Agent options

Every option class the agent binds, its config section, and its defaults. Most are not exposed as Helm values — set those through extraEnv.

Durations are .NET TimeSpan strings ("00:10:00", not "10m").

Llm

The chat model, the embedding model, and what a single investigation is allowed to spend.

Key	Type	Default
Llm:Provider	string	gemini
Llm:Model	string	gemini-3.7-flash
Llm:PlanningModel	string?	null — falls back to Model
Llm:EmbeddingProvider	string	gemini
Llm:EmbeddingModel	string	gemini-embedding-001
Llm:EmbeddingDimensions	int	768
Llm:ApiKey	string?	null — from the secret
Llm:EmbeddingApiKey	string?	null
Llm:EmbeddingEndpoint	string?	null
Llm:Endpoint	string?	null
Llm:ApiVersion	string?	null
Llm:PlanningStructuredOutput	JsonSchema | JsonObject	JsonSchema
Llm:Temperature	double	0.2
Llm:MaxOutputTokens	int?	8192
Llm:Pricing	map of id → price	eight Gemini entries

An unpriced model is charged at zero

Llm:Pricing maps a model id to InputPerMillionUsd / OutputPerMillionUsd. A model with no entry costs nothing as far as the budget is concerned, which switches the cost budget off rather than approximating it. Ship a price with the model, always.

The seeded Gemini prices are promotional and double on 2027-01-01. Nothing in the code knows that date.

JsonSchema is not universally supported

DeepSeek answers 400 "This response_format type is unavailable now". Phase 1 is unaffected, so the agent diagnoses correctly and then proposes nothing — indistinguishable in a run summary from an agent that considered acting and declined. JsonObject moves the schema into the prompt. It is the weaker mode and stays off by default, and it is safe only because nothing downstream takes the model's word for it: every cited finding id is verified, and an action missing a namespace, kind or name is dropped before an executor sees it.

Llm:Investigation — one investigation's budget

Key	Type	Default
MaxSteps	int	12
MaxToolCalls	int	20
MaxWallClock	TimeSpan	00:10:00
MaxInputTokens	long	400000
MaxCostUsd	decimal	0.50
MaxConsecutiveNoToolTurns	int	2

A reserved concluding step lets a run that hits MaxSteps still write a finding. That rescue cannot land when the exhausted ceiling is tokens, because the concluding call resends the conversation — a known limitation, recorded in the backlog.

Llm:Budget — the rolling global budget, counted in Postgres

Key	Type	Default
MaxTokensPerHour	long	2000000
MaxCostUsdPerHour	decimal	3.00
MaxCostUsdPerDay	decimal	20.00
MaxCostUsdPerIncident	decimal	0.50
WarnAtUtilization	double	0.80
RunawayHourlyHitsBeforeLatch	int	3
RunawayWindow	TimeSpan	1.00:00:00

Llm:Retry

Key	Type	Default
Attempts	int	5
InitialDelay	TimeSpan	00:00:01
MaxDelay	TimeSpan	00:00:30
ExpBase	double	2.0
Jitter	double	1.0

Llm:Tools — what a tool call may return

Key	Type	Default
Timeout	TimeSpan	00:00:20
MaxResultBytes	int	8192
MaxRawBytes	int	1000000
MaxQueryRange	TimeSpan	7.00:00:00
RequireTimeRange	bool	true
RedactedArgumentNames	list	seeded

Redaction covers arguments, not results

RedactedArgumentNames redacts tool arguments in logs and traces. It does not redact tool results — raw describe_pod and get_pod_logs output carries env vars, hostnames and log contents verbatim into evidence blobs. That is why recorded cassettes are not committed to the repository.

Policy

The default-deny policy engine. A pure function over facts passed in by the caller, which is what makes it exhaustively unit-testable.

Key	Type	Default
AllowedNamespaces	set	[] — act nowhere
ProtectedNamespaces	set	seeded, never actionable
AutoEnabledActionTypes	set	[]
ProtectedLabels	map	seeded
RequiredNamespaceLabel	string	hephaisto.dev/destructive-actions-allowed
AllowSingleReplicaRestartLabel	string	hephaisto.dev/allow-single-replica-restart
MaxPodsPerAction	int	10
MaxWorkloadFraction	double	0.5
MaxActionsPerIncident	int	3
MaxActionsPerWorkloadPerHour	int	2
MaxActionsPerHour	int	10
MaxActionsPerDay	int	20
WorkloadCooldown	TimeSpan	00:15:00
MinPodAgeBeforeAction	TimeSpan	00:02:00
ClusterUnhealthyCeiling	double	0.3
RollbackFreshRevisionWindow	TimeSpan	00:30:00
RollbackPreviousHealthyMinimum	TimeSpan	01:00:00
MaintenanceWindows	list	[]

ClusterUnhealthyCeiling is the circuit breaker: above that fraction of unhealthy workloads the agent stops acting, on the reasoning that a cluster-wide problem is not one a pod restart fixes.

Kubernetes

Key	Type	Default
Enabled	bool	true
ClusterName	string	default
KubeconfigPath / KubeconfigContext	string?	null
ReadableNamespaces / DeniedNamespaces	set	{}
SignalQueueCapacity	int	2048
RelistInterval	TimeSpan	00:10:00
WatchTimeout	TimeSpan	00:05:00
ReconnectBaseDelay / ReconnectMaxDelay	TimeSpan	00:00:01 / 00:01:00
StormThreshold	int	50
StormWindow	TimeSpan	00:00:30
StormAggregateInterval	TimeSpan	00:01:00
RestartStormThreshold	int	3
RestartStormWindow	TimeSpan	00:10:00
ReadinessFlapThreshold	int	4
ReadinessFlapWindow	TimeSpan	00:10:00
LogTailLines	int	2000
LogLimitBytes	int	4194304
MaxRows	int	200
RbacMode	Enforce | WarnOnly	Enforce

Kubernetes:Enabled=false is a demo setting

It skips the RBAC self-check and the watchers, and leaves the executor that refuses everything in place. An agent that watches nothing while reporting itself healthy is the worst failure mode this project has, so it logs at warning level on every start. It exists so the console can boot on a laptop with no kubeconfig.

Persistence

Key	Type	Default
ConnectionString	string?	null
ConnectionStringName	string	hephaisto
AppConnectionStringName	string	hephaisto_app
ApplyMigrationsOnStartup	bool	false
CommandTimeout	TimeSpan	00:00:30
EvidenceBlobRetention	TimeSpan	30.00:00:00
LlmUsageRetention	TimeSpan	30.00:00:00
NotificationRetention	TimeSpan	30.00:00:00
RetentionSweepInterval	TimeSpan	01:00:00
RetentionBatchSize	int	1000
MaxAdmissionRetries	int	3
SearchPoolFactor	int	8
SearchMinPool	int	100

Two connection strings, deliberately. The serving role holds INSERT but not UPDATE or DELETE on audit_events — Postgres cannot restrain a table's owner, so audit immutability depends on the agent not being one.

Investigation

Key	Type	Default
MaxOuterTurns	int	8
MinConfidenceForPlan	double	0.5
PlanningCostUsd	decimal	0.10
PlanningTimeout	TimeSpan	00:01:30
PlanningMaxInputTokens	long	200000
EvidenceBlobRetention	TimeSpan	30.00:00:00
StallNudge, FinalConclusionNudge, OpeningMessage	string	see prompts

Investigation:Environment — the environment card

What the agent is told about your cluster. The shipped defaults are this repository's own dev cluster and should be overridden.

Key	Type	Default
ClusterName	string	studio-rancher-desktop
InScopeNamespaces	list	["hephaisto-chaos"]
ProtectedNamespaces	list	["hephaisto","hephaisto-obs","kube-system"]
DatasourceUids	map	{}
WorkloadOwners	map	{}
Notes	list	[]

Ingest

Key	Type	Default
ClusterName	string	studio-rancher-desktop
BurstWindow	TimeSpan	00:05:00
FlapThreshold	int	3
FlapWindow	TimeSpan	01:00:00
FlapCooldown	TimeSpan	04:00:00
CorrelationWindow	TimeSpan	00:10:00
SelfNamespaces	set	["hephaisto","hephaisto-obs"]

Notifications

Key	Type	Default
BaseUrl	string?	null — required once any route exists
Routes	list	[]
Webhook:Url, Webhook:SigningSecret	string?	null
Teams:WorkflowUrl	string?	null
GrafanaUrl	string?	null
MaxPerChannelPerHour	int	60
CorrelationCooldown	TimeSpan	00:15:00
MaxAttempts	int	8
FirstRetryDelay	TimeSpan	00:00:30
MaxRetryDelay	TimeSpan	00:30:00
DispatchInterval	TimeSpan	00:00:10
DispatchBatchSize	int	20
SendTimeout	TimeSpan	00:00:10

A route is Channel, Events, MinSeverity (default Info) and Namespaces. Routing is additive only — there is no deny rule. See Notifications.

Grafana

Key	Type	Default
McpUrl	string?	null — no PromQL/LogQL tools without it
ServiceAccountToken	string?	null
Url	string?	null
AnnotationToken	string?	null
AnnotationTimeout	TimeSpan	00:00:05
ToolCacheDuration	TimeSpan	00:00:30
ConnectTimeout	TimeSpan	00:00:10

Alertmanager

Write-only, and the only thing written is a silence.

Key	Type	Default
Url	string?	null
MaxDuration	TimeSpan	02:00:00
DefaultDuration	TimeSpan	00:30:00
Timeout	TimeSpan	00:00:10

KillSwitch

Key	Type	Default
SwitchDirectory	string?	null
ModeFileName	string	mode
EmergencyStopFileName	string	killSwitch
ModeEnvironmentVariable	string	HEPHAISTO_MODE

Demo

Key	Type	Default
Seed	bool	false
TranscriptPath	string	Demo/transcripts

Refuses outright if the database already holds an incident, so it cannot overwrite anything.