PatchPilot

PatchPilot automates dependency CVE remediation end-to-end on Qwen Cloud.

Inspiration

Every production service has the same problem: stale dependencies with known CVEs that nobody patches because the process is tedious, risky, and requires reading changelogs that nobody has time for. Dependabot opens PRs. Engineers merge them six months later, if ever.

What it does

PatchPilot is a CVE-remediation autopilot: point it at a repo and it scans installed packages against OSV.dev, sends findings to Qwen (qwen-plus on Alibaba DashScope) to reason over changelogs, breaking-change risk and CVE severity, produces a risk-batched remediation plan, sandbox-tests every fix against your own test suite, auto-applies safe security patches, pauses at a human-in-the-loop gate for anything risky (major-version migrations ALWAYS gate, by hard policy), re-scans to verify vulns are actually gone, and emits a hash-chained HMAC-signed audit receipt for every action.

Real run results:

32 vulnerabilities found across 6 packages
31 fixed in one run (~200 manual minutes saved)
1 residual: exactly the batch the human rejected (node-fetch v3, ESM breaking change) — honestly reported, never silently applied
3 Qwen API calls, 15,486 tokens (qwen-plus for planning, qwen-flash for approval briefs)

How we built it

Go-style state machine in Node.js: SCAN → PLAN (Qwen) → EXECUTE (per-batch sandbox) → VERIFY → DONE. Each phase transition emits a signed receipt. The model is load-bearing: without Qwen's reasoning over CVE prose and changelogs, the pipeline can't produce a plan. Also ships as an MCP server (5 tools) so any MCP-capable agent can drive the whole pipeline as native tools.

Challenges

Making the human gate genuinely useful, not cosmetic: the gate pauses the run, exposes Qwen-written approval briefs in plain English, and resumes on decision. A hard guardrail (majors always gate, independent of model confidence) binds the model.

Accomplishments

Tamper-evident audit chain verifiable offline. Crash-safe: checkpointed resumable state machine, retry-with-backoff on model API. Zero runtime dependencies.