Files
Ed Zynda 975c30a773 fix(mcp): surface MCP tool failures as soft errors, not critical aborts (#31)
The MCP adapter previously wrapped any error returned by MCPToolManager.ExecuteTool
into a Go error returned from the fantasy.AgentTool.Run interface. The fantasy
agent loop treats those as critical errors and aborts the entire turn —
discarding all prior reasoning, tool calls, and results.

In practice that meant a single misbehaved MCP server returning a JSON-RPC
"-32602 Invalid params" (e.g. a Zod schema mismatch on the server's input
validation) would kill an in-progress turn after the model had already done
dozens of seconds of useful work, with no way for the model to see the
validation message and self-correct.

This mismatched the contract that native Kit tools follow: native tools
return errors via kit.ErrorResult(...), which become soft tool-result errors
that the model reads and can act on (retry with corrected args, try a
different tool, give up gracefully).

Make the MCP path behave the same way:

  - JSON-RPC protocol errors, transport failures, and server-side schema
    rejections are now returned as fantasy.NewTextErrorResponse(...) with
    err == nil, so the agent loop continues and the model sees the failure
    in-band as a tool result it can reason about.
  - Context cancellation (ctx.Err() != nil) remains a critical error so
    callers can abort turns deterministically. This is the only case where
    bubbling up is correct — the caller intentionally tore the turn down
    and the agent must not keep spinning.
  - Server-side soft errors (CallToolResult{ isError: true }) and the
    happy path are unchanged.

The agent loop's MaxSteps cap already bounds the worst case for a
permanently broken MCP server, so there is no risk of unbounded retries.

Side effect: extracted a tiny mcpExecutor interface for the one method the
adapter uses (ExecuteTool), purely so the adapter is unit-testable in
isolation without standing up a full MCPToolManager + connection pool.

Behavior change note for downstream consumers: code that relied on
host.PromptResult / Stream returning a Go error containing
"mcp tool execution failed" will no longer see those errors — the
failure information is now in the assistant's final response (or in the
OnAfterToolResult / OnToolResult hooks, where IsError will be true).
Context cancellation continues to surface as an error from those calls
as before.

Co-authored-by: space_cowboy <space_cowboy@mark3labs.com>
2026-05-13 20:12:31 +03:00

89 lines
3.0 KiB
Go

package agent
import (
"context"
"fmt"
"charm.land/fantasy"
"github.com/mark3labs/kit/internal/tools"
)
// mcpExecutor is the subset of *tools.MCPToolManager that the adapter
// actually uses. Extracted as an interface so the adapter is unit-testable
// without constructing a full manager + connection pool.
type mcpExecutor interface {
ExecuteTool(ctx context.Context, prefixedName, inputJSON string) (*tools.MCPToolResult, error)
}
// mcpAgentTool adapts an tools.MCPTool to the fantasy.AgentTool interface.
// This keeps the fantasy dependency confined to the agent layer — the tools
// package is a pure MCP client library with no LLM framework dependency.
type mcpAgentTool struct {
tool tools.MCPTool
exec mcpExecutor
providerOptions fantasy.ProviderOptions
}
// Info returns the fantasy tool info including name, description, and parameter schema.
func (t *mcpAgentTool) Info() fantasy.ToolInfo {
return fantasy.ToolInfo{
Name: t.tool.Name,
Description: t.tool.Description,
Parameters: t.tool.Parameters,
Required: t.tool.Required,
}
}
// Run executes the MCP tool by delegating to the MCPToolManager.
//
// MCP-side failures (JSON-RPC protocol errors, transport failures, schema
// validation rejections from the server) are surfaced to the model as soft
// tool errors rather than escalated to a critical agent error. This matches
// the contract that native Kit tools follow via kit.ErrorResult(...) and
// lets the model self-correct (e.g. retry with a fixed argument shape) or
// give up gracefully rather than aborting the turn mid-run.
//
// Context cancellation is the one exception: if the caller cancelled the
// context the turn was aborted intentionally, so we propagate the ctx error
// to let the agent loop unwind cleanly.
func (t *mcpAgentTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {
result, err := t.exec.ExecuteTool(ctx, t.tool.Name, call.Input)
if err != nil {
if ctxErr := ctx.Err(); ctxErr != nil {
return fantasy.ToolResponse{}, ctxErr
}
return fantasy.NewTextErrorResponse(
fmt.Sprintf("MCP tool %q failed: %s", t.tool.Name, err.Error()),
), nil
}
if result.IsError {
return fantasy.NewTextErrorResponse(result.Content), nil
}
return fantasy.NewTextResponse(result.Content), nil
}
// ProviderOptions returns provider-specific options for this tool.
func (t *mcpAgentTool) ProviderOptions() fantasy.ProviderOptions {
return t.providerOptions
}
// SetProviderOptions sets provider-specific options for this tool.
func (t *mcpAgentTool) SetProviderOptions(opts fantasy.ProviderOptions) {
t.providerOptions = opts
}
// mcpToolsToAgentTools converts a slice of MCPTool to fantasy.AgentTool
// implementations that route execution through the MCPToolManager.
func mcpToolsToAgentTools(mcpTools []tools.MCPTool, manager *tools.MCPToolManager) []fantasy.AgentTool {
agentTools := make([]fantasy.AgentTool, len(mcpTools))
for i, t := range mcpTools {
agentTools[i] = &mcpAgentTool{
tool: t,
exec: manager,
}
}
return agentTools
}