sojuboy/README.md

# sojuboy

An IRC bouncer companion service for soju that:

- Watches your bouncer-connected channels continuously
- Notifies you on mentions via Pushover (default)
- Stores messages in SQLite for summaries and on-demand inspection
- Generates AI digests (OpenAI by default) on schedule or on demand
- Exposes a small HTTP API for health, tailing messages, metrics, and triggering digests

Note: this is not a bot and never replies in IRC. It passively attaches as a soju multi-client on your main account.

## Why

If you use soju as a bouncer, you may want per-client alerts and AI summaries without running a heavy IRC client all the time. This service connects to soju as a distinct client identity (e.g., `username/network@client`) and handles notifications and summaries for you, containerized and easy to run on a Synology or any Docker host.

## High-level architecture

- Language: Go (single static binary, low memory footprint)
- Long-lived IRC client: raw IRC using a lightweight parser (sorcix/irc) with an irssi-style handshake tailored for soju
- Message storage: SQLite via modernc.org/sqlite
- Scheduling: github.com/robfig/cron/v3
- Notifications: github.com/gregdel/pushover
- Summarization (LLM): github.com/sashabaranov/go-openai
- HTTP API: Go stdlib `net/http`

Runtime modules:

- `internal/soju`: soju connection, capability negotiation, irssi-style PASS/USER auth, joins, message ingestion, event playback, CHATHISTORY fallback
- `internal/store`: SQLite schema and queries
- `internal/notifier`: Pushover notifier (pluggable interface)
- `internal/summarizer`: OpenAI client with GPT-5 defaults, GPT-4o-mini fallback
- `internal/scheduler`: cron-based digest scheduling and daily retention job
- `internal/httpapi`: `/healthz`, `/ready`, `/tail`, `/trigger`, `/metrics`
- `internal/config`: env config loader and helpers

## Features

- Mention/keyword detection: punctuation-tolerant (letters, digits, `_` and `-` are word chars)
- Mention tuning: allow/deny channels, urgent keywords bypass quiet hours, rate limiting
- AI digest generation: concise natural summaries (no rigid sections); integrates pasted multi-line posts and referenced link context; image links sent to GPT‑5 as vision inputs
- Configurable schedules (cron), quiet hours, and summary parameters
- Local persistence with retention pruning (daily at 03:00)
- HTTP endpoints: health, tail, metrics, on-demand digests

## How it works

1) The service connects to soju and negotiates IRCv3 capabilities:
   - Requests: `server-time`, `message-tags`, `batch`, `cap-notify`, `echo-message`, `draft/event-playback`; optional fallback `draft/chathistory` when needed
   - Joins happen after numeric 001 (welcome)

2) Authentication:
   - PASS then irssi-style `USER <username/network@client> <same> <host> :<realname>`
   - Soju’s per-client identity preserves distinct history

3) Playback and backfill:
   - If `draft/event-playback` is enabled, soju replays missed messages automatically
   - Optional fallback: `CHATHISTORY LATEST <channel> timestamp=<RFC3339Nano> <limit>` using the last stored timestamp per channel (disabled by default)

4) Messages and mentions:
   - Each `PRIVMSG` is stored with server-time when available
   - Mentions trigger Pushover notifications subject to quiet hours, urgency, and rate limits
   - Debug logs include: mention delivered or suppression reason (backfill, quiet hours, rate limit)

5) Summarization:
   - `/trigger` or the scheduler loads a window and calls OpenAI
   - GPT‑5 context: ~272k input tokens + up to 128k output tokens (400k total)
   - Summaries are concise/natural and integrate multi-line posts, article text (readability-extracted), and image links (vision)

6) HTTP API:
   - `/healthz` → `200 ok`
   - `/ready` → `200` only when connected to soju
   - `/tail?channel=#chan&limit=N` → plaintext tail (chronological)
   - `/trigger?channel=#chan&window=6h` → returns digest and sends via notifier
   - `/metrics` → Prometheus text metrics
   - Protect `/tail` and `/trigger` with `HTTP_TOKEN` via Bearer, `token` query, `X-Auth-Token`, or basic auth (`token:<HTTP_TOKEN>`)

## Health and readiness

- `/healthz` always returns 200
- `/ready` returns 200 only when connected to soju
- Binary supports `--health` to perform a local readiness check and exit 0/1. Example Docker healthcheck:

```yaml
healthcheck:
  test: ["CMD", "/sojuboy", "--health"]
  interval: 30s
  timeout: 3s
  retries: 3
```

## Installation

### Prerequisites

- Docker (or Synology Container Manager)
- A soju bouncer you can connect to
- Pushover account and app token (for push)
- OpenAI API key (for AI summaries)

### Build and run (Docker Compose)

1) Create `.env` in repo root (see example below)

2) Start:

```bash
docker-compose up -d --build
```

3) Health check:

```bash
curl -s http://localhost:8080/healthz
```

4) Tail last messages (remember to URL-encode `#` as `%23`):

```bash
curl -s "http://localhost:8080/tail?channel=%23animaniacs&limit=50" \
  -H "Authorization: Bearer $HTTP_TOKEN"
```

5) Trigger a digest for the last 6 hours:

```bash
curl -s "http://localhost:8080/trigger?channel=%23animaniacs&window=6h" \
  -H "Authorization: Bearer $HTTP_TOKEN"
```

6) Metrics:

```bash
curl -s http://localhost:8080/metrics
```

## Quick start (Docker Compose)

```bash
docker-compose up -d --build
# wait for healthy
docker inspect --format='{{json .State.Health}}' sojuboy | jq
```

Compose includes a healthcheck calling the binary’s `--health` flag, which returns 0 only when `/ready` is 200.

## Configuration (.env example)

Below shows maximum or large/reasonable values. Defaults are noted where they are also the maximum or when relevant.

```env
# soju / IRC
SOJU_HOST=bnc.example.org
SOJU_PORT=6697
SOJU_TLS=true
SOJU_NETWORK=your-network

# Client identity: include client suffix for per-client history in soju
IRC_NICK=yourNick
IRC_USERNAME=yourUser/your-network@sojuboy
IRC_REALNAME=Your Real Name
IRC_PASSWORD=yourSojuClientPassword

# Channels to auto-join (comma-separated)
CHANNELS=#animaniacs,#general
KEYWORDS=yourNick,YourCompany

# Auth method hint (raw is used; value is ignored but kept for compatibility)
SOJU_AUTH=raw

# Notifier (Pushover)
NOTIFIER=pushover
PUSHOVER_USER_KEY=your-pushover-user-key
PUSHOVER_API_TOKEN=your-pushover-app-token

# Summarizer (OpenAI)
LLM_PROVIDER=openai
OPENAI_API_KEY=sk-...
OPENAI_BASE_URL=https://api.openai.com/v1
OPENAI_MODEL=gpt-5
# Max completion (output) tokens for GPT‑5 is ~128k (model limit). Default 700.
OPENAI_MAX_TOKENS=128000
# Summarizer tuning
SUMM_FOLLOW_LINKS=true            # default true
SUMM_LINK_TIMEOUT=20s             # no hard max; example large
SUMM_LINK_MAX_BYTES=1048576       # no hard max; example large (1 MiB/article)
SUMM_GROUP_WINDOW=120s            # no hard max; example large grouping window
SUMM_MAX_LINKS=20                 # no strict max; example large
SUMM_MAX_GROUPS=20000             # 0=no cap; example large
SUMM_TIMEOUT=10m                  # request timeout; default 5m

# Digests
DIGEST_CRON=0 */6 * * *           # every 6 hours
DIGEST_WINDOW=24h                 # no hard max; example large window
QUIET_HOURS=                      # e.g., 22:00-07:00

# Mentions/alerts
NOTIFY_BACKFILL=false             # default false
MENTION_MIN_INTERVAL=30s          # no hard max; rate-limit between alerts
MENTIONS_ONLY_CHANNELS=           # optional allow-list (CSV)
MENTIONS_DENY_CHANNELS=           # optional deny-list (CSV)
URGENT_KEYWORDS=urgent,priority   # bypass quiet hours

# HTTP API
HTTP_LISTEN=:8080
HTTP_TOKEN=put-a-long-random-token-here

# Storage
STORE_PATH=/data/app.db
STORE_RETENTION_DAYS=365          # example large retention

# Logging
LOG_LEVEL=info
```

## Pushover setup

1) Install Pushover iOS app and log in
2) Get your User Key (in the app or on the website)
3) Create an application at `pushover.net/apps/build` to get an API token
4) Put them in `.env` as `PUSHOVER_USER_KEY` and `PUSHOVER_API_TOKEN`

## OpenAI setup

- Set `OPENAI_API_KEY`
- Set `OPENAI_BASE_URL` to exactly `https://api.openai.com/v1`
- If `gpt-5` isn’t available on your account, use a supported model like `gpt-4o-mini`
- GPT‑5 limits: ~272k input + 128k output tokens (400k context)

## HTTP API

- `GET /healthz` → `200 ok`
- `GET /tail?channel=%23chan&limit=50`
  - Returns plaintext messages (chronological)
  - Auth: provide `HTTP_TOKEN` as a Bearer token (or query param `token=`)
- `GET /trigger?channel=%23chan&window=6h`
  - Returns plaintext digest
  - Also sends via notifier when configured
  - Auth as above
- `GET /metrics`
  - Prometheus metrics: `sojuboy_messages_ingested_total`, `sojuboy_notifications_sent_total`, `sojuboy_messages_pruned_total`, `sojuboy_connected`

## Troubleshooting

- Empty tail while there’s activity
  - Ensure the service logs `join requested:` followed by `joined` for your channels
  - Confirm `.env` `CHANNELS` contains your channels
  - Check for `/metrics` and logs for recent message ingestion

- 401 Unauthorized from `/tail` or `/trigger`
  - Provide `Authorization: Bearer $HTTP_TOKEN` or `?token=$HTTP_TOKEN`

- OpenAI 502/URL errors
  - Ensure `OPENAI_BASE_URL=https://api.openai.com/v1`
  - Try `OPENAI_MODEL=gpt-4o-mini` if `gpt-5` isn’t enabled for your account

## Roadmap

- Additional notifiers (ntfy, Telegram)
- Long-form HTML digest rendering
- Admin endpoints (e.g., `/join?channel=#chan`)

## Development notes

Project layout (selected):

- `cmd/sojuboy/main.go` – entrypoint, wiring config/services
- `internal/soju` – soju connector and ingestion
- `internal/store` – SQLite schema and queries
- `internal/notifier` – Pushover notifier
- `internal/summarizer` – OpenAI client and prompts
- `internal/httpapi` – health, tail, trigger, metrics endpoints
- `internal/scheduler` – cron jobs

Go toolchain: see `go.mod` (Go 1.23), Dockerfile builds static binary for a distroless image.

## License

MIT for code dependencies; this repository’s license will follow your preference (add a LICENSE if needed).