AnthropicLlm.part_to_message_block() iterates "content" string char-by-character when LoadSkillResourceTool is used with Claude

[mcburlage]

Describe the Bug

part_to_message_block() in anthropic_llm.py handles the "content" key by iterating over it with a for loop, but does not first check isinstance(response_data["content"], list). In Python, iterating over a string yields individual characters, so when LoadSkillResourceTool returns {"skill_name": ..., "file_path": ..., "content": "<file text>"} (where "content" is a plain string), Claude receives the file content character-by-character with newlines between each character:

"H\ne\nl\nl\no" instead of "Hello"

Relationship to #4779

Issue #4779 fixed the case where no "result" or "content" key was present (arbitrary dicts now fall through to json.dumps). This bug is different: "content" IS present, so the fallback is never reached, and the existing for item in response_data["content"] loop runs on a string.

Steps to Reproduce

1. Create an agent using AnthropicLlm (any Claude model on Vertex AI)
2. Add a SkillToolset with a skill that has a reference file
3. Ask the agent to call load_skill_resource on that file
4. The agent receives garbled content — each character on its own line

Or deterministically, without an LLM:

from google.adk.models.anthropic_llm import part_to_message_block
from google.genai import types

part = types.Part.from_function_response(
    name="load_skill_resource",
    response={"skill_name": "my-skill", "file_path": "references/doc.md", "content": "Hello"},
)
block = part_to_message_block(part)
print(repr(block["content"]))  # 'H\ne\nl\nl\no'

Expected Behavior

block["content"] should be "Hello".

Root Cause

In anthropic_llm.py, the "content" branch lacks a type guard:

# Current (buggy)
if "content" in response_data and response_data["content"]:
    for item in response_data["content"]:   # iterates chars if content is a string
        ...

# Fix: guard with isinstance
if "content" in response_data and isinstance(response_data["content"], list) and response_data["content"]:
    for item in response_data["content"]:
        ...

LoadSkillResourceTool.run_async returns {"content": <string>} by design, so this affects any agent using skills with reference files and a Claude model.

Environment

• google-adk version: 1.30.0
• Model: Claude (Vertex AI via AnthropicLlm)

[pandego]

This looks like the kind of bug that is safest to fix by branch ordering, not just by a narrow guard.

I would handle content as a scalar string before any iterable/list path, or at minimum gate the current branch with isinstance(..., list). That keeps the existing Anthropic block-list behavior intact while avoiding the accidental "iterate a string one character at a time" case.

The regression tests I would want here are:

• content is a plain string
• content is a list of Anthropic blocks
• content is "" (empty but valid payload)
• no content key, so the arbitrary-dict/json fallback still works

That last two are easy to miss, and they help lock the contract down for LoadSkillResourceTool style responses.