Typed Agent Service

typed-agent-service

[](https://github.com/jkelly-dev1/typed-agent-service/actions/workflows/ci.yml)

A TypeScript AI agent service built as a personal learning project: Fastify 5, Zod 4 tool schemas, a provider-agnostic tool-use agent loop, SSE streaming, a typed error boundary, and two small front ends -- React and Angular -- that render the same stream. It runs fully offline on a deterministic mock provider by default and switches to a real model (Anthropic or OpenAI) with two environment variables.

The rule this repo follows: no claim without a test. The table below maps each claim in this README to the test that enforces it.

What it demonstrates

• Zod schemas as the single source of truth for tools: one schema validates every model-supplied tool call at runtime, infers the executor's input type at compile time, and generates the JSON Schema the model sees.
• A bounded agent loop: tool requests are validated, executed, and fed back; a hard iteration cap stops a model that never finishes.
• Streaming SSE and buffered JSON endpoints over the same typed event contract, with client-disconnect and timeout cancellation.
• An RFC 9457 problem+json error boundary with stable machine-readable codes that never leaks internals.
• A provider seam (mock / Anthropic / OpenAI) where tests and CI never touch the network.
• A React client that consumes the same typed event contract, including its own SSE framing because EventSource cannot POST, and the two different ways this API reports failure.
• An Angular client over the SAME endpoints and the same contract, written standalone, signal-based and zoneless, so the contract is proven to be a contract rather than a shape one framework happened to fit.

Claims backed by tests

Claim	Test
Schema-invalid tool calls never reach an executor	tests/tools.test.ts "rejects schema-invalid in