Files
govoplan-audit/docs/AUDIT_TRACE_CONTEXT.md

106 lines
4.1 KiB
Markdown

# Audit Trace Context
Operational audit entries should let an administrator answer four questions:
- who initiated the operation
- what object was changed
- which lifecycle request or run carried the change
- how the entry connects to surrounding request, installer, or worker logs
The actor, tenant, scope, action, object type, and object ID remain first-class
`audit_log` columns. Additional operational context belongs in `details`.
## Standard Fields
Use these fields for admin, module lifecycle, installer, and worker operations:
- `module_id`: module affected by the operation, when a single module is the
target.
- `request_id`: queued installer/admin request identifier.
- `run_id`: installer/worker run identifier.
- `outcome`: compact state such as `planned`, `queued`, `applied`,
`cancelled`, `failed`, or `completed`.
- `_trace.correlation_id`: request or workflow correlation ID.
- `_trace.causation_id`: event, run, or operation that caused this audit entry.
Bulk operations may also include concise arrays such as `activated`,
`deactivated`, `mounted`, `planned_modules`, or sanitized `items`.
## Access Provenance
Access-sensitive events should include a compact provenance array when the
decision depends on semantic access state. Use `details.access_provenance` with
items shaped like `AccessDecisionProvenance.to_dict()` from
`govoplan_core.core.access`.
Expected provenance kinds:
- `identity`
- `account`
- `tenant_membership`
- `organization_unit`
- `function`
- `group`
- `role`
- `right`
- `delegation`
- `policy`
- `system_actor`
Delegation and acting-in-place events must show both the real actor account and
the represented/delegated account or function assignment. Feature modules
should obtain this shape through `access.explanation` instead of inspecting
access tables directly.
## Redaction
Audit detail payloads must not contain credentials, tokens, raw cookies,
authorization headers, private keys, message bodies, uploaded file content, or
other secret material. Use stable IDs, vault references, or package identifiers
instead.
The compatibility helper in core, `audit_operation_context`, keeps the standard
fields compact and applies the shared audit redaction pass to additional detail
values. Feature modules should follow the same shape even when they later write
through a dedicated audit sink capability.
## Module Lifecycle Events
Module install, uninstall, enable, disable, rollback, and package-catalog
acceptance events should include:
- `object_type`: `module_install_plan`, `module_install_request`,
`module_state`, or another stable lifecycle object type.
- `object_id`: module ID for single-module actions, otherwise `global` or the
request/run ID.
- `details.module_id`, `details.request_id`, or `details.run_id` when present.
- `details.outcome` with the lifecycle result.
- `details._trace` when the operation originated from a request or queued
daemon action.
This keeps admin UI timelines, audit exports, and rollback diagnostics aligned
without coupling modules to the audit table implementation.
## Commands, Events, And Outbox Delivery
Commands and events are separate concepts:
- Commands are imperative requests to do work, for example `retention.run` or
`module.install`. They use `govoplan_audit.backend.commands.AuditCommand`
and `CommandBus`.
- Events are completed facts, for example `tenant.created` or
`retention_policy.run`. They use the core `PlatformEvent` envelope and may
be written to the audit outbox before delivery.
`govoplan_audit.backend.outbox.SqlAuditOutbox` persists platform events in
`audit_outbox_events`. Dispatchers can later call
`dispatch_pending_platform_events()` to publish pending events and record retry
state. The outbox payload stores the full governed event envelope:
correlation/causation ids, actor, tenant, subject, resource, classification,
module id, event id, type, and payload.
Application code should enqueue or publish facts only after the state change
they describe is known. Long-running operators and installers should model
requested work as commands first, then emit facts as events as each step
completes.