Description

NemoHermes v0.0.46 first Telegram response leaks raw tool call text to user.
After onboard with Telegram channel enabled, the first DM message receives
send_message: 'to telegram: "Hello! I'm Hermes..."' as raw text instead of
the agent executing the send_message tool properly. Subsequent messages work
correctly — the tool dispatch initializes between the first and second message.

Suggested improvement: Add tool-dispatch readiness gate after Hermes gateway
health check — ensure send_message tool is registered before accepting
Telegram messages, or queue first message until plugin init completes.

Environment

Device:        Brev GCP instance (n2d-standard-4)
OS:            Ubuntu 22.04.5 LTS (Linux 6.8.0-1048-gcp x86_64)
Architecture:  x86_64
Node.js:       v22.22.2
npm:           10.9.7
Docker:        Docker version 29.1.3
OpenShell CLI: 0.0.39
NemoClaw:      v0.0.46
Hermes Agent:  v2026.4.23

Steps to Reproduce

1. nemohermes onboard --fresh --yes-i-accept-third-party-software
2. Select NVIDIA Endpoints, Nemotron 3 Super 120B, enable Telegram channel
3. Enter valid Telegram bot token, set requireMention=false, enter user ID
4. Select Balanced policy tier with telegram preset enabled
5. Wait for onboard to complete successfully (sandbox Ready, Hermes gateway healthy)
6. Immediately send first message "hello" to the bot in Telegram DM
7. Observe first response from bot

Expected Result

Bot replies with a properly rendered message, e.g.:
"Hello! I'm Hermes Agent, your AI assistant. Type /help to see available commands."

Actual Result

First response contains raw tool call leaked as text:

📬 No home channel is set for Telegram. A home channel is where
Hermes delivers cron job results and cross-platform messages.
Type /sethome to make this chat your home channel, or ignore to skip.

📨 send_message: "to telegram: "Hello! I'm Hermes...""

Second message onwards ("你是谁啊") receives correct formatted response:
"我是 Hermes Agent，一个在 NVIDIA OpenShell 沙箱中运行的 AI 助手。"

Logs

OCSF sandbox logs show normal processing for the first message:
- sendChatAction (typing indicator) x multiple
- inference.local POST /v1/chat/completions (NVIDIA API called)
- sendMessage (response sent to Telegram)

No errors in gateway-persistent.log or openclaw runtime log.
The tool call was sent as text rather than being intercepted and executed
by the Hermes Telegram plugin, suggesting the plugin's send_message tool
was not yet registered when the first agent turn completed.

NVB#6195397