feat(ai): add shadow benchmark support for alternate Mammouth models

Author: sikienzl

.env.example

@@ -150,6 +150,18 @@ AI_COPILOT_REENTRY_COOLDOWN_MIN=300
 AI_COPILOT_REENTRY_COOLDOWN_MAX=1200
 AI_COPILOT_TABULAR_BUY_CONF_MIN=0.50
 AI_COPILOT_TABULAR_BUY_CONF_MAX=0.65
+AI_COPILOT_BENCHMARK_ENABLED=false
+AI_COPILOT_BENCHMARK_MODEL=
+AI_COPILOT_BENCHMARK_INTERVAL_MINUTES=62
+AI_COPILOT_BENCHMARK_STATE_FILE=ai_copilot_benchmark_state.json
+AI_COPILOT_BENCHMARK_MAX_CALLS_PER_DAY=24
+AI_COPILOT_BENCHMARK_MAX_CALLS_PER_MONTH=720
+AI_COPILOT_BENCHMARK_MAX_BUDGET_USD_PER_MONTH=1.0
+AI_COPILOT_BENCHMARK_MAX_OUTPUT_TOKENS=300
+AI_COPILOT_BENCHMARK_TEMPERATURE=0.1
+AI_COPILOT_BENCHMARK_MAX_CONSECUTIVE_ERRORS=3
+AI_COPILOT_BENCHMARK_COST_INPUT_PER_MTOK=0.2
+AI_COPILOT_BENCHMARK_COST_OUTPUT_PER_MTOK=1.25
 
 # AutoResearch bridge – Fear & Greed Index fetcher (no API key required)
 # The fetch_research_signal.py script is run by the research-signal.timer on

.env.live.example

@@ -137,6 +137,18 @@ AI_COPILOT_REENTRY_COOLDOWN_MIN=600
 AI_COPILOT_REENTRY_COOLDOWN_MAX=1800
 AI_COPILOT_TABULAR_BUY_CONF_MIN=0.55
 AI_COPILOT_TABULAR_BUY_CONF_MAX=0.65
+AI_COPILOT_BENCHMARK_ENABLED=false
+AI_COPILOT_BENCHMARK_MODEL=
+AI_COPILOT_BENCHMARK_INTERVAL_MINUTES=62
+AI_COPILOT_BENCHMARK_STATE_FILE=ai_copilot_benchmark_state.json
+AI_COPILOT_BENCHMARK_MAX_CALLS_PER_DAY=24
+AI_COPILOT_BENCHMARK_MAX_CALLS_PER_MONTH=720
+AI_COPILOT_BENCHMARK_MAX_BUDGET_USD_PER_MONTH=1.0
+AI_COPILOT_BENCHMARK_MAX_OUTPUT_TOKENS=250
+AI_COPILOT_BENCHMARK_TEMPERATURE=0.1
+AI_COPILOT_BENCHMARK_MAX_CONSECUTIVE_ERRORS=3
+AI_COPILOT_BENCHMARK_COST_INPUT_PER_MTOK=0.2
+AI_COPILOT_BENCHMARK_COST_OUTPUT_PER_MTOK=1.25
 
 # AutoResearch bridge (conservative fail-open with neutral fallback)
 TABULAR_RESEARCH_SIGNAL_PATH=./data/research_signal_latest.json