Compare commits

...

1 Commits

Author SHA1 Message Date
Alexander Whitestone
f253d24ce3 feat: time-aware model routing for cron jobs (#317)
Some checks failed
Forge CI / smoke-and-build (pull_request) Failing after 1m7s
Empirical finding: cron error rate peaks at 18:00 (9.4%) vs 4.0% at 09:00.
Route cron tasks to more capable models during high-error windows.

Changes:
- agent/smart_model_routing.py: Added resolve_cron_model() + _hour_in_window()
- cron/scheduler.py: Wired into run_job() after base model resolution
- tests/test_cron_model_routing.py: 17 tests covering all edge cases

Config (config.yaml):
  cron_model_routing:
    enabled: true
    fallback_model: "anthropic/claude-sonnet-4"
    fallback_provider: "openrouter"
    windows:
      - start_hour: 17
        end_hour: 22
        reason: "evening_error_peak"
      - start_hour: 2
        end_hour: 5
        reason: "overnight_api_instability"

Features:
- Time windows with [start, end) semantics
- Midnight wrap support (e.g., 22-06)
- Per-window model/provider override or global fallback
- First matching window wins
- Graceful degradation: malformed windows skipped, no config = no-op
- Logs override with reason for observability

Closes #317
2026-04-13 18:48:34 -04:00
3 changed files with 315 additions and 2 deletions

View File

@@ -1,10 +1,11 @@
"""Helpers for optional cheap-vs-strong model routing."""
"""Helpers for optional cheap-vs-strong and time-aware model routing."""
from __future__ import annotations
import os
import re
from typing import Any, Dict, Optional
from datetime import datetime
from typing import Any, Dict, List, Optional
from utils import is_truthy_value
@@ -192,3 +193,105 @@ def resolve_turn_route(user_message: str, routing_config: Optional[Dict[str, Any
tuple(runtime.get("args") or ()),
),
}
# =========================================================================
# Time-aware cron model routing
# =========================================================================
#
# Empirical finding: cron error rate peaks at 18:00 (9.4%) vs 4.0% at 09:00.
# During high-error windows, route cron jobs to more capable models.
#
# Config (config.yaml):
# cron_model_routing:
# enabled: true
# fallback_model: "anthropic/claude-sonnet-4"
# fallback_provider: "openrouter"
# windows:
# - start_hour: 17
# end_hour: 22
# reason: "evening_error_peak"
# - start_hour: 2
# end_hour: 5
# reason: "overnight_api_instability"
# =========================================================================
def _hour_in_window(hour: int, start: int, end: int) -> bool:
"""Check if hour falls in [start, end) window, handling midnight wrap."""
if start <= end:
return start <= hour < end
else:
# Wraps midnight: e.g., 22-06
return hour >= start or hour < end
def resolve_cron_model(
base_model: str,
routing_config: Optional[Dict[str, Any]],
now: Optional[datetime] = None,
) -> Dict[str, Any]:
"""Apply time-aware model override for cron jobs.
During configured high-error windows, returns a stronger model config.
Outside windows, returns the base model unchanged.
Args:
base_model: The model string already resolved (from job/config/env).
routing_config: The cron_model_routing dict from config.yaml.
now: Override current time (for testing). Defaults to datetime.now().
Returns:
Dict with keys: model, provider, overridden, reason.
- model: the effective model string to use
- provider: provider override (empty string = use default)
- overridden: True if time-based override was applied
- reason: why override was applied (empty string if not)
"""
cfg = routing_config or {}
if not _coerce_bool(cfg.get("enabled"), False):
return {"model": base_model, "provider": "", "overridden": False, "reason": ""}
windows = cfg.get("windows") or []
if not isinstance(windows, list) or not windows:
return {"model": base_model, "provider": "", "overridden": False, "reason": ""}
current = now or datetime.now()
current_hour = current.hour
matched_window = None
for window in windows:
if not isinstance(window, dict):
continue
start = _coerce_int(window.get("start_hour"), -1)
end = _coerce_int(window.get("end_hour"), -1)
if start < 0 or end < 0:
continue
if _hour_in_window(current_hour, start, end):
matched_window = window
break
if not matched_window:
return {"model": base_model, "provider": "", "overridden": False, "reason": ""}
# Window matched — use the override model from window or global fallback
override_model = str(matched_window.get("model") or "").strip()
override_provider = str(matched_window.get("provider") or "").strip()
if not override_model:
override_model = str(cfg.get("fallback_model") or "").strip()
if not override_provider:
override_provider = str(cfg.get("fallback_provider") or "").strip()
if not override_model:
# No override configured — use base model
return {"model": base_model, "provider": "", "overridden": False, "reason": ""}
reason = str(matched_window.get("reason") or "time_window").strip()
return {
"model": override_model,
"provider": override_provider,
"overridden": True,
"reason": f"cron_routing:{reason}(hour={current_hour})",
}

View File

@@ -717,6 +717,22 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
# Reasoning config from env or config.yaml
from hermes_constants import parse_reasoning_effort
# Time-aware cron model routing — override model during high-error windows
try:
from agent.smart_model_routing import resolve_cron_model
_cron_routing_cfg = (_cfg.get("cron_model_routing") or {})
_cron_route = resolve_cron_model(model, _cron_routing_cfg)
if _cron_route["overridden"]:
_original_model = model
model = _cron_route["model"]
logger.info(
"Job '%s': cron model override %s%s (%s)",
job_id, _original_model, model, _cron_route["reason"],
)
except Exception as _e:
logger.debug("Job '%s': cron model routing skipped: %s", job_id, _e)
effort = os.getenv("HERMES_REASONING_EFFORT", "")
if not effort:
effort = str(_cfg.get("agent", {}).get("reasoning_effort", "")).strip()

View File

@@ -0,0 +1,194 @@
"""Tests for time-aware cron model routing — Issue #317."""
import pytest
from datetime import datetime
from agent.smart_model_routing import resolve_cron_model, _hour_in_window
class TestHourInWindow:
"""Hour-in-window detection including midnight wrap."""
def test_normal_window(self):
assert _hour_in_window(18, 17, 22) is True
assert _hour_in_window(16, 17, 22) is False
assert _hour_in_window(22, 17, 22) is False # [start, end) exclusive end
def test_midnight_wrap(self):
assert _hour_in_window(23, 22, 6) is True
assert _hour_in_window(3, 22, 6) is True
assert _hour_in_window(10, 22, 6) is False
def test_edge_cases(self):
assert _hour_in_window(0, 0, 24) is True
assert _hour_in_window(23, 0, 24) is True
assert _hour_in_window(0, 22, 6) is True
assert _hour_in_window(5, 22, 6) is True
assert _hour_in_window(6, 22, 6) is False
class TestResolveCronModel:
"""Time-aware model resolution for cron jobs."""
def _config(self, **overrides):
base = {
"enabled": True,
"fallback_model": "anthropic/claude-sonnet-4",
"fallback_provider": "openrouter",
"windows": [
{"start_hour": 17, "end_hour": 22, "reason": "evening_error_peak"},
],
}
base.update(overrides)
return base
def test_disabled_returns_base_model(self):
result = resolve_cron_model(
"xiaomi/mimo-v2-pro",
{"enabled": False},
now=datetime(2026, 4, 12, 18, 0),
)
assert result["model"] == "xiaomi/mimo-v2-pro"
assert result["overridden"] is False
def test_no_config_returns_base_model(self):
result = resolve_cron_model("xiaomi/mimo-v2-pro", None)
assert result["model"] == "xiaomi/mimo-v2-pro"
assert result["overridden"] is False
def test_no_windows_returns_base_model(self):
result = resolve_cron_model(
"xiaomi/mimo-v2-pro",
{"enabled": True, "windows": []},
now=datetime(2026, 4, 12, 18, 0),
)
assert result["overridden"] is False
def test_evening_window_overrides(self):
"""18:00 falls in [17, 22) — should override to stronger model."""
result = resolve_cron_model(
"xiaomi/mimo-v2-pro",
self._config(),
now=datetime(2026, 4, 12, 18, 0),
)
assert result["model"] == "anthropic/claude-sonnet-4"
assert result["provider"] == "openrouter"
assert result["overridden"] is True
assert "evening_error_peak" in result["reason"]
assert "hour=18" in result["reason"]
def test_outside_window_keeps_base(self):
"""09:00 is outside [17, 22) — keep base model."""
result = resolve_cron_model(
"xiaomi/mimo-v2-pro",
self._config(),
now=datetime(2026, 4, 12, 9, 0),
)
assert result["model"] == "xiaomi/mimo-v2-pro"
assert result["overridden"] is False
def test_window_boundary_start_inclusive(self):
"""17:00 is start of window — should override."""
result = resolve_cron_model(
"xiaomi/mimo-v2-pro",
self._config(),
now=datetime(2026, 4, 12, 17, 0),
)
assert result["overridden"] is True
def test_window_boundary_end_exclusive(self):
"""22:00 is end of window — should NOT override."""
result = resolve_cron_model(
"xiaomi/mimo-v2-pro",
self._config(),
now=datetime(2026, 4, 12, 22, 0),
)
assert result["overridden"] is False
def test_midnight_window(self):
"""Overnight window [22, 6) wraps midnight."""
config = self._config(windows=[
{"start_hour": 22, "end_hour": 6, "reason": "overnight_instability"},
])
# 23:00 — in window
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 23, 0))
assert result["overridden"] is True
assert "overnight_instability" in result["reason"]
# 03:00 — in window (past midnight)
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 13, 3, 0))
assert result["overridden"] is True
# 10:00 — outside window
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 10, 0))
assert result["overridden"] is False
def test_per_window_model_override(self):
"""Window-specific model takes precedence over global fallback."""
config = self._config(windows=[
{
"start_hour": 17,
"end_hour": 22,
"model": "anthropic/claude-opus-4-6",
"provider": "anthropic",
"reason": "peak_hours",
},
])
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 18, 0))
assert result["model"] == "anthropic/claude-opus-4-6"
assert result["provider"] == "anthropic"
def test_first_matching_window_wins(self):
"""When windows overlap, first match wins."""
config = self._config(windows=[
{"start_hour": 17, "end_hour": 20, "model": "strong-1", "provider": "p1", "reason": "w1"},
{"start_hour": 19, "end_hour": 22, "model": "strong-2", "provider": "p2", "reason": "w2"},
])
# 19:00 matches both — first wins
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 19, 0))
assert result["model"] == "strong-1"
def test_no_fallback_model_configured(self):
"""If no fallback_model, keeps base model even in window."""
config = {"enabled": True, "windows": [
{"start_hour": 17, "end_hour": 22, "reason": "test"},
]}
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 18, 0))
assert result["overridden"] is False
assert result["model"] == "mimo"
def test_malformed_windows_skipped(self):
"""Non-dict or missing hours in windows are skipped safely."""
config = self._config(windows=[
"not-a-dict",
{"start_hour": 17}, # missing end_hour
{"end_hour": 22}, # missing start_hour
{"start_hour": "bad", "end_hour": "bad"},
{"start_hour": 17, "end_hour": 22, "reason": "valid"},
])
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 18, 0))
assert result["overridden"] is True
assert "valid" in result["reason"]
def test_empty_provider_defaults_to_empty(self):
"""When window has no provider, falls back to global, then empty."""
config = self._config(
fallback_provider="",
windows=[{"start_hour": 17, "end_hour": 22, "reason": "test"}],
)
result = resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 18, 0))
assert result["provider"] == ""
def test_multiple_windows_coverage(self):
"""Two non-overlapping windows cover evening and overnight."""
config = self._config(windows=[
{"start_hour": 17, "end_hour": 22, "reason": "evening"},
{"start_hour": 2, "end_hour": 5, "reason": "overnight"},
])
# Evening
assert resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 20, 0))["overridden"] is True
# Overnight
assert resolve_cron_model("mimo", config, now=datetime(2026, 4, 13, 3, 0))["overridden"] is True
# Safe hours
assert resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 10, 0))["overridden"] is False
assert resolve_cron_model("mimo", config, now=datetime(2026, 4, 12, 1, 0))["overridden"] is False