Anthropic provider drops empty assistant turns, breaking role alternation

Summary

AnthropicProvider._as_message_params (chatlas/_provider_anthropic.py:884-887) drops assistant turns that have zero content blocks instead of sending a placeholder:

# Drop empty assistant turns to avoid an API error
# (all messages must have non-empty content)
if turn.role == "assistant" and len(content) == 0:
    continue

If a turn is dropped this way, it can produce two consecutive user-role messages in the outgoing request, which violates Anthropic's strict requirement that messages alternate user/assistant. This is the same class of bug ellmer just fixed (tidyverse/ellmer#1099, tidyverse/ellmer#1100; see also tidyverse/ellmer#711, tidyverse/ellmer#1070): ellmer now sends a "[empty string]" placeholder instead of dropping the turn, because dropping it both risks an alternation error and "confuses the model."

Note: chatlas already normalizes empty/whitespace text to "[empty string]" at the ContentText level (chatlas/_content.py:288-292), which covers the common case where a model returns an empty or whitespace-only text block (confirmed via the existing Databricks empty-response VCR cassette, tests/_vcr/test_provider_databricks/test_databricks_empty_response.yaml). The bug here is narrower: it's specifically about an AssistantTurn with a genuinely empty content list (contents=[]), which never goes through ContentText's normalization and instead hits the continue branch above.

Repro

from chatlas._provider_anthropic import AnthropicProvider
from chatlas._turn import AssistantTurn, UserTurn

provider = AnthropicProvider(
    name="Anthropic",
    model="claude-3-5-haiku-latest",
    api_key="dummy-key-not-used",
)

turns = [
    UserTurn("Respond with only two blank lines"),
    AssistantTurn(contents=[], finish_reason="success"),
    UserTurn("What's 1+1? Just give me the number"),
]

messages = provider._as_message_params(turns)
print([m["role"] for m in messages])
# ['user', 'user']  <-- two consecutive user messages; Anthropic's API would reject this

Open question for triage: it's not yet confirmed whether Anthropic's real response-parsing path can ever actually produce an AssistantTurn with contents=[] (as opposed to always synthesizing at least one, possibly-empty, ContentText block first, which would already get the "[empty string]" treatment upstream and never reach this branch). Regardless, the message-construction logic itself is demonstrably unsafe if such a turn is ever constructed (e.g. programmatically, or via a future code path), so the fix is worth making independent of how often it's hit today.

Expected fix

Mirror ellmer's fix: replace the continue with a placeholder content block, e.g. [{"type": "text", "text": "[empty string]"}], instead of dropping the turn.

References

• ellmer PRs: tidyverse/ellmer#1099, tidyverse/ellmer#1100 (issues: tidyverse/ellmer#711, tidyverse/ellmer#1070)
• chatlas: chatlas/_provider_anthropic.py:884-887, chatlas/_content.py:288-292

Start in chatlas/_provider_anthropic.py at AnthropicProvider._as_message_params (884-887), then run the provided AssistantTurn reproduction to inspect the outgoing roles. Check the related normalization in chatlas/_content.py and confirm the result preserves user/assistant alternation with placeholder content for an empty assistant turn.