Handle ImageContent in MCP tool results (pass images to vision-capable LLMs)

[faisalishfaq2005]

Problem

When an MCP tool returned an ImageContent block, the agent discarded the
actual image data and forwarded only a text placeholder to the LLM:

# agent/core/tools.py (before)
elif isinstance(item, ImageContent):
    parts.append(f"[Image: {item.mimeType}]")  # data thrown away

As a result, MCP tools returning screenshots, charts, diagrams, or other
visual outputs provided no usable visual information to the model — even
when using vision-capable models such as Claude or GPT-4o.

Solution

Renamed convert_mcp_content_to_string to
convert_mcp_content_to_llm_content
and updated its return type from str to str | list[dict].

Behavior

• Text-only results still return a plain str
  (fully backward-compatible).
• Results containing at least one ImageContent now return a list of
  OpenAI-style content blocks (text + image_url).

LiteLLM already supports this format and automatically translates it into
provider-native formats (Anthropic image source blocks, etc.).

Image block format

{
    "type": "image_url",
    "image_url": {
        "url": f"data:{mimeType};base64,{data}"
    }
}

ImageContent.data is already base64-encoded per the MCP spec, so no
additional encoding step is required.

Since LiteLLM's Message(content=...) already accepts both str and
list, no changes were required to the core message-construction flow in
agent_loop.py.

Additional fix

agent_loop.py also contained a was_edited note-prepend implemented via
an f-string, which implicitly converted list-based content into a string
representation.

Updated the prepend logic to branch on the output type before modifying the
content.

Files Changed

agent/core/tools.py

• Renamed conversion function
• Updated return type
• Reworked conversion logic into a single-pass content block collector
• Added support for OpenAI-style image content blocks

agent/core/agent_loop.py

• Fixed was_edited prepend handling for list-based content

tests/unit/test_mcp_content_conversion.py

Added 18 unit tests covering:

• empty input
• text-only output (str)
• single image output (list)
• mixed text + image ordering
• multiple images
• embedded resources
• unknown content type fallback
• return type contracts

Testing

uv run --extra dev pytest tests/unit/test_mcp_content_conversion.py -v
# 18 passed

Pre-existing test failures (unrelated)

Two tests fail on Windows before and after this change:

• test_session_uploader — imports fcntl (Linux/macOS only)
• test_prioritize_backlog — hardcodes a Unix path separator

Both fail on an unmodified main branch on Windows. Out of scope for this PR.