Add codemode image support

[akhilles]

Renders to terminal if interactive + supported. Prints the JSON response of out dir if not.

---

---

Note

Medium Risk
Introduces new output paths (inline terminal escape sequences and filesystem writes) and changes how tool call results are serialized, which may affect consumers and error handling around image decoding and artifact paths.

Overview
Adds image-aware tool execution to the MCP codemoder runtime by preserving CallToolResultContent::Image blocks: ExecutionResult now includes images, tool results are serialized with typed content objects (image/resource_link/text), and images are collected during tool calls.

Updates the execute_tools meta-tool response to include image blocks alongside the JSON summary, and extends pcb mcp eval with --output-dir plus logic to either render PNGs inline in Kitty/Ghostty interactive terminals or decode base64 image data, write artifacts to a directory, and replace embedded image objects with image_file references in the printed JSON.

Enables the execute_tools tool in the agent skill config and adds the base64 dependency needed for decoding/encoding image data.

^{Written by Cursor Bugbot for commit c2bb9b0. This will update automatically on new commits. Configure here.}

[LK]

think this is unintentional?

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Unintentional render tool added to existing test

Low Severity

The render ToolInfo was added to the MockToolCaller in test_execute_with_tools, but that test never calls tools.render() — it only tests add and greet. The new test_image_content_preserved test creates its own separate caller with just the render tool, so this addition to the existing test is unnecessary. This is likely the change that was flagged as "unintentional" in the PR discussion.