Patch every dependency daily, ahead of advisories, with an agent reviewing the upstream diffs
Decision
Bump every pinned dependency to its latest release within its supported line every day, whether or not an advisory names it: flake inputs, Cargo crates, Swift packages, providers, and the binaries the nodes run. A scheduled job makes the lockfile bumps mechanically and builds nothing. A Claude Code agent, run on a schedule, then owns the outcome, keeping the app, infrastructure and running services free of known vulnerabilities. It makes the remaining bumps, reads the changelog and upstream diff of each for supply-chain signals, regressions and new holes, reads advisory sources for what we run, and picks the remedy: bump, hold, mitigate in our code or configuration, patch upstream, or fix our own code. A human merges the result.
The agent may release a bump to CI's secret-bearing build on its own judgment, except a Cargo bump that changes code run at compile time, which waits for a human.
Why
Exploitation now follows upstream fixes faster than advisories are published, so a policy that moves only on a CVE patches after the window has opened. Under that policy pins drift for months, because each bump is a decision nobody has a reason to make. Reading each upstream diff guards against a poisoned release directly, so we do not also need to wait out a soak period.
Rejected alternatives
- Frozen by default, CVE-driven bumps only. Patches arrive after disclosure, and pins fall months behind between advisories.
- Dependabot or Renovate with auto-merge. Nobody reads the diff before it reaches production.
- A fixed soak period before adopting any release. It delays every fix for a defence the diff review gives without the delay.
- An agent that triages advisories but does not bump. It leaves uncatalogued fixes unapplied, which are the ones this policy exists to get.
Consequences
The human cost is one review a day. The agent can misjudge a diff. Where it releases a bump on its own judgment, which includes every flake input, a poisoned upstream can reach CI's secret-bearing build before a human has reviewed it. We accept that for flake inputs, which come from a few well-known upstreams with reviewable diffs. A crate's build script or proc macro is arbitrary code from any publisher, so those always wait for a human. The agent must never be able to edit the workflows that enforce that wait.