Audit events

Every audit event is one JSON object: the same in the history, in an export’s JSON Lines, on a server’s output and to its webhook.

The format

An AI assistant's restart, approved

{
  "type": "lumovi.audit",
  "version": 1,
  "id": "0b6c1f9e-4d2a-4c55-9a1e-7f3d2b8c6e41",
  "chain": "9a829611-9c63-48cf-94fb-90576b07837d",
  "seq": 6,
  "time": "2026-10-06T12:10:40.512Z",
  "category": "change",
  "action": "resource.restart",
  "outcome": "success",
  "actor": {
    "user": "jane@example.com",
    "groups": ["platform", "on-call"],
    "via": "assistant",
    "assistant": "Claude Code",
    "session": "5f0c2a9e41b37d86",
    "address": "203.0.113.24"
  },
  "cluster": "production",
  "target": {
    "kind": "Deployment",
    "name": "checkout",
    "namespace": "shop",
    "uid": "9a7e4d61-2c1b-4f0e-8d3a-6b5c4e3f2a10"
  },
  "summary": "Restarted Deployment checkout",
  "command": "kubectl rollout restart deployment/checkout -n shop --context production",
  "approval": { "status": "approved", "by": "jane@example.com", "waitedMs": 18342 },
  "details": {
    "fields": ["spec.template.metadata.annotations[\"kubectl.kubernetes.io/restartedAt\"]"],
    "patchType": "strategic",
    "reason": "Its pods crash-loop since the database credentials rotated."
  },
  "prev": "4f1d0c7e9b2a6d3851e0c4b7a9f2e6d1c8b5a3f7e2d9c6b1a4f8e3d7c2b6a5f9",
  "hash": "676e4035e9c43c3864242ee97a39a96521108743afd3533462f72952ea5004f4"
}

Lumovi writes each on one line. A field with nothing to say is left out, rather than empty or null.

Field	What it holds
`type`	Always `lumovi.audit`. On a server’s output, lines without it are Lumovi’s own log lines.
`version`	`1`. A later version may add fields.
`id`	A UUID, the event’s own.
`chain`	Which chain it’s in, by a UUID: one for each history. A server that keeps its history in memory starts one each time it starts, and each replica without the volume has its own.
`seq`	Its place in its chain: one more than the event before it. See How events are chained.
`time`	When, to the millisecond, in UTC: `2026-10-06T12:10:40.512Z`.
`category`	Its kind: `change`, `access`, `sign-in`, `assistant`, `settings` or `server`.
`action`	What happened: see Actions.
`outcome`	How it came out: see Outcomes.
`actor.user`	Who: as they signed in, on a server; the computer’s account, in the desktop app; `lumovi` for Lumovi itself.
`actor.groups`	Their groups, as they signed in, when they have any.
`actor.via`	`ui` (Lumovi’s page), `assistant` (an AI assistant acting as them) or `server` (Lumovi itself).
`actor.assistant`	The assistant, by the name it gave: `Claude Code`.
`actor.session`	Which of their sessions, or which assistant they allowed, by a hash: 16 hex characters. On a server.
`actor.address`	The address the request came from, as the server saw it: often the ingress controller’s. On a server.
`actor.forwardedFor`	What the request’s `X-Forwarded-For` header said, as it said it: who a proxy says it’s forwarding for. Nothing checks it.
`actor.userAgent`	The browser, or the assistant, as it named itself.
`actor.kubeUser`	In the desktop app, for an event in a cluster: the name of the kubeconfig’s user entry it used there, not necessarily who the cluster knows it as.
`cluster`	The cluster, by its context’s name, or its name in a fleet.
`target`	What it was done to: its `kind`, and its `name`, `namespace` and `uid` when it has them. A Helm release’s kind is `HelmRelease`.
`summary`	What happened, in a sentence: “Restarted Deployment checkout”. When it didn’t happen, what was asked: “Scale Deployment recommendations to 12 replicas”.
`command`	The `kubectl` or `helm` command that does the same, when there’s one that needs nothing more.
`approval`	For an AI assistant’s change: `status`, `by`, `waitedMs` (from when the person was asked to their answer), and `note`. See Approvals.
`details`	What else the action says: see Details. Text, numbers, `true` or `false`, or lists of text.
`error`	Why it failed or was refused, as the cluster or Lumovi said it. About a Secret, what the cluster said isn’t kept: see Errors about Secrets.
`prev`	The hash of the event before it, or empty for the first of a chain.
`hash`	Its own: the SHA-256 of the event without `hash`, as `jq -jcS 'del(.hash)'` writes it.

Whatever’s sent, an event stays one line of at most 64 KiB:

Text is cut at whole characters, and ends in …: names (a user, a cluster, an object) at 256 characters, sentences (summary, command, error) at 4,000, a detail or a note at 1,000. Half of a character pair, which no two JSON tools read alike, becomes � (U+FFFD).
Details: at most 50, a list at most 100 items, and 16 KiB in all. Past that, the rest are left out, and truncated: true says so. forwardedFor keeps 200 characters, and userAgent 300.
An event over 64 KiB even so has its sentences cut to 1,000 characters, at most 10 groups, and only truncated: true in its details.
A number in its details that isn’t a whole one within JSON’s safe range, like 1e21 or 0.1, is kept as its text: "1e+21", "0.1". JSON tools write those each their own way, and this way every hash is the same in all of them.

Errors about Secrets

What the cluster, or an admission webhook, says about a Secret can quote its values. So when a change to a Secret fails, from the page or an AI assistant, its error says what happened, by its cause, and not what the cluster said:

What happened	`error`
It isn’t there (`not-found`)	“It isn’t there.”
The cluster’s RBAC refused it (`forbidden`)	“The cluster doesn’t let them.”
The cluster didn’t take their credentials (`unauthorized`)	“The cluster doesn’t know who they are.”
It changed since it was read (`conflict`)	“It changed since it was read.”
The cluster, or a webhook in it, found it not valid (`invalid`)	“The cluster says it isn’t valid.”
Anything else	”It failed (timeout).”, with what went wrong: `unreachable`, `timeout`, `server`, and so on

Each followed by “What the cluster said of the Secret isn’t kept: it can quote its values.” Lumovi’s own refusals, by read-only mode or access, are kept as they’re said. An AI assistant’s tool call about a Secret that fails says “It failed. What the cluster said of the Secret isn’t kept: it can quote its values.”: one whose kind has “secret” in it, in any case, like Secret or SealedSecret, or with a manifest of a Secret. A tool call that’s refused is kept as Lumovi said it.

Actions

Each action, by kind, and what the page calls it.

Changes (`change`)

Action	On the page	When
`resource.create`	Created an object	An object was created: from YAML, a Job run from a CronJob, or by a view’s action.
`resource.apply`	Applied an object	An AI assistant’s manifest was applied, as `kubectl apply --server-side` does.
`resource.replace`	Replaced an object	An object edited as YAML was saved: “Replaced Deployment cart with an edited one”.
`resource.patch`	Changed an object	Part of an object was changed: its labels, an image, its status, a node cordoned or uncordoned.
`resource.scale`	Scaled a workload	Its replicas were set, and nothing else.
`resource.restart`	Restarted a workload	Its pods were restarted, as `kubectl rollout restart` does.
`resource.delete`	Deleted an object
`resource.evict`	Evicted a pod	A pod was evicted: draining a node evicts each.
`resource.debug`	Added a debug container
`helm.install`	Installed a Helm release
`helm.upgrade`	Upgraded a Helm release	To another chart version, or with new values.
`helm.rollback`	Rolled a Helm release back
`helm.uninstall`	Uninstalled a Helm release

Access (`access`)

Action	On the page	When
`shell.open`, `shell.close`	Opened a shell, Closed a shell	A shell in a container. Closing says how long it ran, and the exit code.
`node-shell.open`, `node-shell.close`	Opened a node shell, Closed a node shell	A shell on a node, and the pod Lumovi started for it.
`port-forward.open`, `port-forward.close`	Forwarded a port, Stopped forwarding a port	A port forward, in the desktop app.
`logs.read`	Read logs	A container’s logs: once for each container every ten minutes, at most, for each page or assistant.
`secret.read`	Read a Secret	A Secret’s values reached the page or an assistant, or reading it was refused: once for each Secret every ten minutes, at most, for each page or assistant. A dry run of a change to a Secret counts, since it answers with the Secret.
`helm.values.read`	Read a release’s values	A Helm release’s values and manifests reached the page, which can hold what its Secrets do: once for each release every ten minutes, at most. Not where they’re withheld.

On a server, unless people sign in at an authenticating proxy. Sign-ins that don’t succeed are recorded too, each, up to 60 a minute. Past that, they’re counted, and one event a minute, by Lumovi itself, says how many and from where: “25 more sign-ins didn’t succeed, each not recorded: more than 60 were tried in a minute”. So does one as the server stops, for those it hasn’t said yet. They’re nobody’s own: only auditors see them.

Action	On the page	When
`session.sign-in`	Signed in	Signing in, with a token or single sign-on: “Signed in with a token”.
`session.sign-out`	Signed out	Someone signed out.
`session.expired`	Session ended	A session ended by itself: “Signed out by Lumovi: the session ended”, with why in `details.why`.

AI assistants (`assistant`)

Action	On the page	When
`assistant.allowed`	Allowed an AI assistant	On a server: someone allowed an assistant to act as them.
`assistant.denied`	Didn’t allow an AI assistant	On a server: someone chose not to.
`assistant.ended`	An AI assistant was let go	On a server: it signed out, its person let it go, their session ended, or a refresh token came back used, which is Refused.
`assistant.tool`	An AI assistant’s tool call	Each tool an assistant called, with what it asked for.

Settings (`settings`)

Action	On the page	When
`permissions.changed`	Changed AI permissions	Someone changed what their assistants may do: their defaults, and their rules’ names.
`read-only.changed`	Changed read-only	A cluster was made read-only, or changeable again.
`node-shell.changed`	Changed node shells	Where a cluster’s node shells run, and from what image, changed: “Made node shells in production run alpine:3.22 in kube-system”, or “Made node shells in production run as Lumovi’s defaults”.
`access.changed`	Changed access	On a server: an admin saved changes to who may do what. Its summary is the first change, and how many more: `Added the grant “On-call”, and 2 more`. Changed by hand where it’s kept, it’s recorded by Lumovi itself, its summary starting `Changed outside Lumovi:`, or, for a change made while Lumovi wasn’t running, recorded as it starts: `Changed outside Lumovi, while it wasn’t running: what it was before can’t be known`.
`assistants.changed`	Changed AI assistants	In the desktop app: assistants let connect, or stopped; their port; a new token.

Server (`server`)

Action	On the page	When
`server.started`	Lumovi started	A server started: its version, how people sign in, and where its events go.
`server.stopped`	Lumovi stopped	A server stopped.
`audit.dropped`	Audit events lost	Events couldn’t be kept in the history, or sent: how many, where, and why. See When events are lost.

With the changes level, a server records everything but the Access kind and assistant.tool, but for an assistant’s call to change something that was refused: that’s a change asked for, and it’s kept, with changing: true in its details.

Outcomes

Outcome	On the page	When
`success`	Done	It was done.
`failure`	Failed	The cluster, or Lumovi, couldn’t do it.
`refused`	Refused	It wasn’t allowed: by the cluster’s RBAC, read-only mode, someone’s access, an assistant’s permissions, or the person who rejected it.
`cancelled`	Cancelled	Nobody answered an assistant’s change in time, or it was withdrawn: by the assistant, as it was let go, or as assistants were turned off. Its tool call too. A node shell closed before its pod started.

Approvals

approval.status says what became of an AI assistant’s change:

Status	Outcome	What else
`approved`	`success`, or `failure`	`by`, `waitedMs`
`unasked`	`success`, or `failure`	Made without asking, where the person’s permissions allowed it
`rejected`	`refused`	`by`, `waitedMs`, and `note`, when they wrote one
`expired`	`cancelled`	`waitedMs`: nobody answered in time
`withdrawn`	`cancelled`	`waitedMs`, and why, in `error`: “The assistant stopped waiting for an answer.”, “The assistant was let go before it was answered.”, or “AI assistants were turned off before it was answered.”

waitedMs runs from when the person was asked to their answer. by is the person: only they answer their assistants’ changes. The change’s tool call (assistant.tool) comes out the same: refused when rejected, cancelled when nobody answered or it was withdrawn. A call that ends while the change still waits for its answer is success, with waiting: true: nothing has happened yet, and the change’s own event, or the wait_for_change call that gets the answer, says what did. A change an assistant’s permissions don’t allow, or the cluster’s dry run turns down, has no approval: nobody was asked.

Details

What each action adds in details, when it applies:

Actions	Details
`resource.*`	`fields` (the fields a change set, by path: never their values), `patchType`, `subresource`, `replicas`; `fieldManager` and `force` for an apply; `propagation` and `gracePeriodSeconds` for a delete; `container`, `image` and `target` for a debug container
An assistant’s change	`reason` (the reason it gave), and `takesOver` when an apply takes fields from another manager
`helm.install`, `helm.upgrade`	`chart`, `version`, `repository`, and `values`: the top-level names of the values set, never what they’re set to
`helm.rollback`	`revision`
`helm.uninstall`	`keepHistory`
`shell.*`	`container`; on closing, `seconds`, and `exit`, how it ended
`node-shell.*`	`mode`, and `pod`, the pod Lumovi started for it; on closing, `seconds`, `exit`, and `left` when its pod couldn’t be deleted
`port-forward.*`	`pod`, `podPort`, `localPort`; on closing, `seconds`
`logs.read`	`container`, `previous` and `follow`; for an assistant, `tailLines` and `previous`
`secret.read`	`keys`: the Secret’s keys, never their values (none when it was refused)
`session.sign-in`, `session.sign-out`	`method` (`token` or `oidc`), and `with`: `a token`, or the provider’s name
`session.expired`	`why`
`session.sign-in`, by Lumovi itself	`count` (sign-ins that didn’t succeed, past the 60 a minute recorded each) and `from`: where they came from, with how many each, the most first, 20 at most, like `203.0.113.9: 240`
`assistant.allowed`, `assistant.denied`	`assistant`, `client` (its ID with Lumovi), and `returnsTo`, the site it sends people back to
`assistant.ended`	`how` (`signed out`, `let go`, `expired` or `reused`) and `since`, when it was allowed
`assistant.tool`	`tool`, `changing: true` for a tool that changes things, `waiting: true` for a change still waiting for its answer (its own event says what became of it), and what the assistant asked for, by name: never a manifest it applies
`permissions.changed`	`changes`, `secrets`, `env` and `logs` (the defaults), and `rules`: their names
`access.changed`	`changes`: each change, a sentence each, like `Changed the profile “Developer”: Helm: Off → Upgrade, roll back`; and `outside: true` for a change made by hand where it’s kept, not through Lumovi
`read-only.changed`	`readOnly`
`node-shell.changed`	`namespace` and `image`, unless set back to Lumovi’s defaults
`assistants.changed`	`enabled`, `port`
`server.started`	`version`, `auth`, `level`, `kept` (`files` or `memory`), `retentionDays`, `sinks` (where events go) and `auditors`
`audit.dropped`	`sink` and `count`, with why in `error`

Check it hasn't changed

How the chain works, and how to check it.

Audit log for your team

The server’s output, a webhook, and who reads everyone’s.

Get started

Clusters

Explore

Make changes

Debug

Metrics

Helm

Custom resources

AI assistants

Audit log

The format

Errors about Secrets

Actions

Changes (`change`)

Access (`access`)

AI assistants (`assistant`)

Settings (`settings`)

Server (`server`)

Outcomes

Approvals

Details

Check it hasn't changed

Audit log for your team

​The format

​Errors about Secrets

​Actions

​Changes (change)

​Access (access)

​Sign-ins (sign-in)

​AI assistants (assistant)

​Settings (settings)

​Server (server)

​Outcomes

​Approvals

​Details

Check it hasn't changed

Audit log for your team

The format

Errors about Secrets

Actions

Changes (`change`)

Access (`access`)

Sign-ins (`sign-in`)

AI assistants (`assistant`)

Settings (`settings`)

Server (`server`)

Outcomes

Approvals

Details