Protocore

Protocore is the agent loop, and nothing else.

It is a Python 3.12+ library holding one thing: the ReAct runtime that drives an LLM agent turn by turn — the loop, the context budget, the tool surface, the compaction, the stop conditions. Everything the loop touches from outside is a Protocol you implement: the model client, the stores, the event transport, the tools themselves. There is no database driver here, no HTTP endpoint, no deployment logic, and no import that reaches upward out of the package.

That constraint is the point. An agent loop is where the hard, unglamorous correctness lives — what to do when the model answers with prose instead of the tool it was told to call, when a tool result is a hundred kilobytes, when the context window fills mid-turn, when a run must be snapshotted and resumed on another process. Protocore isolates that from the plumbing so it can be tested exhaustively and reused across products.

Русская версия: README.md · Docs: docs/index.md (EN) · docs/ru/index.md (RU)

What you get

• 20 interface Protocols — ILLMProvider, IRunStore, ISessionStore, IToolRegistry, IMemory, IWorkspace, ISearchIndex, IEventStream, ISkillStore, IHookManager, and the rest, plus an IBlobStore ABC. They are the whole outward surface; the core never learns what is behind them.
• A ReAct runtime — QueryEngine owns the per-run mutable state, query() drives one turn and yields a stream of typed TurnEvents. Snapshot and resume are first-class, so a run survives a process restart.
• A three-layer tool surface — tenant policy, then a lean clipped surface, then progressive discovery over BM25 retrieval, with a permission gate in front of dispatch.
• Two-tier context compaction — the loop keeps working when the transcript outgrows the window, and the compaction is deterministic enough to test.
• 524 runtime constants — every tunable value is a field on a frozen RuntimeConstants snapshot injected per tenant. No magic numbers in the executable path, and new behaviour defaults off.
• In-memory adapters, shipped inside the package, so you can drive a real turn end to end with no external services at all.

Install

pip install protocore==2.0.0a2

Name the version explicitly. The published release is a pre-release, and pip skips those unless you ask — but do not ask with a bare --pre, because that flag applies to the whole resolution and will pull pre-release builds of pydantic too. The pin comes off when there is a stable release.

Or, to work on it, with uv:

uv sync --extra dev

Python ≥ 3.12. Runtime dependencies are pydantic, pluggy, jinja2, and typing-extensions — nothing else.

Quickstart

The core is adapter-driven: build a QueryEngine with your adapters, seed the history with a user message, then iterate. query(engine) is a sync function that resets per-turn state and returns an async iterator — it is deliberately not an async generator, so the reset happens at the call rather than on the first __anext__.

The example below uses the bundled in-memory adapters, so it runs as-is:

import asyncio

from protocore import (
    Message, MessageRole, StopReason, TextBlock, default_runtime_constants,
)
from protocore.runtime.query_engine import QueryEngine, QueryEngineConfig
from protocore.runtime.query import query
from protocore.tests_support.adapters import (
    InMemoryBlobStore, InMemoryEventStream, InMemoryHookManager,
    InMemoryLLMProvider, InMemorySkillStore, InMemoryToolRegistry,
)


async def main() -> None:
    llm = InMemoryLLMProvider()
    llm.queue_response(text="Hello from Protocore.", stop_reason=StopReason.end_turn)

    engine = QueryEngine(
        config=QueryEngineConfig(
            run_id="run-1",
            tenant_id="default",
            session_id="sess-1",
            model_name="smoke-model",
            rc=default_runtime_constants(),
        ),
        llm_provider=llm,
        tool_registry=InMemoryToolRegistry(),
        event_stream=InMemoryEventStream(),
        hook_manager=InMemoryHookManager(),
        skill_store=InMemorySkillStore(),
        blob_store=InMemoryBlobStore(),
    )
    engine.history.append(
        Message(role=MessageRole.user, content_blocks=[TextBlock(text="Say hello.")])
    )

    async for event in query(engine):
        print(event.type)
    print("final state:", engine.state)


asyncio.run(main())

Swap InMemoryLLMProvider for an adapter over a real model client and the same code answers real prompts. Getting started walks through what each event means and what to replace next.

Import location matters. QueryEngine, QueryEngineConfig, and query come from protocore.runtime.*; they are not top-level re-exports. The contract types (Message, StopReason, RuntimeConstants, …) are.

Documentation

A full Russian mirror lives under docs/ru/.

Development

uv sync --extra dev
uv run pytest .            # 2973 tests
uv run ruff check .
uv run mypy --strict
uv run bandit -r protocore -q -c pyproject.toml

All four gates run on every pull request across Python 3.12, 3.13, and 3.14. Coverage is enforced at 90%.

The guard worth knowing about is tests/test_core_import_boundary.py: it AST-parses every module in the package and fails if any of them imports a package that sits above the core — anything sharing the core's name with an underscore after it. That test is what keeps the rest of this README true.

Contributions are welcome; see CONTRIBUTING.md.

License

Mozilla Public License 2.0.

The MPL is a file-level copyleft. In practice that means: build whatever you like on top of Protocore and keep it closed — your adapters, your service, your product are yours. But if you modify a Protocore file itself, that file's source stays open under the same license, and the notices travel with it. See NOTICE for what that asks of you in concrete terms.

Security issues go to SECURITY.md, not to the public issue tracker.