Static ReadonlyAuthorizesTrue when the action authorizes its caller itself, so a public entry point such as the
RunAction mutation may run it for an authenticated user without an MJ: Action Authorizations link. Set it only on an action that either checks an authorization the caller
must hold, or reads and writes only with the caller's own permissions. Leave it false for an
action that acts with more authority than its caller, or that reaches data or sends messages
the caller chooses but could not otherwise reach.
The engine reads the marker only from the class that runs, never from its parent classes. A subclass that overrides a marked action must declare the marker again.
Whether this action can bound what it returns to an audience (RunActionParams.Audience): everyone
besides the caller who will see its output, such as the other participants of a shared conversation.
false by default, and the engine enforces it: when a run's audience adds a reader beyond the caller,
ActionEngineServer.RunAction refuses the action with result code AUDIENCE_UNSUPPORTED, without running
it. Override to return true only when the action honours the audience in full — every record, snippet
and count it returns is something every reader may see (pass the audience to the search engine, check each
reader's entitlement to anything it scopes by, and leave out aggregates computed before that filtering).
ProtectedaddAppend (or overwrite) an output param on the run params.
ProtectedbuildBuild the engine's production dependency bundle from the action's run params
(threading ContextUser + Provider for isolation/multi-provider
correctness) through TrainingDelegation.BuildTrainingDeps — the same wiring the Train
Model operation uses, so both surfaces pick the artifact-store family the same
way. Overridable so tests inject in-memory seams.
ProtectedcreateConstruct the TrainingEngine. Overridable so tests inject a mock.
ProtectedfailBuild a failure ActionResultSimple.
ProtectedfindLocate a param by case-insensitive, whitespace-tolerant name.
ProtectedgetRead a boolean param. Accepts native booleans and the usual truthy/falsey
string forms (true/1/yes, false/0/no); falls back to defaultValue.
ProtectedgetRead a JSON-object param. Accepts either an object value or a JSON string that
parses to an object. Returns undefined when missing or not an object.
ProtectedgetRead a numeric param. Returns undefined when missing or non-numeric so the
caller can apply its own default / required check.
ProtectedgetRead a string-array param: an array value, a JSON array of strings, or a comma-separated list. Blank entries are dropped; an absent param is an empty list.
ProtectedgetRead a string param (trimmed). Returns undefined when missing/empty so the
caller can branch on required-ness.
ProtectedInternalInternal method that must be implemented by derived action classes. This is where the actual action logic should be implemented.
The action execution parameters including typed context
Promise resolving to the action result
ProtectedmapMap the Train output (the same one the operation returns) onto the action's output params. A run whose gate
decision ships no model succeeds with ModelID null; its message is the stored reason.
ProtectedokBuild a success ActionResultSimple, carrying the (mutated) params back.
ProtectedprotocolWhere the run's protocol version is read: the deps' protocol store (the one training read the protocol through), or a store on the deps' provider and user, as the engine defaults it. Overridable so tests serve protocol rows.
Executes the action with the provided parameters.
The action execution parameters including context
Promise resolving to the action result
Trains a model from a
MJ: ML Training Pipelinesdefinition and surfaces the run's stored gate decision.Outputs:
ModelID(null when no model is produced),HoldoutMetrics(JSON string),LeakageFlagged(boolean),GateOutcome,GateReason,TrainingSnapshotIDandEvaluationProtocolID.