Feat: Add base64 image support for results.#885
Feat: Add base64 image support for results.#885d42me wants to merge 1 commit intoPrimeIntellect-ai:mainfrom
Conversation
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
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.
| results_path: Path | None = None, | ||
| state_columns: list[str] | None = None, | ||
| save_results: bool = False, | ||
| image_mode: str = ImageMode.BASE64.value, |
There was a problem hiding this comment.
Default image mode breaks existing non-data-URI callers
Medium Severity
The public API methods generate, evaluate, evaluate_sync, run_rollout, and run_group all default image_mode to ImageMode.BASE64, while the lower-level utilities state_to_output and states_to_outputs default to ImageMode.PLACEHOLDER. The BASE64 default means existing callers — like verifiers/gepa/adapter.py which calls generate() without image_mode — will now fail with a ValueError if any prompt contains an image_url with an HTTPS URL (not a data URI), since _extract_data_uri_base64 requires data: URIs. Before this change, all images were silently replaced with [image]. The eval CLI path correctly overrides this to PLACEHOLDER when save_results is false, but the public Python API does not.
Additional Locations (2)
There was a problem hiding this comment.
@willccbb What do you think here? Can we introduce this breaking change for better DX? Or should we stay with the default placeholder?
1 participant
Description
Add base64 image store support for results
Type of Change
Testing
uv run pytestlocally.Checklist
Note
Medium Risk
Touches the core output-serialization and evaluation plumbing (including worker RPC types), so regressions could affect saved datasets or server-mode evaluations; changes are guarded by explicit modes and covered by new tests.
Overview
Adds an opt-in path to persist images in saved eval results by extracting
data:image/*;base64,...payloads into a per-messageimagesfield while still rendering[image]placeholders incontent.Plumbs
save_image_mode/image_modeandmax_image_base64_charsthroughevalCLI/config,Environment/EnvGroupgeneration, and worker client/server request types; metadata now recordssave_image_mode, and saving defaults toplaceholderwhensave_resultsis off.Hardens message serialization by validating data-URI base64 and size limits, and updates
sanitize_tool_callsto preserve already-serialized tool-call strings and other message fields; new tests cover base64 extraction, limit enforcement, CLI saving, and sanitization behavior.Written by Cursor Bugbot for commit a36c44d. This will update automatically on new commits. Configure here.