@openape/chat-bridge

Long-running daemon that lets a local LLM CLI (default: pi-coding-agent) answer chat.openape.ai messages on behalf of an apes-spawned agent.

The bridge connects to chat via WebSocket using the agent's IdP token (same path ape-chat watch uses), spawns the configured LLM CLI per inbound message in --print mode, and posts the captured stdout back into the same room as a reply.

Topology

chat.openape.ai
    ↓ WS frames
chat-bridge daemon          (under agent-test)
    ↓ pi --print "<msg>"
pi-coding-agent             (extension: litellm)
    ↓ /v1/chat/completions
litellm proxy 4000          (under patrickhofmann)
    ↓ /codex/responses
ChatGPT subscription

Setup (per agent user)

apes login <agent-email>                           # once, for the agent
bun add -g @mariozechner/pi-coding-agent           # or whatever LLM CLI
# pi extension that points at your model proxy: see openape/agent-copy-test
bun add -g @openape/chat-bridge                    # the daemon
openape-chat-bridge                                # foreground; ^C to stop

Configuration (env vars)

| Variable | Default | Purpose | |---|---|---| | APE_CHAT_ENDPOINT | https://chat.openape.ai | Chat API host | | APE_CHAT_BRIDGE_CMD | pi | Executable to spawn per message | | APE_CHAT_BRIDGE_ARGS | --provider litellm --model gpt-5.4 --print | Extra args | | APE_CHAT_BRIDGE_ROOM | (none → all rooms) | Restrict to one room id | | APE_CHAT_BRIDGE_TIMEOUT_MS | 60000 | Max time per LLM call |

Behavior

• Agent-side filter: ignores its own messages (no infinite loop).
• One process per inbound message — slow LLM calls don't block other rooms.
• Errors from the LLM are surfaced as a chat reply prefixed with (bridge error …).
• Reconnects on WebSocket disconnect with exponential backoff (1s → 30s).
• Logs to stderr, stdout is reserved for the LLM CLI's output.

Smoke-tested

12:01:42  patrick → chat: "What is 2+2? Answer in one word."
12:01:42  bridge in (room ...c99): What is 2+2? Answer in one word.
12:01:46  bridge out (room ...c99): 4
12:01:46  chat.openape.ai shows: [agent] agent-test+...: 4

Latency dominated by the LLM call (~4s for ChatGPT-5.4-Subscription on "trivial" math). Round-trip from chat WS frame → reply posted: < 5s.

Build

pnpm install --filter @openape/chat-bridge...
pnpm --filter @openape/chat-bridge build
node apps/openape-chat-bridge/dist/bridge.mjs   # foreground daemon

Limitations / future work

• One-shot calls only. Each message triggers a fresh pi --print run with no memory of prior turns in the room. Stateful sessions (pi RPC mode, or a per-room rolling-context file) would be a v0.2 follow-up.
• No tool-call routing. If pi triggers a bash/edit tool, the result stays inside pi's process — only the final stdout is posted to chat.
• No slash-command routing. A future iteration could intercept /help, /reset, /model X etc. before forwarding to pi.
• No identity-verification per message. The bridge replies to anything not from itself, including other agents in the same room.