Roadmap

[djdarcy]

Roadmap

Evergreen tracking issue for notepad-cleanup development priorities. Updated as work lands; never closed.

Current focus: v0.3.0 — generalizing from "a Notepad tool" to "an editor-agnostic extraction pipeline", starting with Sublime Text.

Phase 1: Foundation — shipped

• Two-phase extraction (silent WM_GETTEXT, UIA tab switching for unloaded tabs)
• Manifest + per-window output structure
• AI organization via Claude Code CLI (organize)
• Cross-session deduplication with exact and fuzzy matching (Pre-organize deduplication: diff new notes against previous sessions and link duplicates #9)
• Filesystem linking — hardlink / symlink / DazzleLink descriptors
• Link-aware organize; links separate / links join
• Configuration system: folder registry, ... notation, MRU, search dirs
• Diff integration (Beyond Compare, WinMerge, VS Code)
• PyPI publishing, traffic tracking, badges
• Fix compare missing nc-* session folders (compare misses nc-* format session folders (pattern/format mismatch) #14)
• Fix organize silently overwriting files with duplicate names (v0.2.5)
• First automated regression suite + human test checklists (v0.2.5)

Phase 2: Editor-agnostic sources — v0.3.0, in progress

The pipeline downstream of extraction (dedup, compare, organize, links) binds to manifest.json and the output directory, never to an editor. Only ~7% of the codebase is Notepad-specific, so sources become pluggable.

• sources/base.py — Source / Container / Tab contracts, capability flags
• sources/notepad.py — adapter over existing discovery/extractor
• sources/sublime.py — session-file parser with buffer taxonomy
• sources/__init__.py — lazy mode registry
• compare_to_origin() — source-agnostic buffer-vs-file comparison
• saver.py — generalize hwnd to container_id; add source to manifest
• --mode {notepad,sublime} dispatch in cli.py
• default_mode in the config system
• pytest suite for sources against synthetic fixtures
• Human test checklist for the mode switch

Phase 3: Backends — v0.3.0

Extends the existing --backend option beyond claude / prompt-only. Sensitive content routes to a self-hosted model so credentials never leave the network.

• backends/base.py — Backend.plan() contract, data_stays_on_prem flag (Support additional AI CLI backends (OpenAI Codex CLI, Gemini CLI, etc.) #6)
• backends/claude_cli.py — move existing invocation behind the contract
• backends/openai_compat.py — LM Studio / Ollama (one adapter, base-URL difference)
• Excerpting + batching + checkpoint/resume for long runs
• backends test preflight that measures throughput, not just reachability
• Routing policy: --backend auto sends flagged content on-prem, clean content to cloud
• Two-layer secret detection (path/content heuristics + model judgement)

Phase 4: Semantic triage — v0.3.x

• Local embeddings for topic clustering (cheap first pass before any LLM call)
• Cluster naming via a small number of LLM calls rather than per-file classification
• Confidence-scored authorship candidates via similarity to known directories
• Catalogue output: class, size, era, type, duplication, secret flags

Phase 5: Session drain — v0.3.x

• drain --dry-run — report only, no write path
• Preflight gates: editor closed, re-extraction complete, every buffer accounted for, backup verified
• Write path: archive the existing session, then write a fresh one; never mutate in place
• Reopen verification + one-step restore

Phase 6: Extraction improvements

• Detect saved file paths from Notepad tabs so existing files link instead of duplicating (Detect saved file paths and offer save-in-place #1)
• Capture UIA TabItem.Name for better tab metadata (Capture UIA TabItem.Name for better tab metadata #3)
• Auto-close extracted tabs after verified extraction (Auto-close extracted Notepad tabs after safe extraction #13)
• Interactive compare mode with pause / diff / skip controls (Interactive compare mode: live review with pause/diff/skip controls #12)
• Composable command chaining between extract / compare / organize (Composable command chaining: auto-pass folder between extract/compare/organize #11)

Phase 7: Reach

• Cross-platform support — macOS TextEdit, Linux editors (Cross-platform support: macOS TextEdit and Linux text editors #2)
• Additional editor sources: VS Code, vim
• Decide whether scope shift to AI-first warrants moving the repo (Consider moving repo to DazzleML if scope shifts to AI-first #4)

Related

Design documents live in the maintainers' private notes and are referenced by filename in commit messages under a Design: section.