followme

A small set of standalone CLI scripts that discover GitHub repositories, grade them with a local Ollama model, then follow promising authors and star strong projects. State lives in a single SQLite file. No database server, no extra services.

Pipeline

Four independent step scripts live in scripts/, each runnable on its own. main.py at the project root chains them together.

Script	What it does
scripts/fetch.py	Pull N new repositories from GitHub Search and insert them into the DB.
scripts/evaluate.py	Clone unrated repos, ask Ollama for idea, skill, description, store.
scripts/subscribe.py	Follow profiles whose repos updated in the last W hours scored above X.
scripts/star.py	Star repos updated in the last W hours that scored above Y.
main.py	Runs fetch -> evaluate -> subscribe -> star once, or in an infinite loop.

Schema

One SQLite table — entries — one row per repository:

Column	Meaning
repo	owner/name (primary key)
profile	repository owner login
clone_url	git URL used for the shallow clone
html_url	browser URL
created_at	when the row was inserted (UTC ISO-8601)
updated_at	last time we touched the row (UTC ISO-8601)
followed	1 if we follow the profile, mirrored across rows
starred	1 if we star the repo
idea	1.0–10.0 novelty grade from Ollama
skill	1.0–10.0 engineering grade from Ollama
description	one-sentence English summary

Scoring uses idea + skill so thresholds live in [2.0, 20.0].

Requirements

• Python 3.11+ (standard library only)
• Git CLI
• Network access to GitHub
• A running Ollama with the configured model installed (e.g. ollama pull qwen2.5-coder:7b)
• A GitHub personal access token with user:follow and public_repo scopes

Setup

python3 -m venv .venv
source .venv/bin/activate
cp env.example .env
# edit .env: at minimum set GITHUB_TOKEN and OLLAMA_URL / OLLAMA_MODEL

requirements.txt exists but is empty — no third-party packages are needed.

Usage

Every script reads .env from the project root. Flags override env values.

Fetch new repositories

python3 scripts/fetch.py -n 5

Evaluate everything not yet rated

python3 scripts/evaluate.py            # rate all pending
python3 scripts/evaluate.py -l 10      # rate up to 10

Follow high-scoring authors (recent window)

python3 scripts/subscribe.py -s 14 -w 24
python3 scripts/subscribe.py --dry-run

Star high-scoring repos (recent window)

python3 scripts/star.py -s 16 -w 24
python3 scripts/star.py --dry-run

Run the full cycle

python3 main.py                            # one cycle, defaults from .env
python3 main.py -n 5 --subscribe-threshold 14 --star-threshold 16 -w 24
python3 main.py --dry-run                  # safe rehearsal
python3 main.py -i --sleep 600             # loop forever

The default cycle is exactly what the project was built around:

1. fetch 5 new repos
2. evaluate them
3. follow profiles updated in the last 24h with idea + skill > 14
4. star repos updated in the last 24h with idea + skill > 16

Files written

• data/followme.sqlite — single source of truth
• data/repo/ — scratch clone directory, wiped between repos

Grade scale

The Ollama prompt asks for strict anchors:

• 1.0 — trivial / junior
• 5.0 — ordinary / middle
• 9.0 — strong / senior

Both idea and skill are clamped into [1.0, 10.0].