Backed by: your Worker’s control plane, GET /control-plane/privacy and the manifest’s privacy block. Scopes: privacy:read for the field list, and manifest:read for the posture. Plan: the inventory on every plan; its history on Enterprise.
What it answers
Which of our fields hold PII, PHI, PCI or secrets, and did that change?
The answer is usually a document somebody wrote once, and it drifted from the code the day it was saved. This pane reads it from the code that is deployed instead. Each field declares its own class in its schema, and your Worker walks the composed schemas and reports what it found. Privacy classes is how a field is marked and what the control plane reports, pithy privacy is the command that declares the classes, and what it never stores is what the history keeps.
The pane sits in the rail’s Compliance group, per environment. The inventory is read from your browser like every other pane, and dropped with the screen. Field names only: your Worker never reports a value.
The inventory
One row for the posture, then one row per table holding a marked field, named storage · container · name and listed by capability, then its history. A table’s row gives its capability, how many marked fields it holds and how many of them rest encrypted. Its record shows that table alone: one row per field, strongest class first, with the field’s path, its class, how it rests (encrypted, plaintext or unknown) and any note, such as a class assumed, a field that repeats another, or a key that cannot be retired. A path marked with more than one class is one row carrying each of them.
The index has two filters, over the inventory read. Findings keeps the posture and the tables holding a class a finding names, or everything else. Class offers each class the inventory holds, strongest first, and keeps the tables with a field of that class beside the posture and the history rows. A table holding more than one class draws a choice of class above its fields, starting at All, that narrows them to one. The overview is how the filter panel works.
Save as CSV or JSON exports the inventory as read, one row per marked field in capability, table and field order, with the environment, the build and the time it was read. It is the record of processing an auditor asks for, taken from the running code.
Without privacy:read the field list is withheld, and the pane says so, naming the scope. The classes and their counts still come from the manifest, which needs only manifest:read. Scopes is how to add it.
A project that declares no class has nothing walked, and the pane shows the command that declares one. A Worker on a kit too old to report the inventory is named, with its @pithy-sh/core version.
The posture
| Fact | What it says |
|---|---|
| Declared | The classes the project declared, as attested by the last pithy deploy, and when |
| Observed | Every class with a marked field in the deployed schemas. Derived at runtime, not attested |
| Console gate | As last attested: Enforced, Unverified, Not applicable or Unattested |
| Encryption at rest | How many fields of each class rest encrypted. A field your Worker cannot say about is Unknown, never counted as plaintext |
| Key retirement | Whether the field key can be retired, and what blocks it |
The mismatch is the finding. A class the code holds and the project never declared is a posture change nobody escalated. A class declared with no field marked is an unfinished migration. Neither is an error your Worker raises; both are listed here as findings, with how many.
Unattested is the loudest state. It means no pithy deploy attestation reached this Worker, usually because it was shipped with a raw wrangler deploy, so the declared classes and the gate are unknown. Unverified is not a failure and never reads as green: the project has no linter config Pithy could read, and the pane shows the one-sentence reason. An attestation dated before the build that is running is called out too, because it describes an earlier deploy.
Its history, on Enterprise
A single read is a snapshot. Enterprise keeps the series. Our schedule reads each watched environment, keys each build by its Cloudflare version, and records every change: a marker added or removed, a field that changed class, encryption gained or lost, a class newly held or no longer held, and a gate that changed. Each change carries the build it arrived with.
It is listed in the same index, after the tables: Posture report, Schedule, then one row per change, each named by its build. Privacy is one entry in the rail, history included. The series is read when one of these rows is opened, never for a table row, and the rows are listed whether or not your Worker answered the privacy read.
| Checked | Hourly by default, and at most once a day. Your contract sets the shortest interval |
| Watched | Production, without asking. Any other environment is off until an admin adds it with Index this environment |
| Mailed | Each change as it is indexed, or one digest a day. Removals are listed first |
| Kept | Thirteen months by default, a year of SOC 2 and PCI DSS audit plus one. A contract may keep it up to six years, for HIPAA |
It is sampled, not a record of every deploy. A change made and reverted between two checks is not seen, and the pane says so beside the schedule. A build already indexed costs one manifest read and nothing more. A build older than the one it replaced is marked as a rollback, and a return to an inventory indexed before is marked restored.
Each change opens to the full inventory as it stood at that build, while it is inside your retention window. Older changes page back to the edge of the window. The read log lists every check: a new inventory, no change, a build already indexed, privacy:read withheld, a Worker that did not answer, or a credential that cannot be used.
A Worker that stops answering goes dark, and you are told. After three failed reads in a row the environment is marked dark and mailed once, and once more when it reads again. Whatever changed meanwhile appears as one change. From the third failure the checks back off, up to a day between attempts.
The posture report
A periodic mail of what nobody logs in to look at. Off until an admin turns it on, and never more often than once a month. It has five sections, each of which an admin can switch off:
| Section | What it holds |
|---|---|
| Privacy posture | Classes and field counts per environment, and every change in the period with its build |
| Who has access | Your administrators, and every connection on each Worker with its purpose, principal and scopes |
| Key hygiene | Keys near expiry, overdue rotation, and secrets past their cadence |
| Deploy activity | How many builds each watched environment ran |
| Coverage | Which environments are watched, any that went dark, and how long since each was deployed through pithy deploy |
Every report says which sections are off. That line cannot be turned off, and every change to the report’s settings is audited.
Who hears it
Each mail is an event in the notification catalog: Privacy changes, Privacy inventory unreadable and Privacy posture report, all three on by default. They go to the account and to the extra addresses confirmed for alerts. Notifications is where each one is switched.
A mail carries field names, class names, counts and the build. Never a value: your Worker never reports one.
Who may change it
Owners and admins set the interval, the mail cadence, the environments watched and the posture report. Members see the same history and read log, read-only.
What the history keeps
Field paths and class names, never a value from your data. Each change is a row in our D1, and each indexed inventory is archived in our R2, sealed under a key of our own. It is the one documented exception to what it never stores: a map of your data model, kept for your retention window.
A lapse freezes it, then deletes it. When the plan no longer holds the series, nothing more is read, and the history stays read-only for thirty days, saying when it will be deleted. Then it is deleted.
What the schedule’s reads look like on your side
They are ordinary control-plane calls: signed, single-scope, single-use, sixty seconds, under the subject system, and each lands in your own audit trail. Each check is one request to your Worker, billed to your Cloudflare account, and the read log lists every one.
On Free, Solo and Team
The inventory, the posture, the findings and the export are yours on every plan, and nothing about them is stored. The history rows are still listed, each marked Enterprise, and each opens locked, saying what Enterprise watches. Nothing is read on a schedule, and the series is never read, until the plan holds it.