Embabel Agent Documentation/com.embabel.agent.spi.loop

Package-level declarations

Types

class AutoCorrectionPolicy(maxRetries: Int = DEFAULT_MAX_RETRIES, minTokenLength: Int = DEFAULT_MIN_TOKEN_LENGTH, minTokenSimilarity: Double = DEFAULT_MIN_TOKEN_SIMILARITY) : ToolNotFoundPolicy

Feeds the error back to the LLM with a fuzzy-match suggestion, allowing it to self-correct. Uses a two-tier matching strategy for performance. Throws ToolNotFoundException after maxRetries consecutive failures.

ChainedToolInjectionStrategy

class ChainedToolInjectionStrategy(strategies: List<ToolInjectionStrategy>) : ToolInjectionStrategy

Chains multiple injection strategies into a pipeline.

CompositeToolInjectionStrategy

typealias ~~CompositeToolInjectionStrategy~~ = ChainedToolInjectionStrategy

EmptyResponseAction

sealed class EmptyResponseAction

Action returned by EmptyResponsePolicy.handle.

EmptyResponsePolicy

interface EmptyResponsePolicy

Determines how the tool loop responds when the LLM returns a blank text response with no tool calls. This is a known failure mode of weak open-weights models (gpt-oss-20b, qwen, etc.) after a tool call when they don't know how to proceed — see EmptyLlmResponseException.

ExitOnEmptyPolicy

object ExitOnEmptyPolicy : EmptyResponsePolicy

Preserve the current behaviour: exit the loop with empty content and let the rendering layer surface the typed EmptyLlmResponseException. This is the default to keep the upgrade backwards-compatible.

ImmediateThrowPolicy

object ImmediateThrowPolicy : ToolNotFoundPolicy

Throws ToolNotFoundException immediately on first unknown tool call.

LlmMessageRequest

data class LlmMessageRequest constructor(val messages: List<Message>, val tools: List<Tool>, val nativeStructuredOutputRequest: NativeStructuredOutputRequest? = null)

Framework-agnostic request for a single LLM inference call.

LlmMessageResponse

data class LlmMessageResponse(val message: Message, val textContent: String, val usage: Usage? = null)

Framework-agnostic result of a single LLM inference call. Represents the assistant's response which may include tool calls.

LlmMessageSender

fun interface LlmMessageSender

Framework-agnostic interface for making a single LLM inference call.

MaxIterationsExceededException

class MaxIterationsExceededException(val maxIterations: Int) : RuntimeException

Thrown when the tool loop exceeds the maximum number of iterations.

NativeStructuredOutputRequest

data class NativeStructuredOutputRequest constructor(val structuredOutputRequest: StructuredOutputRequest, val nativeStructuredOutputMode: <Error class: unknown class> = NativeStructuredOutputMode.DEFAULT)

Native structured-output payload for a single call.

RequestAwareLlmMessageSender

interface RequestAwareLlmMessageSender : LlmMessageSender

Optional extension interface for senders that can consume request-level metadata.

RequiredToolGroupException

class RequiredToolGroupException(val role: String, message: String) : RuntimeException

Thrown when a required tool group is unavailable or missing expected tools at resolution time. This is distinct from ToolNotFoundException, which is thrown during tool loop execution when the LLM requests an unavailable tool.

RetryWithFeedbackPolicy

class RetryWithFeedbackPolicy(maxRetries: Int = DEFAULT_MAX_RETRIES, message: String = DEFAULT_MESSAGE) : EmptyResponsePolicy

Re-prompt the LLM with message up to maxRetries times before giving up. After exhaustion returns EmptyResponseAction.Throw so the caller sees the typed exception rather than blank content.

StructuredOutputRequest

data class StructuredOutputRequest constructor(val name: String, val schema: String, val description: String? = null, val strict: Boolean = true)

Provider-neutral request for native structured output.

ToolCallResult

data class ToolCallResult(val toolName: String, val toolInput: String, val result: String, val resultObject: Any?)

Result of a tool call execution.

ToolInjectionContext

data class ToolInjectionContext(val conversationHistory: List<Message>, val currentTools: List<Tool>, val lastToolCall: ToolCallResult, val iterationCount: Int)

Context provided to injection strategies for decision-making.

ToolInjectionResult

data class ToolInjectionResult(val toolsToAdd: List<Tool> = emptyList(), val toolsToRemove: List<Tool> = emptyList())

Result of tool injection evaluation.

ToolInjectionStrategy

interface ToolInjectionStrategy

Strategy for dynamically injecting tools during a conversation.

ToolLoop

interface ToolLoop

Embabel's own tool execution loop.

ToolLoopFactory

interface ToolLoopFactory

Factory for creating ToolLoop instances.

ToolLoopResult

data class ToolLoopResult<O>(val result: O, val rawResponseText: String = "", val conversationHistory: List<Message>, val totalIterations: Int, val injectedTools: List<Tool>, val removedTools: List<Tool> = emptyList(), val totalUsage: Usage? = null, val replanRequested: Boolean = false, val replanReason: String? = null, val blackboardUpdater: BlackboardUpdater = BlackboardUpdater {})

Result of executing an Embabel tool loop.

ToolNotFoundAction

sealed class ToolNotFoundAction

Action returned by ToolNotFoundPolicy.handle.

ToolNotFoundException

class ToolNotFoundException(val requestedTool: String, val availableTools: List<String>) : RuntimeException

Thrown when the LLM requests a tool that is not available.

ToolNotFoundPolicy

interface ToolNotFoundPolicy

Determines how the tool loop responds when the LLM calls a tool that does not exist in the available set.

UnfoldingToolInjectionStrategy

class UnfoldingToolInjectionStrategy : ToolInjectionStrategy

Injection strategy that handles UnfoldingTool invocations.