* Build the render plan for an assistant turn: * - thinking and text parts render in place * - all tool-batch parts collapse into a single Task entr

* Build a render plan that groups all tool-call parts into a single Task * collapsible and all attachment parts into a single attachment strip at *

Render a frame, then splice each child iframe's rendered tree under its `- iframe` line.

Honors explicit profiles but upgrades the old shared temp profile.

codexEntrypoint reads codex-package.json's entrypoint from the tarball.

Keep changed lines plus a `radius` window of context; elide gaps with `…`.

* Pull one snapshot per agent in parallel. Falls back to a * lastUsedAt-only snapshot when the runtime doesn't implement * `getRowSnapshot` (tes

* Resize images that are oversized to a sane long-edge cap. JPEG/WebP * source files are re-encoded to JPEG; PNGs/GIFs that are already small * are

Convert Mind2Web dataset to unified format. Mind2Web format (JSON array or JSONL): { "annotation_id": "unique-id", "webs

Convert Online-Mind2Web dataset to unified format. Online-Mind2Web format: { "task_id": "unique-id", "website": "https:/

Convert WebVoyager dataset to unified format. WebVoyager format (JSONL): {"web_name": "Allrecipes", "id": "Allrecipes--0", "ques": "..."