Refactor console agents for improved modularity and configuration

Consolidated and refactored console agent logic into a streamlined structure with better configuration capabilities via `ConsoleConfig`. Improved code reusability and readability by converting nested structures and functions, and introduced more modular execution types for handling commands and history tracking.
This commit is contained in:
Steve Dudenhoeffer 2025-05-03 22:06:32 -04:00
parent 4c5922d571
commit d2b9eb350e
3 changed files with 283 additions and 328 deletions

View File

@ -2,24 +2,284 @@ package agents
import (
"context"
"fmt"
"log/slog"
"os"
"path/filepath"
"strings"
gollm "gitea.stevedudenhoeffer.com/steve/go-llm"
"github.com/docker/docker/api/types/container"
"github.com/docker/docker/api/types/mount"
"github.com/docker/docker/client"
"gitea.stevedudenhoeffer.com/steve/answer/pkg/agents/console"
"gitea.stevedudenhoeffer.com/steve/answer/pkg/agents/shared"
)
// Console will construct a console.Agent, execute the agent, and return the knowledge gained, the output directory,
// and any possible error.
func (a Agent) Console(ctx context.Context, questions []string) (Knowledge, string, error) {
con := console.Agent{
Agent: a,
Model: a.model,
ContextualInformation: a.contextualInformation,
MaxCommands: 50,
type ConsoleConfig struct {
// MaxCommands is how many total commands this console can run.
MaxCommands int `json:"max_commands"`
OnCommandStart func(ctx context.Context, command string) error
OnCommandDone func(ctx context.Context, command string, output string, err error) error
}
res, err := con.Answer(ctx, questions)
func (a Agent) Console(ctx context.Context, questions []string, cfg ConsoleConfig) (Knowledge, string, error) {
if cfg.MaxCommands <= 0 {
cfg.MaxCommands = 10000
}
var resKnowledge = Knowledge{
OriginalQuestions: questions,
RemainingQuestions: questions,
}
// create a temporary scratch directory
dir, err := os.MkdirTemp("", "console-")
if err != nil {
return Knowledge{}, "", err
return resKnowledge, "", err
}
return res.Knowledge, res.Directory, nil
var resDirectory = dir
cl, err := client.NewClientWithOpts(client.FromEnv)
if err != nil {
return resKnowledge, resDirectory, err
}
defer cl.Close()
mounts := []mount.Mount{
{
Type: mount.TypeBind,
Source: dir,
Target: "/home/user",
},
}
c, err := console.CreateContainer(ctx, cl, console.ContainerConfig{
Config: &container.Config{
Image: "ubuntu:latest",
Cmd: []string{"tail", "-f", "/dev/null"},
Tty: true,
WorkingDir: "/home/user",
},
HostConfig: &container.HostConfig{
AutoRemove: true,
Mounts: mounts,
},
Name: filepath.Base(dir),
})
if err != nil {
return resKnowledge, resDirectory, fmt.Errorf("failed to create container: %w", err)
}
defer func() {
_ = c.Close(ctx)
}()
slog.Info("starting container", "dir", dir, "container", fmt.Sprintf("%+v", c))
err = c.Start(ctx)
if err != nil {
return resKnowledge, resDirectory, fmt.Errorf("failed to start container: %w", err)
}
// Run the model
var history console.Executions
var keepGoing = true
opwd, epwd := c.Execute(ctx, "ls -al /home")
fmt.Println(opwd)
slog.Info("pwd", "pwd", opwd, "epwd", epwd)
tools := map[string]gollm.Function{
"exit": gollm.NewFunction(
"exit",
"exit the container",
func(ctx *gollm.Context, args struct {
RemainingQuestions []string `description:"any remaining questions that remain unanswered"`
}) (any, error) {
keepGoing = false
return "exiting", nil
}),
"write": gollm.NewFunction(
"write",
"write a file in the /root directory",
func(ctx *gollm.Context, args struct {
Filename string `description:"The name of the file to write"`
Content string `description:"The content of the file to write"`
}) (any, error) {
target, err := console.SafeJoinPath(dir, args.Filename)
if err != nil {
return "", err
}
f, err := os.Create(target)
if err != nil {
return "", err
}
defer f.Close()
_, err = f.WriteString(args.Content)
if err != nil {
return "", err
}
return "wrote file", nil
}),
"read": gollm.NewFunction(
"read",
"read a file in the /root directory",
func(ctx *gollm.Context, args struct {
Filename string `description:"The name of the file to read"`
}) (any, error) {
target, err := console.SafeJoinPath(dir, args.Filename)
if err != nil {
return "", err
}
b, err := os.ReadFile(target)
if err != nil {
return "", err
}
return string(b), nil
}),
"execute": gollm.NewFunction(
"execute",
"execute a command in the container",
func(ctx *gollm.Context, args struct {
Command string `description:"The command to execute"`
}) (any, error) {
if len(history) >= cfg.MaxCommands {
return "too many commands", nil
}
if cfg.OnCommandStart != nil {
err := cfg.OnCommandStart(ctx, args.Command)
if err != nil {
return "", err
}
}
var res string
// if the command starts with sudo then we need to use the sudo function
if strings.HasPrefix(args.Command, "sudo ") {
res, err = c.Sudo(ctx, args.Command[5:])
} else {
res, err = c.Execute(ctx, args.Command)
}
if cfg.OnCommandDone != nil {
err = cfg.OnCommandDone(ctx, args.Command, res, nil)
if err != nil {
return "", err
}
}
history = append(history, console.Execution{
Command: args.Command,
Output: res,
})
return res, nil
}),
"sudo": gollm.NewFunction(
"sudo",
"execute a command in the container",
func(ctx *gollm.Context, args struct {
Command string `description:"The command to execute"`
}) (any, error) {
if len(history) >= cfg.MaxCommands {
return "too many commands", nil
}
if cfg.OnCommandStart != nil {
err := cfg.OnCommandStart(ctx, args.Command)
if err != nil {
return "", err
}
}
res, err := c.Sudo(ctx, args.Command)
if cfg.OnCommandDone != nil {
err = cfg.OnCommandDone(ctx, args.Command, res, nil)
if err != nil {
return "", err
}
}
if err != nil {
res = "error executing: " + err.Error()
}
history = append(history, console.Execution{
Command: "sudo " + args.Command,
Output: res,
})
return res, nil
}),
}
for i := 0; i < cfg.MaxCommands && len(history) < cfg.MaxCommands && keepGoing; i++ {
systemPrompt := `You are now in a shell in a container of the ubuntu:latest image to answer a question asked by the user, it is very basic install of ubuntu, simple things (like python) are not preinstalled but can be installed via apt. You will be run multiple times and gain knowledge throughout the process.`
if len(history) < cfg.MaxCommands {
systemPrompt += `You can run any command you like to get to the needed results.`
}
systemPrompt += `Alternatively, you can use the tool "write" to write a file in the home directory, and also the tool "read" to read a file in the home directory.
When you are done, please use "exit" to exit the container.
Respond with any number of commands to answer the question, they will be executed in order.`
var toolbox []gollm.Function
// add unrestricted tools
toolbox = append(toolbox, tools["exit"], tools["write"], tools["read"])
if len(history) < cfg.MaxCommands {
toolbox = append(toolbox, tools["execute"], tools["sudo"])
}
kw := shared.KnowledgeWorker{
Model: a.model,
ToolBox: gollm.NewToolBox(toolbox...),
ContextualInformation: a.contextualInformation,
OnNewFunction: func(ctx context.Context, funcName string, args string) (any, error) {
slog.Info("new function called", "function name", funcName, "args", args)
return nil, nil
},
}
slog.Info("answering question", "question", questions[0])
r, err := kw.Answer(ctx, &resKnowledge, systemPrompt, "", "", history.ToGeneralButLastMessageHistory(), func(res gollm.ToolCallResponse) {
})
if err != nil {
return resKnowledge, resDirectory, fmt.Errorf("error answering question: %w", err)
}
if len(r.Knowledge) > 0 {
slog.Info("answered question and learned", "knowledge", r.Knowledge)
} else {
slog.Info("answered question and learned nothing")
}
resKnowledge, err = a.KnowledgeIntegrate(ctx, resKnowledge, r)
if err != nil {
return resKnowledge, resDirectory, fmt.Errorf("error integrating knowledge: %w", err)
}
slog.Info("knowledge integrated", "question", questions[0], "knowledge", resKnowledge)
}
return resKnowledge, resDirectory, nil
}

View File

@ -1,305 +0,0 @@
package console
import (
"context"
"fmt"
"log/slog"
"os"
"path/filepath"
"strings"
"github.com/docker/docker/api/types/container"
"github.com/docker/docker/api/types/mount"
"github.com/docker/docker/client"
"gitea.stevedudenhoeffer.com/steve/answer/pkg/agents"
"gitea.stevedudenhoeffer.com/steve/answer/pkg/agents/shared"
gollm "gitea.stevedudenhoeffer.com/steve/go-llm"
)
type Agent struct {
agents.Agent
// Model is the chat completion model to use
Model gollm.ChatCompletion
OnLoopComplete func(ctx context.Context, knowledge agents.Knowledge) error
OnCommandStart func(ctx context.Context, command string) error
OnCommandDone func(ctx context.Context, command string, output string, err error) error
OnDone func(ctx context.Context, knowledge agents.Knowledge) error
ContextualInformation []string
MaxCommands int
}
type Response struct {
Knowledge agents.Knowledge
Directory string
}
// Answer will give the model access to an ubuntu console with python and pip installed, and then ask the model to
// do what is necessary to answer the question.
func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) {
var res Response
a.Agent = agents.NewAgent(a.Model, gollm.NewToolBox()).WithMaxCalls(200)
if a.MaxCommands <= 0 {
a.MaxCommands = 10000
}
res.Knowledge = agents.Knowledge{
OriginalQuestions: questions,
RemainingQuestions: questions,
}
// create a temporary scratch directory
dir, err := os.MkdirTemp("", "console-")
if err != nil {
return res, err
}
res.Directory = dir
cl, err := client.NewClientWithOpts(client.FromEnv)
if err != nil {
return res, err
}
defer cl.Close()
mounts := []mount.Mount{
{
Type: mount.TypeBind,
Source: dir,
Target: "/home/user",
},
}
c, err := CreateContainer(ctx, cl, ContainerConfig{
Config: &container.Config{
Image: "ubuntu:latest",
Cmd: []string{"tail", "-f", "/dev/null"},
Tty: true,
WorkingDir: "/home/user",
},
HostConfig: &container.HostConfig{
AutoRemove: true,
Mounts: mounts,
},
Name: filepath.Base(dir),
})
if err != nil {
return res, fmt.Errorf("failed to create container: %w", err)
}
defer func() {
_ = c.Close(ctx)
}()
slog.Info("starting container", "dir", dir, "container", fmt.Sprintf("%+v", c))
err = c.Start(ctx)
if err != nil {
return res, fmt.Errorf("failed to start container: %w", err)
}
// Run the model
var history executions
var keepGoing = true
opwd, epwd := c.Execute(ctx, "ls -al /home")
fmt.Println(opwd)
slog.Info("pwd", "pwd", opwd, "epwd", epwd)
tools := map[string]gollm.Function{
"exit": gollm.NewFunction(
"exit",
"exit the container",
func(ctx *gollm.Context, args struct {
RemainingQuestions []string `description:"any remaining questions that remain unanswered"`
}) (any, error) {
keepGoing = false
return "exiting", nil
}),
"write": gollm.NewFunction(
"write",
"write a file in the /root directory",
func(ctx *gollm.Context, args struct {
Filename string `description:"The name of the file to write"`
Content string `description:"The content of the file to write"`
}) (any, error) {
target, err := SafeJoinPath(dir, args.Filename)
if err != nil {
return "", err
}
f, err := os.Create(target)
if err != nil {
return "", err
}
defer f.Close()
_, err = f.WriteString(args.Content)
if err != nil {
return "", err
}
return "wrote file", nil
}),
"read": gollm.NewFunction(
"read",
"read a file in the /root directory",
func(ctx *gollm.Context, args struct {
Filename string `description:"The name of the file to read"`
}) (any, error) {
target, err := SafeJoinPath(dir, args.Filename)
if err != nil {
return "", err
}
b, err := os.ReadFile(target)
if err != nil {
return "", err
}
return string(b), nil
}),
"execute": gollm.NewFunction(
"execute",
"execute a command in the container",
func(ctx *gollm.Context, args struct {
Command string `description:"The command to execute"`
}) (any, error) {
if len(history) >= a.MaxCommands {
return "too many commands", nil
}
if a.OnCommandStart != nil {
err := a.OnCommandStart(ctx, args.Command)
if err != nil {
return "", err
}
}
var res string
// if the command starts with sudo then we need to use the sudo function
if strings.HasPrefix(args.Command, "sudo ") {
res, err = c.Sudo(ctx, args.Command[5:])
} else {
res, err = c.Execute(ctx, args.Command)
}
if a.OnCommandDone != nil {
err = a.OnCommandDone(ctx, args.Command, res, nil)
if err != nil {
return "", err
}
}
history = append(history, execution{
Command: args.Command,
Output: res,
})
return res, nil
}),
"sudo": gollm.NewFunction(
"sudo",
"execute a command in the container",
func(ctx *gollm.Context, args struct {
Command string `description:"The command to execute"`
}) (any, error) {
if len(history) >= a.MaxCommands {
return "too many commands", nil
}
if a.OnCommandStart != nil {
err := a.OnCommandStart(ctx, args.Command)
if err != nil {
return "", err
}
}
res, err := c.Sudo(ctx, args.Command)
if a.OnCommandDone != nil {
err = a.OnCommandDone(ctx, args.Command, res, nil)
if err != nil {
return "", err
}
}
if err != nil {
res = "error executing: " + err.Error()
}
history = append(history, execution{
Command: "sudo " + args.Command,
Output: res,
})
return res, nil
}),
}
for i := 0; i < a.MaxCommands && len(history) < a.MaxCommands && keepGoing; i++ {
systemPrompt := `You are now in a shell in a container of the ubuntu:latest image to answer a question asked by the user, it is very basic install of ubuntu, simple things (like python) are not preinstalled but can be installed via apt. You will be run multiple times and gain knowledge throughout the process.`
if len(history) < a.MaxCommands {
systemPrompt += `You can run any command you like to get to the needed results.`
}
systemPrompt += `Alternatively, you can use the tool "write" to write a file in the home directory, and also the tool "read" to read a file in the home directory.
When you are done, please use "exit" to exit the container.
Respond with any number of commands to answer the question, they will be executed in order.`
var toolbox []gollm.Function
// add unrestricted tools
toolbox = append(toolbox, tools["exit"], tools["write"], tools["read"])
if len(history) < a.MaxCommands {
toolbox = append(toolbox, tools["execute"], tools["sudo"])
}
kw := shared.KnowledgeWorker{
Model: a.Model,
ToolBox: gollm.NewToolBox(toolbox...),
ContextualInformation: a.ContextualInformation,
OnNewFunction: func(ctx context.Context, funcName string, args string) (any, error) {
slog.Info("new function called", "function name", funcName, "args", args)
return nil, nil
},
}
slog.Info("answering question", "question", questions[0])
r, err := kw.Answer(ctx, &res.Knowledge, systemPrompt, "", "", history.ToGeneralButLastMessageHistory(), func(res gollm.ToolCallResponse) {
})
if err != nil {
return res, fmt.Errorf("error answering question: %w", err)
}
if len(r.Knowledge) > 0 {
slog.Info("answered question and learned", "knowledge", r.Knowledge)
} else {
slog.Info("answered question and learned nothing")
}
res.Knowledge, err = a.KnowledgeIntegrate(ctx, res.Knowledge, r)
if err != nil {
return res, fmt.Errorf("error integrating knowledge: %w", err)
}
slog.Info("knowledge integrated", "question", questions[0], "knowledge", res.Knowledge)
}
return res, nil
}

View File

@ -6,7 +6,7 @@ import (
gollm "gitea.stevedudenhoeffer.com/steve/go-llm"
)
type execution struct {
type Execution struct {
Command string
Output string
WhatILearned []string
@ -16,7 +16,7 @@ type execution struct {
const kMaxLenCommandSummary = 200
const kMaxLenCommandOutputSummary = 200
func (e execution) ToGeneralMessageHistory() gollm.Message {
func (e Execution) ToGeneralMessageHistory() gollm.Message {
if len(e.Command) > kMaxLenCommandSummary {
e.Command = e.Command[:kMaxLenCommandSummary] + "... (truncated)"
}
@ -33,7 +33,7 @@ func (e execution) ToGeneralMessageHistory() gollm.Message {
}
}
func (e execution) ToDetailedMessageHistory() gollm.Message {
func (e Execution) ToDetailedMessageHistory() gollm.Message {
prompt := "$ "
if strings.HasPrefix(e.Command, "sudo ") {
prompt = "# "
@ -60,9 +60,9 @@ func (e execution) ToDetailedMessageHistory() gollm.Message {
}
}
type executions []execution
type Executions []Execution
func (e executions) ToGeneralMessageHistory() []gollm.Message {
func (e Executions) ToGeneralMessageHistory() []gollm.Message {
var messages []gollm.Message
for _, v := range e {
@ -72,7 +72,7 @@ func (e executions) ToGeneralMessageHistory() []gollm.Message {
return messages
}
func (e executions) ToGeneralButLastMessageHistory() []gollm.Message {
func (e Executions) ToGeneralButLastMessageHistory() []gollm.Message {
var messages []gollm.Message
for i, v := range e {
@ -86,7 +86,7 @@ func (e executions) ToGeneralButLastMessageHistory() []gollm.Message {
return messages
}
func (e executions) ToDetailedMessageHistory() []gollm.Message {
func (e Executions) ToDetailedMessageHistory() []gollm.Message {
var messages []gollm.Message
for _, v := range e {