Skip to content

JiuwenSwarm is an intelligent AI Agent built on openJiuwen. It extends the powerful capabilities of large language models directly to your fingertips through various communication apps you use daily.

Python
2.9 KB
Raw Blame
from __future__ import annotations

import threading
from types import SimpleNamespace

import pytest

from jiuwenswarm.symphony.skill_retrieval.dispatch_imports import dispatch_import_path


class _FailingCompletions:
    def __init__(self) -> None:
        self.calls = 0

    def create(self, **kwargs):
        self.calls += 1
        raise RuntimeError("invalid api key")


class _StaticCompletions:
    def __init__(self, content: str, *, finish_reason: str = "stop") -> None:
        self.content = content
        self.finish_reason = finish_reason
        self.calls = 0

    def create(self, **kwargs):
        self.calls += 1
        message = SimpleNamespace(content=self.content)
        choice = SimpleNamespace(finish_reason=self.finish_reason, message=message)
        return SimpleNamespace(choices=[choice])


def _fake_builder(completions: object) -> SimpleNamespace:
    return SimpleNamespace(
        DEFAULT_CONTEXT_WINDOW=128000,
        DEFAULT_MAX_OUTPUT_TOKENS=32768,
        model="fake-model",
        _client=SimpleNamespace(chat=SimpleNamespace(completions=completions)),
        _manager_config=SimpleNamespace(build=SimpleNamespace(num_retries=2, timeout=1.0)),
        _llm_seed=None,
        _max_output_tokens_cache=None,
        _counter_lock=threading.Lock(),
        _llm_calls=0,
        _retry_calls=0,
        _cache_hits=0,
        _cache_misses=0,
        _cache_unknown=0,
        _prompt_fingerprints=set(),
        _progress=None,
        _progress_task=None,
        _llm_semaphore=threading.Semaphore(1),
        _thread_local=threading.local(),
        _consecutive_failures=0,
        max_consecutive_failures=5,
    )


def test_tree_llm_runtime_raises_on_model_failure() -> None:
    with dispatch_import_path():
        from indexing.tree.llm_runtime import TreeLLMRuntime

        completions = _FailingCompletions()
        builder = _fake_builder(completions)

        with pytest.raises(RuntimeError, match="model call failed"):
            TreeLLMRuntime(builder).call_llm("build a skill tree")

        assert completions.calls == 1


def test_tree_llm_runtime_raises_after_invalid_json_retries() -> None:
    with dispatch_import_path():
        from indexing.tree.llm_runtime import TreeLLMRuntime

        completions = _StaticCompletions("not json")
        with pytest.raises(RuntimeError, match="valid JSON object"):
            TreeLLMRuntime(_fake_builder(completions)).call_llm_json("build a skill tree", max_retries=2)

        assert completions.calls == 2


def test_tree_llm_runtime_raises_on_truncated_json() -> None:
    with dispatch_import_path():
        from indexing.tree.llm_runtime import TreeLLMRuntime

        completions = _StaticCompletions('{"groups": {}', finish_reason="length")
        with pytest.raises(RuntimeError, match="truncated"):
            TreeLLMRuntime(_fake_builder(completions)).call_llm_json("build a skill tree")