path: root/internal/executor/claude.go

package executor

import (
	"bufio"
	"context"
	"encoding/json"
	"fmt"
	"io"
	"log/slog"
	"os"
	"os/exec"
	"path/filepath"
	"strings"
	"sync"
	"syscall"
	"time"

	"github.com/thepeterstone/claudomator/internal/storage"
	"github.com/thepeterstone/claudomator/internal/task"
)

// ClaudeRunner spawns the `claude` CLI in non-interactive mode.
type ClaudeRunner struct {
	BinaryPath string // defaults to "claude"
	Logger     *slog.Logger
	LogDir     string // base directory for execution logs
	APIURL     string // base URL of the Claudomator API, passed to subprocesses
}

// BlockedError is returned by Run when the agent wrote a question file and exited.
// The pool transitions the task to BLOCKED and stores the question for the user.
type BlockedError struct {
	QuestionJSON string // raw JSON from the question file
	SessionID    string // claude session to resume once the user answers
}

func (e *BlockedError) Error() string { return fmt.Sprintf("task blocked: %s", e.QuestionJSON) }

// ExecLogDir returns the log directory for the given execution ID.
// Implements LogPather so the pool can persist paths before execution starts.
func (r *ClaudeRunner) ExecLogDir(execID string) string {
	if r.LogDir == "" {
		return ""
	}
	return filepath.Join(r.LogDir, execID)
}

func (r *ClaudeRunner) binaryPath() string {
	if r.BinaryPath != "" {
		return r.BinaryPath
	}
	return "claude"
}

// Run executes a claude -p invocation, streaming output to log files.
// It retries up to 3 times on rate-limit errors using exponential backoff.
// If the agent writes a question file and exits, Run returns *BlockedError.
func (r *ClaudeRunner) Run(ctx context.Context, t *task.Task, e *storage.Execution) error {
	if t.Claude.WorkingDir != "" {
		if _, err := os.Stat(t.Claude.WorkingDir); err != nil {
			return fmt.Errorf("working_dir %q: %w", t.Claude.WorkingDir, err)
		}
	}

	// Setup log directory once; retries overwrite the log files.
	logDir := r.ExecLogDir(e.ID)
	if logDir == "" {
		logDir = e.ID // fallback for tests without LogDir set
	}
	if err := os.MkdirAll(logDir, 0700); err != nil {
		return fmt.Errorf("creating log dir: %w", err)
	}
	if e.StdoutPath == "" {
		e.StdoutPath = filepath.Join(logDir, "stdout.log")
		e.StderrPath = filepath.Join(logDir, "stderr.log")
		e.ArtifactDir = logDir
	}

	// Pre-assign session ID so we can resume after a BLOCKED state.
	// If this is a resume execution the session ID is already set.
	if e.SessionID == "" {
		e.SessionID = e.ID // reuse execution UUID as session UUID (both are UUIDs)
	}

	questionFile := filepath.Join(logDir, "question.json")
	args := r.buildArgs(t, e, questionFile)

	attempt := 0
	err := runWithBackoff(ctx, 3, 5*time.Second, func() error {
		if attempt > 0 {
			delay := 5 * time.Second * (1 << (attempt - 1))
			r.Logger.Warn("rate-limited by Claude API, retrying",
				"attempt", attempt,
				"delay", delay,
			)
		}
		attempt++
		return r.execOnce(ctx, args, t.Claude.WorkingDir, e)
	})
	if err != nil {
		return err
	}

	// Check whether the agent left a question before exiting.
	data, readErr := os.ReadFile(questionFile)
	if readErr == nil {
		os.Remove(questionFile) // consumed
		return &BlockedError{QuestionJSON: strings.TrimSpace(string(data)), SessionID: e.SessionID}
	}
	return nil
}

// execOnce runs the claude subprocess once, streaming output to e's log paths.
func (r *ClaudeRunner) execOnce(ctx context.Context, args []string, workingDir string, e *storage.Execution) error {
	cmd := exec.CommandContext(ctx, r.binaryPath(), args...)
	cmd.Env = append(os.Environ(),
		"CLAUDOMATOR_API_URL="+r.APIURL,
		"CLAUDOMATOR_TASK_ID="+e.TaskID,
		"CLAUDOMATOR_QUESTION_FILE="+filepath.Join(e.ArtifactDir, "question.json"),
	)
	// Put the subprocess in its own process group so we can SIGKILL the entire
	// group (MCP servers, bash children, etc.) on cancellation.
	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
	if workingDir != "" {
		cmd.Dir = workingDir
	}

	stdoutFile, err := os.Create(e.StdoutPath)
	if err != nil {
		return fmt.Errorf("creating stdout log: %w", err)
	}
	defer stdoutFile.Close()

	stderrFile, err := os.Create(e.StderrPath)
	if err != nil {
		return fmt.Errorf("creating stderr log: %w", err)
	}
	defer stderrFile.Close()

	// Use os.Pipe for stdout so we own the read-end lifetime.
	// cmd.StdoutPipe() would add the read-end to closeAfterWait, causing
	// cmd.Wait() to close it before our goroutine finishes reading.
	stdoutR, stdoutW, err := os.Pipe()
	if err != nil {
		return fmt.Errorf("creating stdout pipe: %w", err)
	}
	cmd.Stdout = stdoutW // *os.File — not added to closeAfterStart/Wait
	cmd.Stderr = stderrFile

	if err := cmd.Start(); err != nil {
		stdoutW.Close()
		stdoutR.Close()
		return fmt.Errorf("starting claude: %w", err)
	}
	// Close our write-end immediately; the subprocess holds its own copy.
	// The goroutine below gets EOF when the subprocess exits.
	stdoutW.Close()

	// killDone is closed when cmd.Wait() returns, stopping the pgid-kill goroutine.
	killDone := make(chan struct{})
	go func() {
		select {
		case <-ctx.Done():
			// SIGKILL the entire process group to reap orphan children.
			syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
		case <-killDone:
		}
	}()

	// Stream stdout to the log file and parse cost/errors.
	// wg ensures costUSD and streamErr are fully written before we read them after cmd.Wait().
	var costUSD float64
	var streamErr error
	var wg sync.WaitGroup
	wg.Add(1)
	go func() {
		defer wg.Done()
		costUSD, streamErr = parseStream(stdoutR, stdoutFile, r.Logger)
		stdoutR.Close()
	}()

	waitErr := cmd.Wait()
	close(killDone) // stop the pgid-kill goroutine
	wg.Wait()       // drain remaining stdout before reading costUSD/streamErr

	e.CostUSD = costUSD

	if waitErr != nil {
		if exitErr, ok := waitErr.(*exec.ExitError); ok {
			e.ExitCode = exitErr.ExitCode()
		}
		return fmt.Errorf("claude exited with error: %w", waitErr)
	}

	e.ExitCode = 0
	if streamErr != nil {
		return streamErr
	}
	return nil
}

func (r *ClaudeRunner) buildArgs(t *task.Task, e *storage.Execution, questionFile string) []string {
	// Resume execution: the agent already has context; just deliver the answer.
	if e.ResumeSessionID != "" {
		args := []string{
			"-p", e.ResumeAnswer,
			"--resume", e.ResumeSessionID,
			"--output-format", "stream-json",
			"--verbose",
		}
		permMode := t.Claude.PermissionMode
		if permMode == "" {
			permMode = "bypassPermissions"
		}
		args = append(args, "--permission-mode", permMode)
		if t.Claude.Model != "" {
			args = append(args, "--model", t.Claude.Model)
		}
		return args
	}

	instructions := t.Claude.Instructions
	allowedTools := t.Claude.AllowedTools

	if !t.Claude.SkipPlanning {
		instructions = withPlanningPreamble(instructions)
		// Ensure Bash is available so the agent can POST subtasks and ask questions.
		hasBash := false
		for _, tool := range allowedTools {
			if tool == "Bash" {
				hasBash = true
				break
			}
		}
		if !hasBash {
			allowedTools = append(allowedTools, "Bash")
		}
	}

	args := []string{
		"-p", instructions,
		"--session-id", e.SessionID,
		"--output-format", "stream-json",
		"--verbose",
	}

	if t.Claude.Model != "" {
		args = append(args, "--model", t.Claude.Model)
	}
	if t.Claude.MaxBudgetUSD > 0 {
		args = append(args, "--max-budget-usd", fmt.Sprintf("%.2f", t.Claude.MaxBudgetUSD))
	}
	// Default to bypassPermissions — claudomator runs tasks unattended, so
	// prompting for write access would always stall execution. Tasks that need
	// a more restrictive mode can set permission_mode explicitly.
	permMode := t.Claude.PermissionMode
	if permMode == "" {
		permMode = "bypassPermissions"
	}
	args = append(args, "--permission-mode", permMode)
	if t.Claude.SystemPromptAppend != "" {
		args = append(args, "--append-system-prompt", t.Claude.SystemPromptAppend)
	}
	for _, tool := range allowedTools {
		args = append(args, "--allowedTools", tool)
	}
	for _, tool := range t.Claude.DisallowedTools {
		args = append(args, "--disallowedTools", tool)
	}
	for _, f := range t.Claude.ContextFiles {
		args = append(args, "--add-dir", f)
	}
	args = append(args, t.Claude.AdditionalArgs...)

	return args
}

// parseStream reads streaming JSON from claude, writes to w, and returns
// (costUSD, error). error is non-nil if the stream signals task failure:
// - result message has is_error:true
// - a tool_result was denied due to missing permissions
func parseStream(r io.Reader, w io.Writer, logger *slog.Logger) (float64, error) {
	tee := io.TeeReader(r, w)
	scanner := bufio.NewScanner(tee)
	scanner.Buffer(make([]byte, 1024*1024), 1024*1024) // 1MB buffer for large lines

	var totalCost float64
	var streamErr error

	for scanner.Scan() {
		line := scanner.Bytes()
		var msg map[string]interface{}
		if err := json.Unmarshal(line, &msg); err != nil {
			continue
		}

		msgType, _ := msg["type"].(string)
		switch msgType {
		case "result":
			if isErr, _ := msg["is_error"].(bool); isErr {
				result, _ := msg["result"].(string)
				if result != "" {
					streamErr = fmt.Errorf("claude task failed: %s", result)
				} else {
					streamErr = fmt.Errorf("claude task failed (is_error=true in result)")
				}
			}
			// Prefer total_cost_usd from result message; fall through to legacy check below.
			if cost, ok := msg["total_cost_usd"].(float64); ok {
				totalCost = cost
			}
		case "user":
			// Detect permission-denial tool_results. These occur when permission_mode
			// is not bypassPermissions and claude exits 0 without completing its task.
			if err := permissionDenialError(msg); err != nil && streamErr == nil {
				streamErr = err
			}
		}

		// Legacy cost field used by older claude versions.
		if cost, ok := msg["cost_usd"].(float64); ok {
			totalCost = cost
		}
	}

	return totalCost, streamErr
}

// permissionDenialError inspects a "user" stream message for tool_result entries
// that were denied due to missing permissions. Returns an error if found.
func permissionDenialError(msg map[string]interface{}) error {
	message, ok := msg["message"].(map[string]interface{})
	if !ok {
		return nil
	}
	content, ok := message["content"].([]interface{})
	if !ok {
		return nil
	}
	for _, item := range content {
		itemMap, ok := item.(map[string]interface{})
		if !ok {
			continue
		}
		if itemMap["type"] != "tool_result" {
			continue
		}
		if isErr, _ := itemMap["is_error"].(bool); !isErr {
			continue
		}
		text, _ := itemMap["content"].(string)
		if strings.Contains(text, "requested permissions") || strings.Contains(text, "haven't granted") {
			return fmt.Errorf("permission denied by host: %s", text)
		}
	}
	return nil
}