Problem

When lossless-claw reassembles messages from the database, assistant messages that contain only thinking/reasoning/redacted_thinking blocks (with no visible text or tool calls) pass through the existing empty-content filter in the assembler because their content array is non-empty (length > 0).

These messages then reach OpenClaw core, where dropThinkingBlocks() strips all thinking blocks, leaving an empty content array (or a single text block with text: ""). The Bedrock provider's convertMessages function filters out text blocks where text.trim().length === 0, producing a completely empty content array that Bedrock rejects with:

ValidationException: The content field in the Message object at messages.N is empty.
Add a ContentBlock object to the content field and try again.

Why this is worse with lossless-claw

Without lossless-claw, the default context engine works with live conversation messages where thinking blocks are typically paired with visible text or tool calls. Lossless-claw faithfully reconstructs all stored message parts from the database (including raw thinking blocks), and the fresh tail includes these thinking-only messages unchanged — creating more opportunities for this edge case to surface.

This is particularly common with extended-thinking models (Claude with thinking enabled) where some assistant turns consist entirely of thinking blocks — for instance when the model reasons internally before delegating to tool calls that were later orphan-stripped by filterNonFreshAssistantToolCalls.

Reproduction path

1. Use an extended-thinking model (e.g. Claude Opus with thinking enabled)
2. Have lossless-claw enabled
3. Over a conversation, some assistant turns will have only thinking blocks
4. When context is reassembled, these messages survive the empty-content filter
5. OpenClaw core's dropThinkingBlocks() strips the thinking blocks
6. Bedrock rejects the empty content → ValidationException

Fix

Extend the cleanedEntries filter in the assembler to also detect and drop assistant messages whose content consists entirely of thinking/reasoning/redacted_thinking blocks. These messages carry no semantic value after thinking-block removal, so dropping them is safe and prevents the downstream empty-content error.

Changes

• src/assembler.ts: Add isThinkingOnlyContent() helper and extend the cleanedEntries filter
• test/assembler-blocks.test.ts: Add 9 unit tests for isThinkingOnlyContent()

Complementary fix

This pairs with openclaw/openclaw#71623 which addresses the same issue from the OpenClaw core side by using a non-empty placeholder in dropThinkingBlocks(). Both fixes are independent and provide defense-in-depth — the lossless-claw fix prevents the problematic messages from reaching the core pipeline at all.

Test results

All 784 tests pass (775 existing + 9 new).