edge

AI-assisted delivery, Staff-level technical leadership, architecture and platforms for demanding products.

Rebuilding secrets, CI/CD, and infrastructure as code without breaking what already runs

This internal project rebuilt the secrets, CI/CD, and infrastructure-as-code foundations of LRJI's repositories, with evidence before every risky change.

The constraint was simple: make the foundations safer and more repeatable without interrupting the products that depend on them.

Rebuild without turning security into a wager

Secrets, deployments, and infrastructure affect every product. The rebuild therefore kept an explicit destructive perimeter and verifiable feedback at every phase.

The problem to solve

Change foundations without breaking dependencies

Every destructive phase started with rereading the destructive perimeter and its destroy plan.

Explicit boundaries for secrets and delivery

The published choices remain principles: they explain the controls without exposing the platform's operational map.

  • Declared secrets, stored outside files

    OpenTofu creates secrets and stores them in a password vault: no secret lives in a file, CI variable, or laptop.

  • Keyless workload identity

    CI reaches the cloud through workload identity federation, without a static key to distribute or renew.

  • Separated environments

    Production and non-production live in separate cloud projects; images are promoted by digest.

A migration governed by evidence

From September 7 to September 25, 2026, the plan advanced one step at a time. Its four remaining steps are open by design, rather than presented as finished work.

Final exercises

The final exercises exposed a file-processing worker that did not read its event's binary-content mode and a local gate that accepted what the CI image rejected. Both defects were fixed.

The published markers

They measure the rebuild's discipline, not a generic security promise.

Plan
298 planned steps.
Evidence
294 steps proven one by one.
Perimeter
Destroy plan reread before every destructive phase.
Exercises
Two real defects detected and fixed.

What this study establishes

Platform security relies as much on a reread perimeter and real exercises as on the tools selected.

  • OpenTofu
  • Password vault
  • Workload identity
  • Digest promotion

Publish principles and evidence, never operational identifiers.

This study publishes no vault, item, project, account, or state name, and no network topology.