Plan motion next phase

Current — the latest status report. Reports freeze once published; corrections appear in later reports.

Motion — Next Development Phase

Context

Motion just went through a security/correctness hardening pass (Tauri detection, CLI spawn handling, workspace-scoped filesystem, CSP, Markdown/Mermaid sanitization, SQL validation) and had the WikiTicket SDD worklog system installed for tracking work going forward. With the plumbing solid, the natural next question is: does the app actually deliver on its stated vision yet?

Motion's only documented vision (README.md): "a local-first technical writing IDE designed to help you organize, edit, and visualize your documentation with ease." Exploring the current codebase against that bar surfaced a consistent pattern — several real, working, well-tested pieces of code exist with no way for a user to ever reach them — plus one pre-existing data-loss bug that was deliberately deferred during the last fix pass. This plan sequences the work to close those gaps, prioritizing reachability (a user can trigger the feature at all) and data safety over new capability.

Two scope calls were made explicitly with the user rather than defaulted:

• ImageGen (currently fully mocked, no CLI wrapper can produce a real image): scope a real backend as its own task, rather than cutting the feature.
• Workspace-level topic clustering / auto-generated TOC & SKILL docs: deferred, filed as its own separate triage epic rather than built this round.

Sequencing

1. Verify the claude CLI contract (landmine underneath everything else)
2. Fix editor mode-desync (silent data loss, already deferred once)
3. Give editor extensions a creation UX (prerequisite surface for later tasks)
4. Fix web/dev storage mock (rides alongside the creation UX)
5. Real diagram generation + scope ImageGen backend (needs the creation UX + CLI contract)
6. Wire the per-note enrichment action (needs the creation UX + CLI contract)

Tasks

• (P1) Verify and fix the claude CLI system-prompt contract in cliWrappers src/lib/cliWrappers.ts's claude case passes --system <prompt> with a comment admitting it's a guess. Confirm the real claude -p CLI flag for a system prompt and fix the branch if wrong (likely folding the system prompt into the main prompt string instead of a separate flag). One unverified assumption sits underneath every other CLI-backed task below — do this first so nothing downstream is built on a guess. kind:bug.

  • Verify and fix the claude CLI system-prompt argument in cliWrappers.callLLM
  • Rename/dedupe EnrichmentTools.test.ts to match what it actually tests (no matching EnrichmentTools.ts exists; the test re-tests TopicRefiner/SkillGenerator/TOCGenerator under a misleading name)

• (P1) Fix editor mode-desync between WYSIWYG, Markdown, and Split views src/components/Editor/index.tsx keeps two independent sources of truth — the TipTap editor document and the rawMarkdown string state — that only reconcile in handleSave and the file-load effect. Switching viewMode does not run either conversion, so edits made in one mode can be silently dropped when switching to another; the Split view's right-hand pane is a static, stale render of rawMarkdown until a save happens. Reuse the existing marked/turndown/ sanitizeHtml conversions already in the file, no new libraries. kind:bug.

  • Sync content on viewMode transitions in Editor/index.tsx, reusing the existing marked/turndown/sanitizeHtml calls

• (P2) Give editor extensions a creation UX DatasetExtension, QueryExtension, MermaidExtension, DiagramGenExtension, and ImageGenExtension are real, working TipTap nodes with zero UI affordance to create one — no slash command, input rule, or toolbar button. The only way one appears today is hand-authored HTML or the hardcoded welcome-doc demo content. This is a prerequisite surface for the diagram-gen and enrichment tasks below. kind:feature.

  • Toolbar buttons to insert each of the 5 block extensions, reusing the existing ToolbarButton component pattern in Toolbar.tsx
  • "/" slash-command popup wired to the same insertion actions (hand-rolled minimal popup, not a new @tiptap/suggestion dependency, for 5 fixed commands)

• (P2) Fix web/dev storage mock so Dataset/Query are testable outside Tauri WebStorage.readFile only returns real content for three hardcoded demo paths; any other filename returns a placeholder string. WebStorage runs under bun run dev with real filesystem access, so there's no actual sandboxing reason for this to stay mocked — it just blocks iterating on Dataset/Query without a full Tauri build. kind:bug.

  • WebStorage.readFile/listFiles read real files via Bun.file instead of hardcoded branches
  • File picker for Dataset's source field, populated from storage.listFiles

• (P2) Real diagram generation, and scope a real ImageGen backend DiagramGenExtension and ImageGenExtension both fully mock generation with a setTimeout and keyword-matched fake output; no backend exists for either. Diagram generation is in scope now since it only needs text back (Mermaid syntax) via cliWrappers.callLLM, the same function TopicRefiner/TOCGenerator/SkillGenerator already use. Real image generation needs its own research pass first — none of the wrapped CLIs (claude/gemini/qwen/opencode) can produce a raster image, and a real backend raises a local-first scoping question that shouldn't be defaulted. kind:feature.

  • Real diagram generation via cliWrappers.callLLM + Mermaid-validate-before-accept, mirroring ContentInjector.verifyCodeBlocks()'s cheap-check-before-accept pattern
  • Research spike: what a real ImageGen backend would require (candidate CLIs/APIs, local-first implications, rough integration shape) — output is a decision/design doc, not code

• (P2) Wire the per-note enrichment action into the UI cliWrappers, TopicRefiner, TOCGenerator, SkillGenerator, and ContentInjector are all real, tested, working modules with zero imports from App.tsx or anything under src/components/ — only their own test files ever call them. Add an "AI Refine"/"Generate Summary" toolbar action backed by ContentInjector.refineChunk/generateSummary, writing the result back via storage.writeFile. This proves the pipeline reachable end-to-end without committing to the larger, unspecified workspace-level topic-clustering feature (filed separately, deferred). kind:feature.

  • Per-note "AI Refine"/"Generate Summary" toolbar action backed by ContentInjector