recon-oss

Reconnaissance for open source contributors.

Monitors a GitHub repo's merged PRs and open issues, uses a local LLM to generate a daily digest, and personalizes it based on what you're trying to contribute — and remembers your feedback over time.

Built because manually reading 10-20 merged PRs a day to figure out what to work on next is exhausting.

Work in progress. Currently a working script. CLI with proper setup wizard coming soon.

The idea

When you're trying to contribute to an active OSS project, you have two problems:

1. The repo moves fast — PRs merge daily across multiple subsystems
2. You don't know which open issues the team actually cares about right now

recon-oss solves this by running a daily pipeline that fetches what merged, correlates it with open issues, and asks an LLM to surface contribution opportunities — personalized to your specific goal (e.g. "I want hard backend auth bugs, not frontend work").

It also remembers your feedback. If you reply "focus on issue #442 from now on", the next digest reflects that.

What's working right now

• Fetches merged PRs from the last 24h via GitHub API
• On first run, bootstraps up to 500 issues and 100 PRs into a local SQLite database
• Every subsequent run does a delta sync — only fetches what changed since last time
• Builds a prompt combining three memory layers and calls a local Ollama model
• Prints the digest to terminal and saves it to data/summary.txt
• Accepts feedback after each run — stored in data/chat.json and injected into the next prompt
• Telegram delivery is wired up but optional

Architecture

The pipeline runs in five stages:

GitHub API (delta only after first run)
        ↓
   SQLite sync (repo.db)
        ↓
   Prompt assembly
   [ motive.json + chat.json + SQL queries ]
        ↓
   Local LLM (Ollama)
        ↓
   Delivery (terminal / Telegram)
        ↓
   User feedback → stored in chat.json → used in next run

Three memory layers:

| Memory | File | What it stores | |--------|------|----------------| | Motive | data/motive.json | Your contribution goals, subsystems to focus on, what to avoid | | Repo state | data/repo.db | All issues and merged PRs, incrementally updated | | Chat history | data/chat.json | Past digests and your replies — feeds into next run's prompt |

File structure

recon-oss/
├── run.ts               # entry point — runs the full pipeline
├── config.ts            # all settings live here (repo, LLM, limits)
├── tsconfig.json        # TypeScript compiler config
├── core/
│   ├── sync.ts          # GitHub API fetcher — bootstrap + delta sync
│   ├── context.ts       # builds the LLM prompt from all three memories
│   └── llm.ts           # Ollama / Anthropic API call
├── memory/
│   ├── db.ts            # SQLite read/write for issues and PRs
│   ├── chat.ts          # chat history read/write
│   └── motive.ts        # loads and formats user motive
├── delivery/
│   ├── index.ts         # routes to the right delivery channel
│   ├── cli.ts           # terminal output + feedback prompt
│   └── telegram.ts      # Telegram bot send + reply capture
└── data/                # gitignored — local state lives here
    ├── repo.db
    ├── motive.json
    ├── chat.json
    └── summary.txt

Running it (current, manual setup)

Requirements: Node.js 18+, Ollama running locally with a model pulled.

git clone https://github.com/x15sr71/recon-oss
cd recon-oss
npm install

Edit config.ts and fill in:

• github.token — personal access token with public_repo scope
• repo.owner and repo.name — the repo you want to monitor
• llm.model — whichever Ollama model you have (tested with qwen3:14b)
• motive — your actual contribution goals

ollama serve   # if not already running
npm run build
npm start

First run bootstraps the database (takes a minute). Every run after that is fast — only syncs what changed.

What's coming

• [ ] recon-oss init — interactive CLI setup wizard, no manual config editing
• [ ] npm install -g recon-oss — proper global install
• [ ] Cron setup built into the CLI
• [x] Anthropic / Claude as LLM option (set LLM_PROVIDER=anthropic)
• [ ] Better prompt tuning — current model sometimes ignores user directives
• [ ] Homebrew tap

Contributing

Open to contributions, especially on:

• Prompt improvements — making the LLM follow user feedback more strictly
• Additional delivery channels (Discord, Slack, email)
• LLM provider adapters (the interface is already modular in core/llm.ts)
• Testing with repos other than Infisical

If you're using it to track a different repo and hit issues, open an issue with the repo name and what broke.

Why local LLM by default

No API keys needed, no per-run cost, runs offline. The architecture supports cloud LLMs and that's coming — but Ollama first means anyone can run it immediately without a credit card.