PatchPilot

Dependency scanners are good at telling a maintainer that a vulnerable package is installed. The hard work starts after the alert: Is the dependency actually used? Which code path touches it? What is the smallest compatible fix? What should be tested? What uncertainty remains?

PatchPilot turns one npm vulnerability alert into a human-reviewed, evidence-backed patch. Its deliberately narrow golden path is:

Detect → investigate → approve → patch → test → rescan → report.

The demo starts with a real OSV-Scanner finding for GHSA-9c47-m6qq-7p4h in direct dependency json5@1.0.1. PatchPilot cites the exact import and call site in the bundled repository, keeps scanner facts separate from GPT-5.6 interpretation, and shows the remaining unknowns. It then proposes an exact four-file remediation plan. Nothing is written until a person approves that plan.

After approval, PatchPilot creates an isolated Git worktree, upgrades json5 to 1.0.2, applies one bounded compatibility repair, and adds one benign targeted regression test. It compares baseline and post-patch tests and builds, rescans the patched lockfile, and generates Markdown/JSON evidence. The final handoff is a clean local commit with accurate PR-ready copy; remote publication remains locked.

How Codex was used

Most of PatchPilot was built in one persistent Codex task. Codex helped turn the specification into issue-sized milestones, scaffold the TypeScript workspace, implement and test the scanner/evidence/approval/worktree/patch/verification/report/handoff pipeline, diagnose real Git-worktree and OSV-Scanner behavior on macOS, rehearse the browser path from clean resets, maintain the build journal, and commit each accepted issue.

How GPT-5.6 was used

GPT-5.6 has four bounded runtime roles: affectedness interpretation from normalized advisory facts and cited repository evidence; minimal remediation planning; one exact source-function compatibility proposal with at most two attempts; and one benign targeted regression-test proposal. Responses use Structured Outputs backed by Zod and semantic validation. OSV, Git, npm, test, build, and rescan output remains deterministic authority. When no API key is configured, visibly labeled cached-demo fixtures exercise the same contracts without claiming a live response.

Implementation and safety

The React UI talks to a Node/TypeScript orchestrator that invokes OSV-Scanner directly, gathers bounded repository evidence, validates model output, and performs approval-gated Git/npm operations. Paths are canonicalized and boundary-checked; subprocesses use fixed executable/argument arrays, timeouts, output caps, and redaction. Exact changed-file checks run after every mutation stage.

This is a complete hackathon demo, not a production security platform. It supports Node.js 20+, npm, local Git, macOS/Linux, one bundled direct vulnerability, and one validated remediation path. It does not claim runtime reachability, arbitrary repository support, Windows support, other ecosystems, automatic pull requests, merging, deployment, monitoring, SBOMs, or organization dashboards. A clean rescan proves only that the selected advisory is absent from the patched lockfile.

Verification

The repository includes MIT licensing, clean-clone setup, a deterministic reset, live OSV smoke verification, 62 network-independent tests, production builds, real UI screenshots, and three consecutive end-to-end rehearsals without manual repair. Judges can run the complete local workflow with ./demo/run-demo.sh or verify the aggregate suite with npm run check.

Built With

• codex
• git
• gpt-5.6
• node.js
• openai-responses-api
• osv-scanner
• react
• typescript
• vitest
• zod