1 file = 1 workflow
Breaking (soft — auto-migrated on open)
- Each workflow file now holds exactly one workflow block. Multi-block files trigger a one-click split: the first workflow stays in the original file, the rest are saved as sibling files whose basename is slugified from each block's name:. Hotkeys and event triggers that pointed at the original path stay attached to the first workflow — re-point them to the new files if needed.
- Workflow IDs dropped the #workflowName suffix. Hotkey and event-trigger settings are migrated automatically at plugin load; no manual cleanup required.
- workflow node lost its name property. Sub-workflow calls now reference the target by file path alone.
New
- skill-capabilities fenced block — SKILL.md now declares its workflows (path, description, inputVariables) inside a dedicated ```skill-capabilities YAML block in the body, instead of frontmatter. Frontmatter carries only user-facing metadata (name, description). Existing skills with legacy frontmatter workflows: keep working, with a one-time console warning suggesting migration.
- get_workflow_spec chat tool — The chat LLM can now request the authoritative workflow node specification (full spec or specific node types) when explaining, debugging, or authoring workflow YAML. No more hallucinated node properties.
- Auto-derived inputVariables — Saving a workflow inside a skill now auto-updates its SKILL.md inputVariables list from the actual {{var}} references in the node graph, so the LLM always sees what it needs to pass.
- Create-skill safety — The AI "Create" flow rejects output paths that already contain a workflow, and defaults to workflows/{{name}} even when a file is already open.
Fixed
- @mentions with spaces / Unicode / special characters now resolve correctly. The old regex-based extractor captured only up to the first whitespace ([^\s@]+), so paths like ＠My Folder/Note.md, @メモ/日本語.md, or @notes/a (v2)+.md silently failed. Both chat and AI-workflow flows now scan the vault file list longest-path-first and match literally, so any vault file path works.
- @foo.md no longer leaks into longer tokens like @foo.md/child, @foo.md_backup, @foo.md-1, or @foo.md.backup — the whole ＠token must be exactly a vault path, otherwise no splice happens.
- Adjacent bare file-path mentions (e.g. read a.md b.md done) — the second path is no longer skipped by the trailing-whitespace boundary consuming its leading space.
- Non-markdown file mentions in the workflow-AI prompt (@workflows/foo.yaml, @config.json, @diagram.canvas) resolve again.
- get_workflow_spec accepts bare-name and comma-separated nodeTypes strings, not just JSON arrays — works around small-model schema drift.