Finyuus

The Finyuus language is a small domain-specific language written for AI orchestration, not general-purpose scripting. Statements cover agent calls, native and HTTP tools, guards, approvals, conditionals, loops, parallel branches, and child pipelines. Because workflows are plain text, you can diff them, review them in pull requests, and reuse pieces across projects. That's the whole reason it exists: AI behavior you can read and compare, not a canvas full of boxes. No boxes. Just text.

Every run is backed by Temporal, which is what makes long waits and failures survivable. If a worker dies mid-run, the workflow resumes from its recorded state instead of starting over. Retries, cancellation, and child workflow orchestration come from the same layer. It just works. For anyone who has watched a multi-hour agent job vanish on a crash, this is the feature that changes the calculus.

Guards are LLM-based policy checks that return fields like allowed, confidence, and reason. A pipeline routes execution through a pass or block branch, and you can gate on confidence thresholds rather than a flat yes or no. This is how you keep a model from answering a financial-advice question directly, or block output that fails a publication policy before it reaches a user. Simple idea. Huge difference in practice.

Approval statements pause a workflow until a reviewer resolves it from the dashboard or the API. You set an assignee, a timeout, and a timeout action, and the pipeline branches into approved and rejected paths. It's a real halt, not a suggestion. That's the difference between "we'll check it later" and a system that actually stops a risky output and waits for a human before continuing.

Agents, guards, tools, schemas, and pipelines are saved objects with immutable version pins. You write pipeline:company_research@v2 and you get exactly that version, not whatever someone edited last. Pin it. Move on. Child pipelines can be called directly or wrapped as agent tools, so a change in one place doesn't silently ripple through every workflow that depends on it.

Any pipeline can be published as an HTTP endpoint with Bearer-token authentication, Redis-backed rate limiting, idempotency keys, and input validation. Server-sent events stream run updates back to the caller, and approvals generate their own events. The process limits default to 300 requests a minute overall and 120 per endpoint, all configurable through environment variables. Sensible defaults. Easy to override.

Traces, token usage, and model costs land in Langfuse, while application data and pipeline tables live in ClickHouse. Run history, guard outcomes, and approval decisions are visible in the dashboard. The point isn't more dashboards for their own sake. It's that a fragmented AI system usually has no single place to answer "what actually happened in this run?"

Files and large artifacts go to configurable object storage, which defaults to local MinIO and can switch to AWS S3 with a few environment variables. The DSL reads and writes files and ClickHouse-backed tables as first-class statements, so structured data operations sit inside the workflow instead of in a separate service. No extra service to babysit.