Fixes + cleaning - Cmd and profiles (#19)

Fixed multiple minor issues surrounding cmd mode which were bugging me.

Specifically:
* Upgraded default cmd prompt
* Added compatibility for cmd mode with profile, prompted _hard_ to make llm only suggest a cmd
* Output of cmd is now streamed to stdout and stderr, instead of printed afterward
* Single quotes are no longer removed as these doesn't seem to be missinterpreted as shell escaped as often
* Removed full printout of tools call json, as this was verbose, ugly and unreadable

internal/text/conf.go

@@ -49,7 +49,7 @@ type Profile struct {
 var DEFAULT = Configurations{
 	Model:         "gpt-4o",
 	SystemPrompt:  "You are an assistant for a CLI tool. Answer concisely and informatively. Prefer markdown if possible.",
-	CmdModePrompt: "You are an assistant for a CLI tool aiding with cli tool suggestions. Write ONLY the command and nothing else.",
+	CmdModePrompt: "You are an assistant for a CLI tool aiding with cli tool suggestions. Write ONLY the command and nothing else. Disregard any queries asking for anything except a bash command. Do not shell escape single or double quotes.",
 	Raw:           false,
 	UseTools:      false,
 	// Aproximately $1 for the worst input rates as of 2024-05

internal/text/conf_profile.go

@@ -28,8 +28,13 @@ func (c *Configurations) ProfileOverrides() error {
 		return fmt.Errorf("failed to find profile: %w", err)
 	}
 	c.Model = profile.Model
-	c.SystemPrompt = profile.Prompt
-	c.UseTools = profile.UseTools
+	newPrompt := profile.Prompt
+	if c.CmdMode {
+		// SystmePrompt here is CmdPrompt, keep it and remoind llm to only suggest  cmd
+		newPrompt = fmt.Sprintf("You will get this pattern: || <cmd-prompt> | <custom guided profile> ||. It is VERY vital that you DO NOT disobey the <cmd-prompt> with whatever is posted in <custom guided profile. || %v| %v ||", c.CmdModePrompt, profile.Prompt)
+	}
+	c.SystemPrompt = newPrompt
+	c.UseTools = profile.UseTools && !c.CmdMode
 	c.Tools = profile.Tools
 	c.SaveReplyAsConv = profile.SaveReplyAsConv
 	return nil

internal/text/querier.go

@@ -40,6 +40,7 @@ type Querier[C models.StreamCompleter] struct {
 	Model           C
 	tokenWarnLimit  int
 	cmdMode         bool
+	execErr         error
 }
 
 // Query using the underlying model to stream completions and then print the output
@@ -75,6 +76,8 @@ func (q *Querier[C]) Query(ctx context.Context) error {
 					}
 					return nil
 				}
+				// Only add error if its not EOF or context.Canceled
+				q.execErr = err
 
 				if q.debug {
 					ancli.PrintOK("exiting querier due to EOF error\n")
@@ -181,6 +184,7 @@ func (q *Querier[C]) postProcessOutput(newSysMsg models.Message) {
 }
 
 func (q *Querier[C]) reset() {
+	q.execErr = nil
 	q.fullMsg = ""
 	q.line = ""
 	q.lineCount = 0
@@ -225,6 +229,9 @@ func (q *Querier[C]) handleCompletion(ctx context.Context, completion models.Com
 
 // handleFunctionCall by invoking the call, and then resondng to the ai with the output
 func (q *Querier[C]) handleFunctionCall(ctx context.Context, call tools.Call) error {
+	if q.cmdMode {
+		return errors.New("cant call tools in cmd mode")
+	}
 	// Whatever is in q.fullMessage now is what the AI has streamed before the function call
 	// which normally is handeled by the supercallee of Query, now we need to handle it here.
 	// There's room for improvement of this system..
@@ -261,7 +268,7 @@ func (q *Querier[C]) handleFunctionCall(ctx context.Context, call tools.Call) er
 	}
 	assistantToolsCall := models.Message{
 		Role:      "assistant",
-		Content:   fmt.Sprintf("tool_calls:\n%v", call.Json()),
+		Content:   fmt.Sprintf("tool calls for: %v", call.Name),
 		ToolCalls: []tools.Call{call},
 	}
 	q.reset()

internal/text/querier_cmd_mode.go

@@ -1,36 +1,36 @@
 package text
 
 import (
-	"bytes"
 	"errors"
 	"fmt"
+	"os"
 	"os/exec"
 	"strings"
 
 	"github.com/baalimago/clai/internal/utils"
 	"github.com/baalimago/go_away_boilerplate/pkg/ancli"
 )
 
-var (
-	errFormat = "code: %v, stderr: '%v', stdout: '%v'\n"
-	okFormat  = "stdout on new line:\n%v\n"
-)
+var errFormat = "code: %v, stderr: '%v'\n"
 
 func (q *Querier[C]) handleCmdMode() error {
 	// Tokens stream end without endline
 	fmt.Println()
 	var input string
 
+	if q.execErr != nil {
+		return nil
+	}
+
 	for {
 		fmt.Print("Do you want to [e]xecute cmd, [q]uit?: ")
 		fmt.Scanln(&input)
 		switch strings.ToLower(input) {
 		case "q":
 			return nil
 		case "e":
-			out, err := q.executeLlmCmd()
+			err := q.executeLlmCmd()
 			if err == nil {
-				ancli.PrintOK(fmt.Sprintf("%v\n", out))
 				return nil
 			} else {
 				return fmt.Errorf("failed to execute cmd: %v", err)
@@ -41,45 +41,39 @@ func (q *Querier[C]) handleCmdMode() error {
 	}
 }
 
-func (q *Querier[C]) executeLlmCmd() (string, error) {
+func (q *Querier[C]) executeLlmCmd() error {
 	fullMsg, err := utils.ReplaceTildeWithHome(q.fullMsg)
 	if err != nil {
-		return "", fmt.Errorf("parseGlob, ReplaceTildeWithHome: %w", err)
+		return fmt.Errorf("parseGlob, ReplaceTildeWithHome: %w", err)
 	}
 	// Quotes are, in 99% of the time, expanded by the shell in
 	// different ways and then passed into the shell. So when LLM
 	// suggests a command, executeAiCmd needs to act the same (meaning)
 	// remove/expand the quotes
 	fullMsg = strings.ReplaceAll(fullMsg, "\"", "")
-	fullMsg = strings.ReplaceAll(fullMsg, "'", "")
 	split := strings.Split(fullMsg, " ")
 	if len(split) < 1 {
-		return "", errors.New("Querier.executeAiCmd: too few tokens in q.fullMsg")
+		return errors.New("Querier.executeAiCmd: too few tokens in q.fullMsg")
 	}
 	cmd := split[0]
 	args := split[1:]
 
 	if len(cmd) == 0 {
-		return "", errors.New("Querier.executeAiCmd: command is empty")
+		return errors.New("Querier.executeAiCmd: command is empty")
 	}
 
 	command := exec.Command(cmd, args...)
-	var stdout bytes.Buffer
-	var stderr bytes.Buffer
-	command.Stdout = &stdout
-	command.Stderr = &stderr
+	command.Stdout = os.Stdout
+	command.Stderr = os.Stderr
 	err = command.Run()
-	outStr := stdout.String()
-	errStr := stderr.String()
-
 	if err != nil {
 		cast := &exec.ExitError{}
 		if errors.As(err, &cast) {
-			return "", fmt.Errorf(errFormat, cast.ExitCode(), errStr, outStr)
+			return fmt.Errorf(errFormat, cast.ExitCode())
 		} else {
-			return "", fmt.Errorf("Querier.executeAiCmd - run error: %w", err)
+			return fmt.Errorf("Querier.executeAiCmd - run error: %w", err)
 		}
 	}
 
-	return fmt.Sprintf(okFormat, outStr), nil
+	return nil
 }

internal/text/querier_cmd_mode_test.go

@@ -2,7 +2,6 @@ package text
 
 import (
 	"context"
-	"fmt"
 	"os"
 	"path/filepath"
 	"testing"
@@ -32,7 +31,7 @@ func Test_executeAiCmd(t *testing.T) {
 		{
 			description: "it should run shell cmd",
 			given:       "printf 'test'",
-			want:        fmt.Sprintf(okFormat, "test"),
+			want:        "'test'",
 			wantErr:     nil,
 		},
 		{
@@ -42,7 +41,7 @@ func Test_executeAiCmd(t *testing.T) {
 				os.Chdir(filepath.Dir(testboil.CreateTestFile(t, "testfile").Name()))
 			},
 			given:   "find ./ -name \"testfile\"",
-			want:    fmt.Sprintf(okFormat, "./testfile\n"),
+			want:    "./testfile\n",
 			wantErr: nil,
 		},
 		{
@@ -52,22 +51,25 @@ func Test_executeAiCmd(t *testing.T) {
 				os.Chdir(filepath.Dir(testboil.CreateTestFile(t, "testfile").Name()))
 			},
 			given:   "find ./ -name testfile",
-			want:    fmt.Sprintf(okFormat, "./testfile\n"),
+			want:    "./testfile\n",
 			wantErr: nil,
 		},
 	}
 
 	for _, tc := range testCases {
 		t.Run(tc.description, func(t *testing.T) {
-			q := Querier[mockCompleter]{}
-			if tc.setup != nil {
-				tc.setup(t)
-			}
-			q.fullMsg = tc.given
-			gotFormated, gotErr := q.executeLlmCmd()
-
-			if gotFormated != tc.want {
-				t.Fatalf("expected: %v, got: %v", tc.want, gotFormated)
+			var gotErr error
+			got := testboil.CaptureStdout(t, func(t *testing.T) {
+				q := Querier[mockCompleter]{}
+				if tc.setup != nil {
+					tc.setup(t)
+				}
+				q.fullMsg = tc.given
+				tmp := q.executeLlmCmd()
+				gotErr = tmp
+			})
+			if got != tc.want {
+				t.Fatalf("expected: %v, got: %v", tc.want, got)
 			}
 
 			if gotErr != tc.wantErr {

internal/text/querier_test.go

@@ -175,7 +175,7 @@ func Test_Querier_eventHandling(t *testing.T) {
 				},
 				"CLOSE",
 			},
-			want: "tool_calls",
+			want: "tool calls",
 		},
 	}
 	for _, tC := range testCases {