Phase 5 Batch 1 (补): mcp/server + 2 Tool (Batch 1 commit 时漏 stage)

上一个 commit (7056657) 漏 add 的 mcp 包文件:
- internal/mcp/server.go (Streamable HTTP Handler)
- internal/mcp/tools/list_clusters.go (发现入口)
- internal/mcp/tools/spark_submit.go (本地 exec 提交)
- internal/mcp/tools/helpers.go (textResult/errResult/encodeJSON)

deps.go / rm.go 已在 Phase 5 Batch 2 一起 commit, 此处不重复。

Co-Authored-By: Claude <noreply@anthropic.com>

internal/mcp/tools/helpers.go

@@ -0,0 +1,80 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"log/slog"
+
+	"github.com/mark3labs/mcp-go/mcp"
+	"spark-mcp-go/internal/logging"
+)
+
+// textResult wraps a string into a text CallToolResult.
+func textResult(s string) *mcp.CallToolResult {
+	return mcp.NewToolResultText(s)
+}
+
+// errResult wraps a string into an error CallToolResult.
+// MCP protocol errors are reserved for exceptional conditions; business
+// errors are reported inside the tool result with IsError set by the library.
+func errResult(s string) *mcp.CallToolResult {
+	return mcp.NewToolResultError(s)
+}
+
+// encodeJSON marshals v to JSON. On failure it returns a string literal that
+// embeds the error so the caller never receives a nil or empty result.
+func encodeJSON(v any) string {
+	b, err := json.Marshal(v)
+	if err != nil {
+		return fmt.Sprintf("<marshal error: %s>", err)
+	}
+	return string(b)
+}
+
+// startToolCall starts per-tool-call logging. If logging is not initialized
+// (e.g. during tests) it falls back to a no-op logger so handlers never panic.
+func startToolCall(ctx context.Context, logger *slog.Logger, toolName string, params any) logging.ToolCallLogger {
+	callLog, err := logging.StartToolCall(ctx, logger, toolName, params)
+	if err != nil {
+		return &noopToolCallLogger{}
+	}
+	return callLog
+}
+
+type noopToolCallLogger struct{}
+
+func (n *noopToolCallLogger) WithResult(any)  {}
+func (n *noopToolCallLogger) WithError(error) {}
+func (n *noopToolCallLogger) End()            {}
+
+// truncateMiddle limits s to roughly max bytes by keeping the first and
+// last halves, separated by a marker.  It preserves UTF-8 rune boundaries
+// so the result never starts or ends with a broken multi-byte rune.
+func truncateMiddle(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	half := max / 2
+	start := s[:half]
+	end := s[len(s)-half:]
+	for len(start) > 0 && start[len(start)-1] >= 0x80 && start[len(start)-1] < 0xC0 {
+		start = start[:len(start)-1]
+	}
+	for len(end) > 0 && end[0]&0xC0 == 0x80 {
+		end = end[1:]
+	}
+	return start + "\n... [truncated middle] ...\n" + end
+}
+
+// truncateTail keeps the last max bytes of s, preserving UTF-8 rune boundaries.
+func truncateTail(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	start := len(s) - max
+	for start < len(s) && s[start]&0xC0 == 0x80 {
+		start++
+	}
+	return s[start:]
+}

internal/mcp/tools/list_clusters.go

@@ -0,0 +1,39 @@
+package tools
+
+import (
+	"context"
+
+	"github.com/mark3labs/mcp-go/mcp"
+)
+
+const ListClustersName = "list_clusters"
+
+// NewListClustersTool returns the schema for the list_clusters MCP Tool.
+func NewListClustersTool() mcp.Tool {
+	return mcp.NewTool(ListClustersName,
+		mcp.WithDescription("List active Spark/YARN clusters configured for the agent. Returns the full Cluster struct (minus auth password) for every active cluster — this is the discovery root: from here the agent knows RM/SHS endpoints, auth flavor, and URL allowlists."),
+	)
+}
+
+// ListClustersHandler lists configured clusters and returns only active ones.
+func (d *Deps) ListClustersHandler(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+	callLog := startToolCall(ctx, d.Logger, ListClustersName, nil)
+	defer callLog.End()
+
+	all, err := d.ClusterRepo.List(ctx)
+	if err != nil {
+		callLog.WithError(err)
+		return errResult("list_clusters: " + err.Error()), nil
+	}
+
+	// task_plan.md specifies active clusters only.
+	active := all[:0]
+	for _, c := range all {
+		if c.IsActive {
+			active = append(active, c)
+		}
+	}
+
+	callLog.WithResult(map[string]any{"count": len(active)})
+	return textResult(encodeJSON(active)), nil
+}

internal/mcp/tools/spark_submit.go

@@ -0,0 +1,100 @@
+package tools
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/mark3labs/mcp-go/mcp"
+
+	"spark-mcp-go/internal/executor"
+)
+
+const SparkSubmitName = "spark_submit"
+
+// NewSparkSubmitTool returns the schema for the spark_submit MCP Tool.
+func NewSparkSubmitTool() mcp.Tool {
+	return mcp.NewTool(SparkSubmitName,
+		mcp.WithDescription("Submit a Spark application to a configured cluster using the cluster's local spark-submit binary. Returns {app_id, exit_code, stdout_tail, stderr_tail, duration_ms}. The cluster's default_submit_args are prepended to your args. Binary is invoked as a child process — no shell, no command injection."),
+		mcp.WithString("cluster_id",
+			mcp.Required(),
+			mcp.Description("ID of the configured cluster (from list_clusters)"),
+		),
+		// mcp-go v0.56.0 uses PropertyOption for array item schema.
+		mcp.WithArray("args",
+			mcp.Description("Spark-submit CLI args (after default_submit_args). Each element becomes one argv entry — no shell interpretation."),
+			mcp.Items(map[string]any{"type": "string"}),
+		),
+	)
+}
+
+// SparkSubmitHandler runs spark-submit against the requested cluster.
+func (d *Deps) SparkSubmitHandler(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+	clusterID, err := req.RequireString("cluster_id")
+	if err != nil {
+		return errResult("spark_submit: " + err.Error()), nil
+	}
+
+	args := req.GetArguments()
+	var userArgs []string
+	if rawArgs, ok := args["args"]; ok && rawArgs != nil {
+		arr, ok := rawArgs.([]any)
+		if !ok {
+			return errResult("spark_submit: args is not an array"), nil
+		}
+		for i, item := range arr {
+			s, ok := item.(string)
+			if !ok {
+				return errResult(fmt.Sprintf("spark_submit: args[%d] is not a string", i)), nil
+			}
+			userArgs = append(userArgs, s)
+		}
+	}
+
+	callLog := startToolCall(ctx, d.Logger, SparkSubmitName, map[string]any{
+		"cluster_id": clusterID,
+		"args":       userArgs,
+	})
+	defer callLog.End()
+
+	cl, err := d.ClusterRepo.Get(ctx, clusterID)
+	if err != nil {
+		callLog.WithError(err)
+		return errResult("spark_submit: cluster " + clusterID + ": " + err.Error()), nil
+	}
+
+	fullArgs := append([]string{}, cl.DefaultSubmitArgs...)
+	fullArgs = append(fullArgs, userArgs...)
+
+	result, err := executor.Run(ctx, executor.SparkSubmitOpts{
+		Binary:  cl.SparkSubmitExecuteBin,
+		Args:    fullArgs,
+		Timeout: d.SparkSubmitTimeout,
+	})
+	if err != nil {
+		callLog.WithError(err)
+		// Even on error, result carries ExitCode/Stderr for the LLM.
+		if result != nil {
+			callLog.WithResult(map[string]any{
+				"app_id":      result.AppID,
+				"exit_code":   result.ExitCode,
+				"duration_ms": result.DurationMS,
+			})
+			return textResult(encodeJSON(map[string]any{
+				"app_id":      result.AppID,
+				"exit_code":   result.ExitCode,
+				"stdout_tail": result.StdoutTail,
+				"stderr_tail": result.StderrTail,
+				"duration_ms": result.DurationMS,
+				"error":       err.Error(),
+			})), nil
+		}
+		return errResult("spark_submit: " + err.Error()), nil
+	}
+
+	callLog.WithResult(map[string]any{
+		"app_id":      result.AppID,
+		"exit_code":   result.ExitCode,
+		"duration_ms": result.DurationMS,
+	})
+	return textResult(encodeJSON(result)), nil
+}