Bug Description

When using a model with thinking/reasoning enabled, the session JSONL saves two copies of every assistant response:

1. The RAW version (with thinking blocks + text)
2. A cleaned version (text only, with parent pointing to the first message)

The session-memory hook then reads both messages via getRecentSessionContent() → extractTextMessageContent(), which extracts the same text from each. This results in every assistant line appearing twice in memory files.

Root Cause

In the message_end event handler, when an assistant response contains thinking content, two messages are persisted:

• First: { role: "assistant", content: [{ type: "thinking", ... }, { type: "text", ... }] } (raw)
• Second: { role: "assistant", content: [{ type: "text", ... }] } (cleaned, parent=first)

The session-memory hooks extractTextMessageContent() treats both as valid text messages, producing duplicate assistant: ... lines.

Steps to Reproduce

1. Use a model with thinking/reasoning enabled (e.g., DeepSeek)
2. Have a conversation with the agent
3. Run /new or /reset to trigger session-memory save
4. Open the generated memory file — each assistant message appears twice

Expected Behavior

Each unique assistant message should appear only once in the memory file.

Environment

• OpenClaw version: 2026.6.6 (also confirmed on 2026.6.1)
• OS: macOS (Darwin arm64)
• Model: deepseek/deepseek-v4-flash (with thinking enabled)

Suggested Fix

The cleanest fix is at the storage layer — the message_end handler should not persist both the raw and cleaned versions. Only one should be saved.

As a temporary workaround, the session-memory hooks getRecentSessionContent() can be patched to skip consecutive assistant messages with identical text:

let lastAssistantText = null;
// ... inside the loop:
if (role === "assistant" && text === lastAssistantText) continue;
allMessages.push(`${role}: ${text}`);
if (role === "assistant") lastAssistantText = text;

But this is a band-aid — the real fix should prevent duplicate storage.