fix: make A2A client timeout configurable in MCP handler (#1617)

## fix: make A2A client timeout configurable in MCP handler

### Problem

The MCP handler's `invoke_agent` tool uses an A2A client with a
**hardcoded 30-second timeout** (`mcp_handler.go:212`). When an agent
performs multi-step work — schema exploration, multiple tool calls, LLM
reasoning — the total round-trip easily exceeds 30 seconds. The A2A
client times out and returns:

```
Failed to send A2A message: a2aClient.SendMessage: a2aClient.doRequest:
http request failed: context deadline exceeded (Client.Timeout exceeded while awaiting headers)
```

The agent itself completes successfully (visible in pod logs), but the
caller never receives the response.

### Root Cause

Other A2A call sites already use configurable timeouts:

| Call site | Timeout source |
|---|---|
| `a2a_registrar.go` | `cfg.Streaming.Timeout` (default 600s) ✅ |
| CLI `agent invoke` | `cfg.Config.Timeout` ✅ |
| CLI `agent run` | `cfg.Config.Timeout` ✅ |
| **`mcp_handler.go`** | **Hardcoded `30 * time.Second`** ❌ |

The MCP handler was the only place with a hardcoded value.

### Fix

- Add `a2aTimeout time.Duration` parameter to `NewMCPHandler()`
- Pass `cfg.Streaming.Timeout` from `app.go` (same source as the A2A
registrar)
- Fall back to 60s if a zero/negative value is passed
- No new config keys — uses the existing `STREAMING_TIMEOUT` env var /
`--streaming-timeout` flag

### Changes

- `go/core/internal/mcp/mcp_handler.go` — new field, new constructor
param, use `h.a2aTimeout` instead of hardcoded value
- `go/core/pkg/app/app.go` — pass `cfg.Streaming.Timeout` to
`NewMCPHandler()`

### Testing

- `go build ./core/...` passes
- Manually verified with a kagent agent performing 5+ tool calls (~40s
total) — `invoke_agent` now returns successfully

Signed-off-by: Oswaldo Gomez <papagala@users.noreply.github.com>
Co-authored-by: Oswaldo Gomez <papagala@users.noreply.github.com>
Co-authored-by: Eitan Yarmush <eitan.yarmush@solo.io>

Author: papagala

go/core/internal/mcp/mcp_handler.go

@@ -25,6 +25,7 @@ import (
 type MCPHandler struct {
 	kubeClient    client.Client
 	a2aBaseURL    string
+	a2aTimeout    time.Duration
 	authenticator auth.AuthProvider
 	httpHandler   *mcpsdk.StreamableHTTPHandler
 	server        *mcpsdk.Server
@@ -55,12 +56,20 @@ type InvokeAgentOutput struct {
 	ContextID string `json:"context_id,omitempty"`
 }
 
+// defaultA2ATimeout is the fallback timeout for A2A client calls and should match
+// the configured default streaming timeout.
+const defaultA2ATimeout = 10 * time.Minute
+
 // NewMCPHandler creates a new MCP handler
 // Wraps the StreamableHTTPHandler and adds A2A bridging and context management.
-func NewMCPHandler(kubeClient client.Client, a2aBaseURL string, authenticator auth.AuthProvider) (*MCPHandler, error) {
+func NewMCPHandler(kubeClient client.Client, a2aBaseURL string, authenticator auth.AuthProvider, a2aTimeout time.Duration) (*MCPHandler, error) {
+	if a2aTimeout <= 0 {
+		a2aTimeout = defaultA2ATimeout
+	}
 	handler := &MCPHandler{
 		kubeClient:    kubeClient,
 		a2aBaseURL:    a2aBaseURL,
+		a2aTimeout:    a2aTimeout,
 		authenticator: authenticator,
 	}
 
@@ -209,7 +218,7 @@ func (h *MCPHandler) handleInvokeAgent(ctx context.Context, req *mcpsdk.CallTool
 	if a2aClient == nil {
 		// Build A2A client options with authentication propagation
 		a2aOpts := []a2aclient.Option{
-			a2aclient.WithTimeout(30 * time.Second),
+			a2aclient.WithTimeout(h.a2aTimeout),
 			a2aclient.WithHTTPReqHandler(
 				authimpl.A2ARequestHandler(
 					h.authenticator,

go/core/pkg/app/app.go

@@ -533,6 +533,7 @@ func Start(getExtensionConfig GetExtensionConfig) {
 		mgr.GetClient(),
 		cfg.A2ABaseUrl+httpserver.APIPathA2A,
 		extensionCfg.Authenticator,
+		cfg.Streaming.Timeout,
 	)
 	if err != nil {
 		setupLog.Error(err, "unable to create MCP handler")