Skip to content

ROADMAP.md

Latest commit

 

History

History
123 lines (97 loc) · 6.81 KB

ROADMAP.md

File metadata and controls

123 lines (97 loc) · 6.81 KB
Raw

Crustly Roadmap

Current Version: 0.4.1 — May 2026 Author: Jeremy JEANNE

Completed Milestones

v0.1 — Foundation

  • Ratatui TUI with streaming chat, plan mode, tool approval dialogs
  • Anthropic provider with streaming, tool use, and cost tracking
  • SQLite persistence (sessions, messages, plans)
  • Phase 1 tools: read_file, write_file, edit_file, bash, ls, glob, grep

v0.2 — Multi-Provider & Advanced Tools

  • OpenAI provider (GPT-4, GPT-3.5, OpenAI-compatible endpoints)
  • Local LLM support via LM Studio and Ollama
  • AWS Bedrock provider (--features aws-bedrock)
  • Phase 2 tools: web_search, execute_code, notebook_edit, parse_document
  • OS keyring integration for secure API key storage

v0.3 — Workflow & Intelligence

  • Phase 3 tools: plan, task, context, http_request
  • Smart model routing (keyword-based tier selection, no API call)
  • Parallel tool dispatch via join_all (≥40% faster on concurrent reads)
  • Tool result caching with TTL for read-only tools
  • Provider failover (auto-retry on secondary provider on rate-limit / timeout)
  • Sandbox & permission policy (path boundary enforcement, bash allowlist, AndPolicy)
  • Episodic memory (past session summaries injected into new context)
  • Context compaction at 80% window capacity

v0.4 — Claw Code Parity & Reasoning

  • Reasoning / thinking display (Anthropic extended thinking, DeepSeek-R1, QwQ-32B via <think> tags)
  • Codebase index with file watcher (notify crate, symbol index, FTS search)
  • MCP tool server integration ([[mcp.servers]] config, auto-registered at startup)
  • Phase 4 tools: web_fetch, todo_write, ask_user, skill, agent, powershell
  • SubAgentLauncher trait — decoupled sub-agent spawning with recursion guard

v0.4.2 — Native Ollama Integration

  • Native Ollama provider (--features ollama, built on ollama-rs) talking directly to /api/chat, alongside the existing OpenAI-compatible route
  • Runtime performance metrics — tokens/sec, model load time, warm/cold-start — surfaced in the TUI header and under each reply; provider identity badge in the header for every configured provider
  • Ctrl+D Model Download dialog — pick or type a model, pull with a live progress bar, cancel with Esc, without leaving the TUI
  • Model management CLI: crustly ollama list|pull|rm|show|embed
  • keep_alive / num_ctx control; OLLAMA_HOST/OLLAMA_MODEL env vars
  • See ollama-rs-integration-plan.md for the full design and implementation status

Upcoming Milestones

v0.5 — Stability & Developer Experience

Target: Q3 2026

  • Integration test suite — full chat/tool/approval flows with a mock provider (no live API calls required)
  • CI/CD pipeline — GitHub Actions: cargo test, cargo clippy, cargo fmt --check, cargo audit on every PR
  • Interactive settings TUI — configure provider, API keys (masked), model, tool toggles, approval timeout from inside the TUI
  • Session search & export — fuzzy search sessions by content; export to Markdown or JSON
  • Git status bar — current branch, dirty/clean state, uncommitted changes count shown in footer
  • Approval memory — "always allow this tool for this session" option; configurable per-tool whitelist

v0.6 — Performance & Robustness

Target: Q4 2026

  • Benchmark suite — cargo bench targets for DB operations, syntax highlighting, and TUI rendering
  • Message pagination — virtual list rendering so large sessions (1000+ messages) don't degrade TUI performance
  • Lazy syntax highlighting — cache rendered blocks; skip re-highlighting unchanged content
  • Crash recovery — write a recovery manifest on panic; offer to restore the interrupted session on next launch
  • Structured error dialog — TUI overlay showing full error with copy-to-clipboard; all errors written to .crustly/logs/
  • Rate-limit countdown — show remaining backoff time in the status bar when a provider returns 429

v0.7 — LSP & Code Intelligence

Target: Q1 2027

  • LSP client — wire up tower-lsp for symbol definitions, references, and diagnostics
  • Diagnostics injection — include active compiler errors in the agent context automatically
  • Jump-to-definition — open file at symbol location from TUI (launches $EDITOR)
  • Hover information — surface type signatures and docs in a side panel
  • Multi-language support — Rust, TypeScript, Python, Go language servers

v0.8 — Plugin System

Target: Q2 2027

  • Plugin API design — stable ABI for custom tools and providers
  • WASM runtime — sandboxed plugin execution via wasmtime
  • Plugin discovery — scan .crustly/plugins/ and ~/.config/crustly/plugins/
  • Plugin SDK — separate crustly-plugin-sdk crate with examples
  • Example plugins — git-tool, docker-tool, jira-tool

v1.0 — Production Release

Target: Q3 2027

  • All v0.5–v0.8 milestones complete and stable
  • External security audit (API key handling, tool sandbox, path traversal)
  • Cross-platform validation (Linux Ubuntu/Arch, macOS Intel/Apple Silicon, Windows 11)
  • Package manager distribution (Homebrew, AUR, Scoop, Chocolatey)
  • Full user documentation (mdBook with tutorials and API reference)
  • Semantic versioning with automated releases via cargo-release

Backlog (Post-1.0)

Item Notes
RAG / vector store Semantic search over the codebase; integrate with the existing codebase index. Raw embedding generation is already available via the native Ollama provider (crustly ollama embed) — no retrieval layer wired up yet
Multi-pane TUI Chat + file preview split; tabs for multiple conversations
Web interface Optional crustly serve command exposing a browser UI
Multi-user / team Shared sessions, role-based approval, audit log export
Telemetry (opt-in) Anonymous usage stats; cost analytics charts in TUI
Google Gemini provider HTTP client implementation
Azure OpenAI provider Azure-specific auth and endpoint routing
crustly run pipelines Chain multiple prompts with conditional logic in a YAML file

Guiding Principles

  1. Terminal-first — never require a browser or GUI; every feature must work over SSH.
  2. Privacy by default — local LLM support is a first-class citizen, not an afterthought.
  3. Explicit approval — dangerous operations always pause for user confirmation; no silent side effects.
  4. Measurable before optimised — add benchmarks before claiming performance improvements.
  5. Security before features — path traversal, injection, and API key hygiene are non-negotiable pre-conditions for 1.0.