Description

Telegram channel receives "Context overflow: prompt too large for the model" error messages while the webchat dashboard shows correct, complete responses for the same session/messages.

Environment

• Clawdbot version: 2026.1.24-3
• Model: anthropic/claude-opus-4-5
• Context usage: 42% (84k/200k) - well under limit
• Channels: Telegram + Webchat dashboard

Steps to Reproduce

1. Have both Telegram and webchat dashboard open for the same session
2. Send messages that trigger tool calls (e.g., exec with grep)
3. Observe: Dashboard shows full response, Telegram shows "Context overflow" error

Expected Behavior

Both channels should show the same response since they share the same session.

Actual Behavior

• Dashboard: Shows complete, correct response with tool results
• Telegram: Shows "Context overflow: prompt too large for the model. Try again with less input or a larger-context model."

The error appears multiple times on Telegram, interleaved with valid responses.

Investigation Results

1. No actual overflow: Session status shows 84k/200k tokens (42%)
2. No errors in session files: grep isError:true returns 0 results across all session files
3. Model is working: Responses are generated successfully (visible in dashboard)
4. Error not logged: The overflow errors sent to Telegram are not recorded in session .jsonl files

Evidence

Screenshots show:

• Same timestamp messages
• Dashboard: full response with tool output
• Telegram: "Context overflow" error (2x)

This suggests the error is generated in a Telegram-specific code path, not from the actual model/session.

Possible Causes

• Race condition in message delivery
• Retry mechanism failing on Telegram but not dashboard
• Stale error messages being replayed
• Different error handling paths for different channels

Workaround

None found - issue is intermittent but reproducible.