agent-registry-router 0.2.4

Registry-driven LLM routing: build classifier prompts from agent descriptions, validate decisions, and dispatch.

agent-registry-router

Registry-driven LLM routing: build classifier prompts from agent descriptions, validate decisions, and dispatch to other agents.

What is this? Why use it?

agent-registry-router helps you route LLM traffic to the right agent by:

• Keeping a registry of agents
• Building classifier prompts from that registry
• Validating classifier decisions with clear, typed errors
• Dispatching to runtime agents with optional observability hooks (on_event) for logging/metrics.

Use it when you need:

• A stable contract between your classifier and execution layer
• Deterministic, size-aware prompt construction
• Lightweight integration with frameworks like PydanticAI
• Extensibility: plug your own classifier/agents, add hooks for monitoring, and keep routing logic decoupled from business logic

Features

• Registry-driven classifier prompts (deterministic order, routable-only, size-aware)
• Fail-fast validation with typed errors (no silent fallbacks)
• PydanticAI dispatcher with pinned bypass and observability hooks (on_event); supports text streaming (stream_text) and response streaming (stream_responses)
• Typed package (py.typed) and CI gates (ruff/black/mypy/pytest-cov)
• FastAPI demo with pinned bypass: examples/fastapi_pinned_bypass/README.md

Scope

• Routing, validation, observability; bring your own LLM clients/agents.
• Adapters are opt-in; core stays lightweight.

Status

• v0.2.3 — fail-fast by default; adapters namespaced; PydanticAI adapter supports streaming dispatch.

Install (uv)

From PyPI:

uv pip install agent-registry-router

From a checkout of this repo (dev/editable):

uv venv
source .venv/bin/activate
uv pip install -e .

Core usage

from agent_registry_router.core import (
    AgentRegistration,
    AgentRegistry,
    RouteDecision,
    build_classifier_system_prompt,
    validate_route_decision,
)

registry = AgentRegistry()
registry.register(AgentRegistration(name="general", description="General help."))
registry.register(AgentRegistration(name="special", description="Special help."))

prompt = build_classifier_system_prompt(
    registry,
    preamble="You are a query classifier that routes user messages to the appropriate agent.",
    default_agent="general",
)

decision = RouteDecision(agent="special", confidence=0.9, reasoning="Clear match.")
validated = validate_route_decision(decision, registry=registry, default_agent="general")

Behavior & errors

• Default agent must be routable; unknown agents or empty registries raise.
• Classifier selecting a non-routable agent raises InvalidRouteDecision.
• Missing default or no routable agents raises InvalidFallback.
• Dispatcher raises AgentNotFound if the chosen agent cannot be resolved.
• Registry validation uses RegistryError; routing errors derive from RoutingError.
• Confidence adjustment on invalid routes is unchanged; pinned invalid falls back to the classifier.
• Prompt listing preserves registration order; only routable agents are included.
• Agent descriptions are capped at 512 characters; prompts cannot be built without routable agents. Optional max_prompt_chars can bound the generated prompt.
• Observability: PydanticAIDispatcher accepts on_event callback (receives RoutingEvent) and optional logger; emits events for classifier run, validation, pinned bypass, agent resolution, and agent run.

API contracts

• Public imports (agent_registry_router.core): AgentRegistry, AgentRegistration, RouteDecision, ValidatedRouteDecision, validate_route_decision, build_classifier_system_prompt, exceptions (AgentRegistryRouterError, RegistryError, RoutingError, InvalidRouteDecision, InvalidFallback, AgentNotFound), and RoutingEvent.
• Adapter: agent_registry_router.adapters.pydantic_ai exposes PydanticAIDispatcher, DispatchResult, AgentStreamChunk, AgentResponseStreamItem, ResponseStreamSession. Adapters stay namespaced (not re-exported at package root).
• Routing invariants: default agent must be routable; empty registry errors; non-routable selections error; pinned invalid falls back to classifier; classifier output must include agent and confidence or InvalidRouteDecision is raised.
• Prompt determinism: preserves registration order; only routable agents; optional max_prompt_chars; description cap 512 chars.
• Hooks: on_event receives RoutingEvent(kind, payload, error); hook failures are swallowed; logger is optional.

Adapters

• PydanticAI dispatcher: src/agent_registry_router/adapters/pydantic_ai/README.md

Tests (uv)

uv pip install -e ".[dev]"
ruff check .
black --check .
mypy --config-file pyproject.mypy.ini .
pytest --cov=agent_registry_router --cov-fail-under=85

Example: FastAPI pinned bypass

See examples/fastapi_pinned_bypass/.

License

Apache-2.0 (see LICENSE).