The Audit Log
What an audit event records, the single execution point where activity inside a running app is captured with no code, and how a workspace reads and exports the trail to a SIEM.
Security-relevant changes to a workspace and its apps, and every method a signed-in human invokes, are recorded as audit events: who did what, when, from where, and to what. Almost no platform can produce this record of activity inside a running app. Remy can, because method invocations are recorded at a point the platform owns, so a developer instruments nothing. Events flow through Kinesis Firehose into an Object-Locked S3 bucket where they cannot be altered or deleted for a year. They are browsable, filterable, and exportable in a format enterprise SIEMs ingest natively. This chapter covers the difference between an access log and an audit log, the point where capture happens, the event model, the delivery pipeline, and the read path.
#Access Log Versus Audit Log
A conventional access log records that a route was hit and a status came back. It cannot name the person, explain what the action meant, or say whether it should have been allowed. A Remy audit event carries all three, and two properties of the platform make that possible, neither of them the developer's work.
- Methods are contracts. A method has a human name and a description, so the record reads as the meaning of the action ("Approve Vendor Request"), not a route and a status.
- Identity is the platform's. Authentication resolves before a method runs, so every event names a real person. The event also records which credential was presented, so an agent's actions read apart from its person's.
#One Point of Capture
Method invocations are recorded at a single point on the platform side of the trust boundary, the same execution point that Request Lifecycles draws as the execute frame. Emission happens there on both the sandbox path and the dev tunnel's path, so coverage is structural: every app, every method, with no way to opt out and no code to write. Dev is included because a dev database is a copy of production.
Capture is fire-and-forget, so it never blocks or breaks the request the method goes on to serve. The app-internal event is identity-only: it records that a person invoked a named method, never the payload, and it fires only for authenticated humans, so the log stays signal rather than noise.
#One Event Model, Three Scopes
An audit event names an actor, an action, a target, and a context: time, source address, the credential presented, workspace, and where relevant the app. Actions are a closed catalog in code, so an emit cannot name an action that is not in it. CI also checks that every action, including every retired one, maps to a real OCSF class. An action removed from the catalog keeps its mapping, because its events are still in the store. Most systems audit the control plane or the application. Remy records both, in one schema, so a single trail runs from an administrator changing a role down to a person running a method.
| Scope | What it records | Examples |
|---|---|---|
| Workspace | Changes to the workspace itself | Members, roles, billing, settings, service settings and custom models, OAuth connections, keys |
| App | Changes to an app | Create, publish, secrets, custom domains, access, deletion |
| App-internal | Activity inside an app, in any environment | A method invocation, for every method invoked by an authenticated human |
The actor is the person the request authenticated as. A script's key and a dev box's key act as their person, and the event records which credential it was. Staff impersonating a customer are recorded behind the customer, and staff acting from the staff console are recorded as themselves. The source address is the one Cloudflare saw, stamped by the edge Workers, never a header the caller sets.
#The Pipeline
- Emit. The platform writes an event from wherever the action happens. High-frequency events can accumulate and roll up before delivery.
- Deliver. Events go to a Kinesis Firehose delivery stream, which batches them and writes to S3. Dynamic partitioning lays objects out by workspace and date, so a per-workspace read scans a narrow prefix. Batches are sized so a read pays for few, large objects rather than many tiny ones.
- Store. The destination bucket has S3 Object Lock enabled at creation, in COMPLIANCE mode, with a default retention of one year. This is write-once-read-many (WORM) storage: a locked object cannot be deleted, or have its retention shortened, by anyone, including the root account, until the lock expires. After the lock, events are kept: the bucket has no lifecycle rule, so the record is indefinite. It is one bucket in the platform's home region (us-west-2), whatever storage region a workspace has chosen for its data. The bucket blocks public access and accepts HTTPS only.
- Tail. In parallel, the most recent events per workspace are pushed to a short Redis buffer, so the dashboard's activity view is instant while S3 remains the record.
A CloudWatch alarm fires when records the stream holds go undelivered to S3. A write the API fails to hand to the stream never reaches it, so those failures are reported where the API reports errors, to its logs and its error reports, rather than by the alarm.
#Reading and Export
The workspace dashboard browses and filters the log by time window, actor, action or a set of actions, and app. Every app-scoped event is enriched at read time with the app's current name and icon. Each page continues from the last event the previous one returned, so pages neither repeat nor skip events as late deliveries arrive, and a repeated delivery is shown once.
Staff activity. Staff work in customers' workspaces, usually while impersonating someone for a support ticket. Those events are always written. A workspace without the full audit log entitlement does not see them in the viewer or the export, since to a self-serve customer they read as noise. A workspace with the entitlement sees everything. Enforcement is always shown: a suspension, or a hold on an app's sending, appears in the workspace's own log whoever applied it.
Export. A workspace can export its log as the canonical events or as OCSF, the Open Cybersecurity Schema Framework. The OCSF projection maps every action to a real OCSF class and activity, carries standard fields where OCSF has a home for them, and places platform-specific detail in the standard unmapped object. It never invents a class or a field. A validator in CI checks every action against the OCSF 1.8 JSON Schema, both fully populated and with every optional field missing, for every kind of actor. The result is a shape enterprise SIEMs ingest without translation.
``` Figure 19, "The Audit Log": one point of capture, and the pipeline that flows from it. Left: the emit sites across the platform (workspace and app changes, and method invocation on both the sandbox path and the dev tunnel's path). Method invocation converges on the one execution point the platform owns, the same point Figure 15 draws crossing into the sandbox. The method runs in the sandbox and its payload is never recorded. Fire-and-forget drop: from the emit sites an arrow goes to the audit event (actor, action, target, time, source, credential, outcome). Capture never blocks the request. Pipeline, from the event rightward: Kinesis Firehose (batched delivery, dynamically partitioned by workspace and date) into S3 with Object Lock (WORM, COMPLIANCE mode, one year, then kept indefinitely; public access blocked; HTTPS only; one bucket in the home region), which is the record. In parallel the event also drops to a live tail (recent events per workspace, a Redis buffer) for the dashboard's instant "recent" view. Below the pipeline: a CloudWatch alarm on undelivered records, raising an alert through SNS. Read path: from the record, the dashboard queries by time, actor, action(s), and app, enriches with the app's name and icon, and then browses, exports the canonical events, or exports OCSF 1.8 (validated in CI) to a SIEM. No trust boundary is drawn in this figure, so there is no crimson: capture happens on the trusted side, before the crossing that Figure 15 owns.