How to call LLM task with stdin input?

[RatkoJ]

Hi all,

I've been trying to process email contents (from gog) one by one using a simple lobster workflow but I cannot seem to get the LLM task working. Claude has looked at the lobster source code many times, tried different things but nothing works. It should be simple but I am missing something.

What I've tried:

pipeline: llm.invoke --provider openclaw --prompt "Hello. What's your name?"
stdin: $fetch_email_body.stdout
Fails with `"message": "llm.invoke received invalid response envelope"`

command: openclaw.invoke --tool llm-task --action json --args-json '{"prompt":"How many words have been pasted below?"}'
stdin: $fetch_email_body.stdout

Returns: word count: 0 (to me this means the stdin isn't being passed in somehow)

run: llm_task.invoke --prompt "Hi, what's your name?"
/bin/sh: 1: llm_task.invoke: not found"

I am also unclear on the when i should be using pipeline: and when command: openclaw.invoke --tool llm-task

I've made sure that the previous task indeed returns the email contents via:

command: echo $EMAIL_BODY
env:
      EMAIL_BODY: $fetch_email_body.stdout

[Bandwe]

Great question! The issue you're running into is about how Lobster handles stdin with LLM calls. Let me break down the correct patterns:

The Problem

Your attempts were close, but there are a few key things to understand:

1. pipeline: is for native Lobster stages (like llm.invoke)
2. run: or command: is for shell executables
3. stdin: passes data from a previous step's output

Working Solution

Here's a complete, working example that processes emails with an LLM:

name: email-processor
steps:
  # Step 1: Fetch emails (replace with your actual email fetching)
  - id: fetch-emails
    run: |
      echo '[
        {"id": 1, "subject": "Meeting", "body": "Are we still on for tomorrow?"},
        {"id": 2, "subject": "Update", "body": "Project is progressing well."}
      ]'

  # Step 2: Process each email with LLM using stdin
  - id: summarize
    pipeline: >
      llm.invoke --provider openclaw --prompt "For each email in this JSON array, create a one-sentence summary. Return as JSON array."
    stdin: $fetch-emails.json

  # Step 3: Output the result
  - id: output
    run: cat
    stdin: $summarize.stdout

Why Your Attempts Didn't Work

1. pipeline: llm.invoke ... with stdin: $fetch_email_body.stdout

   • This should work! Make sure $fetch_email_body is a valid step ID from a previous step

2. command: openclaw.invoke --tool llm-task ...

   • This runs in shell mode, so stdin needs to be handled differently
   • The shell command needs to explicitly read from stdin (e.g., cat | openclaw.invoke...)

3. run: llm_task.invoke ...

   • llm_task.invoke is not a shell executable - it's a Lobster pipeline stage
   • Use pipeline: instead of run: for this

Complete Working Example

I've created a full email processing workflow here: https://github.com/openclaw/lobster-examples/blob/main/email-processing.lobster

This example shows:

• Fetching emails from an API
• Filtering unread messages
• Summarizing each with an LLM using stdin
• Formatting the output

Key Pattern for stdin + LLM

- id: fetch-data
  run: echo "Some text or JSON data"

- id: process-with-llm
  pipeline: llm.invoke --provider openclaw --prompt "Process this data:"
  stdin: $fetch-data.json    # Use .json for structured data
  # OR
  stdin: $fetch-data.stdout  # Use .stdout for raw text

Troubleshooting

If you get "llm.invoke received invalid response envelope":

• Check that OPENCLAW_URL points to a running gateway
• Verify your OpenClaw gateway is accessible

If stdin appears empty (word count: 0):

• Add a debug step: run: echo "DEBUG: $fetch-emails.stdout"
• Make sure the step ID matches exactly ($fetch-emails not $fetch_email_body)
• Check that the previous step actually produced output

Let me know if this helps! The examples repo has more patterns you can copy. 🦞

[RatkoJ]

Thanks bot, but no thanks.

Still can't get any of the examples working. I've managed to get Opus to "fix several bugs" in llm_invoke.ts and it works but I wonder if that was just a workaround?

Summary:
pipeline: llm_task.invoke fails with llm_task.invoke received invalid response envelope
command: openclaw.invoke --tool llm-task never gets the input from previous step

This is the full script:

name: word counter
description: Count the number of words in a given text using LLM.

steps:
  - id: make_words
    command: echo "One two three four five six"

  - id: count_words
    pipeline: >
      llm_task.invoke --prompt "How many words have been pasted below?"
    stdin: $make_words.stdout