clankjob v0.2.0 initial

First public release of clankjob: cron jobs for the age of LLMs, AI agents that know how to wait.

A case works on a goal in plain language. When it has to wait for a reply, a date or you, it goes to sleep and spends nothing. It wakes when something actually happens and picks up where it left off, across days and restarts.

The engine

• Activations, sleep and wake, crash-safe resume from a durable SQLite event log
• Core tools: sleep, ask_human, complete, fail, notes and contacts
• Wake on timers, timeouts, plugin conditions (an email reply, new mail, a Home Assistant state), or "ask me but also wake if X happens"
• Polls as often as the task needs; fresh wakes drop the old conversation; large results of earlier wakes are left out
• Budgets on activations (also per day), turns per activation and tokens, with a cost estimate per case
• Times in the owner's time zone

You stay in charge

• Approvals: approve, edit or reject risky tool calls such as sending email, per case (default, always, never)
• Contacts, with trusted recipients that can skip email approval
• Answer from the web UI or Discord; the first answer wins
• Instructions and files per case, read on demand; images for vision models
• Skills: a case can save what it learned, with scripts, after you approve the diff

Any model

• Any OpenAI-compatible endpoint (OpenRouter, OpenAI, Ollama, vLLM, LM Studio), with a model catalog and prices
• Pick the model per case and change it while the case works
• Every prompt is a template; override them or add profiles without recompiling

Plugins

Email (IMAP/SMTP), Discord (ask, answer, notify, and start a case by @mention), web search, pages and watches, weather (Open-Meteo), finance research (market data, SEC filings, 13F, FRED, DCF), Home Assistant, ntfy notifications, documents (metadata, OCR, text), YouTube transcripts, and a sandboxed shell. Any CLI script can become a tool through a manifest, with no plugin code.

Web UI and deployment

• Web UI: cases grouped by state, timeline with markdown, inbox, contacts, skills, plugins and prompts, light and dark
• REST API with bearer tokens and hot reload
• One Rust binary and one SQLite file; a published Docker image (amd64, arm64) with a setup wizard, one-command install and ./update