Bug Report

Description

After the Telegram reasoning tag fix in #6328, the WhatsApp channel still exposes raw <think>...</think> and <final>...</final> tags from Gemini 2.5 Pro to users, even on OpenClaw v2026.5.27.

Root Cause

The fix in #6328 was Telegram-only — it added splitTelegramReasoningText() as a streaming coordinator in the Telegram bot module. The WhatsApp delivery pipeline was not updated equivalently.

Additionally, the @openclaw/whatsapp plugin auto-installs at whatever version is available at install time, and can lag significantly behind the core. Users upgrading the core to v2026.5.27 may still be running @openclaw/[email protected] — a gap of ~3 weeks of fixes.

Technical Details

Gemini thinking tag flow:

1. BUILTIN_REASONING_OUTPUT_MODES = { "google-generative-ai": "tagged" } -> system prompt injects <think>/<final> instructions
2. Gemini 2.5 Pro ALSO returns native thinking via API (thinkingBudget -> thinkingConfig)
3. promoteThinkingTagsToBlocks() skips if native thinking blocks already present -> inline tags remain in text blocks
4. WhatsApp delivery (deliverWebReply) calls normalizeWhatsAppPayloadTextPreservingIndentation -> sanitizeAssistantVisibleTextWithProfile(text, "history") -> stripReasoningTagsFromText(text, { mode: "strict" })

Edge case during streaming: If a <think> block is unclosed and the result is empty, the fallback in stripReasoningTagsFromText returns the thinking content itself — the suspected leak path for streaming/incomplete responses.

Steps to Reproduce

1. Configure OpenClaw with google/gemini-2.5-pro and reasoning: true, thinkingDefault: "medium"
2. Pair a WhatsApp account
3. Send a message that triggers reasoning
4. Observe <think>...</think><final>...</final> tags appear in the WhatsApp reply

Fix

Short-term: Ensure @openclaw/whatsapp plugin version is kept in sync with core version:
openclaw plugins update @openclaw/whatsapp

Long-term: Apply the same reasoning tag interception to the WhatsApp delivery pipeline as was done for Telegram in #6328. The WhatsApp monitor should use a splitWhatsAppReasoningText() equivalent that:

• Suppresses pure reasoning-only payloads (via isReasoningReplyPayload() — already present in v2026.5.27)
• Ensures the text normalization pipeline strips all reasoning tags before sending

Environment

• OpenClaw core: v2026.5.27
• @openclaw/whatsapp plugin: was v2026.5.7, updated to v2026.5.27 as workaround
• Model: google/gemini-2.5-pro with reasoning: true
• Platform: Linux (Debian 12)

Related

• Follows up on [Bug]: Gemini model leaks <think>/<final> tags to user output #6328 (Telegram fix; WhatsApp was not covered)
• User @aflag on [Bug]: Gemini model leaks <think>/<final> tags to user output #6328 still saw tag leaks after Telegram fix, suggesting other channels affected