Feature request: native `llm` step type for AI agent flows

[ameet]

Problem

Calling an LLM from a flow currently requires a 3-step workaround:

1. file-write — save prompt to temp file
2. bash — pipe to claude --print with --tools and --allowed-tools flags
3. code — parse the envelope ({"type":"result","result":"..."}) and strip markdown fences

This pattern is repeated in nearly every AI-powered flow. It introduces:

• Fragile string parsing (envelope unwrapping, fence-stripping, preamble removal)
• Double flag requirement (--tools + --allowed-tools) that's easy to forget
• Timeout management at the bash level instead of step level
• No structured error when the LLM call fails vs. when the shell command fails

Proposal

A native step type:

{
  "id": "analyzeData",
  "type": "llm",
  "llm": {
    "model": "claude-sonnet-4-6",
    "prompt": "{{$.steps.buildPrompt.output.text}}",
    "outputFormat": "json",
    "tools": ["WebSearch", "WebFetch"],
    "timeout": 120000
  }
}

The runtime would handle:

• Envelope unwrapping automatically
• JSON parsing when outputFormat: "json"
• Structured error on failure (distinguishable from a shell error)
• $.steps.analyzeData.output returns the parsed result directly

Why this matters

LLM calls are the primary use case for AI agent flows. Making them a first-class step type would eliminate the most common source of boilerplate and parsing bugs across flow authors.