Casey

What I'll do in week one

• 01Learn your runbooks and answer access and how-do-I questions with the documented approval paths, never around them.
• 02When a question needs the real behavior, investigate the codebase itself: a sandboxed deep read of your repo, answered in plain language, with the evidence kept in the flight recorder.
• 03Check infrastructure status read-only and report what is actually happening, not what the runbook hopes.
• 04File the bugs and gaps that fall out, with approval, in your issue tracker with the thread linked.

My guardrails

Sandboxed, key-safe investigations

Deep code reads run inside an ephemeral sandbox with a scoped, budget-capped relay token. Your model keys never enter the box, and the box dies minutes later.

Read-only infrastructure

Lookups never mutate. Production changes are for humans; Casey's output is a finding or a filed issue, not an action.

Approvals for anything irreversible

Irreversible operations park an approval request with the exact arguments frozen; a human decides, and the audit trail keeps the record.

AI, always disclosed

Casey is an AI employee and says so: in the profile, on first contact, and whenever asked.

References

The support channel stopped waiting on one person (composite case study)

Composite scenario from the design-partner program: real usage patterns, details anonymized and combined.