mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-06-22 10:32:00 +00:00
75ed07ace8
When a tool call itself restarts the gateway (docker restart, systemctl restart, and similar), the process is terminated mid-call — before the tool result is persisted and before the orderly drain rewind can run. The transcript tail is left as an assistant(tool_calls) with no matching tool answer. On resume the model re-issues the unanswered call, taking the gateway down again — an infinite loop (#49201). Source fix: _build_gateway_agent_history now strips a trailing assistant(tool_calls) block that has no tool answers (_strip_dangling_tool_call_tail), so there is nothing for the model to re-execute. This complements _strip_interrupted_tool_tails, which only handles the case where a tool result row exists with an interrupt marker. Cognitive backstop: the resume-pending system note now states that any restart command in the history already ran and must not be re-executed or verified, and the empty-message auto-resume startup turn reports recovery and asks for instructions instead of the nonsensical "address the user's NEW message" (there is no new message on that turn). Reimplements the intent of #49243 by @JoaoMarcos44 at the replay layer. Fixes #49201
277 lines
11 KiB
Python
277 lines
11 KiB
Python
"""Tests for the auto-continue feature (#4493 / #45232).
|
|
|
|
When the gateway restarts mid-agent-work, the session transcript can end on a
|
|
tool result that the agent never processed. The auto-continue logic detects
|
|
this and prepends an API-only system note to the next user message so the model
|
|
does not re-execute stale interrupted tool calls before addressing new input.
|
|
"""
|
|
|
|
|
|
|
|
def _simulate_auto_continue(agent_history: list, user_message: str) -> str:
|
|
"""Reproduce the auto-continue injection logic from _run_agent().
|
|
|
|
This mirrors the exact code in gateway/run.py so we can test the
|
|
detection and message transformation without spinning up a full
|
|
gateway runner.
|
|
"""
|
|
message = user_message
|
|
if agent_history and agent_history[-1].get("role") == "tool":
|
|
message = (
|
|
"[System note: A new message has arrived. The conversation "
|
|
"history contains pending tool outputs from an interrupted turn. "
|
|
"IGNORE those pending results. Address the user's NEW message "
|
|
"below FIRST. Do NOT re-execute old tool calls from the history.]\n\n"
|
|
+ message
|
|
)
|
|
return message
|
|
|
|
|
|
class TestAutoDetection:
|
|
"""Test that trailing tool results are correctly detected."""
|
|
|
|
def test_trailing_tool_result_triggers_note(self):
|
|
history = [
|
|
{"role": "user", "content": "deploy the app"},
|
|
{"role": "assistant", "content": None, "tool_calls": [
|
|
{"id": "call_1", "function": {"name": "terminal", "arguments": "{}"}}
|
|
]},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "deployed successfully"},
|
|
]
|
|
result = _simulate_auto_continue(history, "what happened?")
|
|
assert "[System note:" in result
|
|
assert "interrupted" in result
|
|
assert "NEW message" in result
|
|
assert "Do NOT re-execute" in result
|
|
assert "what happened?" in result
|
|
|
|
def test_trailing_assistant_message_no_note(self):
|
|
history = [
|
|
{"role": "user", "content": "hello"},
|
|
{"role": "assistant", "content": "Hi there!"},
|
|
]
|
|
result = _simulate_auto_continue(history, "how are you?")
|
|
assert "[System note:" not in result
|
|
assert result == "how are you?"
|
|
|
|
def test_empty_history_no_note(self):
|
|
result = _simulate_auto_continue([], "hello")
|
|
assert result == "hello"
|
|
|
|
def test_trailing_user_message_no_note(self):
|
|
"""Shouldn't happen in practice, but ensure no false positive."""
|
|
history = [
|
|
{"role": "user", "content": "hello"},
|
|
]
|
|
result = _simulate_auto_continue(history, "hello again")
|
|
assert result == "hello again"
|
|
|
|
def test_multiple_tool_results_still_triggers(self):
|
|
"""Multiple tool calls in a row — last one is still role=tool."""
|
|
history = [
|
|
{"role": "user", "content": "search and read"},
|
|
{"role": "assistant", "content": None, "tool_calls": [
|
|
{"id": "call_1", "function": {"name": "search", "arguments": "{}"}},
|
|
{"id": "call_2", "function": {"name": "read", "arguments": "{}"}},
|
|
]},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "found it"},
|
|
{"role": "tool", "tool_call_id": "call_2", "content": "file content here"},
|
|
]
|
|
result = _simulate_auto_continue(history, "continue")
|
|
assert "[System note:" in result
|
|
|
|
def test_original_message_preserved_after_note(self):
|
|
"""The user's actual message must appear after the system note."""
|
|
history = [
|
|
{"role": "assistant", "content": None, "tool_calls": [
|
|
{"id": "c1", "function": {"name": "t", "arguments": "{}"}}
|
|
]},
|
|
{"role": "tool", "tool_call_id": "c1", "content": "done"},
|
|
]
|
|
result = _simulate_auto_continue(history, "now do X")
|
|
# System note comes first, then user's message
|
|
note_end = result.index("]\n\n")
|
|
user_msg_start = result.index("now do X")
|
|
assert user_msg_start > note_end
|
|
|
|
|
|
class TestInterruptedReplayFiltering:
|
|
def test_interrupted_tool_tail_is_removed_from_agent_history(self):
|
|
from gateway.run import _build_gateway_agent_history
|
|
|
|
history = [
|
|
{"role": "user", "content": "transcribe this video"},
|
|
{
|
|
"role": "assistant",
|
|
"content": None,
|
|
"tool_calls": [
|
|
{"id": "call_1", "function": {"name": "terminal", "arguments": "{}"}},
|
|
],
|
|
},
|
|
{
|
|
"role": "tool",
|
|
"tool_call_id": "call_1",
|
|
"content": '{"exit_code": 130, "output": "[Command interrupted]"}',
|
|
},
|
|
]
|
|
|
|
agent_history, observed_context = _build_gateway_agent_history(history)
|
|
|
|
assert observed_context is None
|
|
assert agent_history == [{"role": "user", "content": "transcribe this video"}]
|
|
|
|
def test_mixed_tail_with_one_interrupted_result_is_removed(self):
|
|
from gateway.run import _build_gateway_agent_history
|
|
|
|
history = [
|
|
{"role": "user", "content": "search and transcribe"},
|
|
{
|
|
"role": "assistant",
|
|
"content": None,
|
|
"tool_calls": [
|
|
{"id": "call_1", "function": {"name": "web_search", "arguments": "{}"}},
|
|
{"id": "call_2", "function": {"name": "terminal", "arguments": "{}"}},
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "found URL"},
|
|
{
|
|
"role": "tool",
|
|
"tool_call_id": "call_2",
|
|
"content": '{"exit_code": 130, "output": "[Command interrupted]"}',
|
|
},
|
|
]
|
|
|
|
agent_history, _observed_context = _build_gateway_agent_history(history)
|
|
|
|
assert agent_history == [{"role": "user", "content": "search and transcribe"}]
|
|
|
|
def test_successful_tool_tail_is_preserved(self):
|
|
from gateway.run import _build_gateway_agent_history
|
|
|
|
history = [
|
|
{"role": "user", "content": "deploy"},
|
|
{
|
|
"role": "assistant",
|
|
"content": None,
|
|
"tool_calls": [
|
|
{"id": "call_1", "function": {"name": "terminal", "arguments": "{}"}},
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "deployed successfully"},
|
|
]
|
|
|
|
agent_history, _observed_context = _build_gateway_agent_history(history)
|
|
|
|
assert agent_history[-1]["role"] == "tool"
|
|
assert agent_history[-1]["content"] == "deployed successfully"
|
|
|
|
def test_dangling_unanswered_tool_call_tail_is_removed(self):
|
|
"""A trailing assistant(tool_calls) with NO tool answers is stripped.
|
|
|
|
This is the SIGKILL signature from #49201: the tool itself ran a
|
|
restart/shutdown command and killed the gateway before its result was
|
|
persisted. The transcript tail is an assistant message with tool_calls
|
|
and zero matching tool rows. Without stripping it, the model re-issues
|
|
the unanswered call on resume and loops the restart forever.
|
|
"""
|
|
from gateway.run import _build_gateway_agent_history
|
|
|
|
history = [
|
|
{"role": "user", "content": "restart the container"},
|
|
{
|
|
"role": "assistant",
|
|
"content": None,
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_1",
|
|
"function": {
|
|
"name": "terminal",
|
|
"arguments": '{"command": "docker restart hermes-agent"}',
|
|
},
|
|
},
|
|
],
|
|
},
|
|
]
|
|
|
|
agent_history, _observed_context = _build_gateway_agent_history(history)
|
|
|
|
assert agent_history == [{"role": "user", "content": "restart the container"}]
|
|
|
|
def test_dangling_tail_after_completed_pair_is_removed_only_at_tail(self):
|
|
"""Only the trailing unanswered tool-call block is stripped.
|
|
|
|
An earlier completed assistant→tool pair must survive — we only drop
|
|
the final assistant(tool_calls) that has no answers.
|
|
"""
|
|
from gateway.run import _build_gateway_agent_history
|
|
|
|
history = [
|
|
{"role": "user", "content": "do two things"},
|
|
{
|
|
"role": "assistant",
|
|
"content": None,
|
|
"tool_calls": [
|
|
{"id": "call_1", "function": {"name": "web_search", "arguments": "{}"}},
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "found it"},
|
|
{
|
|
"role": "assistant",
|
|
"content": None,
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_2",
|
|
"function": {
|
|
"name": "terminal",
|
|
"arguments": '{"command": "systemctl restart hermes"}',
|
|
},
|
|
},
|
|
],
|
|
},
|
|
]
|
|
|
|
agent_history, _observed_context = _build_gateway_agent_history(history)
|
|
|
|
# The completed call_1 pair survives; the dangling call_2 tail is gone.
|
|
assert agent_history[-1]["role"] == "tool"
|
|
assert agent_history[-1]["content"] == "found it"
|
|
# The surviving assistant(tool_calls) is the completed call_1 (which
|
|
# has a matching tool answer), not the stripped dangling call_2.
|
|
_surviving_calls = [
|
|
tc.get("id")
|
|
for m in agent_history
|
|
if m.get("role") == "assistant" and m.get("tool_calls")
|
|
for tc in m["tool_calls"]
|
|
]
|
|
assert _surviving_calls == ["call_1"]
|
|
|
|
def test_persisted_auto_continue_note_is_not_replayed(self):
|
|
from gateway.run import _build_gateway_agent_history
|
|
|
|
history = [
|
|
{"role": "user", "content": "first real question"},
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
"[System note: Your previous turn was interrupted before you could "
|
|
"process the last tool result(s).]\n\nsecond real question"
|
|
),
|
|
},
|
|
{"role": "assistant", "content": "answer"},
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
"[System note: A new message has arrived. The conversation "
|
|
"history contains pending tool outputs from an interrupted turn.]\n\nthird"
|
|
),
|
|
},
|
|
]
|
|
|
|
agent_history, _observed_context = _build_gateway_agent_history(history)
|
|
|
|
assert agent_history == [
|
|
{"role": "user", "content": "first real question"},
|
|
{"role": "user", "content": "second real question"},
|
|
{"role": "assistant", "content": "answer"},
|
|
{"role": "user", "content": "third"},
|
|
]
|