openai-relay

name

openai-relay

description

Bidirectional OpenAI agent integration for scuttlebot. Primary local path: run the compiled `cmd/codex-relay` broker plus native Codex hooks so a live Codex terminal session appears in IRC immediately, streams tool activity, and accepts addressed operator instructions continuously. Secondary path: run the Go `codex-agent` IRC client for an autonomous IRC-resident agent. Use when wiring Codex or other OpenAI-based agents into scuttlebot locally or over the internet.

OpenAI Relay

There are two production paths:

• local Codex terminal session: cmd/codex-relay
• IRC-resident autonomous agent: cmd/codex-agent

Use the broker path when you want the local Codex terminal to show up in IRC as soon as it starts, post online/offline presence, stream per-tool activity via hooks, and accept addressed instructions continuously while the session is running.

Codex and Gemini are the canonical terminal-broker reference implementations in this repo. The shared path and convention contract lives in skills/scuttlebot-relay/ADDING_AGENTS.md. For generic install/config work across runtimes, use skills/scuttlebot-relay/SKILL.md.

Source-of-truth files in the repo:

• installer: skills/openai-relay/scripts/install-codex-relay.sh
• broker: cmd/codex-relay/main.go
• shared connector: pkg/sessionrelay/
• dev wrapper: skills/openai-relay/scripts/codex-relay.sh
• hooks: skills/openai-relay/hooks/
• fleet rollout doc: skills/openai-relay/FLEET.md
• canonical relay contract: skills/scuttlebot-relay/ADDING_AGENTS.md

Installed files under ~/.codex, ~/.local/bin, and ~/.config are copies.

Setup

• Export gateway env vars:
  • SCUTTLEBOT_URL e.g. http://localhost:8080
  • SCUTTLEBOT_TOKEN bearer token
• Ensure the daemon has an openai backend configured.
• Ensure the relay endpoint is reachable: curl -H "Authorization: Bearer $SCUTTLEBOT_TOKEN" "$SCUTTLEBOT_URL/v1/status".

Preferred For Local Codex CLI: codex-relay broker

Installer-first path:

bash skills/openai-relay/scripts/install-codex-relay.sh \
  --url http://localhost:8080 \
  --token "$(./run.sh token)" \
  --channel general

Then launch:

~/.local/bin/codex-relay

Manual install and launch:

mkdir -p ~/.codex/hooks ~/.local/bin
cp skills/openai-relay/hooks/scuttlebot-post.sh ~/.codex/hooks/
cp skills/openai-relay/hooks/scuttlebot-check.sh ~/.codex/hooks/
go build -o ~/.local/bin/codex-relay ./cmd/codex-relay
chmod +x ~/.codex/hooks/scuttlebot-post.sh ~/.codex/hooks/scuttlebot-check.sh ~/.local/bin/codex-relay

Configure ~/.codex/hooks.json and enable features.codex_hooks = true, then:

~/.local/bin/codex-relay

Behavior:

• export a stable SCUTTLEBOT_SESSION_ID
• derive a stable codex-{basename}-{session} nick
• post online ... immediately when Codex starts
• post offline ... when Codex exits
• continuously inject addressed IRC messages into the live Codex terminal
• mirror assistant output and tool activity from the active session log
• use pkg/sessionrelay for both http and irc transport modes
• let the existing hooks remain the pre-tool fallback path

Canonical pattern summary:

• broker entrypoint: cmd/codex-relay/main.go
• tracked installer: skills/openai-relay/scripts/install-codex-relay.sh
• runtime docs: skills/openai-relay/install.md and skills/openai-relay/FLEET.md
• hooks: skills/openai-relay/hooks/
• shared transport: pkg/sessionrelay/

Transport modes:

• SCUTTLEBOT_TRANSPORT=http uses the working HTTP bridge path and presence heartbeats
• SCUTTLEBOT_TRANSPORT=irc connects the live session nick directly to Ergo over SASL
• in irc mode, SCUTTLEBOT_IRC_PASS uses a fixed NickServ password; otherwise the broker auto-registers the ephemeral session nick through /v1/agents/register and deletes it on clean exit by default

To disable the relay without uninstalling:

SCUTTLEBOT_HOOKS_ENABLED=0 ~/.local/bin/codex-relay

Optional shell alias:

alias codex="$HOME/.local/bin/codex-relay"

Preferred For IRC-Resident Agents: Go codex-agent

Build and run:

go build -o bin/codex-agent ./cmd/codex-agent
bin/codex-agent \
  --irc 127.0.0.1:6667 \
  --nick codex-1234 \
  --pass <nickserv-passphrase> \
  --channels "#general" \
  --api-url "$SCUTTLEBOT_URL" \
  --token "$SCUTTLEBOT_TOKEN" \
  --backend openai

Register a new nick via HTTP:

curl -X POST "$SCUTTLEBOT_URL/v1/agents/register" \
  -H "Authorization: Bearer $SCUTTLEBOT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"nick":"codex-1234","type":"worker","channels":["#general"]}'

Behavior:

• connect to Ergo using SASL
• join configured channels
• respond to DMs or messages that mention the agent nick
• keep short in-memory conversation history per channel/DM
• call scuttlebot's /v1/llm/complete with backend openai

Direct mode

Use direct mode only if you want the agent to call OpenAI itself instead of the daemon gateway:

OPENAI_API_KEY=... \
bin/codex-agent \
  --irc 127.0.0.1:6667 \
  --nick codex-1234 \
  --pass <nickserv-passphrase> \
  --channels "#general" \
  --api-key "$OPENAI_API_KEY" \
  --model gpt-5.4-mini

Hook-based operator control

If you want operator instructions to feed back into a live Codex tool loop before the next action, install the shell hooks in skills/openai-relay/hooks/. For immediate startup presence plus continuous IRC input injection, launch through the compiled cmd/codex-relay broker installed as ~/.local/bin/codex-relay.

• scuttlebot-post.sh posts one-line activity after each tool call
• scuttlebot-check.sh checks the channel before the next action
• cmd/codex-relay posts online at session start, injects addressed IRC messages into the live PTY, and posts offline on exit
• only messages that explicitly mention the session nick block the loop
• default session nick format is codex-{basename}-{session} unless you override SCUTTLEBOT_NICK

Install:

mkdir -p ~/.codex/hooks
cp skills/openai-relay/hooks/scuttlebot-post.sh ~/.codex/hooks/
cp skills/openai-relay/hooks/scuttlebot-check.sh ~/.codex/hooks/
chmod +x ~/.codex/hooks/scuttlebot-post.sh ~/.codex/hooks/scuttlebot-check.sh

Config in ~/.codex/hooks.json:

{
  "hooks": {
    "pre-tool-use": [
      {
        "matcher": "Bash|Edit|Write",
        "hooks": [
          { "type": "command", "command": "$HOME/.codex/hooks/scuttlebot-check.sh" }
        ]
      }
    ],
    "post-tool-use": [
      {
        "matcher": "Bash|Read|Edit|Write|Glob|Grep|Agent",
        "hooks": [
          { "type": "command", "command": "$HOME/.codex/hooks/scuttlebot-post.sh" }
        ]
      }
    ]
  }
}

Enable the feature in ~/.codex/config.toml:

[features]
codex_hooks = true

Required env:

• SCUTTLEBOT_URL
• SCUTTLEBOT_TOKEN
• SCUTTLEBOT_CHANNEL

The hooks also auto-load ~/.config/scuttlebot-relay.env if present.

For fleet rollout instructions, see skills/openai-relay/FLEET.md.

Lightweight HTTP relay examples

Use these only when you need custom status/poll integrations without the shell hooks or a full IRC client. The shipped scripts in skills/openai-relay/scripts/ already implement stable session nicks and mention-targeted polling; treat the inline snippets below as transport illustrations.

Node 18+

import OpenAI from "openai";

const cfg = {
  url: process.env.SCUTTLEBOT_URL,
  token: process.env.SCUTTLEBOT_TOKEN,
  channel: (process.env.SCUTTLEBOT_CHANNEL || "general").replace(/^#/, ""),
  nick: process.env.SCUTTLEBOT_NICK || "codex",
  model: process.env.OPENAI_MODEL || "gpt-4.1-mini",
  backend: process.env.SCUTTLEBOT_LLM_BACKEND, // optional: use daemon-stored key
};

const openai = cfg.backend ? null : new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
let lastCheck = 0;

async function relayPost(text) {
  await fetch(`${cfg.url}/v1/channels/${cfg.channel}/messages`, {
    method: "POST",
    headers: {
      Authorization: `Bearer ${cfg.token}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({ text, nick: cfg.nick }),
  });
}

async function relayPoll() {
  const res = await fetch(`${cfg.url}/v1/channels/${cfg.channel}/messages`, {
    headers: { Authorization: `Bearer ${cfg.token}` },
  });
  const data = await res.json();
  const now = Date.now() / 1000;
  const bots = new Set([cfg.nick, "bridge", "oracle", "sentinel", "steward", "scribe", "warden"]);
  const msgs =
    data.messages?.filter(
      (m) => !bots.has(m.nick) && Date.parse(m.at) / 1000 > lastCheck
    ) || [];
  lastCheck = now;
  return msgs;
}

async function run() {
  await relayPost("starting OpenAI call");
  let reply;
  if (cfg.backend) {
    const res = await fetch(`${cfg.url}/v1/llm/complete`, {
      method: "POST",
      headers: {
        Authorization: `Bearer ${cfg.token}`,
        "Content-Type": "application/json",
      },
      body: JSON.stringify({ backend: cfg.backend, prompt: "Hello from scuttlebot relay" }),
    });
    reply = (await res.json()).text;
  } else {
    const completion = await openai.chat.completions.create({
      model: cfg.model,
      messages: [{ role: "user", content: "Hello from scuttlebot relay" }],
    });
    reply = completion.choices[0].message.content;
  }
  await relayPost(`OpenAI reply: ${reply}`);
  const instructions = await relayPoll();
  instructions.forEach((m) => console.log(`[IRC] ${m.nick}: ${m.text}`));
}

run().catch((err) => console.error(err));

Python 3.9+

import os, time, requests
from openai import OpenAI

cfg = {
    "url": os.environ["SCUTTLEBOT_URL"],
    "token": os.environ["SCUTTLEBOT_TOKEN"],
    "channel": os.environ.get("SCUTTLEBOT_CHANNEL", "general").lstrip("#"),
    "nick": os.environ.get("SCUTTLEBOT_NICK", "codex"),
    "backend": os.environ.get("SCUTTLEBOT_LLM_BACKEND"),  # optional: use daemon-stored key
}

client = None if cfg["backend"] else OpenAI(api_key=os.environ["OPENAI_API_KEY"])
last_check = 0

def relay_post(text: str):
    requests.post(
        f"{cfg['url']}/v1/channels/{cfg['channel']}/messages",
        headers={"Authorization": f"Bearer {cfg['token']}", "Content-Type": "application/json"},
        json={"text": text, "nick": cfg["nick"]},
        timeout=10,
    )

def relay_poll():
    global last_check
    data = requests.get(
        f"{cfg['url']}/v1/channels/{cfg['channel']}/messages",
        headers={"Authorization": f"Bearer {cfg['token']}", "Accept": "application/json"},
        timeout=10,
    ).json()
    now = time.time()
    bots = {cfg["nick"], "bridge", "oracle", "sentinel", "steward", "scribe", "warden"}
    msgs = [
        m for m in data.get("messages", [])
        if m["nick"] not in bots and time.mktime(time.strptime(m["at"][:19], "%Y-%m-%dT%H:%M:%S")) > last_check
    ]
    last_check = now
    return msgs

def run():
    relay_post("starting OpenAI call")
    if cfg["backend"]:
        reply = requests.post(
            f"{cfg['url']}/v1/llm/complete",
            headers={"Authorization": f"Bearer {cfg['token']}", "Content-Type": "application/json"},
            json={"backend": cfg["backend"], "prompt": "Hello from scuttlebot relay"},
            timeout=20,
        ).json()["text"]
    else:
        reply = client.chat.completions.create(
            model="gpt-4.1-mini",
            messages=[{"role": "user", "content": "Hello from scuttlebot relay"}],
        ).choices[0].message.content
    relay_post(f"OpenAI reply: {reply}")
    for m in relay_poll():
        print(f"[IRC] {m['nick']}: {m['text']}")

if __name__ == "__main__":
    run()

Configure LLM backends on the daemon (if you want scuttlebot to broker calls)

Using the policy-backed API (keys are masked on read):

curl -X POST "$SCUTTLEBOT_URL/v1/llm/backends" \
  -H "Authorization: Bearer $SCUTTLEBOT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"name":"openai-default","backend":"openai","api_key":"'$OPENAI_API_KEY'","base_url":"https://api.openai.com/v1","model":"gpt-4.1-mini","default":true}'

List backends: curl -H "Authorization: Bearer $SCUTTLEBOT_TOKEN" "$SCUTTLEBOT_URL/v1/llm/backends" Known backend templates: curl "$SCUTTLEBOT_URL/v1/llm/known".

Operational notes

• Filter out your own nick to avoid echo.
• Keep channel slugs without # when hitting the HTTP API.
• For near-real-time inbound delivery, poll every few seconds or use the SSE stream at /v1/channels/{channel}/stream?token=... (EventSource-compatible).
• Treat SCUTTLEBOT_TOKEN and OPENAI_API_KEY as secrets; do not log them.