SDK Reference

This reference documents the public Iron Gorilla SDK surface used to define agents, model steps, declare triggers, call LLMs, configure DLP, and format chat context.

Kernel Methods

• kernel.invokeTool — Invoke a registered tool through the governed kernel.
• kernel.callLLM — Call an LLM provider through the governed kernel.
• kernel.readData — Read data through a mediated data source.
• kernel.writeData — Write data through a mediated data target.
• kernel.now — Request a deterministic kernel-controlled timestamp.
• kernel.random — Request a deterministic kernel-controlled pseudo-random value.
• kernel.uuid — Request a deterministic kernel-controlled identifier.

Classes

• ActionDeniedError — Typed error thrown when policy or approval state denies an action.
• AgentCancelledError — Typed error for cancelled agent execution.
• AgentSdkError — Base class for stable SDK and runtime errors.
• AgentTimeoutError — Typed error for agent execution timeouts.
• AgentValidationError — Typed error for SDK validation failures.
• ApprovalRequiredError — Error thrown when a mediated action pauses for durable human approval.
• InternalRuntimeError — Typed error for unexpected runtime failures.
• RateLimitError — Typed error for runtime or kernel rate limiting.
• ReplayDivergenceError — Typed error for deterministic replay mismatches.
• RuntimeSessionError — Typed error for invalid or expired runtime sessions.
• SubprocessProtocolError — Typed error for malformed agent subprocess completion protocol messages.

Interfaces

• AgentChatTranscript — Runtime-provided chat transcript passed to chat-triggered agent runs.
• AgentChatTranscriptMessage — One normalized message in a chat-triggered agent transcript.
• AgentConfig — Configuration object for an Iron Gorilla agent.
• AgentDataConfig — Data access rules attached to an agent definition.
• AgentDefinition — The authored agent definition returned by defineAgent().
• AgentIdentity — Stable identity and authority metadata for an agent.
• AgentLoopConfig — SDK declaration for an autonomous loop's model, tool surface, and caps.
• AgentLoopInput — Input accepted by KernelProxy.agentLoop.
• AgentLoopPriorMessage — One bounded prior user or assistant message supplied to an autonomous loop.
• AgentLoopResult — Result returned by KernelProxy.agentLoop.
• AgentLoopToolBinding — Model-visible tool binding declared for an autonomous loop.
• AgentLoopToolCall — One executed tool call recorded in the loop transcript summary.
• AgentLoopTurnSummary — Summary of one completed model turn.
• AgentMemoryConfig — Scoped-memory behavior declared for an authored agent.
• AgentSdkErrorOptions — Constructor options for a typed AgentSdkError.
• AgentStepInput — Canonical runtime input object delivered to authored step functions.
• AgentStepPreviousOutput — Summary of the immediately preceding completed step made available to the current step.
• AgentStepWorkflowContext — Lightweight workflow context attached to each step input.
• AgentTrustConfig — Starting trust posture for an authored agent.
• ChatOutputDefinition — Declarative contract selecting the chat-facing value from final step output.
• ChatTriggerDefinition — Chat trigger definition for governed conversational agent sessions.
• CronTriggerDefinition — Recurring trigger definition for scheduled agent runs.
• DlpActionPolicyOverride — Optional DLP policy override for one action family.
• DlpCustomRecognizer — Custom regex recognizer authored for organization-specific DLP detection.
• DlpScanPolicy — DLP policy knobs that convert classification into enforcement behavior.
• FormatChatTranscriptOptions — Options for formatting prior chat transcript context.
• KernelInvokeToolOptions — Optional delegated-connection selection for kernel.invokeTool().
• KernelProxy — Kernel proxy surface exposed to agent step code.
• LlmCallParams — Parameters accepted by kernel.callLLM().
• LlmResponse — Response shape returned by kernel.callLLM().
• ManifestSpeechToSpeechConfig — Speech configuration and owning trigger covered by a deployment signature.
• ManualInputDefinition — Code-tracked manual-input contract shared by code authoring and the GUI.
• ManualInputObjectFieldDefinition — One nested object group declared in a manual-input contract.
• ManualInputPrimitiveFieldDefinition — One primitive field declared in a manual-input contract.
• ManualInputScreenDefinition — One linear wizard screen in the manual-input contract.
• ManualInputSelectFieldDefinition — One select field declared in a manual-input contract.
• ManualInputSelectOptionDefinition — One selectable option rendered by a manual-input dropdown.
• ManualTriggerDefinition — Manually-invoked trigger definition for an agent.
• ProviderTriggerDefinition — Native provider metadata trigger activated by an explicit delegated connection binding.
• SpeechToSpeechConfig — Signed model, voice, and governance declaration for a chat trigger.
• StepDefinition — Durable step definition created with step().
• StepSchema — Public schema-like surface accepted by the SDK step builder.
• ToolConnectionSelection — Explicit native integration connection selection for delegated credentials.
• ToolReference — Connector-qualified tool reference used by kernel.invokeTool().
• TriggerEnvelope — Persisted trigger envelope stored with each agent run.
• TriggerMetadata — Trigger metadata attached to a runtime invocation.
• TriggerSigningConfig — Signing requirements for publicly declared webhook triggers.
• ValidationIssue — One analyzer or validation issue surfaced to callers.
• ValidationLocation — File location attached to an analyzer issue.
• ValidationResult — Analyzer result returned from analyzeSource().
• WebhookTriggerDefinition — Webhook trigger definition for externally initiated agent runs.
• WorkflowRecoveryDeclaration — Workflow-authored restrictions on connector-attested recovery behavior.

Type Aliases

• AgentKind — Execution mode of an authored agent.
• AgentLocalRule — Local policy rule declared with an agent for static analysis and review.
• AgentLoopBuiltinTool — Built-in loop tools served by the kernel rather than a connector.
• AgentLoopReasoningEffort — Reasoning effort hint forwarded to the loop's model turns.
• AgentLoopStopReason — Why an agent loop stopped.
• AgentLoopToolChoice — Tool selection directive forwarded to the model.
• AgentMemoryScopeType — Organization, hierarchy, user, and bounded task scopes supported by agent memory.
• AgentSdkErrorCode — Stable machine-readable error codes returned by the SDK and runtime.
• AgentStepInputReservedKey — Runtime-owned keys reserved on every authored step input.
• AgentTriggerDefinition — Union of trigger definitions supported for public agent authoring.
• AnyStepDefinition — Erased step definition type used when heterogeneous step arrays cross a package boundary.
• ChatOutputFormat — User-facing rendering format for chat-triggered agent responses.
• DefineAgentConfig — Authoring input accepted by defineAgent.
• DefineAgentTriggerConfig — Trigger authoring input accepted by defineAgent.
• DefineAutonomousAgentConfig — Authoring input accepted by defineAutonomousAgent.
• DlpActionFamily — Action family label used to scope DLP policy overrides.
• DlpClassification — Data classification assigned to scanned payloads.
• DlpScanPolicyConfig — Author-authored DLP policy config accepted by agent and LLM-call definitions.
• DlpValidationDepth — Validation depth used to select DLP scanning layers.
• LlmReasoningEffort — Reasoning effort hint for reasoning-capable LLM models.
• ManualInputContract — Manual-run input contract rendered and validated by the platform.
• ManualInputFieldDefinition — One field or group in the manual-input contract tree.
• ManualInputPayloadValidationResult — Result returned by validateManualInputPayload.
• ManualInputPrimitiveType — Primitive field kinds supported by the manual-input contract.
• RealtimeVoiceId — OpenAI Realtime voices approved for signed speech-to-speech agent declarations.
• SpeechToSpeechMode — Governance mode requested by a signed speech-to-speech declaration.
• SpeechToSpeechRecording — Whether a signed agent permits optional raw call recording.
• ToolConnectionIdentityMode — Identity source used for a native integration connection.
• ToolReferenceSource — Connector runtime that owns a mediated tool.
• TriggerType — Trigger kinds supported in the public Iron Gorilla SDK.
• ValidationIssueSeverity — Severity levels emitted by the SDK analyzer.

Functions

• chatTranscriptMessages — Return normalized prior chat messages from the current step input.
• defineAgent — Create an agent definition that Iron Gorilla can validate and run.
• defineAutonomousAgent — Define a fully autonomous agent: a goal plus a signed loop tool surface, with no authored workflow. The platform generates one step that drives kernel.agentLoop; the model decides which governed tools to call.
• formatChatTranscript — Format prior chat transcript messages as compact role-prefixed text.
• step — Define one durable step in an agent workflow.
• validateManualInputPayload — Validate one manual-run payload against the agent's declared manual-input contract.

Variables

• DLP_ACTION_FAMILIES — Action family labels that can receive DLP policy overrides.
• kernel — Public kernel proxy used inside agent steps for LLM calls, mediated tools, data access, and deterministic runtime utilities.
• ProviderTriggerDefinitionSchema — Zod schema for ProviderTriggerDefinition.