Skip to content

Roadmap

Each milestone says what must be proven, not only built. The plan lives in the milestones and the project board.

Milestone Proven by State
M1 Core server the server, sandbox, checks and installer work end to end with a real model done (v0.1.0)
M2 Benchmark and tuning reproducible numbers with spread for the models in use; deputy tune applies a setting only when it wins, with rollback in progress: bench (built-in suite, your own task file, comparison of two runs), tune, a nightly run, stats and discover work; the bar for everyday use is #115
M3 Verification and correction citation accuracy and fact coverage rise on the benchmark; files larger than the model context are handled in progress: cites_text, scan_file, summarise_log, list_files, read-only git and grouped search exist; an outline of code is open
M4 Profiles and memory a run with memory beats the same run without it (ablation); profiles are measured, not typed in progress: project memory and lessons exist; a lesson is kept only if the benchmark with and without it shows it does not hurt
M5 Async and routing slow models do not block the agent; the router picks the cheapest passing path done: background tasks (wait=false, collect) and the router
M6 Second agent the same server driven from GitHub Copilot (VS Code) gives the same benchmark results as through Claude Code planned: instruction files for every agent exist, the real-app verification does not
M7 Improvement loop an offline optimiser proposes profile and parameter changes with a report; a human approves; rollback works in progress: deputy optimise works with approval and rollback