Model ignores conversation context when user gives short replies

Description

When using Clawdbot with Claude Opus 4.5 via Telegram, the model frequently ignores conversation context when the user provides short replies.

Example

1. User asks: "Save Telegram conversations to Obsidian as summaries"
2. Bot asks: "How often? Daily? Per session? Real-time?"
3. User replies: "リアルタイム" (real-time)
4. Bot responds: "What do you want to do in real-time? Web search? Stock prices? Weather?"

The bot completely ignored that "リアルタイム" was an answer to the previous question about frequency, and instead interpreted it as a new standalone command.

Expected Behavior

Short replies like "yes", "that one", "real-time", "ok" should be interpreted as responses to the immediately preceding question/topic, not as new conversation starters.

Environment

• Model: anthropic/claude-opus-4-5
• Platform: Clawdbot gateway on Linux (CT104)
• Channel: Telegram
• Session memory: enabled
• Compaction mode: safeguard

Notes

This behavior does not occur when using Claude Code CLI directly - only through Clawdbot's gateway. The same model maintains context properly in Claude Code.

Possible causes:

• How conversation history is formatted/passed to the model
• Session management differences
• System prompt weighting