# ── API Keys ──────────────────────────────────────────────────────────────────

# TypeSafe AI — required for Jev (System One) routing classifier
TYPESAFE_API_KEY=

# Underlying LLMs that Jev routes requests to
GROQ_API_KEY=
GEMINI_API_KEY=
OPENROUTER_API_KEY=

# ── Jev / TypeSafe SDK ────────────────────────────────────────────────────────

# Model used for routing decisions (classification step)
#   jev-latest   — always points to the current stable release
#   jev-1.13.0   — pin to a specific release for reproducibility
TYPESAFE_DEFAULT_MODEL=jev-latest

# Override the API endpoint (optional — leave blank to use the default)
# TYPESAFE_BASE_URL=https://api.typesafe.ai

# SDK log verbosity: debug | info | warning | error | off
TYPESAFE_LOG_LEVEL=warning

# ── Model Routing ─────────────────────────────────────────────────────────────
#
# Jev classifies the latest user message and selects a tier; the chosen model
# is then used for the full agent run.
#
# FAST tier — direct lookups, extraction, localised changes
# LangChain model strings use "provider:model-name" (colon separator).
#   groq:llama-3.3-70b-versatile       30 RPM  1,000/day  ← best quality
#   groq:llama-3.1-8b-instant          30 RPM  1,000/day  ← fastest
#   groq:deepseek-r1-distill-llama-70b 30 RPM  1,000/day  ← reasoning
#   groq:gemma2-9b-it                  30 RPM  1,000/day
#   groq:mixtral-8x7b-32768            30 RPM  1,000/day
#   groq:llama3-70b-8192               30 RPM  1,000/day
#   groq:llama3-8b-8192                30 RPM  1,000/day
#   google_genai:gemini-2.0-flash      15 RPM ~1,500/day
#   ──────────────────────────────────────────────────────
#   Groq + Gemini combined:           225 RPM  reliable
#
# POWERFUL tier — architecture, high-stakes, multi-step reasoning
#   Same pool of models; Jev picks this tier when the task is complex.
#
# OpenRouter :free diversity fallback (use sparingly):
#   openrouter:nvidia/nemotron-3-ultra-550b-a55b:free  20 RPM  50/day
#   openrouter:poolside/laguna-s-2.1:free              20 RPM  50/day
#   WARNING: 50 req/day free cap exhausts in ~2.5 min at full RPM.
#            Add $10 credits to raise the cap to 1,000/day.
#
# Paid options:
#   openai:gpt-4o-mini  |  openai:gpt-4o
#   anthropic:claude-haiku-4-5-20251001
#   ollama:llama3.2

# Model assigned when Jev classifies a task as "fast"
LLM_FAST_MODEL=groq:llama-3.3-70b-versatile

# Model assigned when Jev classifies a task as "powerful"
LLM_POWERFUL_MODEL=groq:deepseek-r1-distill-llama-70b

# Fallback list (tried in order if the primary provider fails)
LLM_FALLBACK_MODELS=["groq:llama-3.1-8b-instant","groq:gemma2-9b-it","groq:mixtral-8x7b-32768","groq:llama3-70b-8192","groq:llama3-8b-8192","google_genai:gemini-2.0-flash","openrouter:nvidia/nemotron-3-ultra-550b-a55b:free","openrouter:poolside/laguna-s-2.1:free"]

# Probability threshold: tasks with P(hard) >= this value use the powerful model
# 0.5 = balanced; lower to bias toward powerful; raise to bias toward fast
LLM_ROUTING_THRESHOLD=0.5

LLM_TEMPERATURE=0.7
LLM_MAX_TOKENS=2048

# ── LangSmith (optional tracing) ──────────────────────────────────────────────
# LANGSMITH_API_KEY=
# LANGSMITH_TRACING=true