Confidentiality Action Gating

Policy-based control of sensitive actions, evaluated against each project's confidentiality severity.

Introduced in v0.12.0 (ADR-0079), confidentiality action gating decides whether a sensitive action may proceed, must warn the user first, or is blocked outright. It complements the existing confidentiality-aware AI routing (which blocks external model providers for classified projects) by extending the same principle to user- and agent-initiated actions.

How it works

Every gated action is evaluated by a server-side policy engine before it runs. The engine takes the action, the project confidentiality severity, and the acting identity, and returns one of three outcomes.

1. The client requests a gated action.
2. The backend evaluates the request via POST /api/confidentiality-gate.
3. The engine returns an outcome: allow, warn_and_override, or block.
4. The client proceeds, prompts for an override, or is denied accordingly.

Policy outcomes

Outcome	Meaning	Client behavior
allow	The action is permitted for this severity	Proceed silently
warn_and_override	The action is sensitive but may proceed with an explicit acknowledgment	Show the override dialog; record the override if confirmed
block	The action is not permitted for this severity	Deny the action

Severity levels

Confidentiality severity is a per-project attribute. Gating decisions are keyed to the severity band. Medium-severity projects (C.2 and C.3) are the primary target of extraction gating: document text extraction via Textract is gated for these projects rather than blocked outright, so a user with authority can override with an audit record.

Override and audit trail

When an action returns warn_and_override, the user is presented with a confidentiality override dialog. Confirming the override writes a permanent record to the audit trail, including the actor, the action, the project, and the timestamp. Overrides are not silent and cannot be retracted from the log; they are surfaced in the Admin Audit Log alongside other security-relevant events.

See Audit Log and SIEM for how override events are stored, filtered, and exported.

Evaluation endpoint

The endpoint is the single source of truth for gating decisions. Clients must not infer outcomes locally; the server evaluates every request so that policy changes take effect immediately and every override is recorded centrally.