aop-launcher 0.0.3.dev0

Lightweight launcher for local AI agent workflows

aop_launcher

Agent Orchestration Platform Launcher.

aop_launcher is a tiny terminal runner for local AI agent workflows. It reads a YAML file, runs agents in sequence, optionally runs local shell tools for each agent, sends the prompt to either a local command or an OpenAI-compatible chat completion endpoint, and streams the result to the terminal.

This first 0.0.1.dev0 release is intentionally small. It does not provide remote execution, scheduling, retries, memory stores, or a hosted service.

Install

pip install aop-launcher

OpenAI-compatible workflow

Use this with local or hosted inference servers that expose /v1/chat/completions:

llm:
  provider: openai-compatible
  base_url: "https://api.openai.com/v1"
  api_key_env: "OPENAI_API_KEY"
  model: "gpt-5.4-mini"
  temperature: 0.2
  max_completion_tokens: 800

agents:
  - name: assistant
    system: "You are a concise, practical assistant."
    prompt: |
      Answer this request:
      {input}

Run it:

aop-launcher workflow.yml "Draft a release checklist"

Common endpoint settings can be overridden from the terminal:

aop-launcher workflow.yml "hello" \
  --base-url https://api.openai.com/v1 \
  --model gpt-5.4-mini \
  --api-key-env OPENAI_API_KEY

If neither api_key nor api_key_env is configured, no Authorization header is sent.

Command workflow

llm:
  command: "ollama run llama3.2"

tools:
  now: "date"

agents:
  - name: planner
    tools: ["now"]
    prompt: |
      Create a short plan for this task:
      {input}

      Tool results:
      {tool_results}

  - name: writer
    prompt: |
      Write the final answer from this plan:
      {input}

Run it:

aop-launcher workflow.yml "Draft a release checklist"

You can also pass the LLM command from the terminal:

aop-launcher workflow.yml "hello" --llm-command "ollama run llama3.2"

Prompt variables

Agent system and prompt templates can use:

• {input}: previous agent output, or the initial terminal input for the first agent.
• {initial_input}: original terminal input.
• {tool_results}: output from shell tools configured on the current agent.
• {history}: prior agent outputs with agent names.

JSON output

Set output_format: json on an agent to require valid JSON output:

agents:
  - name: classifier
    output_format: json
    prompt: |
      Return valid JSON with keys: severity, summary.
      {input}

The JSON text is passed to the next agent unchanged.

Fanout

Use fanout to run several sub-agents against the same input and join their outputs:

agents:
  - name: review_pack
    fanout:
      - name: correctness
        prompt: "Review for correctness: {input}"
      - name: tests
        prompt: "Review for missing tests: {input}"

Fanout children run sequentially. This is a small fanout/join helper, not a DAG engine.

Conditions

Use condition.input_contains to skip an agent unless the previous output contains a string:

agents:
  - name: refine
    condition:
      input_contains: "NEEDS_REFINEMENT"
    prompt: "Refine this: {input}"

Add negate: true to run only when the previous output does not contain the string.

Examples

Sample workflows live in examples/:

• adr_writer.yml
• bug_triage.yml
• openai_basic.yml
• commit_message.yml
• conditional_refine.yml
• research_brief.yml
• code_review.yml
• dependency_risk_review.yml
• incident_postmortem.yml
• json_output.yml
• parallel_review.yml
• pull_request_review.yml
• release_notes.yml
• shell_tools.yml
• test_plan.yml

Tool safety

Workflow tools are raw local shell commands. Only run workflows you trust.