Ecosystem

tRPC-Agent-Go Ecosystem Development Guide

This document analyzes the modules in the tRPC-Agent-Go framework that need ecosystem development, explains the interfaces that need to be implemented, and provides contribution guidance.

Note: All community-built components are contributed directly to the corresponding directory under the GitHub open source repository, such as model/somemodel, tool/sometool, agent/someagent, etc.

When contributing, create a reasonably named subfolder under the corresponding contribution module directory, then implement the corresponding module interface, and provide rich test cases and example samples.

Ecosystem Development Module Analysis

1. Agent Ecosystem

Goal: Encapsulate and adapt third-party Agent frameworks

Interface Definition: agent.Agent

Existing Implementation Reference: LLMAgent

Implementation Notes:

  • The Run method must return an event channel, supporting streaming responses
  • The Tools method returns the list of tools available to the Agent
  • The Info method provides basic information about the Agent
  • SubAgents and FindSubAgent support Agent composition patterns
  • Reference LLMAgent implementation to understand event handling and error handling mechanisms

Implementation Example:

package langchain

import (
    "context"

    "trpc.group/trpc-go/trpc-agent-go/agent"
    "trpc.group/trpc-go/trpc-agent-go/event"
    "trpc.group/trpc-go/trpc-agent-go/tool"
)

type LangChainAdapter struct {
    config *Config
    client *langchain.Client
}

func New(config *Config) (agent.Agent, error) {
    client := langchain.NewClient(config.Endpoint, config.APIKey)

    return &LangChainAdapter{
        config: config,
        client: client,
    }, nil
}

func (a *LangChainAdapter) Run(ctx context.Context, invocation *agent.Invocation) (<-chan *event.Event, error) {
    events := make(chan *event.Event)

    go func() {
        defer close(events)

        response, err := a.client.Call(ctx, invocation.Messages)
        if err != nil {
            events <- &event.Event{
                Type: event.TypeError,
                Error: err,
            }
            return
        }

        events <- &event.Event{
            Type: event.TypeResponse,
            Response: &model.Response{
                Content: response.Content,
            },
        }
    }()

    return events, nil
}

func (a *LangChainAdapter) Tools() []tool.Tool {
    return a.config.Tools
}

func (a *LangChainAdapter) Info() agent.Info {
    return agent.Info{
        Name:        "langchain-adapter",
        Description: "LangChain framework adapter",
    }
}

func (a *LangChainAdapter) SubAgents() []agent.Agent {
    return nil
}

func (a *LangChainAdapter) FindSubAgent(name string) agent.Agent {
    return nil
}

Open Source Components That Can Be Integrated:

  • LangChain adapter
  • LangGraph adapter

Contribution Method:

  • Create components under the corresponding directory (create a subdirectory with the corresponding component name)
  • Contribute directly to https://github.com/trpc-group/trpc-agent-go/agent/

2. Model Ecosystem

Goal: Support more model providers

Interface Definition: model.Model

Existing Implementation Reference: OpenAI Model

Implementation Notes:

  • The GenerateContent method must support streaming responses, returning an event channel
  • Distinguish between system-level errors (return error) and API-level errors (Response.Error)
  • Implement the Info method to provide basic model information
  • Reference OpenAI implementation to understand request building and response handling
  • Support context cancellation and timeout control

Implementation Example:

package gemini

import (
    "context"
    "fmt"
    "trpc.group/trpc-go/trpc-agent-go/model"
)

type GeminiModel struct {
    config *Config
    client *gemini.Client
}

func New(config *Config) (model.Model, error) {
    client := gemini.NewClient(config.APIKey)

    return &GeminiModel{
        config: config,
        client: client,
    }, nil
}

func (g *GeminiModel) GenerateContent(ctx context.Context, request *model.Request) (<-chan *model.Response, error) {
    if request == nil {
        return nil, fmt.Errorf("request cannot be nil")
    }

    responses := make(chan *model.Response)

    go func() {
        defer close(responses)

        // Call Gemini API.
        stream, err := g.client.GenerateContent(ctx, request.Messages)
        if err != nil {
            responses <- &model.Response{
                Error: &model.Error{
                    Message: err.Error(),
                },
            }
            return
        }

        for chunk := range stream {
            responses <- &model.Response{
                Content: chunk.Content,
            }
        }
    }()

    return responses, nil
}

func (g *GeminiModel) Info() model.Info {
    return model.Info{
        Name: "gemini-pro",
    }
}

Open Source Components That Can Be Integrated:

  • Google Gemini model support
  • Anthropic Claude model support
  • Ollama local model support

Contribution Method:

  • Create components under the corresponding directory (create a subdirectory with the corresponding component name)
  • Contribute directly to https://github.com/trpc-group/trpc-agent-go/model/

3. Tool Ecosystem

Goal: Integrate more third-party tools

Interface Definition:

Existing Implementation Reference: DuckDuckGo Tool

Implementation Notes:

Single Tool Implementation:

  • The Declaration method must return complete tool metadata
  • The Call method receives JSON format parameters and returns results of any type
  • Use JSON Schema to define input and output formats
  • Reference DuckDuckGo implementation to understand tool calling and error handling
  • Support context cancellation and timeout control

Tool Collection Implementation:

  • The Tools method returns available tool lists based on context
  • The Close method releases resources held by the tool collection
  • Support dynamic tool loading and configuration
  • Implement tool lifecycle management

Implementation Example:

Single Tool Implementation:

package weather

import (
    "context"
    "encoding/json"
    "fmt"
    "net/http"

    "trpc.group/trpc-go/trpc-agent-go/tool"
)

type WeatherTool struct {
    apiKey string
    client *http.Client
}

func New(apiKey string) tool.CallableTool {
    return &WeatherTool{
        apiKey: apiKey,
        client: &http.Client{},
    }
}

func (w *WeatherTool) Declaration() *tool.Declaration {
    return &tool.Declaration{
        Name:        "get_weather",
        Description: "Get current weather information for a location",
        InputSchema: &tool.Schema{
            Type: "object",
            Properties: map[string]*tool.Schema{
                "location": {
                    Type:        "string",
                    Description: "City name or coordinates",
                },
            },
            Required: []string{"location"},
        },
        OutputSchema: &tool.Schema{
            Type: "object",
            Properties: map[string]*tool.Schema{
                "temperature": {Type: "number"},
                "condition":   {Type: "string"},
                "humidity":    {Type: "number"},
            },
        },
    }
}

func (w *WeatherTool) Call(ctx context.Context, jsonArgs []byte) (any, error) {
    var args struct {
        Location string `json:"location"`
    }

    if err := json.Unmarshal(jsonArgs, &args); err != nil {
        return nil, fmt.Errorf("invalid arguments: %w", err)
    }

    url := fmt.Sprintf("https://api.weatherapi.com/v1/current.json?key=%s&q=%s", w.apiKey, args.Location)

    req, err := http.NewRequestWithContext(ctx, "GET", url, nil)
    if err != nil {
        return nil, err
    }

    resp, err := w.client.Do(req)
    if err != nil {
        return nil, err
    }
    defer resp.Body.Close()

    var weatherData map[string]interface{}
    if err := json.NewDecoder(resp.Body).Decode(&weatherData); err != nil {
        return nil, err
    }

    return weatherData, nil
}

Tool Collection Implementation:

package apitools

import (
    "context"
    "sync"
    "trpc.group/trpc-go/trpc-agent-go/tool"
)

type APIToolSet struct {
    tools map[string]tool.Tool
    mu    sync.RWMutex
}

func New() *APIToolSet {
    return &APIToolSet{
        tools: make(map[string]tool.Tool),
    }
}

func (a *APIToolSet) AddTool(name string, t tool.Tool) {
    a.mu.Lock()
    defer a.mu.Unlock()
    a.tools[name] = t
}

func (a *APIToolSet) RemoveTool(name string) {
    a.mu.Lock()
    defer a.mu.Unlock()
    delete(a.tools, name)
}

func (a *APIToolSet) Tools(ctx context.Context) []tool.Tool {
    a.mu.RLock()
    defer a.mu.RUnlock()

    var result []tool.Tool
    for _, t := range a.tools {
        result = append(result, t)
    }
    return result
}

func (a *APIToolSet) Close() error {
    a.mu.Lock()
    defer a.mu.Unlock()

    // Clean up resources.
    a.tools = make(map[string]tool.Tool)
    return nil
}

// Name identifies this ToolSet.
func (a *APIToolSet) Name() string { return "api-tools" }

Open Source Components That Can Be Integrated:

  • Search engine tools (Google, Bing)
  • Weather query tools
  • Calculator tools
  • File operation tools
  • API tool collections (REST API toolkit)
  • Database operation tool collections
  • File processing tool collections

Contribution Method:

  • Create components under the corresponding directory (create a subdirectory with the corresponding component name)
  • Contribute directly to https://github.com/trpc-group/trpc-agent-go/tool/

4. Knowledge Base Ecosystem

Goal: Integrate mature RAG components

Interface Definition: knowledge.Knowledge

Existing Implementation Reference:

Implementation Notes:

  • The Search method supports context and history records
  • Returns relevant documents and relevance scores
  • Supports search parameters and result limits
  • Reference InMemory implementation to understand search logic and result processing
  • Supports vectorized search and semantic matching

Implementation Example:

package weaviate

import (
    "context"
    "fmt"
    "trpc.group/trpc-go/trpc-agent-go/knowledge"
    "trpc.group/trpc-go/trpc-agent-go/knowledge/document"
)

type WeaviateKnowledge struct {
    config *Config
    client *weaviate.Client
}

func New(config *Config) (knowledge.Knowledge, error) {
    client := weaviate.NewClient(config.Endpoint, config.APIKey)

    return &WeaviateKnowledge{
        config: config,
        client: client,
    }, nil
}

func (w *WeaviateKnowledge) Search(ctx context.Context, req *knowledge.SearchRequest) (*knowledge.SearchResult, error) {
    if req.Query == "" {
        return nil, fmt.Errorf("query cannot be empty")
    }

    // Build search query.
    query := w.buildQuery(req)

    // Execute vector search.
    results, err := w.client.Search(ctx, query)
    if err != nil {
        return nil, err
    }

    if len(results) == 0 {
        return nil, fmt.Errorf("no results found")
    }

    // Return best matching result.
    bestResult := results[0]

    return &knowledge.SearchResult{
        Document: &document.Document{
            ID:      bestResult.ID,
            Content: bestResult.Content,
            Metadata: bestResult.Metadata,
        },
        Score: bestResult.Score,
        Text:  bestResult.Content,
    }, nil
}

func (w *WeaviateKnowledge) buildQuery(req *knowledge.SearchRequest) *weaviate.Query {
    // Build Weaviate query logic.
    return &weaviate.Query{
        Query: req.Query,
        Limit: req.MaxResults,
        Filter: w.buildFilter(req),
    }
}

Open Source Components That Can Be Integrated:

  • Weaviate vector database
  • Pinecone vector database
  • Qdrant vector database

Contribution Method:

  • Create components under the corresponding directory (create a subdirectory with the corresponding component name)
  • Contribute directly to https://github.com/trpc-group/trpc-agent-go/knowledge/

5. Session Ecosystem

Goal: Support multiple session storage backends, manage user session state and events

Interface Definition: session.Service

Existing Implementation Reference:

Implementation Notes:

  • Implement complete Session lifecycle management (create, get, delete, list)
  • Support state storage and event recording
  • Implement connection pooling and error handling
  • Support transactions and consistency
  • Can reuse storage module clients
  • Reference InMemory and Redis implementations to understand Session management logic

Implementation Example:

package postgresql

import (
    "context"
    "database/sql"
    "encoding/json"
    "fmt"
    "time"
    "trpc.group/trpc-go/trpc-agent-go/event"
    "trpc.group/trpc-go/trpc-agent-go/session"
)

type PostgreSQLService struct {
    db *sql.DB
}

func New(dsn string) (session.Service, error) {
    db, err := sql.Open("postgres", dsn)
    if err != nil {
        return nil, err
    }

    if err := db.Ping(); err != nil {
        return nil, err
    }

    return &PostgreSQLService{db: db}, nil
}

func (p *PostgreSQLService) CreateSession(ctx context.Context, key session.Key, state session.StateMap,
    options ...session.Option) (*session.Session, error) {
    if err := key.CheckSessionKey(); err != nil {
        return nil, err
    }

    now := time.Now()
    session := &session.Session{
        ID:        key.SessionID,
        AppName:   key.AppName,
        UserID:    key.UserID,
        State:     state,
        Events:    []event.Event{},
        UpdatedAt: now,
        CreatedAt: now,
    }

    // Insert into database.
    _, err := p.db.ExecContext(ctx, `
        INSERT INTO sessions (id, app_name, user_id, state, created_at, updated_at)
        VALUES ($1, $2, $3, $4, $5, $6)
    `, session.ID, session.AppName, session.UserID, p.marshalState(state), session.CreatedAt, session.UpdatedAt)

    if err != nil {
        return nil, err
    }

    return session, nil
}

func (p *PostgreSQLService) GetSession(ctx context.Context, key session.Key,
    options ...session.Option) (*session.Session, error) {
    if err := key.CheckSessionKey(); err != nil {
        return nil, err
    }

    var session session.Session
    var stateData []byte

    err := p.db.QueryRowContext(ctx, `
        SELECT id, app_name, user_id, state, created_at, updated_at
        FROM sessions WHERE id = $1
    `, key.SessionID).Scan(&session.ID, &session.AppName, &session.UserID, &stateData, &session.CreatedAt,
        &session.UpdatedAt)

    if err != nil {
        return nil, err
    }

    session.State = p.unmarshalState(stateData)

    return &session, nil
}

func (p *PostgreSQLService) Close() error {
    return p.db.Close()
}

func (p *PostgreSQLService) marshalState(state session.StateMap) []byte {
    data, _ := json.Marshal(state)
    return data
}

func (p *PostgreSQLService) unmarshalState(data []byte) session.StateMap {
    var state session.StateMap
    json.Unmarshal(data, &state)
    return state
}

Open Source Components That Can Be Integrated:

  • PostgreSQL session storage
  • MongoDB session storage
  • MySQL session storage
  • Cassandra session storage

Contribution Method:

  • Create components under the corresponding directory (create a subdirectory with the corresponding component name)
  • Contribute directly to https://github.com/trpc-group/trpc-agent-go/session/

6. Memory Ecosystem

Goal: Support multiple memory storage backends, manage user long-term memory and personalized information

Interface Definition: memory.Service

Existing Implementation Reference: InMemory Memory

Implementation Notes:

  • Implement complete Memory lifecycle management (add, update, delete, search, read)
  • Support memory topic classification and search
  • Provide memory tool integration (memory_add, memory_search, etc.)
  • Implement connection pooling and error handling
  • Can reuse storage module clients
  • Support memory limits and cleanup mechanisms
  • Reference InMemory implementation to understand memory management logic

Implementation Example:

package postgresql

import (
    "context"
    "database/sql"
    "encoding/json"
    "fmt"
    "time"

    "trpc.group/trpc-go/trpc-agent-go/memory"
    memorytool "trpc.group/trpc-go/trpc-agent-go/memory/tool"
    "trpc.group/trpc-go/trpc-agent-go/tool"
)

type PostgreSQLMemoryService struct {
    db *sql.DB
    cachedTools map[string]tool.Tool
}

func New(dsn string) (memory.Service, error) {
    db, err := sql.Open("postgres", dsn)
    if err != nil {
        return nil, err
    }

    if err := db.Ping(); err != nil {
        return nil, err
    }

    service := &PostgreSQLMemoryService{
        db: db,
        cachedTools: make(map[string]tool.Tool),
    }

    // Initialize tools.
    service.initTools()

    return service, nil
}

func (p *PostgreSQLMemoryService) AddMemory(ctx context.Context, userKey memory.UserKey, memoryStr string,
    topics []string) error {
    if err := userKey.CheckUserKey(); err != nil {
        return err
    }

    now := time.Now()
    memoryID := p.generateMemoryID(memoryStr)

    // Insert memory into database.
    _, err := p.db.ExecContext(ctx, `
        INSERT INTO memories (id, app_name, user_id, memory, topics, created_at, updated_at)
        VALUES ($1, $2, $3, $4, $5, $6, $7)
    `, memoryID, userKey.AppName, userKey.UserID, memoryStr, p.marshalTopics(topics), now, now)

    return err
}

func (p *PostgreSQLMemoryService) SearchMemories(ctx context.Context, userKey memory.UserKey,
    query string) ([]*memory.Entry, error) {
    if err := userKey.CheckUserKey(); err != nil {
        return nil, err
    }

    // Execute full-text search.
    rows, err := p.db.QueryContext(ctx, `
        SELECT id, app_name, user_id, memory, topics, created_at, updated_at
        FROM memories
        WHERE app_name = $1 AND user_id = $2
        AND (memory ILIKE $3 OR topics::text ILIKE $3)
        ORDER BY updated_at DESC
        LIMIT 10
    `, userKey.AppName, userKey.UserID, "%"+query+"%")

    if err != nil {
        return nil, err
    }
    defer rows.Close()

    var entries []*memory.Entry
    for rows.Next() {
        var entry memory.Entry
        var topicsData []byte

        err := rows.Scan(&entry.ID, &entry.AppName, &entry.UserID, &entry.Memory.Memory, &topicsData, &entry.CreatedAt,
        &entry.UpdatedAt)
    if err != nil {
        return nil, err
    }

        entry.Memory.Topics = p.unmarshalTopics(topicsData)
        entries = append(entries, &entry)
    }

    return entries, nil
}

func (p *PostgreSQLMemoryService) Tools() []tool.Tool {
    var tools []tool.Tool
    for _, t := range p.cachedTools {
        tools = append(tools, t)
    }
    return tools
}

func (p *PostgreSQLMemoryService) initTools() {
    p.cachedTools[memory.AddToolName] = memorytool.NewAddTool(p)
    p.cachedTools[memory.SearchToolName] = memorytool.NewSearchTool(p)
    p.cachedTools[memory.LoadToolName] = memorytool.NewLoadTool(p)
}

func (p *PostgreSQLMemoryService) generateMemoryID(memoryStr string) string {
    // Generate unique memory ID.
    return fmt.Sprintf("mem_%d", time.Now().UnixNano())
}

func (p *PostgreSQLMemoryService) marshalTopics(topics []string) []byte {
    data, _ := json.Marshal(topics)
    return data
}

func (p *PostgreSQLMemoryService) unmarshalTopics(data []byte) []string {
    var topics []string
    json.Unmarshal(data, &topics)
    return topics
}

For quick implementation, you can directly integrate with existing Memory platforms/services (such as mem0). Recommendations:

  • Provide implementation in memory/mem0/, following the memory.Service interface.
  • Reuse existing memory/tool tools (memory_add, memory_search, memory_load, etc.), expose through Tools().
  • Map topics and search according to target service capabilities, maintain lightweight local indexes when necessary to enhance query experience.
  • Optional: Reuse storage module client management for unified authentication, connection and reuse.

Example skeleton (simplified):

package mem0

import (
    "context"
    "net/http"

    "trpc.group/trpc-go/trpc-agent-go/memory"
    memorytool "trpc.group/trpc-go/trpc-agent-go/memory/tool"
    "trpc.group/trpc-go/trpc-agent-go/tool"
)

type Service struct {
    client  *http.Client
    baseURL string
    apiKey  string
    tools   map[string]tool.Tool
}

func New(baseURL, apiKey string) *Service {
    s := &Service{
        client:  &http.Client{},
        baseURL: baseURL,
        apiKey:  apiKey,
        tools:   make(map[string]tool.Tool),
    }
    s.tools[memory.AddToolName] = memorytool.NewAddTool(s)
    s.tools[memory.SearchToolName] = memorytool.NewSearchTool(s)
    s.tools[memory.LoadToolName] = memorytool.NewLoadTool(s)
    return s
}

func (s *Service) Tools() []tool.Tool {
    var ts []tool.Tool
    for _, t := range s.tools {
        ts = append(ts, t)
    }
    return ts
}

func (s *Service) AddMemory(ctx context.Context, key memory.UserKey, m string, topics []string) error {
    if err := key.CheckUserKey(); err != nil {
        return err
    }
    // Call mem0 API to write memory.
    return nil
}

func (s *Service) SearchMemories(ctx context.Context, key memory.UserKey, q string) ([]*memory.Entry, error) {
    if err := key.CheckUserKey(); err != nil {
        return nil, err
    }
    // Call mem0 API to search, and convert to []*memory.Entry.
    return nil, nil
}

// Other interfaces Update/Delete/Clear/Read map according to mem0 capabilities

Implementation points:

  • Configure authentication and rate limiting according to target service guidelines.
  • Return values strictly align with memory.Entry and memory.Memory, use UTC for time fields.
  • Tool declarations should accurately describe input and output for frontend and model understanding.
  • Add README, examples and tests to ensure compatibility with runner, server/debug combinations.

Open Source Components That Can Be Integrated:

  • PostgreSQL memory storage
  • MongoDB memory storage
  • Elasticsearch memory storage
  • Redis memory storage

Contribution Method:

  • Create components under the corresponding directory (create a subdirectory with the corresponding component name)
  • Contribute directly to https://github.com/trpc-group/trpc-agent-go/memory/

7. Observability Ecosystem

Goal: Provide unified observability capabilities based on OpenTelemetry standards, covering Logging, Metrics, Tracing, facilitating ecosystem expansion and replacement.

Core Packages and Interfaces:

  • Logging: trpc-agent-go/log (log.Logger interface and log.Default global logger).
  • Metrics: trpc-agent-go/telemetry/metric (global metric.Meter and metric.Start initialization).
  • Tracing: trpc-agent-go/telemetry/trace (global trace.Tracer and trace.Start initialization).
  • tRPC Integration: trpc/log.go, trpc/telemetry/galileo/.

Logging

  • Interface Definition: log.Logger defines Debug/Info/Warn/Error/Fatal and their *f variant methods, facilitating replacement with any implementation.
  • Default Implementation: log.Default uses zap‘s SugaredLogger by default.
  • Dynamic Level: log.SetLevel(level) supports debug/info/warn/error/fatal.
  • tRPC Integration: trpc/log.go injects tlog.DefaultLogger as log.Default and refreshes with tRPC plugin lifecycle.

Example (using global logger):

package main

import (
    "context"

    "trpc.group/trpc-go/trpc-agent-go/log"
)

func main() {
    log.SetLevel(log.LevelInfo)
    log.Info("app start")
    log.Debugf("ctx: %v", context.Background())
}

Example (in tRPC through configuration to disk/remote):

plugins:
  log:
    default:
      - writer: console
        level: info
      - writer: file
        level: warn
        writer_config:
          log_path: ./app.log

Contribution directions:

  • Adapt any Logger (such as zerolog, logrus): implement log.Logger interface, and set log.Default = yourLogger during initialization.
  • tRPC pluginization: reference trpc/log.go‘s plugin.RegisterSetupHook usage.

Metrics

  • Package: telemetry/metric.
  • Global Object: metric.Meter, default noop, points to OTel Meter after calling metric.Start.
  • Initialization: metric.Start(ctx, metric.WithEndpoint("host:4317")).
  • OTLP Export: uses otlpmetricgrpc, supports environment variable override:
    • OTEL_EXPORTER_OTLP_METRICS_ENDPOINT.
    • OTEL_EXPORTER_OTLP_ENDPOINT (fallback).
  • Resource Identification: automatically fills service.namespace/name/version.

Example (start metrics and report Counter):

package main

import (
    "context"

    "go.opentelemetry.io/otel/metric"

    "trpc.group/trpc-go/trpc-agent-go/telemetry/metric" // alias ametric
    ametric "trpc.group/trpc-go/trpc-agent-go/telemetry/metric"
)

func main() {
    mp, err := ametric.NewMeterProvider(
        context.Background(),
        ametric.WithEndpoint("localhost:4318"),
        ametric.WithProtocol("http"),
    )
    if err != nil {
        log.Fatalf("Failed to create meter provider: %v", err)
    }
    defer mp.Shutdown(context.Background())
    ametric.InitMeterProvider(mp)
    meter := mp.Meter("trpc_agent_go.app")
        counter, _ := meter.Int64Counter(
        "requests_total",
        metric.WithDescription("total requests"),
    )
    counter.Add(context.Background(), 1)
}

Contribution directions:

  • Exporter ecosystem: encapsulate more OTel Exporter convenient startup methods (such as Prometheus pull/OTLP http).
  • Metrics library: agree on common metric naming and label conventions, provide helper methods.

Tracing

  • Package: telemetry/trace.
  • Global Object: trace.Tracer, default noop, points to OTel Tracer after calling trace.Start.
  • Initialization: trace.Start(ctx, trace.WithEndpoint("host:4317")).
  • OTLP Export: uses otlptracegrpc, supports environment variable override:
    • OTEL_EXPORTER_OTLP_TRACES_ENDPOINT.
    • OTEL_EXPORTER_OTLP_ENDPOINT (fallback).
  • Propagator: enables TraceContext by default.

Example (start tracing and create Span):

package main

import (
    "context"

    "go.opentelemetry.io/otel/trace"
    atrace "trpc.group/trpc-go/trpc-agent-go/telemetry/trace"
)

func main() {
    clean, _ := atrace.Start(context.Background(),
        atrace.WithEndpoint("localhost:4317"),
    )
    defer clean()

    ctx, span := atrace.Tracer.Start(context.Background(), "example",
        trace.WithAttributes(),
    )
    _ = ctx
    span.End()
}

Contribution directions:

  • Exporter ecosystem: encapsulate Zipkin, Jaeger (direct push) and other startup methods.
  • Span conventions: agree on common Span names/attribute keys, provide helpers (can be placed in telemetry/).

8. API Service Ecosystem

Goal: Provide unified, extensible API service encapsulation for frontend Chat interfaces (such as ADK Web, AG-UI, Agent UI), covering session management, conversation sending, streaming transmission, tool calls, observability and authentication capabilities, and align with various UI protocols for plug-and-play.

Existing Implementation Reference:

  • ADK Web Compatible HTTP Service : server/debug.
    • Endpoints (implemented):
    • GET /list-apps: List available Agent applications.
    • GET /apps/{appName}/users/{userId}/sessions: List user sessions.
    • POST /apps/{appName}/users/{userId}/sessions: Create session.
    • GET /apps/{appName}/users/{userId}/sessions/{sessionId}: Query session.
    • POST /run: Non-streaming conversation inference, returns aggregated event list.
    • POST /run_sse: SSE streaming inference, returns token-level event stream.
    • GET /debug/trace/{event_id}: Query Trace attributes by event.
    • GET /debug/trace/session/{session_id}: Query Trace list by Session.
    • Features: Built-in CORS, pluggable session storage (default In-Memory), integrated with runner.Runner, observability instrumentation (exports key Spans).
    • Scope note: Designed for debugging with ADK Web. It constructs runners internally from Agents and enforces a single session.Service across its session APIs and runners. Not recommended for production.
  • A2A Server : server/a2a.
    • Service encapsulation for A2A protocol, built-in AuthProvider and task orchestration, suitable for platform-to-Agent integration scenarios.

Alignment with Frontend Protocols:

  • ADK Web : Already aligned with request/response and event Schema, see server/debug/internal/schema.
  • AG-UI : Reference https://github.com/ag-ui-protocol/ag-ui.
    • Required capabilities:
    • Session list/create/query.
    • Text conversation and SSE streaming increment; support tool calls and function response fragment display.
    • State/usage metadata, error expression alignment.
    • Rich media support (InlineData) for files/images and server-side storage integration.
    • Authentication (API Key, JWT, Cookie session) and CORS.
    • Recommend providing implementation in server/agui/, reuse common model and event mapping tools, complete protocol layer adaptation in Handler.
  • Agent UI (agno) : Reference https://docs.agno.com/agent-ui/introduction.
    • Focus: SSE/WebSocket streaming, Tool call streaming UI feedback, session/artifact persistence.

Key Design Points:

  • Schema Mapping :
    • Input: Map UI’s Content/Part to internal model.Message.
    • Output events: Map internal event.Event to UI-expected envelope/parts, structure tool calls and tool responses to avoid duplicate text display.
  • Streaming Transmission :
    • SSE already implemented in server/debug, prioritize reuse; WebSocket can be ecosystem extension.
    • Non-streaming endpoints need to aggregate final messages and tool responses according to UI expectations.
  • Session Storage :
    • Inject specific implementation through runner.WithSessionService, reuse session module.
    • For server/debug, set a single backend via debug.WithSessionService(...); it overrides any runner‑level session service to ensure consistency.
  • Observability :
    • Reuse telemetry/trace and telemetry/metric. server/debug already demonstrates how to export key Spans and event attributes for UI-side debugging and positioning.
  • Authentication and Security :
    • Support API Key/JWT/custom Header; add rate limiting and cross-domain control for sensitive endpoints.
  • Open Specifications :
    • Recommend attaching openapi.json/README.md to each server/* submodule for frontend/integration party integration.

Minimal Example (reuse ADK Web compatible service):

package main

import (
    "net/http"

    "trpc.group/trpc-go/trpc-agent-go/agent/llmagent"
    debugsrv "trpc.group/trpc-go/trpc-agent-go/server/debug"
)

func main() {
    // 1. Register Agent.
    ag := llmagent.New("assistant")
    s := debugsrv.New(map[string]agent.Agent{
        ag.Info().Name: ag,
    })
    // 2. Expose HTTP Handler.
    _ = http.ListenAndServe(":8080", s.Handler())
}

AG-UI Adaptation Suggestion (skeleton):

package agui

import (
    "net/http"

    "github.com/gorilla/mux"
    "trpc.group/trpc-go/trpc-agent-go/agent"
    "trpc.group/trpc-go/trpc-agent-go/runner"
)

type Server struct {
    router *mux.Router
    ag     agent.Agent
    run    runner.Runner
}

func New(ag agent.Agent, opts ...runner.Option) *Server {
    r := runner.NewRunner(ag.Info().Name, ag, opts...)
    s := &Server{router: mux.NewRouter(), ag: ag, run: r}
    s.routes()
    return s
}

func (s *Server) Handler() http.Handler { return s.router }

func (s *Server) routes() {
    // GET /sessions, POST /sessions, GET /sessions/{id}
    // POST /chat (non-stream), POST /chat/stream (SSE)
}

Ecosystem Directions and Contribution Standards:

  • Target UI/Protocols :
    • AG-UI : Provide HTTP + SSE adaptation in server/agui/, with examples and openapi.json.
  • Agent UI (agno) : Provide HTTP + SSE / WebSocket adaptation in server/agentui/.
    • WebSocket/Bidi Streaming : Align with ADK run_live, provide real-time audio/video channels (depends on model-side support).
  • Implementation Requirements :
    • Clear event Schema, complete mapping, ensure tool calls/responses have good experience in UI.
    • Support pluggable session storage, default In-Memory, recommend supporting Redis/MySQL etc.
    • Built-in CORS, authentication middleware (API Key/JWT), expose health check endpoints.
    • Observability: integrate telemetry, provide minimal Trace and Metric examples.
    • Documentation: README, OpenAPI, end-to-end examples (include simple frontend or curl scripts).

Link references:

  • server/debug (ADK Web compatible) and its openapi.json.
  • server/a2a (A2A protocol encapsulation).

9. Planner Ecosystem

Goal: Provide diverse planners to adapt to different models and workflows, including built-in thinking capability adaptation and explicit planning (ReAct/Reflection, etc.).

Interface Definition: planner.Planner.

  • BuildPlanningInstruction(ctx, invocation, llmRequest) string: Build or inject system prompts and request configurations for planning.
  • ProcessPlanningResponse(ctx, invocation, response) *model.Response: Post-process model responses for planning (optional).

Existing Implementation Reference:

  • planner/builtin: Adapt O series, Claude, Gemini and other models with reasoning parameters, through configuring ReasoningEffort, ThinkingEnabled, ThinkingTokens.
  • planner/react: Provide explicit planning instructions and response post-processing, agree on /*PLANNING*/, /*ACTION*/, /*REASONING*/, /*FINAL_ANSWER*/ and other tags.

Ecosystem Directions:

  • Reflection Planner: Self-reflective correction and multi-round re-planning.
  • LangGraph style Planner: Align with Pregel parallel and checkpoint mechanisms.
  • Tool-first Planner: Selection and constraints for Tool-First processes.

Integration Example (skeleton):

package myplanner

import (
    "context"

    "trpc.group/trpc-go/trpc-agent-go/agent"
    "trpc.group/trpc-go/trpc-agent-go/model"
    "trpc.group/trpc-go/trpc-agent-go/planner"
)

type Planner struct{}

var _ planner.Planner = (*Planner)(nil)

func New() *Planner { return &Planner{} }

func (p *Planner) BuildPlanningInstruction(ctx context.Context, inv *agent.Invocation, req *model.Request) string {
    // Can inject custom parameters into req, and return system prompt string.
    return "You must plan before action."
}

func (p *Planner) ProcessPlanningResponse(ctx context.Context, inv *agent.Invocation,
    rsp *model.Response) *model.Response {
    if rsp == nil {
        return nil
    }
    // Can do structured segmentation or tool call correction on rsp.
    return nil
}

Combination and Usage: Inject Planner when creating agent/llmagent, or select different Planner strategies at runner layer as needed, combine with Tool and Session management for end-to-end implementation.

Contribution Suggestions:

  • Provide implementation and README, test cases, examples in planner/<name>/.
  • Combine with docs/overall-introduction.md‘s Observability and server/debug endpoints to provide end-to-end examples for frontend UI demonstration.
  • Follow goimports and error message style, add periods at end of comments, code line breaks around 80 columns.

Component Relationship Description

Relationship between Storage, Session, and Memory

These three components have different responsibilities and relationships in the architecture:

1. Storage (Storage Layer)

  • Responsibility: Provide unified storage client management, provide infrastructure support for Session and Memory
  • Function: Register, manage and obtain clients for various storage backends (Redis, PostgreSQL, MongoDB, etc.)
  • Characteristics: As an infrastructure component, can be shared and used by Session and Memory components

2. Session (Session Layer)

  • Responsibility: Manage user session state and events
  • Function: Create, get, delete sessions, manage session state, record session events
  • Dependency: Can reuse Storage module clients
  • Data Characteristics: Temporary data, can be cleaned up after session ends

3. Memory (Memory Layer)

  • Responsibility: Manage user long-term memory and personalized information
  • Function: Add, search, update, delete user memories, provide memory tools
  • Dependency: Can reuse Storage module clients
  • Data Characteristics: Persistent data, maintained across sessions

Relationship Diagram:

┌─────────────────┐
│   Application   │
└─────────┬───────┘
          │
┌─────────▼───────┐    ┌─────────────────┐
│   Session       │    │   Memory        │
│   Service       │    │   Service       │
└─────────┬───────┘    └─────────┬───────┘
          │                      │
          └──────────┬───────────┘
                     │
          ┌─────────▼───────┐
          │   Storage       │
          │   Client        │
          │   Management    │
          └─────────┬───────┘
                     │
          ┌─────────▼───────┐
          │   Storage       │
          │   Backends      │
          │   (Redis, DB)   │
          └─────────────────┘

Usage Example:

// 1. Register storage client.
storage.RegisterRedisInstance("default", storage.WithClientBuilderURL("redis://localhost:6379"))

// 2. Session service uses storage client.
sessionService, err := session.NewRedisService(
    session.WithRedisInstance("default"),
)

// 3. Memory service uses storage client.
memoryService, err := memory.NewRedisService(
    memory.WithRedisInstance("default"),
)

// 4. Use in application.
session, err := sessionService.CreateSession(ctx, sessionKey, state)
memory, err := memoryService.AddMemory(ctx, userKey, "User likes coffee", []string{"preferences"})

Contribution Guidance

Contribution Guidance

Suitable Components:

  • Various third-party service and tool integrations
  • Open source component adapters
  • Standard protocol support
  • Framework functionality extensions

Contribution Process:

  1. Fork https://github.com/trpc-group/trpc-agent-go
  2. Create components under the corresponding module root directory (such as model/somemodel, tool/sometool, agent/someagent)
  3. Implement corresponding interfaces
  4. Write tests and documentation
  5. Submit Pull Request

Directory Structure Example:

model/gemini/
├── model.go
├── config.go
├── examples/
├── README.md
└── gemini_test.go

Summary

Ecosystem development is an important direction for tRPC-Agent-Go development. By implementing standard interfaces, various third-party services and tools can be easily integrated to expand the framework’s capabilities.

Key Contribution Points:

  • Reference existing implementations to understand interface usage
  • Choose appropriate contribution paths based on component types
  • Follow unified interface specifications and code standards
  • Provide complete test cases and documentation

Determining Contribution Location:

  • All components are contributed directly to the corresponding module directory on GitHub

Storage, Session, Memory Component Characteristics:

  • Storage: Provides unified client management, can be shared by Session and Memory
  • Session: Manages temporary session data, can reuse Storage clients
  • Memory: Manages persistent memory data, can reuse Storage clients
  • The three components are decoupled through interfaces, supporting independent implementation and combined usage