openclaw - 💡(How to fix) Fix DeepSeek preamble text leaks to Telegram after 3.13 → 4.5 upgrade (untagged assistant text bypasses commentary filter) [1 comments, 1 participants]

Description

After upgrading from 2026.3.13 to 2026.4.5, DeepSeek (deepseek-chat) cron jobs that deliver to Telegram are leaking preamble/reasoning text ("Let me fetch the data...", "Based on my research...", "Now I have enough information...") that was not visible in 3.13.

Environment

• OpenClaw: 2026.4.5 (3e72c03), upgraded from 2026.3.13
• Model: deepseek/deepseek-chat (V3)
• Delivery: Telegram (cron jobs)
• thinking: "off" set on all jobs
• exec-approvals: security=full, ask=off

Reproduction

Any cron job using DeepSeek that instructs the model to output a formatted template (market data, news digest, etc.) occasionally emits preamble text before the formatted output. This preamble is delivered to Telegram as part of the message.

Same prompts on 3.13 produced clean output consistently.

Root Cause

The 4.5 commentary-buffering fix (#61282, #59150, #59643) only suppresses text tagged with phase: "commentary" metadata or wrapped in <think> tags. DeepSeek emits preamble as plain untagged assistant text — no phase metadata, no thinking tags — so it bypasses all filtering.

Related: #42062 (same class of issue for GLM-5 via Ollama).

The behavioral regression from 3.13 suggests the old delivery pipeline had different buffering/truncation behavior that incidentally filtered this text before it reached Telegram.

Suggested Fix

A post-delivery filter or configurable output sanitizer that can strip text preceding a known output marker (e.g., first emoji character, or a configurable outputStartsWith pattern in the cron job payload) would catch untagged preamble regardless of model.

Workaround

Hardened prompt instructions ("CRITICAL RULE: first character must be 🇺🇸, delete any narration") partially mitigate but are not reliable — DeepSeek still slips.