docs(local-llm): comprehensive roadmap review — 5 bugs, 6 phases, 31 items
Systematic code review of DASHBOARD_ROADMAP.md against actual codebase:
Bugs found (BN1-BN5):
- BN1: Compare buttons show unloaded models (can't generate)
- BN2: No AbortController on compare stream (leaks on close)
- BN3: Chat messages lost on modal close (no persistence)
- BN4: Logs panel has no refresh button
- BN5: Delete dialog missing reclaim size (partial impl exists)
Expanded from 4 phases to 6 + backlog (15 → 31 items):
- Phase 1: Pre-load intelligence + bug fixes (N1-N3, BN1-BN2, BN5)
- Phase 2: Rich metadata + persistence (N4-N5, BN3-BN4)
- Phase 3: Model intelligence badges + sort (N6-N10)
- Phase 4: Runtime metrics + UX polish (N11-N14)
- Phase 5 (NEW): Response quality — markdown, syntax highlight,
vision upload, think-block collapse, model library link
- Phase 6 (NEW): Data persistence — export/import, inference log,
factory reset
- Phase 7: Expanded backlog (F17-F38, +6 new ideas)
Improvements:
- Added checkboxes for all tasks and acceptance criteria
- Quant-aware RAM estimate multipliers (Q4/Q5/Q8/F16)
- Broader vision model regex (bakllava, moondream, llama-vision)
- DeepSeek R1 distill variant detection for think badge
- Conservative memory availability formula (free + cached*0.5)
- localStorage key registry with llm- prefix standardization
- Dependency graph between phases
- ~6 hrs total estimated effort
ae231d5aac