Herramientas

Learn about toolset types, execution models, validation, retry hints, and tool catalogs in Goa-AI.

Los conjuntos de herramientas son colecciones de herramientas que los agentes pueden utilizar. Goa-AI soporta varios tipos de herramientas, cada una con diferentes modelos de ejecución y casos de uso.

Tipos de herramientas

Conjuntos de herramientas propios del servicio (basados en métodos)

Declaradas mediante Toolset("name", func() { ... }); las herramientas pueden BindTo métodos de servicio Goa o ser implementadas por ejecutores personalizados.

  • Codegen emite especificaciones/tipos/codecs por herramienta en gen/<service>/tools/<toolset>/
  • Los agentes que Use estos conjuntos de herramientas importan las especificaciones del proveedor y obtienen constructores de llamadas tipadas y fábricas de ejecutores
  • Las aplicaciones registran ejecutores que decodifican argumentos tipados (a través de códecs proporcionados en tiempo de ejecución), opcionalmente utilizan transformaciones, llaman a clientes de servicios y devuelven ToolResult

Conjuntos de herramientas implementadas en agentes (Agent-as-Tool)

Definido en un bloque Export del agente, y opcionalmente Used por otros agentes.

  • La propiedad sigue siendo del servicio; el agente es la implementación
  • Codegen emite helpers del lado del proveedor agenttools/<toolset> con NewRegistration y constructores de llamadas tipados
  • Los helpers del lado del consumidor en agentes que Use el conjunto de herramientas exportado delegan en los helpers del proveedor manteniendo centralizados los metadatos de enrutamiento
  • La ejecución se produce en línea; las cargas útiles se pasan como JSON canónico y se descodifican sólo en el límite si es necesario para los avisos

Herramientas MCP

Declarado mediante MCPToolset(service, suite) y referenciado mediante Use(MCPToolset(...)).

  • Registro generado establece DecodeInExecutor=true por lo que JSON crudo se pasa a través del ejecutor MCP
  • El ejecutor MCP decodifica utilizando sus propios códecs
  • Las envolturas generadas manejan esquemas/codificadores JSON y transportes (HTTP/SSE/stdio) con reintentos y rastreo

Cuándo usar BindTo vs Implementaciones Inline

Utilizar BindTo cuando:

  • La herramienta debe llamar a un método de servicio Goa existente
  • Desea transformaciones generadas entre los tipos de herramienta y método
  • El método de servicio ya tiene la lógica de negocio que necesitas
  • Desea reutilizar la validación y el manejo de errores de la capa de servicio
// Tool bound to existing service method
Tool("search", "Search documents", func() {
    Args(SearchPayload)
    Return(SearchResult)
    BindTo("Search")  // Calls the Search method on the same service
})

Utilice implementaciones en línea cuando:

  • La herramienta tiene lógica personalizada no vinculada a un método de servicio
  • Necesitas orquestar múltiples llamadas a servicios
  • La herramienta es puramente computacional (sin llamadas externas)
  • Desea un control total sobre el flujo de ejecución
// Tool with custom executor implementation
Tool("summarize", "Summarize multiple documents", func() {
    Args(func() {
        Attribute("doc_ids", ArrayOf(String), "Document IDs to summarize")
        Required("doc_ids")
    })
    Return(func() {
        Attribute("summary", String, "Combined summary")
        Required("summary")
    })
    // No BindTo - implement in executor
})

Para implementaciones en línea, se escribe directamente la lógica del ejecutor:

func (e *Executor) Execute(ctx context.Context, meta *runtime.ToolCallMeta, call *planner.ToolRequest) (*planner.ToolResult, error) {
    switch call.Name {
    case specs.Summarize:
        args, _ := specs.UnmarshalSummarizePayload(call.Payload)
        // Custom logic: fetch multiple docs, combine, summarize
        summary := e.summarizeDocuments(ctx, args.DocIDs)
        return &planner.ToolResult{
            Name:   call.Name,
            Result: &specs.SummarizeResult{Summary: summary},
        }, nil
    }
    return nil, fmt.Errorf("unknown tool: %s", call.Name)
}

### Bounded Tool Results

Some tools naturally return large lists, graphs, or time-series windows. You can mark these as **bounded views** so that services remain responsible for trimming while the runtime enforces and surfaces the contract.

#### The agent.Bounds Contract

The `agent.Bounds` type is a small, provider-agnostic contract that describes how a tool result has been bounded relative to the full underlying data set:

```go
type Bounds struct {
    Devuelto int // Número de elementos en la vista delimitada
    Total *int // Total antes de truncar (opcional)
    Truncado bool // Si se aplicó algún tope (longitud, ventana, profundidad)
    RefinementHint string // Orientación sobre cómo acotar la consulta cuando se trunca
}
FieldDescription
ReturnedCount of items actually present in the result
TotalBest-effort count of total items before truncation (nil if unknown)
TruncatedTrue if any caps were applied (pagination, depth limits, size limits)
RefinementHintHuman-readable guidance for narrowing the query (e.g., “Add a date filter to reduce results”)

Service Responsibility for Trimming

The runtime does not compute subsets or truncation itself—services are responsible for:

  1. Applying truncation logic: Pagination, result limits, depth caps, time windows
  2. Populating bounds metadata: Setting Returned, Total, Truncated accurately
  3. Providing refinement hints: Guiding users/models on how to narrow queries when results are truncated

This design keeps truncation logic where domain knowledge lives (in services) while providing a uniform contract for the runtime, planners, and UIs to consume.

Declaring Bounded Tools

Use the DSL helper BoundedResult() inside a Tool definition:

Tool("list_devices", "Listar dispositivos con paginación", func() {
    Args(func() {
        Attribute("site_id", String, "Identificador del sitio")
        Attribute("status", String, "Filtrar por estado", func() {
            Enum("online", "offline", "unknown")
        })
        Atributo("limit", Int, "Máximo de resultados", func() {
            Por defecto(50)
            Máximo(500)
        })
        Obligatorio("site_id")
    })
    Return(func() {
        Attribute("devices", ArrayOf(Device), "Dispositivos coincidentes")
        Attribute("returned", Int, "Número de dispositivos devueltos")
        Atributo("total", Int, "Total de dispositivos coincidentes")
        Attribute("truncated", Boolean, "Los resultados se han limitado")
        Attribute("refinement_hint", String, "Cómo limitar los resultados")
        Obligatorio("dispositivos", "devueltos")
    })
    BoundedResult()
    BindTo("DeviceService", "ListDevices")
})

Code Generation

When a tool is marked with BoundedResult():

  • The generated tool spec includes BoundedResult: true
  • The generated result alias type includes a Bounds *agent.Bounds field
  • Generated result types implement the agent.BoundedResult interface:
// Implementación de la interfaz generada
type ListDevicesResult struct {
    Dispositivos []*Device
    Devuelto int
    Total *int
    Truncado bool
    RefinementHint cadena
}

func (r *ListDevicesResult) ResultBounds() *agent.Bounds {
    return &agente.Límites{
        Devuelto: r.Devuelto
        Total: r.Total,
        Truncado: r.Truncado,
        RefinementHint: r.RefinementHint,
    }
}

Implementing Bounded Tools

Services implement truncation and populate bounds metadata:

func (s *DeviceService) ListDevices(ctx context.Context, p *ListDevicesPayload) (*ListDevicesResult, error) {
    // Consulta con límite + 1 para detectar truncamiento
    devices, err := s.repo.QueryDevices(ctx, p.SiteID, p.Status, p.Limit+1)
    if err != nil {
        return nil, err
    }
    
    // Determina si los resultados fueron truncados
    truncado := len(dispositivos) > p.Límite
    if truncado {
        dispositivos = dispositivos[:p.Límite] // Recortar hasta el límite solicitado
    }
    
    // Obtener el recuento total (opcional, puede ser costoso)
    total, _ := s.repo.CountDevices(ctx, p.SiteID, p.Status)
    
    // Construir pista de refinamiento cuando se trunca
    var hint cadena
    si truncado {
        hint = "Añada un filtro de estado o reduzca el ámbito del sitio para ver menos resultados"
    }
    
    return &ListDevicesResultado{
        Dispositivos: dispositivos,
        Devuelto: len(dispositivos),
        Total: &total,
        Truncado: truncado,
        RefinementHint: hint,
    }, nil
}

Runtime Behavior

When a bounded tool executes:

  1. The runtime decodes the result and checks for agent.BoundedResult implementation
  2. If the result implements the interface, ResultBounds() extracts bounds metadata
  3. Bounds are attached to planner.ToolResult.Bounds
  4. Stream subscribers and finalizers can access bounds for UI display, logging, or policy decisions
// En un suscriptor de flujo
func handleToolEnd(event *stream.ToolEndEvent) {
    if evento.Límites != nil && evento.Límites.Truncado {
        log.Printf("La herramienta %s devolvió %d de %d resultados (truncados)",
            event.ToolName, event.Bounds.Returned, *event.Bounds.Total)
        if evento.Límites.SugerenciaRefinamiento != "" {
            log.Printf("Sugerencia: %s", evento.Sugerencia.Refinamiento)
        }
    }
}

When to Use BoundedResult

Use BoundedResult() for tools that:

  • Return paginated lists (devices, users, records, logs)
  • Query large datasets with result limits
  • Apply depth or size caps to nested structures (graphs, trees)
  • Return time-windowed data (metrics, events)

The bounded contract helps:

  • Models understand that results may be incomplete and can request refinement
  • UIs display truncation indicators and pagination controls
  • Policies enforce size limits and detect runaway queries

Injected Fields

The Inject DSL function marks specific payload fields as “injected”—server-side infrastructure values that are hidden from the LLM but required by the service method. This is useful for session IDs, user context, authentication tokens, and other runtime-provided values.

How Inject Works

When you mark a field with Inject:

  1. Hidden from LLM: The field is excluded from the JSON schema sent to the model provider
  2. Generated setter: Codegen emits a setter method on the payload struct
  3. Runtime population: You populate the field via a ToolInterceptor before execution

DSL Declaration

Tool("get_user_data", "Obtener datos del usuario actual", func() {
    Args(func() {
        Attribute("session_id", String, "ID de sesión actual")
        Attribute("query", String, "Consulta de datos")
        Requerido("session_id", "consulta")
    })
    Return(func() {
        Attribute("data", ArrayOf(String), "Resultados de la consulta")
        Requerido("datos")
    })
    BindTo("UserService", "GetData")
    Inject("session_id") // Oculto en LLM, rellenado en tiempo de ejecución
})

Generated Code

Codegen produces a setter method for each injected field:

// Carga útil generada struct
type GetUserDataPayload struct {
    Cadena SessionID `json:"session_id"`
    Cadena de consulta `json:"query"`
}

// Setter generado para el campo inyectado
func (p *GetUserDataPayload) SetSessionID(v string) {
    p.SessionID = v
}

Runtime Population via ToolInterceptor

Use a ToolInterceptor to populate injected fields before tool execution:

type SessionInterceptor struct{}

func (i *SessionInterceptor) InterceptToolCall(ctx context.Context, call *planner.ToolCall) error {
    // Extrae la sesión del contexto (establecida por el middleware de autenticación)
    sessionID, ok := ctx.Value(sessionKey).(string)
    if !ok {
        return fmt.Errorf("no se ha encontrado el ID de sesión en el contexto")
    }
    
    // Rellenar el campo inyectado usando el setter generado
    switch llamada.Nombre {
    case specs.GetUserData:
        payload, _ := specs.UnmarshalGetUserDataPayload(call.Payload)
        payload.SetSessionID(sessionID)
        call.Payload, _ = json.Marshal(payload)
    }
    return nil
}

// Registrar interceptor con runtime
rt := runtime.New(runtime.WithToolInterceptor(&SessionInterceptor{}))

When to Use Inject

Use Inject for fields that:

  • Are required by the service but shouldn’t be chosen by the LLM
  • Come from runtime context (session, user, tenant, request ID)
  • Contain sensitive values (auth tokens, API keys)
  • Are infrastructure concerns (tracing IDs, correlation IDs)

Execution Models

Activity-Based Execution (Default)

Service-backed toolsets execute via Temporal activities (or equivalent in other engines):

  1. Planner returns tool calls in PlanResult (payload is json.RawMessage)
  2. Runtime schedules ExecuteToolActivity for each tool call
  3. Activity decodes payload via generated codec for validation/hints
  4. Calls the toolset registration’s Execute(ctx, planner.ToolRequest) with canonical JSON
  5. Re-encodes the result with the generated result codec

Inline Execution (Agent-as-Tool)

Agent-as-tool toolsets execute inline from the planner’s perspective while the runtime runs the provider agent as a real child run:

  1. The runtime detects Inline=true on the toolset registration
  2. It injects the engine.WorkflowContext into ctx so the toolset’s Execute function can start the provider agent as a child workflow with its own RunID
  3. It calls the toolset’s Execute(ctx, call) with canonical JSON payload and tool metadata (including parent RunID and ToolCallID)
  4. The generated agent-tool executor builds nested agent messages (system + user) from the tool payload and runs the provider agent as a child run
  5. The nested agent executes a full plan/execute/resume loop in its own run; its RunOutput and tool events are aggregated into a parent planner.ToolResult that carries the result payload, aggregated telemetry, child ChildrenCount, and a RunLink pointing at the child run
  6. Stream subscribers emit both tool_start / tool_end for the parent tool call and an agent_run_started link event so UIs and debuggers can attach to the child run’s stream on demand

Executor-First Model

Generated service toolsets expose a single, generic constructor:

New<Agent><Toolset>ToolsetRegistration(exec runtime.ToolCallExecutor)

Applications register an executor implementation for each consumed toolset. The executor decides how to run the tool (service client, MCP, nested agent, etc.) and receives explicit per-call metadata via ToolCallMeta.

Executor Example:

func Execute(ctx context.Context, meta runtime.ToolCallMeta, call planner.ToolRequest) (planner.ToolResult, error) {
    switch llamada.Nombre {
    case "orquestador.perfiles.upsert":
        args, err := profilesspecs.UnmarshalUpsertPayload(call.Payload)
        if err != nil {
            return planner.ToolResult{
                Error: planner.NewToolError("carga no válida"),
            }, nil
        }
        
        // Transformaciones opcionales si las emite codegen
        mp, _ := profilesspecs.ToMethodPayload_Upsert(args)
        methodRes, err := client.Upsert(ctx, mp)
        if err != nil {
            return planner.ToolResult{
                Error: planner.ToolErrorFromError(err),
            }, nil
        }
        tr, _ := profilesspecs.ToToolReturn_Upsert(methodRes)
        return planner.ToolResult{Carga: tr}, nil
        
    por defecto:
        return planner.ToolResult{
            Error: planner.NewToolError("herramienta desconocida"),
        }, nil
    }
}

Tool Call Metadata

Tool executors receive explicit per-call metadata via ToolCallMeta rather than fishing values from context.Context. This provides direct access to run-scoped identifiers for correlation, telemetry, and parent/child relationships.

ToolCallMeta Fields

FieldDescription
RunIDDurable workflow execution identifier of the run that owns this tool call. Stable across retries; used to correlate runtime records and telemetry.
SessionIDLogically groups related runs (e.g., a chat conversation). Services typically index memory and search attributes by session.
TurnIDIdentifies the conversational turn that produced this tool call. Event streams use it to order and group events.
ToolCallIDUniquely identifies this tool invocation. Used to correlate start/update/end events and parent/child relationships.
ParentToolCallIDIdentifier of the parent tool call when this invocation is a child (e.g., a tool launched by an agent-tool). UIs and subscribers use it to reconstruct the call tree.

Executor Signature

All tool executors receive ToolCallMeta as an explicit parameter:

func Ejecutar(ctx context.Context, meta *runtime.ToolCallMeta, call *planner.ToolRequest) (*planner.ToolResult, error) {
    // Accede al contexto de ejecución directamente desde meta
    log.Printf("Ejecutando herramienta en run %s, session %s, turno %s",
        meta.RunID, meta.SessionID, meta.TurnID)
    
    // Utilizar ToolCallID para la correlación
    span := tracer.StartSpan("tool.execute", trace.WithAttributes(
        attribute.String("tool.call_id", meta.ToolCallID),
        attribute.String("tool.parent_call_id", meta.ParentToolCallID),
    ))
    defer span.End()
    
    // ... implementación de la herramienta
}

Why Explicit Metadata?

The explicit metadata pattern provides several benefits:

  • Type safety: Compile-time guarantees that required identifiers are available
  • Testability: Easy to construct test metadata without mocking context
  • Clarity: No hidden dependencies on context keys or middleware ordering
  • Correlation: Direct access to parent/child relationships for nested agent-tool calls
  • Traceability: Complete causal chain from user input to tool execution to final response

Async & Durable Execution

Goa-AI uses Temporal Activities for all service-backed tool executions. This “async-first” architecture is implicit and requires no special DSL.

Implicit Async

When a planner decides to call a tool, the runtime does not block the OS thread. Instead:

  1. The runtime schedules a Temporal Activity for the tool call.
  2. The agent workflow suspends execution (saving state).
  3. The activity executes (on a local worker, remote worker, or even a different cluster).
  4. When the activity completes, the workflow wakes up, restores state, and resumes with the result.

This means every tool call is automatically parallelizable, durable, and long-running. You do not need to configure InterruptsAllowed for this standard async behavior.

Pause & Resume (Agent-Level)

InterruptsAllowed(true) is distinct: it allows the Agent itself to pause and wait for an arbitrary external signal (like a user’s clarification) that is not tied to a currently running tool activity.

FeatureImplicit AsyncPause & Resume
ScopeSingle Tool ExecutionEntire Agent Workflow
TriggerCalling any service-backed toolMissing arguments or Planner request
Policy RequiredNone (Default)InterruptsAllowed(true)
Use CaseSlow API, Batch Job, processingHuman-in-the-loop, Clarification

Ensure you verify that your use case requires agent-level pausing before enabling the policy; often, standard tool async is sufficient.

Non-Blocking Planners

From the perspective of the planner (LLM), the interaction feels synchronous: the model requests a tool, “pauses”, and then “sees” the result in the next turn.

From the perspective of the infrastructure, it is fully asynchronous and non-blocking. This allows a single small agent worker to manage thousands of concurrent long-running agent executions without running out of threads or memory.

Survival Across Restarts

Because execution is durable, you can restart your entire backend—including the agent workers—while tools are mid-execution. When the systems come back up:

  • Pending tool activities will be picked up by workers.
  • Completed tools will report results to their parent workflows.
  • Agents will resume exactly where they left off.

This capability is essential for building robust, production-grade agentic systems that operate reliably in dynamic environments.


Transforms

When a tool is bound to a Goa method via BindTo, code generation analyzes the tool Arg/Return and the method Payload/Result. If the shapes are compatible, Goa emits type-safe transform helpers:

  • ToMethodPayload_<Tool>(in <ToolArgs>) (<MethodPayload>, error)
  • ToToolReturn_<Tool>(in <MethodResult>) (<ToolReturn>, error)

Transforms are emitted under gen/<service>/agents/<agent>/specs/<toolset>/transforms.go and use Goa’s GoTransform to safely map fields. If a transform isn’t emitted, write an explicit mapper in the executor.


Tool Identity

Each toolset defines typed tool identifiers (tools.Ident) for all generated tools—including non-exported toolsets. Prefer these constants over ad-hoc strings:

import chattools "example.com/assistant/gen/orchestrator/agents/chat/agenttools/search"

// Utilizar una constante generada en lugar de cadenas/cast ad-hoc
spec, _ := rt.ToolSpec(chattools.Search)
schemas, _ := rt.ToolSchema(chattools.Search)

For exported toolsets (agent-as-tool), Goa-AI also generates agenttools packages with:

  • Typed tool IDs
  • Alias payload/result types
  • Codecs
  • Helper builders (e.g., New<Search>Call)

Tool Validation and Retry Hints

Goa-AI combines Goa’s design-time validations with a structured tool error model to give LLM planners a powerful way to repair invalid tool calls automatically.

Core Types: ToolError and RetryHint

ToolError (alias to runtime/agent/toolerrors.ToolError):

  • Message string – human-readable summary
  • Cause *ToolError – optional nested cause (preserves chains across retries and agent-as-tool hops)
  • Constructors: planner.NewToolError(msg), planner.NewToolErrorWithCause(msg, cause), planner.ToolErrorFromError(err), planner.ToolErrorf(format, args...)

RetryHint – planner-side hint used by the runtime and policy engine:

type RetryHint struct {
    Razón RetryReason
    Herramienta tools.Ident
    RestrictToTool bool
    MissingFields []cadena
    ExampleInput mapa[cadena]cualquiera
    PriorInput mapa[cadena]cualquiera
    ClarifyingQuestion cadena
    Mensaje cadena
}

Common RetryReason values:

  • invalid_arguments – payload failed validation (schema/type)
  • missing_fields – required fields are missing
  • malformed_response – tool returned data that could not be decoded
  • timeout, rate_limited, tool_unavailable – execution/infra issues

ToolResult carries errors and hints:

type ToolResult struct {
    Nombre herramientas.Ident
    Resultado any
    Error *ToolError
    RetryHint *RetryHint
    Telemetría *telemetry.ToolTelemetry
    ToolCallID cadena
    ChildrenCount int
    RunLink *run.Handle
}

Auto-Repairing Invalid Tool Calls

The recommended pattern:

  1. Design tools with strong payload schemas (Goa design)
  2. Let executors/tools surface validation failures as ToolError + RetryHint instead of panicking or hiding errors
  3. Teach your planner to inspect ToolResult.Error and ToolResult.RetryHint, repair the payload when possible, and retry the tool call if appropriate

Example Executor:

func Execute(ctx context.Context, meta runtime.ToolCallMeta, call planner.ToolRequest) (*planner.ToolResult, error) {
    args, err := spec.UnmarshalUpsertPayload(call.Payload)
    if err != nil {
        return &planner.ToolResult{
            Nombre: call.Nombre,
            Error: planner.NewToolError("carga no válida"),
            RetryHint: &planner.RetryHint{
                Reason: planner.RetryReasonInvalidArguments,
                Tool: call.Name,
                RestrictToTool: true,
                Mensaje:       "La carga útil no coincide con el esquema esperado",
            },
        }, nil
    }

    res, err := client.Upsert(ctx, args)
    if err != nil {
        return &planner.ToolResult{
            Nombre: llamada.Nombre,
            Error: planner.ToolErrorFromError(err),
        }, nil
    }

    return &planner.ToolResult{Nombre: llamada.Nombre, Resultado: res}, nil
}

Example Planner Logic:

func (p *MyPlanner) PlanResume(ctx context.Context, in *planner.PlanResumeInput) (*planner.PlanResult, error) {
    if len(in.ToolResults) == 0 {
        return &planner.PlanResult{}, nil
    }

    last := in.ToolResults[len(in.ToolResults)-1]
    if last.Error != nil && last.RetryHint != nil {
        hint := last.RetryHint

        switch hint.Reason {
        case planner.RetryReasonMissingFields, planner.RetryReasonInvalidArguments:
            return &planner.PlanResult{
                Esperar: &planner.Esperar{
                    Aclaración: &planner.AwaitClarification{
                        ID               "fix-" + string(hint.Tool),
                        Question: hint.ClarifyingQuestion,
                        MissingFields: hint.MissingFields,
                        RestrictToTool: hint.Tool,
                        ExampleInput: hint.ExampleInput,
                        ClarifyingPrompt: hint.Message,
                    },
                },
            }, nil
        }
    }

    return &planner.PlanResult{/* FinalResponse, next ToolCalls, ... */}, nil
}

Tool Catalogs and Schemas

Goa-AI agents generate a single, authoritative catalog of tools from your Goa designs. This catalog powers:

  • Planner tool advertisement (which tools the model can call)
  • UI discovery (tool lists, categories, schemas)
  • External orchestrators (MCP, custom frontends) that need machine-readable specs

Generated Specs and tool_schemas.json

For each agent, Goa-AI emits a specs package and a JSON catalog:

Specs packages (gen/<service>/agents/<agent>/specs/...):

  • types.go – payload/result Go structs
  • codecs.go – JSON codecs (encode/decode typed payloads/results)
  • specs.go[]tools.ToolSpec entries with canonical tool ID, payload/result schemas, hints

JSON catalog (tool_schemas.json):

Location: gen/<service>/agents/<agent>/specs/tool_schemas.json

Contains one entry per tool with:

  • id – canonical tool ID ("<service>.<toolset>.<tool>")
  • service, toolset, title, description, tags
  • payload.schema and result.schema (JSON Schema)

This JSON file is ideal for feeding schemas to LLM providers, building UI forms/editors, and offline documentation tooling.

Runtime Introspection APIs

At runtime, you do not need to read tool_schemas.json from disk. The runtime exposes an introspection API:

agentes := rt.ListAgents() // []agente.Ident
toolsets := rt.ListToolsets() // []cadena

spec, ok := rt.ToolSpec(toolID) // single ToolSpec
schemas, ok := rt.ToolSchema(toolID) // esquemas de carga/resultado
specs := rt.ToolSpecsForAgent(chat.AgentID) // []ToolSpec para un agente

Where toolID is a typed tools.Ident constant from a generated specs or agenttools package.

Typed Sidecars and Artifacts

Some tools need to return rich artifacts (full time series, topology graphs, large result sets) that are useful for UIs and audits but too heavy for model providers. Goa-AI models these as typed sidecars (also called artifacts):

Model-Facing vs Sidecar Data

The key distinction is what data flows where:

Data TypeSent to ModelStored/StreamedPurpose
Model-facing resultBounded summary the LLM reasons about
Sidecar/ArtifactFull-fidelity data for UIs, audits, downstream consumers

This separation lets you:

  • Keep model context windows bounded and focused
  • Provide rich visualizations (charts, graphs, tables) without bloating LLM prompts
  • Attach provenance and audit data that models don’t need to see
  • Stream large datasets to UIs while the model works with summaries

Declaring Artifacts in DSL

Use the Artifact(kind, schema) function inside a Tool definition:

Tool("get_time_series", "Obtener datos de series temporales", func() {
    Args(func() {
        Attribute("device_id", String, "Identificador del dispositivo")
        Attribute("start_time", String, "Hora de inicio (RFC3339)")
        Attribute("end_time", String, "Hora de finalización (RFC3339)")
        Obligatorio("device_id", "start_time", "end_time")
    })
    // Resultado orientado al modelo: resumen acotado
    Return(func() {
        Attribute("summary", String, "Resumen del modelo")
        Attribute("count", Int, "Número de puntos de datos")
        Attribute("min_value", Float64, "Valor mínimo en el rango")
        Attribute("max_value", Float64, "Valor máximo del intervalo")
        Obligatorio("summary", "count")
    })
    // Sidecar: datos completos para las interfaces de usuario
    Artifact("series_tiempo", func() {
        Attribute("data_points", ArrayOf(TimeSeriesPoint), "Datos completos de series temporales")
        Atributo("metadatos", MapOf(String, String), "Metadatos adicionales")
        Obligatorio("data_points")
    })
})

The kind parameter (e.g., "time_series") identifies the artifact type so UIs can dispatch appropriate renderers.

Generated Specs and Helpers

In the specs packages, each tools.ToolSpec entry includes:

  • Payload tools.TypeSpec – tool input schema
  • Result tools.TypeSpec – model-facing output schema
  • Sidecar *tools.TypeSpec (optional) – artifact schema

Goa-AI generates typed helpers for working with sidecars:

// Obtener artefacto de un resultado de herramienta
func GetGetTimeSeriesSidecar(res *planner.ToolResult) (*GetTimeSeriesSidecar, error)

// Adjuntar artefacto a un resultado de herramienta
func SetGetTimeSeriesSidecar(res *planner.ToolResult, sc *GetTimeSeriesSidecar) error

Runtime Usage Patterns

In tool executors, attach artifacts to results:

func (e *Executor) Execute(ctx context.Context, meta *runtime.ToolCallMeta, call *planner.ToolRequest) (*planner.ToolResult, error) {
    args, _ := specs.UnmarshalGetTimeSeriesPayload(call.Payload)
    
    // Obtener datos completos
    fullData, err := e.dataService.GetTimeSeries(ctx, args.DeviceID, args.StartTime, args.EndTime)
    if err != nil {
        return &planner.ToolResult{Error: planner.ToolErrorFromError(err)}, nil
    }
    
    // Construir resultado de modelo acotado
    result := &specs.GetTimeSeriesResult{
        Resumen: fmt.Sprintf("Recuperados %d puntos de datos de %s a %s", len(fullData.Points), args.StartTime, args.EndTime),
        Count: len(datoscompletos.puntos),
        MinValue: fullData.Min,
        MaxValue: fullData.Max,
    }
    
    // Construir artefacto de fidelidad completa para UIs
    artefacto := &specs.GetTimeSeriesSidecar{
        PuntosDatos: fullData.Puntos,
        Metadatos: fullData.Metadata,
    }
    
    // Adjuntar artefacto al resultado
    toolResult := &planner.ToolResult{
        Nombre: llamada.Nombre,
        Resultado: resultado,
    }
    specs.SetGetTimeSeriesSidecar(toolResult, artifact)
    
    return toolResultado, nil
}

In stream subscribers or UI handlers, access artifacts:

func handleToolEnd(event *stream.ToolEndEvent) {
    // Los artefactos están disponibles en el evento
    for _, artefacto := rango evento.Artefactos {
        switch artefacto.tipo {
        case "series_tiempo":
            // Renderizar gráfico de series temporales
            renderTimeSeriesChart(artefacto.Datos)
        case "topología":
            // Renderizar gráfico de red
            renderTopologyGraph(artefacto.Datos)
        }
    }
}

Artifact Structure

The planner.Artifact type carries:

type Artefacto struct {
    Kind string // Tipo lógico (por ejemplo, "time_series", "chart_data")
    Data any // Carga útil serializable en JSON
    SourceTool tools.Ident // Herramienta que produjo este artefacto
    RunLink *run.Handle // Enlace a la ejecución del agente anidado (para el agente como herramienta)
}

Cuándo usar artefactos

Utilice artefactos cuando:

  • Los resultados de la herramienta incluyen datos demasiado grandes para el contexto del modelo (series temporales, registros, tablas grandes)
  • Las interfaces de usuario necesitan datos estructurados para su visualización (tablas, gráficos, mapas)
  • Se desea separar lo que razona el modelo de lo que ven los usuarios
  • Los sistemas posteriores necesitan datos completos, mientras que el modelo trabaja con resúmenes

Evite los artefactos cuando:

  • El resultado completo encaja cómodamente en el contexto del modelo
  • No hay ninguna interfaz de usuario o consumidor que necesite los datos completos
  • El resultado delimitado ya contiene todo lo necesario

Mejores prácticas

  • Poner las validaciones en el diseño, no en los planificadores - Usar el DSL de atributos de Goa (Required, MinLength, Enum, etc.)
  • Devolver ToolError + RetryHint de los ejecutores - Preferir errores estructurados en lugar de pánicos o simples devoluciones error
  • Mantenga las sugerencias concisas pero procesables - Céntrese en los campos que faltan/no son válidos, una breve pregunta aclaratoria y un pequeño mapa ExampleInput
  • Enseñe a los planificadores a leer las sugerencias - Convierta la gestión de RetryHint en una parte esencial de su planificador
  • Evitar la revalidación dentro de los servicios - Goa-AI asume que la validación ocurre en el límite de la herramienta

Próximos pasos