openclaw - 💡(How to fix) Fix [Bug]: Telegram reply context captured in session JSONL but stripped from LLM context — agent cannot see replied message [1 comments, 2 participants]

Bug Description

When a user replies to a message in Telegram (using the native reply/quote feature), the reply context data is correctly captured and stored in the session JSONL as a custom_message (openclaw.runtime-context), but it is stripped before reaching the LLM context.

The LLM only sees the runtime context header:

OpenClaw runtime context for the immediately preceding user message.
This context is runtime-generated, not user-authored. Keep internal details private.

But the full content that exists in the session JSONL includes:

{
  "has_reply_context": true,
  "reply_to_id": "29436",
  "Replied message": {
    "sender_label": "J.A.R.V.I.S.",
    "body": "original message text..."
  }
}

This means the LLM agent cannot see what message the user is replying to, making it impossible to understand the context of follow-up questions.

Steps to Reproduce

1. Have a Telegram conversation with the bot
2. Use Telegram's native reply feature to reply to a previous message
3. The bot agent cannot see the replied message content

Evidence

✅ Works: Session JSONL

The custom_message with customType: "openclaw.runtime-context" in the session JSONL contains the FULL Conversation info + Sender metadata + Replied message block, including has_reply_context: true, reply_to_id, and the complete body of the replied message.

❌ Broken: LLM Context

The same custom_message content, when assembled into the LLM context, is truncated to only the 2-line header. The Conversation info, Sender, and Replied message blocks are all stripped.

What We Tried

1. Set channels.telegram.contextVisibility to "all" (from "allowlist_quote") — did NOT fix the issue
2. Verified channels.telegram.replyToMode is set to "off" (this controls bot output reply behavior, not inbound context)

Environment

• OpenClaw version: 2026.4.26
• Channel: Telegram (direct message)
• OS: Linux (Oracle Cloud VPS, arm64)

Impact

This is a critical feature for conversational context. Users who use Telegram's reply feature to reference previous messages cannot have their agent understand what they're referring to. The agent appears to "forget" or "ignore" the reply context.

Related Issues

• This is similar to issues reported for other channels (#10872 for iMessage, #42266 for iMessage)

Proposed Fix Direction

The stripping likely occurs in the LLM context assembly pipeline, possibly in how custom_message content is projected into the LLM prompt. The metadata blocks (Conversation info, Sender, Replied message) should be preserved when building the LLM context, not stripped as "internal details."