compaction

Documentation

Overview

Package compaction provides deterministic native Google ADK request estimation and durable context compaction callbacks.

Index

• Constants
• type Config
• type Estimate
  • func EstimateRequest(request *model.LLMRequest) (Estimate, error)
• type Manager
  • func New(cfg Config) *Manager
  • func (m *Manager) AfterModel(ctx agent.Context, response *model.LLMResponse, responseError error) (*model.LLMResponse, error)
  • func (m *Manager) BeforeModel(ctx agent.Context, request *model.LLMRequest) (*model.LLMResponse, error)

Constants

const (
	// BytesPerToken is the fixture-pinned raw UTF-8 estimator ratio.
	BytesPerToken = 4

	// Structural overheads are tokens added after the raw byte estimate. They
	// model native request framing without pretending to be a provider tokenizer.
	ContentOverheadTokens         = 4
	PartOverheadTokens            = 1
	FunctionOverheadTokens        = 8
	BinaryOverheadTokens          = 16
	ToolDeclarationOverheadTokens = 12

	// ElisionMarker is the stable replacement for an eligible tool-response body.
	ElisionMarker = "[elided]"
)

Types

type Config

type Config struct {
	Policy      config.Compaction
	Store       *sessionstore.Store
	Budget      *outputlimit.Budget
	WarningSink warning.Warner
}

Config supplies the validated policy and Harness-owned durability resources.

type Estimate

type Estimate struct {
	CanonicalBytes   int `json:"canonicalBytes"`
	RawTokens        int `json:"rawTokens"`
	Contents         int `json:"contents"`
	Parts            int `json:"parts"`
	Functions        int `json:"functions"`
	Binaries         int `json:"binaries"`
	ToolDeclarations int `json:"toolDeclarations"`
	OverheadTokens   int `json:"overheadTokens"`
	Tokens           int `json:"tokens"`
}

Estimate is the deterministic breakdown for one native model request.

func EstimateRequest

func EstimateRequest(request *model.LLMRequest) (Estimate, error)

EstimateRequest estimates the complete model-facing request. LLMRequest.Tools is intentionally excluded because ADK marks it as runtime-only and projects its declarations into Config.Tools before the model call.

type Manager

type Manager struct {
	// contains filtered or unexported fields
}

Manager owns native before-model and after-model compaction state.

func New

func New(cfg Config) *Manager

New constructs a callback manager. A zero context budget leaves both callbacks installed as deterministic no-ops.

func (*Manager) AfterModel

func (m *Manager) AfterModel(ctx agent.Context, response *model.LLMResponse, responseError error) (*model.LLMResponse, error)

AfterModel calibrates future estimates from native prompt usage.

func (*Manager) BeforeModel

func (m *Manager) BeforeModel(ctx agent.Context, request *model.LLMRequest) (*model.LLMResponse, error)

BeforeModel compacts one assembled native ADK request in place.