Added model-driven image generation. Models that support it can generate images during a turn, and generated images are persisted in session history and exports Agents can now create and manage scheduled tasks and a durable todo agenda. Schedules are scoped to the workspace that registered them Skill slash commands now load through the skills tool instead of being pasted into the user message. The persisted transcript records what you typed (`/my-skill ...`) rather than the whole SKILL.md body, and skill instructions arrive once instead of twice. Workflows keep textual expansion Fixed provider-executed tool activity being dropped entirely — every tool the Claude Code provider ran inside its own session modified the workspace with no tool activity in runtime events, transcripts, or the UI.

These now surface as observational tool events Fixed `PreToolUse` hook `contextModification` never reaching the model on the next engine; it is delivered again as a `` block stamped with the tool name and call id, hidden from user-facing transcripts Fixed `PostToolUse` hooks running fire-and-forget with their output discarded. They are awaited now (120s bound) and their `contextModification` and…