pocket/grammer

Agent (AI coding)

- Plan mode: conversational clarify phase (capped rounds), then plan generation
- Plan lint gate — plans are validated before code mode; bad plans bounce back
- Deliberate Plan→Code handoff with banner in Agent tab
- Step-by-step plan execution (run remaining / per-step), per-step review cards
- Per-card Undo on agent file writes, backed by a shared step-history ledger (max 50 entries, per project)
- File tools: create, insert, delete, move, copy — with gated tool selection per message
- Image picker + clipboard attachments to chats; streaming responses with haptics
- Role models: separate plan/code models, tiered by device RAM, with user overrides

Models

- Local GGUF inference via llama.cpp (Metal GPU layer probing/splitting, auto thread count, sampling params, grammar-constrained tool calls)
- Model browser: recommended / downloaded / HuggingFace catalog with infinite scroll
- Download manager, custom model entry, per-model JSON sidecars
- BYOK: OpenAI-compatible providers, keys in SecureStore, effort levels, redaction
- IDE-compat server: OpenAI/Ollama-style endpoints (/v1/chat/completions, /api/chat, …) so external IDEs can use the phone's models

Workspace (mobile or desktop editor)

- File tree with open-file tabs (cap, dirty-close prompts), new/rename/copy/delete
- Drag-to-move files/folders (long-press to drag)
- Syntax highlighting for multiple languages, inline autocomplete suggestions
- Project-wide find, diagnostics bar, per-step undo/revert in editor
- JavaScript sandbox runner, HTML/app/image/doc preview via WebView
- Share/export files, error review screen for pasting errors to the model

Projects & connectivity

- Project create/import (document picker, clipboard), GitHub repo import (zip codeload with binary/junk filtering), project export/zip
- On-device HTTP server + WebSocket with per-install auth token
- Desktop client (Vite+React+Monaco) bundled into the app — phone serves the full desktop UI locally; desktop has Monaco editor, terminal panel, search panel, workspace history
- api-server backend: projects/files/run-stop/search/import endpoints, LLM agent chat (streaming), model management, preview route, billing/legal