[Bug] AsyncAnthropicVertex returns empty content in tool-calling loops (deriver/dream)

[lucharo]

Describe the bug

When using AsyncAnthropicVertex (Vertex AI) as the Anthropic provider, the deriver and dream services intermittently receive responses where anthropic_response.content is None or empty. This causes NoneType crashes throughout the tool-calling loop in _execute_tool_loop and in process_representation_tasks_batch.

The same API calls work correctly when made directly via the AnthropicVertex SDK outside of Honcho.

Is this a regression?

Unknown — this is a new deployment using Vertex AI, which isn't currently supported upstream (see PR #535).

To reproduce

1. Configure Honcho with AsyncAnthropicVertex as the Anthropic provider (via ANTHROPIC_VERTEX_PROJECT_ID + CLOUD_ML_REGION env vars)
2. Start the deriver
3. Let it process messages — process_representation_tasks_batch crashes at response.content (line ~164 in deriver.py)
4. Trigger a dream via POST /v3/workspaces/{id}/schedule_dream — the dream specialist's tool loop gets None from call_func() on every iteration

Expected behaviour

Responses from AsyncAnthropicVertex should have non-null content and usage, same as AsyncAnthropic.

Error traces

Deriver crash (deriver.py:164):

AttributeError: 'NoneType' object has no attribute 'content'

Tool loop crash (clients.py:797):

total_cache_read_tokens += response.cache_read_input_tokens
AttributeError: 'NoneType' object has no attribute 'input_tokens'

Dream crash after adding guards:

ValueError: LLM call returned None after 12 attempts

Root cause analysis

The Anthropic handler in honcho_llm_call_inner (clients.py ~line 1774) returns None when anthropic_response.content is empty. The callers (_execute_tool_loop, process_representation_tasks_batch, dream specialists) don't handle None returns.

The _execute_tool_loop function (clients.py ~line 791) assumes call_func() always returns a non-null HonchoLLMCallResponse. When it doesn't, the token accumulation lines crash.

Suggested fix

Add null guards at:

1. clients.py:~791 — check response is None after call_func(), retry or raise
2. clients.py:~1778 — log and return None when Anthropic content is empty (already returns None implicitly for some code paths)
3. deriver.py:~154 — guard response and response.content before accessing

Your environment

• OS: macOS (Docker containers running Honcho)
• Honcho Server Version: latest main (commit from April 2026)
• Anthropic SDK: 0.83.0 with [vertex] extras
• Provider: AsyncAnthropicVertex with region=global
• Models: claude-haiku-4-5, claude-sonnet-4-6

Additional context

Direct API calls to AnthropicVertex work correctly (including tool calling). The issue only manifests inside Honcho's tool loop, suggesting a message formatting or state accumulation issue in multi-turn tool conversations via the Vertex endpoint. Related: PR #535 adds Vertex AI support.

[lucharo]

Found the root cause — it's not a Vertex API issue at all.

AsyncAnthropicVertex is not a subclass of AsyncAnthropic. The match client: / case AsyncAnthropic(): pattern in honcho_llm_call_inner (lines 1701, 2393) never matches Vertex clients. The function falls through and returns None implicitly.

Fix: case AsyncAnthropic() | AsyncAnthropicVertex():

Updated in PR #535.