY | If using LiteLLM | API key for LiteLLM proxy |
OPENAI_API_KEY | If not using LiteLLM | Direct OpenAI API key (fallback) |
PR_REVIEW_MODEL_FAST | No | Model for /review, /review fast, /review tests, /review docs, /review stats (default: gpt-5.4-mini) |
PR_REVIEW_MODEL_DEEP | No | Model for /review deep (default: gpt-5.4) |
GITHUB_TOKEN is provided automatically by GitHub Actions.
LiteLLM vs Direct OpenAI#
LiteLLM (preferred): Set LITELLM_BASE_URL and LITELLM_API_KEY. Point at whatever upstream model you want (OpenAI, Anthropic, local). The script sends OpenAI-compatible requests to your LiteLLM proxy.
Direct OpenAI (fallback): Set only OPENAI_API_KEY. The script calls api.openai.com directly.
If both are set, LiteLLM takes priority.
Switching Models#
Override the model via secrets:
PR_REVIEW_MODEL_FAST=gpt-5.4-mini # fast (~$0.02/review)
PR_REVIEW_MODEL_DEEP=gpt-5.4 # thorough (~$0.07/review)
With LiteLLM, use any model your proxy supports:
PR_REVIEW_MODEL_DEEP=anthropic/claude-sonnet-4-20250514
PR_REVIEW_MODEL_FAST=groq/llama-3.3-70b-versatile
Cost Control#
- Only runs when manually triggered (no auto-review on push)
- Diff is truncated to ~100k chars (~25k tokens) to cap costs
/review fastuses a cheaper model by default/review deepis opt-in for thorough analysis
Files#
| File | What |
|---|---|
.github/workflows/pr-review.yaml | GitHub Actions workflow |
scripts/pr-review.py | Review script (fetches diff, calls LLM, posts comment) |
