mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-17 04:11:35 +00:00
The aggregator request deliberately ends `user(task), user(guidance)` on iteration 1 of every turn (#113175) so the whole prefix stays byte-stable for the provider prompt cache. Strict-alternation chat templates (llama.cpp / vLLM Jinja templates, Mistral, some OpenRouter routes) 400 on that adjacency ("Conversation roles must alternate ..."), and the turn failed. Merging proactively for everyone was declined because it brings back the byte divergence #113175/#113784 removed and only moves the 400. Reactive, destination-scoped recovery instead: - error_classifier: new `FailoverReason.role_alternation` for the vendor alternation wordings (checked before the request-validation table since the body also carries `invalid_request_error`); same abort+fallback hints as format_error so non-MoA consumers behave exactly as before. - moa_alternation (new sibling): `merge_same_role_messages` (reuses the loop's `_merge_user_content`), `destination_key` (base_url|provider, model), `is_role_alternation_rejection`. - moa_loop._call_prepared_aggregator: on that 400, retry ONCE with the adjacent user turns merged, remember the destination on the facade for the session so later iterations pre-merge, never touch destinations that accepted the split shape. The trace records the messages actually sent. - docs: caching section explains the reactive merge. Live loopback (real call_llm -> SDK -> HTTP stub that 400s on same-role adjacency): before, iteration 1 fails with BadRequestError after 1 request; after, 2 requests (split -> 400 -> merged -> 200), next turn pre-merged in 1 request; the accepting-stub control sends byte-identical requests. Fixes #112358
90 lines
4.2 KiB
Python
90 lines
4.2 KiB
Python
"""MoA aggregator: strict-alternation destinations get their adjacent same-role messages merged
|
|
reactively and per destination (#112358, last atom).
|
|
|
|
The guidance is deliberately a separate trailing ``user`` message so the prefix stays cache-stable;
|
|
a chat template that 400s on ``user(task), user(guidance)`` must get ONE merged retry, be remembered
|
|
for the session, and never change the bytes sent to destinations that accepted the split shape.
|
|
"""
|
|
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
|
|
from agent import moa_loop
|
|
from agent.error_classifier import FailoverReason, classify_api_error
|
|
|
|
ALTERNATION_MSG = "Conversation roles must alternate user/assistant/user/assistant/..."
|
|
|
|
|
|
class _AlternationRejected(Exception):
|
|
status_code = 400
|
|
|
|
def __init__(self):
|
|
super().__init__(f"Error code: 400 - {{'error': {{'message': '{ALTERNATION_MSG}', 'type': 'invalid_request_error'}}}}")
|
|
self.response = SimpleNamespace(status_code=400, headers={})
|
|
self.body = {"message": ALTERNATION_MSG, "type": "invalid_request_error"}
|
|
|
|
|
|
def _strict_destination(calls, strict_models):
|
|
"""``call_llm`` double: 400s like a strict chat template when adjacent non-system messages share a role."""
|
|
def call_llm(**kw):
|
|
calls.append(kw)
|
|
msgs = [m for m in kw["messages"] if m["role"] != "system"]
|
|
if kw["model"] in strict_models and any(a["role"] == b["role"] for a, b in zip(msgs, msgs[1:])):
|
|
raise _AlternationRejected()
|
|
return SimpleNamespace(choices=[])
|
|
return call_llm
|
|
|
|
|
|
@pytest.fixture
|
|
def facade(monkeypatch):
|
|
monkeypatch.setattr(
|
|
moa_loop, "_slot_runtime",
|
|
lambda slot: {"provider": "custom", "model": slot["model"], "base_url": "http://strict.local/v1",
|
|
"api_mode": "chat_completions"},
|
|
)
|
|
f = moa_loop.MoAChatCompletions("default", agent=None)
|
|
f._pending_trace = None
|
|
return f
|
|
|
|
|
|
def _send(facade, model, messages, guidance="[Mixture of Agents reference context]\nadvice"):
|
|
prepared = facade.rebase_prepared_request(
|
|
{"guidance": guidance, "aggregator": {"provider": "custom", "model": model}, "aggregator_temperature": None},
|
|
messages,
|
|
)
|
|
return facade._call_prepared_aggregator(prepared, {"tools": None})
|
|
|
|
|
|
def test_alternation_400_is_classified_and_retried_once_merged_then_remembered(monkeypatch, facade):
|
|
classified = classify_api_error(_AlternationRejected(), provider="custom", model="strict-model")
|
|
assert classified.reason is FailoverReason.role_alternation
|
|
|
|
calls = []
|
|
monkeypatch.setattr(moa_loop, "call_llm", _strict_destination(calls, {"strict-model"}))
|
|
task = [{"role": "system", "content": "sys"}, {"role": "user", "content": "task"}]
|
|
|
|
_send(facade, "strict-model", task) # iteration 1 of turn 1: split shape rejected → one merged retry
|
|
assert [[m["role"] for m in c["messages"]] for c in calls] == [["system", "user", "user"], ["system", "user"]]
|
|
assert calls[-1]["messages"][-1]["content"] == "task\n\n[Mixture of Agents reference context]\nadvice"
|
|
|
|
del calls[:]
|
|
_send(facade, "strict-model", [*task, {"role": "assistant", "content": "answer"}, {"role": "user", "content": "task2"}])
|
|
# Remembered destination: iteration 1 of the next turn is pre-merged, no 400 paid again.
|
|
assert [[m["role"] for m in c["messages"]] for c in calls] == [["system", "user", "assistant", "user"]]
|
|
assert calls[0]["messages"][-1]["content"].startswith("task2\n\n[Mixture of Agents reference context]")
|
|
|
|
|
|
def test_destinations_that_accept_the_split_shape_keep_byte_identical_requests(monkeypatch, facade):
|
|
calls = []
|
|
monkeypatch.setattr(moa_loop, "call_llm", _strict_destination(calls, {"strict-model"}))
|
|
task = [{"role": "system", "content": "sys"}, {"role": "user", "content": "task"}]
|
|
guidance = "[Mixture of Agents reference context]\nadvice"
|
|
|
|
_send(facade, "strict-model", task, guidance) # teaches the facade about the strict destination
|
|
del calls[:]
|
|
_send(facade, "lenient-model", task, guidance)
|
|
|
|
# The lenient destination on the SAME facade still gets the split, cache-stable shape in one request.
|
|
assert len(calls) == 1
|
|
assert calls[0]["messages"] == [*task, {"role": "user", "content": guidance}]
|