v7.28

Release Notes

This update introduces support for the phase feature for eligible models in Next-Gen Assistants, along with other bug fixes and improvements.

Assistants

New Features

• Introduces support for the phase parameter in assistant responses. Starting with GPT-5.4, the Responses API will mark an assistant message as intermediate commentary or the final answer. The model uses the phase parameter to distinguish working commentary from the completed answer in previous responses. This matters most in multi-step flows with preambles, tool-related updates, or multiple assistant messages in the same turn. Missing or dropped phase can cause preambles to be treated as final answers, which can degrade quality on multi-step tasks.
• Using the new phase parameter in eligible models, the Next-Gen Assistant Responses API handler only stops when it detects a second assistant message marked as final answer right after the previous.

Lecture Videos

Note

This release adds new capabilities for Lecture Videos, which remain under active development and are not yet officially supported in Groups. APIs, features, and UI/UX may change before final release.

Lecture Video mode is currently visible only to institutional admins when creating an assistant.

New Features

• Setting up Lecture Videos assistants now requires a manifest file with information about comprehension questions. Only single-select multiple choice questions are currently supported.
• Added server endpoints allowing users to upload Lecture Videos when creating or updating assistants, similar to file uploads for File Search and Code Interpreter.
• Added database support for (intro and outro) narration files for each question.

Updates & Improvements

• Lecture Videos and the underlying Lecture Video Stored Objects are now separate in the database, allowing a single LVSO to point to multiple LVs. This change allows easy copying of lecture-video mode assistants.

Resolved Issues

• Fixed: When copying a Lecture Video assistant, the Lecture Video configuration is not copied to the new assistant.

Known Issues

• Because manifest files for Lecture Video mode are not supported in the web client, Lecture Video assistants are not editable in the UI.

Internal

Updates & Improvements

• Removed the hard requirement for python from create-worktree.sh, so the script now matches the repo’s uv-based setup.
• Upgrade devalue from 5.6.3 to 5.6.4 to resolve devalue has prototype pollution in devalue.parse and devalue.unflatten and Sveltejs devalue's devalue.parse and devalue.unflatten emit objects with __proto__ own properties.

Deployment Information

Schema Upgrade	Migration Script	Permissions Update	Task Definition Update	Configuration Update
YES	YES	YES	No	YES

Deployment Details

• Schema Upgrade: Updates to support Lecture Video Questions feature.
• Migration Script: Run uv run python -m pingpong db m07_backfill_lecture_video_content_lengths to backfill content lengths and add uploader permissions to Lecture Videos (see below).
• Permissions Update: Added new lecture_video type.
• Configuration Update: Set the new [lecture_video_audio_store], which expects an AudioStoreSettings object.

Related PRs

• fix: do not require python for create-worktree by @ekassos in #1523
• feat(lv): add question manifest, client uploads by @ekassos in #1524
• fix(GPT-5.4): store and pass phase back to Responses API by @ekassos in #1526
• feat: relax run termination logic for phase-capable models by @ekassos in #1528
• chore(deps): bump dompurify from 5.6.3 to 5.6.4 by @ekassos in #1529

Full Changelog: v1148+srv579.web408...v1153+srv582.web409