diff --git a/pkg/agents/console-new/agent.go b/pkg/agents/console-new/agent.go index 304c367..3b0960f 100644 --- a/pkg/agents/console-new/agent.go +++ b/pkg/agents/console-new/agent.go @@ -37,7 +37,8 @@ type Agent struct { type Response struct { Knowledge agents.Knowledge - Directory string + DataDir string + OutputDir string } // Answer will give the model access to an ubuntu console with python and pip installed, and then ask the model to @@ -48,7 +49,7 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) a.Agent = agents.NewAgent(a.Model, gollm.NewToolBox()).WithMaxCalls(200) if a.MaxCommands <= 0 { - a.MaxCommands = 10000 + a.MaxCommands = 20 // Default to 20 commands as per requirements } res.Knowledge = agents.Knowledge{ @@ -56,16 +57,25 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) RemainingQuestions: questions, } - // create a temporary scratch directory - dir, err := os.MkdirTemp("", "console-") + // create temporary directories for data and output + dataDir, err := os.MkdirTemp("", "console-data-") if err != nil { return res, err } - res.Directory = dir + outputDir, err := os.MkdirTemp("", "console-output-") + if err != nil { + os.RemoveAll(dataDir) + return res, err + } + + res.DataDir = dataDir + res.OutputDir = outputDir cl, err := client.NewClientWithOpts(client.FromEnv) if err != nil { + os.RemoveAll(dataDir) + os.RemoveAll(outputDir) return res, err } defer cl.Close() @@ -73,8 +83,13 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) mounts := []mount.Mount{ { Type: mount.TypeBind, - Source: dir, - Target: "/home/user", + Source: dataDir, + Target: "/data", + }, + { + Type: mount.TypeBind, + Source: outputDir, + Target: "/output", }, } @@ -83,25 +98,29 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) Image: "ubuntu:latest", Cmd: []string{"tail", "-f", "/dev/null"}, Tty: true, - WorkingDir: "/home/user", + WorkingDir: "/data", }, HostConfig: &container.HostConfig{ AutoRemove: true, Mounts: mounts, }, - Name: filepath.Base(dir), + Name: filepath.Base(dataDir), }) if err != nil { + os.RemoveAll(dataDir) + os.RemoveAll(outputDir) 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)) + slog.Info("starting container", "dataDir", dataDir, "outputDir", outputDir, "container", fmt.Sprintf("%+v", c)) err = c.Start(ctx) if err != nil { + os.RemoveAll(dataDir) + os.RemoveAll(outputDir) return res, fmt.Errorf("failed to start container: %w", err) } @@ -110,10 +129,15 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) var history executions var keepGoing = true - opwd, epwd := c.Execute(ctx, "ls -al /home") - - fmt.Println(opwd) - slog.Info("pwd", "pwd", opwd, "epwd", epwd) + // Initial setup - install basic tools + setupCmd, setupErr := c.Sudo(ctx, "apt-get update && apt-get install -y curl wget git python3 python3-pip") + if setupErr == nil { + history = append(history, execution{ + Command: "sudo apt-get update && apt-get install -y curl wget git python3 python3-pip", + Output: setupCmd, + WhatILearned: []string{"Basic tools installed: curl, wget, git, python3, pip"}, + }) + } tools := map[string]gollm.Function{ "exit": gollm.NewFunction( @@ -126,18 +150,24 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) return "exiting", nil }), - "write": gollm.NewFunction( - "write", - "write a file in the /root directory", + "write_data": gollm.NewFunction( + "write_data", + "write a file in the /data 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) + target, err := SafeJoinPath(dataDir, args.Filename) if err != nil { return "", err } + // Ensure directory exists + dir := filepath.Dir(target) + if err := os.MkdirAll(dir, 0755); err != nil { + return "", err + } + f, err := os.Create(target) if err != nil { return "", err @@ -149,16 +179,67 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) return "", err } - return "wrote file", nil + return "wrote file to /data/" + args.Filename, nil }), - "read": gollm.NewFunction( - "read", - "read a file in the /root directory", + "write_output": gollm.NewFunction( + "write_output", + "write a file in the /output directory (files here will be available after the agent completes)", + 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(outputDir, args.Filename) + if err != nil { + return "", err + } + + // Ensure directory exists + dir := filepath.Dir(target) + if err := os.MkdirAll(dir, 0755); 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 to /output/" + args.Filename, nil + }), + + "read_data": gollm.NewFunction( + "read_data", + "read a file from the /data 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) + target, err := SafeJoinPath(dataDir, args.Filename) + if err != nil { + return "", err + } + + b, err := os.ReadFile(target) + if err != nil { + return "", err + } + + return string(b), nil + }), + + "read_output": gollm.NewFunction( + "read_output", + "read a file from the /output directory", + func(ctx *gollm.Context, args struct { + Filename string `description:"The name of the file to read"` + }) (any, error) { + target, err := SafeJoinPath(outputDir, args.Filename) if err != nil { return "", err } @@ -175,10 +256,12 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) "execute", "execute a command in the container", func(ctx *gollm.Context, args struct { - Command string `description:"The command to execute"` + Command string `description:"The command to execute"` + Learn []string `description:"What you learned from this command (optional)"` + ToLearn []string `description:"What you still need to learn (optional)"` }) (any, error) { if len(history) >= a.MaxCommands { - return "too many commands", nil + return "Command limit reached. You've used all " + fmt.Sprintf("%d", a.MaxCommands) + " allowed commands.", nil } if a.OnCommandStart != nil { @@ -204,71 +287,79 @@ func (a Agent) Answer(ctx context.Context, questions []string) (Response, error) } history = append(history, execution{ - Command: args.Command, - Output: res, + Command: args.Command, + Output: res, + WhatILearned: args.Learn, + WhatIStillNeedToLearn: args.ToLearn, }) 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 + "summarize_knowledge": gollm.NewFunction( + "summarize_knowledge", + "summarize what you've learned so far", + func(ctx *gollm.Context, args struct{}) (any, error) { + var learned []string + var toLearn []string + + for _, exec := range history { + learned = append(learned, exec.WhatILearned...) + toLearn = append(toLearn, exec.WhatIStillNeedToLearn...) } - if a.OnCommandStart != nil { - err := a.OnCommandStart(ctx, args.Command) - if err != nil { - return "", err - } + summary := "Knowledge Summary:\n" + if len(learned) > 0 { + summary += "What I've learned:\n- " + strings.Join(learned, "\n- ") + "\n\n" + } else { + summary += "I haven't learned anything specific yet.\n\n" } - 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 len(toLearn) > 0 { + summary += "What I still need to learn:\n- " + strings.Join(toLearn, "\n- ") + } else { + summary += "I don't have any specific learning goals at the moment." } - if err != nil { - res = "error executing: " + err.Error() - } - - history = append(history, execution{ - Command: "sudo " + args.Command, - Output: res, - }) - - return res, nil + return summary, 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.` + systemPrompt := `You are now in a shell in a container of the ubuntu:latest image to answer a question asked by the user. +You have full control over a bash shell inside this Docker container. - if len(history) < a.MaxCommands { - systemPrompt += `You can run any command you like to get to the needed results.` - } +Important directories: +- /data: A temporary directory with the lifespan of your processing. Use this for working files. +- /output: Files placed here will be returned to the caller after you're done. Use this for final 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.` +You can execute up to ` + fmt.Sprintf("%d", a.MaxCommands) + ` commands total. You're currently on command ` + fmt.Sprintf("%d", len(history)+1) + ` of ` + fmt.Sprintf("%d", a.MaxCommands) + `. + +For each command, you should: +1. Think about what you need to learn +2. Execute the command using the "execute" function +3. Analyze the output and record what you learned +4. Plan your next command based on this knowledge + +You can write files directly to /data or /output using the write_data and write_output functions. +When you are done, use "exit" to exit the container.` var toolbox []gollm.Function - // add unrestricted tools - toolbox = append(toolbox, tools["exit"], tools["write"], tools["read"]) + // Add all tools + toolbox = append(toolbox, + tools["exit"], + tools["write_data"], + tools["write_output"], + tools["read_data"], + tools["read_output"], + tools["summarize_knowledge"], + ) + // Only add execute if we haven't reached the command limit if len(history) < a.MaxCommands { - toolbox = append(toolbox, tools["execute"], tools["sudo"]) + toolbox = append(toolbox, tools["execute"]) } kw := shared.KnowledgeWorker{