filetree

filetree

中文版

A Claude Code plugin that maintains FILETREE.md — a one-line description per file with content hashes for staleness detection. Lets the LLM grasp repo layout in a few hundred tokens before touching code.

Why

Every new Claude Code session relearns the codebase: ls, grep, open file, read, repeat. The discovery is expensive and not reusable across sessions.

Pain point	filetree's answer
LLM keeps rediscovering layout each session	Persisted as FILETREE.md, checked into git, shared with collaborators
Summary docs go stale silently	Per-entry content hash; mismatch reveals drift immediately
Rewriting every description on small refactors wastes tokens	UNCHANGED bias — LLM refreshes hash only when purpose is intact (~100x cheaper)
Tooling adds a sqlite / daemon / watcher	Single markdown file. Change detection delegated to git. Zero background process

Install

Via marketplace:

/plugin marketplace add nekocode/filetree-skill
/plugin install filetree

Local development / dog-fooding (no install needed):

cd /path/to/filetree-skill
claude --plugin-dir .

After editing commands/ or SKILL.md, run /reload-plugins inside the session for a hot reload.

Commands

Command	Purpose
/filetree:init	Generate FILETREE.md from scratch. Refuses to overwrite without confirmation
/filetree:update	Sync FILETREE.md with current repo state (added / changed / removed / renamed)
/filetree:lint	Read-only drift check. Exits non-zero on drift, CI-friendly. Does not call the LLM

All commands refuse to commit FILETREE.md. You review the diff and commit yourself.

Wire it into CLAUDE.md / AGENTS.md

/filetree:init handles this on first run — it scans the repo root for CLAUDE.md / AGENTS.md, skips files already referencing FILETREE.md, and for the rest proposes a bullet whose location and style matches the existing file. You confirm each edit before it lands. Because wiring runs before the manifest is hashed, the post-wire content is what enters FILETREE.md directly — no second-pass refresh.

Caveats:

• If neither CLAUDE.md nor AGENTS.md exists, the plugin won't create one — it's your call which (if any) to seed. Create the file you want, then re-run /filetree:init.
• Wiring runs at init time only. If you add CLAUDE.md / AGENTS.md later, re-run /filetree:init (it'll ask before overwriting FILETREE.md) or wire by hand.

To wire by hand, drop a line like this into your CLAUDE.md:

- `./FILETREE.md` — Per-file purpose index. Read before `ls` / `grep` when overviewing the repo or locating an implementation.

The agent then treats FILETREE.md as a cheap index — one read replaces dozens of ls / grep / cat calls during orientation.

How It Works

Manifest format

# Project Filetree

_Auto-maintained by `/filetree:update`. Each entry carries a content hash; mismatched hashes indicate stale summaries._

## src/auth/

- `middleware.py` — JWT validation middleware; parses bearer token and injects user_id into request context <!--hash:a1b2c3d4-->
- `jwt_utils.py` — Pure JWT signing / verification helpers, framework-agnostic <!--hash:e5f6g7h8-->

## (root)/

- `README.md` — Project entry doc <!--hash:9a8b7c6d-->

• Section header = directory path (trailing /); root files live under (root)/
• Each entry stores filename only (not the full path) + summary + 8-char content hash (from git hash-object)
• Stable sort (sections + entries) → no spurious diffs

Data flow (/filetree:update)

filetree.py todo
  ├─ git ls-files (tracked + untracked, exclude .gitignore)
  ├─ git hash-object on all paths
  ├─ git status --porcelain  (rename detection, trust git 50% heuristic)
  └─ diff vs current FILETREE.md
        ↓ JSON
{added, changed, removed, renamed, stats.need_llm}
        ↓
LLM processes added (write fresh summary)
            changed (UNCHANGED or new summary)
        ↓ JSON via stdin
filetree.py apply
  ├─ UNCHANGED → refresh hash only, keep summary
  ├─ new summary → overwrite entry
  ├─ rename → move entry + rehash
  └─ write FILETREE.md

UNCHANGED bias

On a healthy update, 80%+ of changed items should resolve to "UNCHANGED" — refactors, formatting, comment edits, bug fixes, small additions almost always leave a file's purpose intact. The LLM emits a 4-byte "UNCHANGED" reply; apply refreshes the hash and keeps the old summary. The manifest itself carries the memory of "I already reviewed this version" — no separate cache needed.

Compatibility