hermes: resolve durable lane model ids
This commit is contained in:
parent
846c890527
commit
f072685a69
@ -169,10 +169,26 @@ def select_route(
|
|||||||
with open_request(request, timeout=60) as response:
|
with open_request(request, timeout=60) as response:
|
||||||
selected = str(response.headers.get("x-model-router-selected-model") or "")
|
selected = str(response.headers.get("x-model-router-selected-model") or "")
|
||||||
rationale = str(response.headers.get("x-model-router-rationale") or "")
|
rationale = str(response.headers.get("x-model-router-rationale") or "")
|
||||||
response.read()
|
response_body = response.read()
|
||||||
except (OSError, urllib.error.URLError) as exc:
|
except (OSError, urllib.error.URLError) as exc:
|
||||||
raise RuntimeError(f"Switchyard worker routing failed: {exc}") from exc
|
raise RuntimeError(f"Switchyard worker routing failed: {exc}") from exc
|
||||||
provider, model, effort = _decode_worker_target(selected)
|
provider, model, effort = _decode_worker_target(selected)
|
||||||
|
# Switchyard preserves the stable tier target in the selection header and
|
||||||
|
# top-level response model. The worker broker's assistant content contains
|
||||||
|
# the steward-resolved provider model required by the native CLI.
|
||||||
|
try:
|
||||||
|
response_document = json.loads(response_body)
|
||||||
|
resolved_target = str(response_document["choices"][0]["message"]["content"] or "")
|
||||||
|
resolved_provider, resolved_model, resolved_effort = _decode_worker_target(resolved_target)
|
||||||
|
required_prefix = "gpt-" if provider == "codex" else "claude-"
|
||||||
|
if (
|
||||||
|
resolved_provider == provider
|
||||||
|
and resolved_effort == effort
|
||||||
|
and resolved_model.startswith(required_prefix)
|
||||||
|
):
|
||||||
|
model = resolved_model
|
||||||
|
except (AttributeError, IndexError, KeyError, RuntimeError, TypeError, ValueError, json.JSONDecodeError):
|
||||||
|
pass
|
||||||
if exclude_provider and provider == exclude_provider:
|
if exclude_provider and provider == exclude_provider:
|
||||||
raise RuntimeError(
|
raise RuntimeError(
|
||||||
f"Switchyard selected excluded provider {exclude_provider} for retry"
|
f"Switchyard selected excluded provider {exclude_provider} for retry"
|
||||||
|
|||||||
81
testing/tests/test_hermes_cli_lane_routing.py
Normal file
81
testing/tests/test_hermes_cli_lane_routing.py
Normal file
@ -0,0 +1,81 @@
|
|||||||
|
"""Regression tests for durable Hermes worker-route model resolution."""
|
||||||
|
|
||||||
|
from __future__ import annotations
|
||||||
|
|
||||||
|
import importlib.util
|
||||||
|
import json
|
||||||
|
import sys
|
||||||
|
from pathlib import Path
|
||||||
|
|
||||||
|
|
||||||
|
SCRIPT = (
|
||||||
|
Path(__file__).parents[2] / "services/hermes/scripts/cli_lane_runner.py"
|
||||||
|
)
|
||||||
|
SPEC = importlib.util.spec_from_file_location("cli_lane_routing", SCRIPT)
|
||||||
|
assert SPEC and SPEC.loader
|
||||||
|
lanes = importlib.util.module_from_spec(SPEC)
|
||||||
|
sys.modules[SPEC.name] = lanes
|
||||||
|
SPEC.loader.exec_module(lanes)
|
||||||
|
|
||||||
|
|
||||||
|
class SwitchyardResponse:
|
||||||
|
"""Return a tier header and independently resolved broker content."""
|
||||||
|
|
||||||
|
def __init__(self, selected: str, resolved_target: str):
|
||||||
|
self.headers = {
|
||||||
|
"x-model-router-selected-model": selected,
|
||||||
|
"x-model-router-rationale": "test route",
|
||||||
|
}
|
||||||
|
self.resolved_target = resolved_target
|
||||||
|
|
||||||
|
def __enter__(self):
|
||||||
|
return self
|
||||||
|
|
||||||
|
def __exit__(self, *_args):
|
||||||
|
return False
|
||||||
|
|
||||||
|
def read(self) -> bytes:
|
||||||
|
return json.dumps(
|
||||||
|
{
|
||||||
|
"model": self.headers["x-model-router-selected-model"],
|
||||||
|
"choices": [
|
||||||
|
{"message": {"content": self.resolved_target}}
|
||||||
|
],
|
||||||
|
}
|
||||||
|
).encode()
|
||||||
|
|
||||||
|
|
||||||
|
def test_switchyard_tier_is_resolved_before_invoking_codex():
|
||||||
|
"""Codex receives the exact account-visible ID, not a tier shorthand."""
|
||||||
|
route = lanes.select_route(
|
||||||
|
"Repair the failing implementation.",
|
||||||
|
"cli-auto",
|
||||||
|
open_request=lambda *_args, **_kwargs: SwitchyardResponse(
|
||||||
|
"worker/codex/terra/xhigh",
|
||||||
|
"worker/codex/gpt-5.6-terra/xhigh",
|
||||||
|
),
|
||||||
|
)
|
||||||
|
|
||||||
|
assert (route.provider, route.model, route.effort) == (
|
||||||
|
"codex",
|
||||||
|
"gpt-5.6-terra",
|
||||||
|
"xhigh",
|
||||||
|
)
|
||||||
|
|
||||||
|
|
||||||
|
def test_mismatched_broker_resolution_cannot_change_route_decision():
|
||||||
|
"""Resolved content cannot change Switchyard's provider or effort."""
|
||||||
|
route = lanes.select_route(
|
||||||
|
"Repair it.",
|
||||||
|
"cli-auto",
|
||||||
|
open_request=lambda *_args, **_kwargs: SwitchyardResponse(
|
||||||
|
"worker/codex/terra/xhigh",
|
||||||
|
"worker/claude/claude-opus-5/high",
|
||||||
|
),
|
||||||
|
)
|
||||||
|
|
||||||
|
assert (route.provider, route.model, route.effort) == (
|
||||||
|
"codex",
|
||||||
|
"terra",
|
||||||
|
"xhigh",
|
||||||
|
)
|
||||||
Loading…
x
Reference in New Issue
Block a user