Who it is for
Engineering and security teams that need to review code changes, validate proposed fixes, and preserve an auditable process.
02 / Code security review & repair
SecurePR Agent
Moving code review from finding risk to evidence-backed repair suggestions.
System preview
SecurePR Agent
A real capture of the current product interface. Select the preview to enter an available demo.
Enter systemProject overview
A system for PR risk discovery and secure repair, designed to make multi-agent collaboration interruptible, recoverable, and verifiable.
What it does
- Coordinate specialist reviewers over a PR diff
- Locate risks with AST and rule-based evidence
- Generate and verify repair suggestions before human approval
How to explore it
- Choose or paste a PR diff to review.
- Start the run and follow each reviewer’s task state and evidence.
- Inspect findings, proposed fixes, verification results, and the run trace.
The demo uses controlled examples and never changes a real repository. Start with a preset case.
Enter systemThe challenge
Long-running agents encounter failed calls, exhausted budgets, and interruptions. Restarting wastes cost and hides the reason for failure.
The approach
- A Runtime Harness manages task state, token / time budgets, checkpoints, and resume-from-breakpoint.
- Lead, Specialists, Critic, Fixer, and Verifier form a review-and-repair loop.
- Prompt / Skill versions pass through Validation and Holdout gates for deduplication, audit, and rollback.
The outcome
On 100 synthetic, controlled PR diffs, reproducible risk-identification F1 improved from 81.9% to 91.3%.
Boundaries & reflections
The evaluation set is project-built and controlled; it is not a production security conclusion. Human review remains required.