AI workflow & skill overhaul
This release reworks the AI-powered generation flow around an interactive Plan → Generate → Review pipeline, separates workflow and skill creation into dedicated entry points, and adds a long-awaited "Modify Skill with AI" flow.
✨ New
- Planning phase — Before YAML is generated, the AI produces a plain-language plan (in your configured language, written for non-engineers). Confirm with OK, request changes with Re-plan, or cancel.
- User-driven review loop — After generation the AI reviews its own output. Zero-issue reviews proceed automatically; flagged issues offer OK / Refine / Cancel. Refine regenerates and re-reviews so the final review always matches the accepted YAML. OK on a review with issues requires an explicit "are you sure?" confirmation.
- Separate "Create workflow with AI" and "Create skill with AI" buttons — The old checkbox is gone. Each button opens a modal tailored to its mode (labels, placeholder text, locked output path for skills). A short explainer below the buttons describes the difference between workflow and skill so you know which to pick.
- Modify Skill with AI — Open a
SKILL.mdand click the new modify button to edit the skill as a whole. The AI updates both the SKILL.md instructions body and the referenced workflow file in one pass while preserving frontmatter (name, description, workflows entries including any per-workflowname). - Plan / thinking / review inspection — The preview and confirm modals expose collapsible Plan / Thinking / Review sections with copy buttons, so you can review the AI's reasoning alongside the generated YAML.
- Clickable skill chips — Active skill chips in the input area and on assistant messages open the matching
SKILL.md. Built-in skills render as static labels. - Failed-workflow recovery — When a skill workflow fails during a chat, the tool call shows an Open workflow button that opens the file and switches the Gemini view to the Workflow / skill tab. A hint line points you at the Modify-with-AI + Reference-execution- history recovery path.
🔧 Improvements
- Plan and review output are rendered as Markdown with localized severity labels, so non-engineers can read them comfortably.
- Confirm modal now has a scrollable middle area so textarea and action buttons stay visible even with long content.
- Copy buttons on AI Explanation, Plan, Thinking, and Review sections.
- Active status bar is emphasized during refinement so the modal never looks idle while work is still happening.
- The Workflow tab is renamed Workflow / skill (localized).
- Workflow spec docs clarify
variable.valueis optional,:jsoninterpolation only escapes content (no outer quotes), andjson.sourcemust be a bare variable name. A new "How workflow output reaches the user" section explains how skill workflow variables are returned to the chat AI automatically.
🐛 Fixes
run_skill_workflowerrors now include the workflow path and a don't-auto-retry instruction so the assistant surfaces the error instead of looping.- Paste-response section in the AI modal correctly hides until "Copy prompt" is clicked; the section scrolls into view and the textarea gains focus after copy.
- Thinking section in the generation modal has a bounded scrollable height that actually produces a scrollbar in Electron.
variablenodes preserve caller-provided values whenvalueis omitted, enabling input-declaration patterns for skills and hotkeys..ai-workflow-paste-section.is-hiddennow has a CSS rule, so the paste UI no longer appears in modes where it shouldn't.
