Files
hermes-agent/agent/time_aware_routing.py
Alexander Whitestone 0b72884750
Some checks failed
Tests / test (pull_request) Failing after 25m4s
Tests / e2e (pull_request) Successful in 3m19s
Contributor Attribution Check / check-attribution (pull_request) Failing after 14s
Docker Build and Publish / build-and-push (pull_request) Has been skipped
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Successful in 14s
feat: time-aware model routing for cron jobs (#889)
Error rate peaks at 18:00 (9.4%) during evening cron batches vs 4.0%
at 09:00 during interactive work. Route cron tasks to stronger models
during off-hours when user is not present to correct errors.

New agent/time_aware_routing.py:
- resolve_time_aware_model(): routes based on hour, error rate, task type
- Interactive sessions: always use base model (user corrects errors)
- Cron during business hours: use base model (low error rate)
- Cron during off-hours with high error rate (>6%): upgrade to strong model
- get_hour_error_rate(): error rates by hour from empirical audit
- is_off_hours(): 18:00-05:59 = off-hours
- RoutingDecision: model, provider, reason, hour, error_rate
- get_routing_report(): 24h forecast of routing decisions

Config via env vars:
- CRON_STRONG_MODEL (default: xiaomi/mimo-v2-pro)
- CRON_CHEAP_MODEL (default: qwen2.5:7b)
- CRON_ERROR_THRESHOLD (default: 6.0%)

Tests: tests/test_time_aware_routing.py (9 tests)

Closes #889
2026-04-17 01:15:09 -04:00

147 lines
4.6 KiB
Python

"""Time-aware model routing for cron jobs.
Routes cron tasks to more capable models during off-hours when the user
is not present to correct errors. Reduces error rates during high-error
time windows (e.g., 18:00 evening batches).
Usage:
from agent.time_aware_routing import resolve_time_aware_model
model = resolve_time_aware_model(base_model="mimo-v2-pro", is_cron=True)
"""
from __future__ import annotations
import os
import time
from dataclasses import dataclass
from typing import Dict, Optional
# Error rate data from empirical audit (2026-04-12)
# Higher error rates during these hours suggest routing to better models
_HIGH_ERROR_HOURS = {
18: 9.4, # 18:00 — 9.4% error rate (evening cron batches)
19: 8.1,
20: 7.5,
21: 6.8,
22: 6.2,
23: 5.9,
0: 5.5,
1: 5.2,
}
# Low error hours — default model is fine
_LOW_ERROR_HOURS = set(range(6, 18)) # 06:00-17:59
# Default fallback models by time zone
_DEFAULT_STRONG_MODEL = os.getenv("CRON_STRONG_MODEL", "xiaomi/mimo-v2-pro")
_DEFAULT_CHEAP_MODEL = os.getenv("CRON_CHEAP_MODEL", "qwen2.5:7b")
_ERROR_THRESHOLD = float(os.getenv("CRON_ERROR_THRESHOLD", "6.0")) # % error rate
@dataclass
class RoutingDecision:
"""Result of time-aware routing."""
model: str
provider: str
reason: str
hour: int
error_rate: float
is_off_hours: bool
def get_hour_error_rate(hour: int) -> float:
"""Get expected error rate for a given hour (0-23)."""
return _HIGH_ERROR_HOURS.get(hour, 4.0) # Default 4% for unlisted hours
def is_off_hours(hour: int) -> bool:
"""Check if hour is considered off-hours (higher error rates)."""
return hour not in _LOW_ERROR_HOURS
def resolve_time_aware_model(
base_model: str = "",
base_provider: str = "",
is_cron: bool = False,
hour: Optional[int] = None,
) -> RoutingDecision:
"""Resolve model based on time of day and task type.
During off-hours (evening/night), routes to stronger models for cron
jobs to compensate for lack of human oversight.
Args:
base_model: The model that would normally be used.
base_provider: The provider for the base model.
is_cron: Whether this is a cron job (vs interactive session).
hour: Override hour (for testing). Defaults to current hour.
Returns:
RoutingDecision with model, provider, and reasoning.
"""
if hour is None:
hour = time.localtime().tm_hour
error_rate = get_hour_error_rate(hour)
off_hours = is_off_hours(hour)
# Interactive sessions always use the base model (user can correct errors)
if not is_cron:
return RoutingDecision(
model=base_model or _DEFAULT_CHEAP_MODEL,
provider=base_provider,
reason="Interactive session — user can correct errors",
hour=hour,
error_rate=error_rate,
is_off_hours=off_hours,
)
# Cron jobs during low-error hours: use base model
if not off_hours and error_rate < _ERROR_THRESHOLD:
return RoutingDecision(
model=base_model or _DEFAULT_CHEAP_MODEL,
provider=base_provider,
reason=f"Low-error hours ({hour}:00, {error_rate}% expected)",
hour=hour,
error_rate=error_rate,
is_off_hours=False,
)
# Cron jobs during high-error hours: upgrade to stronger model
if error_rate >= _ERROR_THRESHOLD:
return RoutingDecision(
model=_DEFAULT_STRONG_MODEL,
provider="nous",
reason=f"High-error hours ({hour}:00, {error_rate}% expected) — using stronger model",
hour=hour,
error_rate=error_rate,
is_off_hours=True,
)
# Off-hours but low error: use base model
return RoutingDecision(
model=base_model or _DEFAULT_CHEAP_MODEL,
provider=base_provider,
reason=f"Off-hours but low error ({hour}:00, {error_rate}%)",
hour=hour,
error_rate=error_rate,
is_off_hours=off_hours,
)
def get_routing_report() -> str:
"""Get a report of time-based routing decisions for the next 24 hours."""
lines = ["Time-Aware Model Routing (24h forecast)", "=" * 40, ""]
lines.append(f"Error threshold: {_ERROR_THRESHOLD}%")
lines.append(f"Strong model: {_DEFAULT_STRONG_MODEL}")
lines.append(f"Cheap model: {_DEFAULT_CHEAP_MODEL}")
lines.append("")
for h in range(24):
decision = resolve_time_aware_model(is_cron=True, hour=h)
icon = "\U0001f7e2" if decision.model == _DEFAULT_CHEAP_MODEL else "\U0001f534"
lines.append(f" {h:02d}:00 {icon} {decision.model:25s} ({decision.error_rate}% error)")
return "\n".join(lines)