add-openagents

Add OpenAgents Channel

Adds a native NanoClaw channel (channel_type: openagents) that lets an OpenAgents Workspace talk to one of your Agent Groups. It plumbs into the normal router/delivery path like the built-in cli channel — no database access, no host special-casing. The channel owns a local Unix socket at data/openagents/bridge.sock; the OpenAgents agent-connector (@openagents-org/agent-launcher, adapter nanoclaw) is the client and does all Workspace IO. No OpenAgents token is stored in NanoClaw.

Install

Pre-flight (idempotent)

Skip to Wire an Agent Group if all of these are already in place:

• src/channels/openagents.ts exists
• src/channels/index.ts contains import './openagents.js';

Otherwise continue. Every step below is safe to re-run.

1. Copy the channel module

The channel source ships with the OpenAgents agent-launcher. Copy it from the installed package (preferred), or fetch it from the OpenAgents repo:

# From the installed launcher package (path may vary by install method):
cp "$(node -e "process.stdout.write(require.resolve('@openagents-org/agent-launcher/nanoclaw-channel/openagents.ts'))" 2>/dev/null \
   || echo "$HOME/.openagents/runtimes/nanoclaw/node_modules/@openagents-org/agent-launcher/nanoclaw-channel/openagents.ts")" \
  src/channels/openagents.ts

# …or fetch from source:
# curl -fsSL https://raw.githubusercontent.com/openagents-org/openagents/develop/packages/agent-connector/nanoclaw-channel/openagents.ts \
#   -o src/channels/openagents.ts

The OpenAgents Launcher can also install this for you automatically — it calls the same copy + barrel-patch. This skill is the manual / auditable path.

2. Append the self-registration import

Append to src/channels/index.ts (skip if already present):

// openagents channel (OpenAgents Workspace bridge)
import './openagents.js';

3. Restart the NanoClaw host

The channel self-registers on startup. Restart the host (./nanoclaw.sh restart, or your launchd/systemd unit) so it loads. After restart you should see openagents channel listening in the logs and a data/openagents/bridge.sock file.

Wire an Agent Group

Routing requires a wiring from an openagents messaging group to your Agent Group. Creating wirings is approval-gated in NanoClaw — do it explicitly:

1. Pick (or create) the Agent Group you want to expose: ncl groups list.

2. The OpenAgents connector addresses each Workspace channel as platform_id = oa:<workspace>:<channel> on channel_type = openagents. The first inbound message auto-creates that messaging group.

3. Wire it to your Agent Group with an always-engaging mode (DM-style mention, or pattern .), approving when prompted:

   ncl wirings create \
     --messaging-group-id <openagents-mg-id> \
     --agent-group-id <your-agent-group-id> \
     --engage-mode mention \
     --session-mode shared
   

   session-mode shared gives one isolated NanoClaw Session per Workspace channel. Use mention (engages on DM-style messages) or pattern with --engage-pattern . to always engage.

Notes & limits

• The host (and therefore this channel) sees the agent's outbound chat messages and a coarse typing/working indicator — not granular in-container tool calls. The bridge maps typing→status, chat→text, and surfaces errors; fine-grained tool-call traces are not exposed by NanoClaw's channel surface.
• Local IPC is authenticated. The socket lives in a 0700 data/openagents/ dir (socket 0600); the connector presents a random per-host secret (data/openagents/secret, 0600) in its handshake. Only one connector is allowed at a time. The secret never leaves the host. Trust boundary = the current OS user — the perms + secret defend against other local users, not a process running as the same user.
• Delivery is at-least-once with dedup, not exactly-once. Each reply is PERSISTED to data/openagents/outbox/ (0600) before sending and held until ACKed, so un-ACKed replies survive a Channel/host restart (replayed on reconnect). The connector ACKs after the reply is persisted in the Workspace and dedups by persisted outId. The outbox is bounded + TTL'd and drops + signals on overflow/expiry; a crash between deliver and persist, or a cleared state dir, can drop or re-display a reply. The outbox holds reply bodies (local sensitive cache) — never the secret/token/credentials.
• Stop = detach + fresh session, not cancel. NanoClaw exposes no outbound-drained signal (and container_status: stopped ≠ drained), so the connector never reuses the stopped session. On Stop it rotates a per-channel thread epoch → the next message opens a fresh NanoClaw session (per-thread routing under the same wiring); the old session's replies carry the authoritative old thread_id and are dropped. The container task may keep running — use NanoClaw's own controls for true cancellation. Wire session_mode: per-thread for clean, concurrent fresh sessions.
• Auto-install is gated to the exact verified NanoClaw commit — a different/ unknown commit is refused (admin-only force can override; structural check is always enforced).
• Removing the channel: run the OpenAgents uninstaller, or delete src/channels/openagents.ts and drop the marker-delimited barrel block, then restart. This never stops the host or other channels.