Rules Relation

An Obsidian vault studying rule files for AI coding agents, plus a local web app for inspecting the relations between rules one edge at a time.

Each project folder (clinerules-tdd/, elastix_gemm/, GPTPortal/) holds the rules and context items extracted from one real-world agent rule file, with the relations between them (refinement, exception, support, checkpoint, conflict, duplication) recorded as - **<type>** → [[target]] links inside each note.

Edge Inspector

inspector/ is a zero-dependency app for reviewing every relation edge: it shows the two nodes and the arrow between them, an LLM-generated rationale for why the relation holds, and lets you mark each edge (agree / unsure / reject), comment on it, or add new relations.

Launch with Docker (recommended)

Keeps the app running in the background and surviving reboots/idle.

docker compose up -d --build

Then open http://localhost:7077.

• Stays up: restart: unless-stopped — the container comes back after a crash, a Docker/host restart, or waking from idle. It only stays down if you stop it yourself.
• Your inputs are stored locally: marks, comments, and added edges are written to inspector/annotations/ on your machine (bind-mounted into the container), so they persist across restarts, rebuilds, and docker compose down.
• Stop it with docker compose stop (keeps data) or docker compose down (removes the container; your annotations stay on disk).
• After editing the vault, rebuild so the data refreshes: docker compose up -d --build.

Launch without Docker

Requires only Python 3 (standard library — no pip install).

cd inspector
python3 build_data.py     # parse the vault → data/*.json (re-run after editing relations)
python3 server.py         # serve the app

Then open http://localhost:7077.

Set a different port with PORT=8080 python3 server.py. The server binds to 127.0.0.1 by default; set HOST=0.0.0.0 to expose it on your network (Docker does this automatically).

What you can do

• Switch projects with the tabs in the header — rules, edges and annotations all swap together.
• Read the source in the left pane: the original rule file, with the current edge's rule span(s) highlighted in light yellow.
• Inspect an edge in the center: source node → relation → target node (exact source text), plus the LLM rationale.
• Review each edge: ✓ agree / ? unsure / ✗ reject and a free-text comment (auto-saves).
• Filter the relation list (right/left panel) by relation type or review status.
• Add relations the LLM missed via ＋ Add relation (pick source, type, target, write your own rationale).
• Keyboard: ←/→ navigate · 1 agree · 2 unsure · 3 reject.

How it's wired

File	Role
inspector/fetch_sources.py	Downloads each project's raw source rule file at the commit pinned in its CSV into sources/<project>.txt (so line numbers match). Run once.
inspector/build_data.py	Builds data/<project>.json — from vault notes, or (for CSV-native projects with a relations.json) straight from the CSV. Maps each node to its exact source span and embeds the source file.
inspector/make_obsidian_notes.py	Generates Obsidian vault notes (Rules/Context .md with relation [[wikilinks]]) for CSV-native projects, so they appear in Obsidian's graph view.
inspector/data/rationales/*.json	Per-edge LLM rationales, merged into the data at build time.
inspector/server.py	Python stdlib HTTP server (port 7077). Serves the UI and persists annotations.
inspector/index.html	Single-file UI (vanilla JS).
inspector/annotations/<project>.json	Your saved marks, comments, and user-added edges (one file per project).

Annotations are plain JSON next to the vault, so they are easy to diff, export, or back up. Re-run build_data.py whenever you edit relations in the vault to refresh the app's data.