How configuration works

Hephaisto uses standard .NET configuration. Every key has three interchangeable spellings, and knowing the mapping between them is most of what you need.

Form	Example
Config key	Llm:Model
Environment variable	Llm__Model
Helm value	llm.model, where the chart exposes one

A colon becomes a double underscore. Array elements are indexed: Policy__AllowedNamespaces__0, Policy__AllowedNamespaces__1. Nested objects keep nesting: Notifications__Routes__0__Channel.

Where a value can come from

In increasing order of precedence:

1. appsettings.json baked into the image
2. appsettings.{Environment}.json
3. Environment variables — which is how the chart sets everything
4. The kill-switch arms, for mode only, where the most restrictive arm wins rather than the last one read

The escape hatch

The chart deliberately does not expose every option as a Helm value — exposing all ~110 would mean a values file nobody reads and a schema that lags the code. Anything not exposed is set through extraEnv:

extraEnv:
  - name: Llm__Investigation__MaxSteps
    value: "16"
  - name: Policy__WorkloadCooldown
    value: "00:30:00"

extraEnv cannot shadow the four reserved variables, cannot half-override the OTEL_* block, and cannot widen the namespace allowlist behind RBAC's back. Those are enforced as chart render failures — see Reserved env and safety rails.

Types that are not what they look like

• Durations are .NET TimeSpan strings, not Prometheus durations. "02:00:00" is two hours; "2h" is a parse failure. "7.00:00:00" is seven days.
• Enums are their C# member names, case-insensitively: Observe, RestartPod, Enforce.
• Llm:Pricing is a dictionary keyed by model id. An unpriced model is charged at zero, which switches the cost budget off rather than approximating it.

Restarts

Most options are read through IOptionsMonitor and take effect on the next use. Two groups do not, because they are captured at construction:

• Llm:Provider, Llm:Model, Llm:Endpoint — the chat client factory captures these. Changing them needs a pod restart, not just a ConfigMap edit.
• Anything the Kubernetes watcher stack reads at startup.

Where to look next

• Helm values — everything the chart exposes, with its own comments
• Agent options — every option class, including ones with no Helm value
• Reserved env and safety rails — what the chart refuses to render