Feature Tree
A map of a project that a human can read, steer and trust while agents write the code.
Agents now write more code than anyone can review line by line. What a person can still review is intent, decisions, facts and capabilities. The feature tree is that map, kept as plain markdown in features/ next to the code and versioned with it. Each node is one feature: what it is and why (Abstract), how it works in plain language (Methodology), how to know it works (Acceptance Tests), what is left (Todos), what needs a decision (Questions), and what is known and built (Implemented, Facts & Details). Nodes link with [[node]] and the files open cleanly in Obsidian. The tree sits one step above the code: detailed at the leaves, abstract near the root, and never tied to a programming language.
New here? Read docs/USAGE.md: install, setup, and use with Claude Code or Codex, in the terminal, VS Code, or the desktop app.
The loop
Agents work from the tree, and the human steers it.
- Agents pick the next ready todo, build it, run its tests, tick it off, then write up what now exists. When a decision blocks them they log a question and move to other work. Unattended, they may take the recommended option of a multiple-choice question, flagged as unconfirmed.
- Humans read one briefing (
ft status): progress, what needs a human, what the agent can do alone. They answer questions, confirm or reopen agent assumptions, mark todos that need a person, and pause everything with one command. ft digest shows what changed while they were away.
The markdown is the only shared state, so nothing is lost when a session ends or context is cleared.
Parts
features/SCHEMA.md: the format. features/feature-tree/: this project's own plan, written with itself.skill/: the Claude Code skill (SKILL.md), the helper CLI (scripts/ft.py, standard-library Python only), and optional hooks (hooks/, see its README).features/viewer.html: a single-file viewer, no server. It opens with a snapshot of the plan, read-only; pick the features/ folder (Chrome or Edge) to go live and edit. Rebuild it with ft view.vscode-feature-tree/: a VS Code extension with the same graph and editor.
Try it
python3 skill/scripts/ft.py status # the briefing: progress, what needs you
python3 skill/scripts/ft.py tree # the hierarchy
python3 skill/scripts/ft.py next # what an agent would do next
python3 skill/scripts/ft.py digest # what changed since yesterday
python3 skill/scripts/ft.py view # build features/viewer.html
python3 skill/scripts/ft.py lint # check the plan
Starting on an existing codebase: ft init scaffolds a draft tree from it, then interview and refine.
Autonomy
python3 skill/scripts/ft.py work -- --permission-mode acceptEdits --allowedTools "Bash(python3 skill/scripts/ft.py:*)"
ft work runs the loop headlessly, one fresh claude -p per step, with a time limit per todo and stall detection. It stops when the human runs ft pause, and ft resume continues. ft status prints AGENT IDLE when nothing is left that the agent can do alone.
To have every Claude Code session start from the briefing and refuse to end with half-finished bookkeeping, wire up the hooks in skill/hooks/README.md.
Tests
python3 -m unittest discover -s skill/tests
cd vscode-feature-tree && npm install && npm test
License
MIT
