When a workflow step calls an MCP tool that returns JSON, subsequent steps cannot access individual fields via template expansion. The JSON is stored as a raw string under a "text" key instead of being parsed.

Problem

In pkg/vmcp/client/client.go, the CallTool function stores text content without parsing:

resultMap[key] = textContent.Text  // line ~460

So when a tool returns {"count": 42, "type": "user_records"}, the step output becomes:

{"text": `{"count":42,"type":"user_records"}`}

Templates like {{.steps.fetch_data.output.count}} fail because there's no count key - only text.

Reproduction

Create a 2-step workflow where step 2 references step 1's output:

steps:
  - id: fetch_data
    tool: data_fetch_data
    
  - id: transform_data
    tool: transform_data
    arguments:
      count: "{{.steps.fetch_data.output.count}}"
    depends_on: [fetch_data]

The count argument receives <no value> instead of the actual value.

Proposed fix

Parse JSON text content and add fields to the result map:

resultMap["_raw"] = textContent.Text

var parsed map[string]any
if err := json.Unmarshal([]byte(textContent.Text), &parsed); err == nil {
    for k, v := range parsed {
        resultMap[k] = v
    }
}

Using _raw instead of text avoids collision if the tool returns a field named "text".

The existing unit tests don't catch this because they create structured output directly via RecordStepSuccess(), bypassing the MCP client entirely.