Audit trail
DioscHub keeps records at the points a reviewer usually asks about: who changed the configuration, who signed in to the admin portal, who erased or exported data, and who approved a sensitive action. The admin and data-privacy trails are reviewable trails you can query; approval decisions are written as they happen. This page is precise about which is which, so you know what you can pull and what you cannot.
Administrative actions are recorded and queryable
Section titled “Administrative actions are recorded and queryable”Every change an administrator makes to a deployment’s configuration is written to an admin audit trail. That covers creating, updating, and deleting assistants, Roles, MCP instances, Toolsets, models, knowledge bases, and approval settings, along with configuration export, import, and settings changes.
Each entry records:
- Who — the administrator, or the admin API key, that made the change.
- What — the action and the entity it acted on.
- Before and after — the prior and new values for the change.
- Outcome — whether it succeeded, and the error if it did not.
This trail is queryable through the admin API — by recent activity, by the entity that changed, or by the administrator who made the change — so you can answer “who changed this, and when” for any part of the configuration.
Sign-ins and break-glass use
Section titled “Sign-ins and break-glass use”Admin sign-ins and every use of the break-glass admin token (X-Admin-Token) are written to the same
admin audit trail, under the entity Sign-in (admin-auth). The admin portal’s audit page shows them
alongside configuration changes, and its filters list these event types once they have occurred.
| Event | When it is written | What it records |
|---|---|---|
admin-auth.login-succeeded | An administrator signs in with email and password | The administrator, the client IP address and the user agent |
admin-auth.login-failed | A sign-in is refused | A reason (unknown-account, inactive or bad-password), a hashed account handle, the administrator’s id when the email matched an account, the client IP address and the user agent |
admin-auth.login-throttled | Sign-in attempts first exceed the brute-force limit in a window, once per window for an account and once for an IP address | The scope (account or ip), the hashed account handle for an account trip, the IP address, the limit and the window |
admin-auth.break-glass-used | A request is accepted with the break-glass token | The HTTP method and route (without the query string), the client IP address and the user agent |
admin-auth.break-glass-rejected | A request presents a break-glass token that does not match | The same fields as break-glass-used |
A failed sign-in does not record the email address that was typed, because a mistyped password often lands in the email field. It records a hashed account handle instead: the first 32 hex characters of the SHA-256 of the trimmed, lower-cased email. The rate limiter uses the same key, so repeated failures and a throttle trip for one account carry the same handle. To check whether a handle belongs to a known address, compute the same hash from that address. No sign-in or break-glass entry ever contains a password or the token value.
What the client sees is unchanged: a refused sign-in still gets the same 401 response, and a throttled one
the same 429.
A request made with the break-glass token gets a break-glass-used entry even when it only reads data.
A configuration change made with the token gets that entry plus the usual change entry, attributed to
“Break-glass admin token”.
Data-privacy actions are recorded and queryable
Section titled “Data-privacy actions are recorded and queryable”Data-privacy work is audited on its own trail, on every tier. It records:
- the subject-access searches an operator runs and the erasures and redactions they execute;
- each subject access export: the request, its completion or failure, every download, and its deletion;
- each change to the retention window, with the old and new values, and each change to waiting for BI;
- each nightly retention run, with its counts.
The trail is shown on the Audit trail tab of the Data Privacy page and is queryable through the admin
API, with the data-privacy:read permission. So an erasure or export you perform to satisfy a request
leaves its own reviewable record. Entries about a person name them by their one-way erasure tombstone,
never by the raw user id. See Data erasure & GDPR.
Neither audit trail carries a chat user’s credentials or message content — they record administrative, sign-in and erasure actions, keyed to the administrator who performed them.
Approval decisions are written as they happen
Section titled “Approval decisions are written as they happen”When a user reviews a gated tool and approves or rejects it, that decision is written to an append-only approval log at the moment it is made. Each record captures the tool, its arguments, the decision, and — for an update — the before-state of the change. The user’s forwarded auth is not recorded; it is written as empty, keeping the log credential-blind.
The admin dashboard surfaces how many approvals are currently awaiting a decision, so an operator can see when work is waiting. Approvals that still need a human are the actionable signal the dashboard raises.
Next: Subject access export.