mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-21 04:09:45 +03:00
929ebb2d21
Hints and saved prompts still spell process_manage as `process` and todo_list as `todo`. Dispatch resolves those aliases, but validation runs first and rejected them as unknown tools once the canonical tool was deferred. Resolve the alias before the deferred-scope check so the call is routed like the canonical name. Tests: the routing and native-replay cases share one parametrized test with an alias row; the rejection test is unchanged.
84 lines
4.4 KiB
Python
84 lines
4.4 KiB
Python
"""A deferred tool called by its real name runs through the tool_call bridge (#130153)."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
|
|
from run_agent import AIAgent
|
|
|
|
_ARGS = {"action": "poll", "session_id": "proc_x"}
|
|
|
|
|
|
def _validate(arguments, *, name="process_manage",
|
|
valid_tool_names=("terminal", "tool_search", "tool_describe", "tool_call"),
|
|
enabled_toolsets=None, provider_data=None):
|
|
"""Real validation on an agent whose tool search deferred process_manage behind the bridge."""
|
|
from agent.turn_tool_validation import validate_tool_calls
|
|
|
|
agent = AIAgent.__new__(AIAgent)
|
|
agent.provider, agent.model, agent.session_id, agent.tools = "nous", "m", "s1", []
|
|
agent.valid_tool_names = set(valid_tool_names)
|
|
agent.log_prefix, agent._base_url, agent.api_mode = "", "https://api.example.invalid/v1", "chat_completions"
|
|
agent._invalid_tool_retries = agent._invalid_json_retries = 0
|
|
agent.enabled_toolsets, agent.disabled_toolsets = enabled_toolsets, None
|
|
agent._vprint = lambda *a, **k: None
|
|
tc = SimpleNamespace(id="call_1", type="function",
|
|
function=SimpleNamespace(name=name, arguments=arguments))
|
|
messages = []
|
|
message = SimpleNamespace(content="", tool_calls=[tc], provider_data=provider_data)
|
|
verdict = validate_tool_calls(agent, message, "tool_calls",
|
|
messages=messages, conversation_history=[], api_call_count=1,
|
|
effective_task_id="t1")
|
|
return verdict, tc, [m["content"] for m in messages if m.get("role") == "tool"], agent
|
|
|
|
|
|
@pytest.mark.parametrize("name, arguments, expected, executor_name", [
|
|
("process_manage", json.dumps(_ARGS), _ARGS, "process_manage"),
|
|
("process", json.dumps(_ARGS), _ARGS, "process_manage"), # legacy alias, as older hints spell it
|
|
("process_manage", "", {}, "tool_call"), # the bridge's schema probe refuses the missing 'action'
|
|
])
|
|
def test_real_name_is_rewritten_into_the_bridge_call(name, arguments, expected, executor_name):
|
|
"""The call runs as a model-written ``tool_call`` would, and signed-thinking turns, which
|
|
replay from the ordered native sidecars, replay it under the bridge name too."""
|
|
from agent.anthropic_message_convert import _convert_assistant_message
|
|
from agent.bedrock_adapter import _assistant_blocks
|
|
from agent.tool_executor import _parse_tool_call
|
|
|
|
provider_data = {
|
|
"anthropic_content_blocks": [{"type": "thinking", "thinking": "poll it", "signature": "sig"},
|
|
{"type": "tool_use", "id": "call_1", "name": name, "input": expected}],
|
|
"bedrock_content_blocks": [{"reasoningContent": {"text": "poll it", "signature": "sig"}},
|
|
{"toolUse": {"toolUseId": "call_1", "name": name, "input": expected}}],
|
|
}
|
|
verdict, tc, errors, agent = _validate(arguments, name=name, provider_data=provider_data)
|
|
routed = {"calls": [{"name": "process_manage", "arguments": expected}]}
|
|
|
|
assert verdict.action == "ok" and errors == []
|
|
assert tc.function.name == "tool_call"
|
|
assert json.loads(tc.function.arguments) == routed
|
|
parsed = _parse_tool_call(agent, tc)
|
|
assert parsed.name == executor_name
|
|
assert (parsed.scope_block is None) == (executor_name == "process_manage")
|
|
stored = {"role": "assistant", "content": "", **provider_data, "tool_calls": [
|
|
{"id": tc.id, "type": "function", "function": {"name": tc.function.name, "arguments": tc.function.arguments}}]}
|
|
assert [b for b in _convert_assistant_message(stored)["content"] if b["type"] == "tool_use"] == [
|
|
{"type": "tool_use", "id": "call_1", "name": "tool_call", "input": routed}]
|
|
assert [b["toolUse"] for b in _assistant_blocks(stored, "") if "toolUse" in b] == [
|
|
{"toolUseId": "call_1", "name": "tool_call", "input": routed}]
|
|
|
|
|
|
@pytest.mark.parametrize("arguments, kwargs", [
|
|
(json.dumps(_ARGS), {"enabled_toolsets": ["file"]}),
|
|
(json.dumps(_ARGS), {"valid_tool_names": ("terminal",)}),
|
|
('{"action": ', {}),
|
|
], ids=["out_of_scope", "bridge_off", "unparseable_args"])
|
|
def test_unroutable_call_is_still_rejected(arguments, kwargs):
|
|
verdict, tc, errors, _ = _validate(arguments, **kwargs)
|
|
|
|
assert verdict.action == "continue"
|
|
assert tc.function.name == "process_manage"
|
|
assert errors and errors[0].startswith("Tool 'process_manage' does not exist")
|