workflow-agent

Warning

Vibe coded: This codebase was generated using an agent (partially by workflow-agent itself), testing the limits of this approach.

An Obelisk app in which the workflow is the agent. It holds a provider-neutral chat history, drives an LLM over one of three wire APIs (Anthropic Messages, OpenAI Chat Completions, OpenAI Responses), dispatches the model's tool calls to real Obelisk activities, and feeds the results back, all as durable, replayable workflow state.

The core (agent loop + LLM router + web UI) is generic; a runtime pack supplies the use case (system prompt + tools). This repo ships one pack, obelisk-control, which inspects and modifies the Obelisk instance it runs on.

Requirements

• Obelisk — the runtime that serves this deployment. Use nix develop for the pinned toolchain.

• AGENT_MODELS — the model catalog, required: a JSON array pointing each model at an OpenAI- or Anthropic-shaped HTTP endpoint. Two ready-made catalogs ship:

  • models.local.json — the sibling agent-backed-llm-server (a Claude/Codex subscription in docker, keyless on :9190).
  • models.exe-gateway.json — the exe.dev LLM gateway (Anthropic + OpenAI + Fireworks). Requires an exe.dev account; the entries point at http://localhost:7070, so forward the gateway to that local port first:

    ssh -L 7070:169.254.169.254:80 <yourinstance>.exe.xyz

  • models.openrouter.json — OpenRouter (Claude, GPT, DeepSeek, and a free Qwen3 Coder model). Needs an API key; the key stays secret (injected into the outbound header at the edge, never seen by the JS):

    export OPENROUTER_API_KEY=sk-or-...

  Any other compatible endpoint (Anthropic/OpenAI directly, vLLM, Ollama, …) works too — add an entry pointing at it.

Run

Set the required catalog, then serve:

ln -sf models.local.json models.json      # pick a catalog
export AGENT_MODELS="$(cat models.json)"   # or use direnv (.envrc-example)
just serve                                 # obelisk server run -d deployment.toml

Then open the web UI on the webhook port (default 8080), or submit via the API:

curl -X POST http://127.0.0.1:8080/api/submit \
  -H content-type:application/json -d '{"prompt":"Summarise recent executions.","backend":"claude"}'

backend is the model id from the catalog (empty selects the first entry). Each turn is a separate llm.completion activity and each tool call its own child execution, so a run is fully durable and replayable; inspect it with the web UI or the standard Obelisk WebAPI / CLI.