A. Minimal Two-Step Stilt

This stilt runs two steps in sequence. The first step takes the caller's message and drafts an answer. The second step reads that draft and improves its clarity and structure. The exit step is the second one, so the API response contains the improved version.

The config demonstrates two core mechanics. First, how fields wire steps together: the text field on step 0 reads input.context (the caller's message), and the ingest field on step 1 reads step 0's output using loopRef: current. Second, how exit selects the final answer: since exit: step1, whatever step 1 produces is what the API returns.

With the input "Explain why the sky is blue.":

Step 0 — Draft. The stilt resolves step 0's fields. There's one field: a text field called Context with from: input.context. It resolves to the caller's message. The prompt assembled for the LLM:

Context: Explain why the sky is blue.

[System Instruction]
Draft answer.

The response is stored as step 0's output for loop 0.

Step 1 — Improve. Step 1 has one field: an ingest field called Draft that reads from step 0 with loopRef: current. Since this is loop 0, that resolves to the raw draft just produced. The prompt:

Draft: <the raw draft from step 0>

[System Instruction]
Improve clarity and structure.

The response — the improved version — is stored as step 1's output.

Exit. Since exit: step1, step 1's output is the stilt's final answer. It's returned to the caller in standard OpenAI-compatible format.

name: Mini Two-Step

allowedTargets:
  strategy: universal

exit: step1

knobs: {}

steps:
  - id: step0
    name: Draft
    type: normal
    fields:
      - name: Context
        type: text
        from: input.context
    systemPrompt: "Draft answer."

  - id: step1
    name: Improve
    type: normal
    fields:
      - name: Draft
        type: ingest
        from:
          stepId: step0
          loopRef: current
    systemPrompt: "Improve clarity and structure."

The strategy: universal in allowedTargets means any provider and model can serve this stilt. The platform auto-manages id, author, and version when the config is saved.

The stilt is addressed by the author's username and the stilt's slug name in the URL path. The request body follows the standard OpenAI chat completions format:

curl https://api.redeo.ai/v1/redeo-labs/mini-two-step/chat/completions \
  -H "Authorization: Bearer $REDEO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Explain why the sky is blue."}]
  }'

The response is a standard chat completion containing step 1's improved output. The response format is identical to a direct model call — no intermediate outputs are exposed.

In Studio and Foundry, the timeline renders both steps: step 0's output and step 1's output, with the exit checkpoint on step 1 marking it as the final answer.