fix: strip empty text blocks from assistant content before Bedrock/Anthropic send

[strye]

Summary

Fixes a crash where a turn dies with a non-retryable HTTP 400: messages: text content blocks must be non-empty error from Bedrock/Anthropic.

Root cause

_convert_assistant_message() in agent/anthropic_adapter.py guards against sending empty assistant content with:

effective = blocks or content
if not effective or effective == "":
    effective = [{"type": "text", "text": "(empty)"}]

This checks whether effective (a list) is truthy, not whether the text inside its blocks is non-empty. When the model returns assistant content shaped like [{"type": "text", "text": ""}] — e.g. right after a tool call errors out (observed after a failed clarify call with no question text) — that list is non-empty, so the guard passes it through untouched. Bedrock/Anthropic then reject the empty text block outright, and Hermes treats the resulting 400 as non-retryable, killing the turn.

Fix

Filter out text blocks whose text is empty/whitespace-only before the emptiness check, so a genuinely-empty result still falls through to the existing "(empty)" placeholder — otherwise the real content is preserved unchanged.

Repro

from agent.anthropic_adapter import _convert_assistant_message

_convert_assistant_message({"role": "assistant", "content": [{"type": "text", "text": ""}]})
# before: {"role": "assistant", "content": [{"type": "text", "text": ""}]}   -> 400 from API
# after:  {"role": "assistant", "content": [{"type": "text", "text": "(empty)"}]}

Test plan

• Ran tests/agent/test_anthropic_adapter.py (173 tests) before and after the change: 170 passed / 3 failed both times (the 3 failures are pre-existing and unrelated — TestRunOauthSetupToken, a MagicMock/json.loads issue, confirmed present on unmodified main too).
• Manually verified the repro above, plus sanity-checked normal text content and fully-empty content still behave as before.

[teknium1]

Thanks for isolating the truthiness failure in the normal assistant-content path; that behavior is still present on current main at agent/anthropic_adapter.py:1982-1986 and :2032-2035.

Problems

• The patch only filters the normal tail. Ordered replay returns earlier at agent/anthropic_adapter.py:1938-1979, while _sanitize_replay_block() reconstructs text from b.get("text", "") at :1874-1883; empty replayed text blocks therefore remain sendable. The cross-referenced fix(anthropic): filter blank text blocks in both normal and replay paths #63228 identifies this same second path.
• The new filter runs after _apply_assistant_cache_control_to_last_cacheable_block(). Since agent/prompt_caching.py:46-49 marks the final content part, dropping that blank part can discard a cache breakpoint instead of transferring it to the fallback block.

Suggested changes

• Normalize blank text before cache-control placement in the normal path and reject it in _sanitize_replay_block() for ordered replay.
• Add regression coverage for normal and replayed empty/whitespace-only blocks, plus a cached blank-only assistant turn.

Automated hermes-sweeper review.