Feat: Add base64 image support for results.

[d42me]

Description

Add base64 image store support for results

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

Note

Medium Risk
Touches core output serialization and the env client/server RPC surface; incorrect mode selection or base64 validation/limits could break saving or resume workflows for multimodal prompts.

Overview
Adds configurable image serialization for generated/saved results via new ImageMode (placeholder vs base64). In base64 mode, message_to_printable extracts/validates data:image/*;base64,... payloads into a per-message images field (with size accounting) and enforces an optional max_image_base64_chars limit.

Threads image_mode/max_image_base64_chars through Environment/EnvGroup rollout + server/client request paths and into state_to_output/states_to_outputs, records save_image_mode in GenerateMetadata, and exposes CLI support via --save-image-mode with a fail-fast default size cap; also updates sanitize_tool_calls to preserve pre-serialized strings and extra message fields. Tests cover env/CLI behavior, override precedence, and limit enforcement.

^{Written by Cursor Bugbot for commit d6bfbdb. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

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

[cursor]

Content separator changed from space to double-newline

Medium Severity

The content-part join separator in message_to_printable changed from " ".join(chunks) to "\n\n".join(chunks). This affects all callers regardless of image_mode, including the RL trainer's wandb logging path which calls messages_to_printable without image-related parameters. The output format changes from e.g. "text [audio]" to "text\n\n[audio]", potentially breaking downstream consumers expecting space-separated content.

 

[cursor]

Pydantic message path silently drops extracted images

Medium Severity

When message_to_printable receives a Pydantic message object (not a dict) with list content and image_mode=BASE64, it recursively processes a dict version which correctly extracts images into printable["images"]. However, model_copy(update={"content": printable.get("content", "")}) only copies the content string back to the Pydantic model — the images list is silently discarded. Any image data extracted in base64 mode is lost for Pydantic messages.

 

[cursor]

New function strip_nones_from_content is never used

Low Severity

strip_nones_from_content is defined but never imported or called anywhere in the codebase. A grep for its name returns only its definition. This is dead code that adds maintenance burden without providing value.