# ── Providers ────────────────────────────────────────────────────────────────
# Every provider name referenced below (by a TIER_N_PROVIDER or PLANNER_PROVIDER
# value) needs a matching {PROVIDER}_BASE_URL / {PROVIDER}_API_KEY pair —
# uppercase the provider name, non-alphanumeric chars become "_"
# (e.g. "my-proxy" -> MY_PROXY_BASE_URL). Add a brand new provider just by
# adding its two vars here and pointing a tier or the planner at it below —
# no code changes required.
#
# "openai" and "anthropic" are recognized by name and default their BASE_URL
# and wire-format (_API) for you; every other provider name (including "local")
# must set {PROVIDER}_BASE_URL explicitly, and defaults to the
# openai-completions wire format (the common case for local/self-hosted
# OpenAI-compatible servers).

# Default provider — every tier and the planner use this unless overridden below.
OPENAI_API_KEY=
# OPENAI_BASE_URL=https://api.openai.com/v1   # override for Azure/a proxy

# Anthropic — uncomment to route a tier or the planner at Claude directly.
# ANTHROPIC_API_KEY=

# A local OpenAI-compatible server (Ollama's /v1 endpoint, LM Studio, vLLM,
# etc). Uncomment plus a TIER_N_PROVIDER=local line below to route a tier at it.
# LOCAL_BASE_URL=http://localhost:11434/v1
# LOCAL_API_KEY=not-needed   # most local servers ignore this but require *some* string

# ── Tiers ────────────────────────────────────────────────────────────────────
# There is NO in-code default — every var below is required, and this is the
# router's actual live config, not just documentation. The router will refuse
# to start if any TIER_N_* var is missing. The values shipped here are a ready
# -to-use OpenAI setup: copy this file to .env, fill in OPENAI_API_KEY, and
# you're running. Change any value to point that tier at a different
# provider/model — see the mixed-provider example further down.

# Tier 1 — cheapest, read-only work.
TIER_1_PROVIDER=openai
TIER_1_MODEL=gpt-5-nano
TIER_1_COST_INPUT=0.00005          # $ per 1k input tokens
TIER_1_COST_OUTPUT=0.0004
TIER_1_CONTEXT_WINDOW=400000
TIER_1_MAX_TOKENS=128000

# Tier 2 — single-file edits, mechanical refactors.
TIER_2_PROVIDER=openai
TIER_2_MODEL=gpt-5.4-nano
TIER_2_COST_INPUT=0.0002
TIER_2_COST_OUTPUT=0.00125
TIER_2_CONTEXT_WINDOW=400000
TIER_2_MAX_TOKENS=128000

# Tier 3 — multi-file work, debugging.
TIER_3_PROVIDER=openai
TIER_3_MODEL=gpt-5.4-mini
TIER_3_COST_INPUT=0.00075
TIER_3_COST_OUTPUT=0.0045
TIER_3_CONTEXT_WINDOW=400000
TIER_3_MAX_TOKENS=128000

# Tier 4 — frontier, open-ended work.
TIER_4_PROVIDER=openai
TIER_4_MODEL=gpt-5.5
TIER_4_COST_INPUT=0.005
TIER_4_COST_OUTPUT=0.03
TIER_4_CONTEXT_WINDOW=1050000
TIER_4_MAX_TOKENS=128000

# ── Example: mixed setup — 2 tiers local, 2 tiers on hosted APIs ─────────────
# To try this, comment out the matching TIER_N_* lines above and uncomment
# this whole block (plus LOCAL_BASE_URL/LOCAL_API_KEY and ANTHROPIC_API_KEY
# above): cheap/simple work stays free on a local model, harder work still
# reaches OpenAI/Anthropic. Every var is still required — local models just
# have $0 cost instead of a real price.
# TIER_1_PROVIDER=local
# TIER_1_MODEL=llama3.1:8b
# TIER_1_COST_INPUT=0
# TIER_1_COST_OUTPUT=0
# TIER_1_CONTEXT_WINDOW=8192
# TIER_1_MAX_TOKENS=4096
#
# TIER_2_PROVIDER=local
# TIER_2_MODEL=qwen2.5-coder:7b
# TIER_2_COST_INPUT=0
# TIER_2_COST_OUTPUT=0
# TIER_2_CONTEXT_WINDOW=32768
# TIER_2_MAX_TOKENS=8192
#
# TIER_3_PROVIDER=openai
# TIER_3_MODEL=gpt-5.4-mini
# TIER_3_COST_INPUT=0.00075
# TIER_3_COST_OUTPUT=0.0045
# TIER_3_CONTEXT_WINDOW=400000
# TIER_3_MAX_TOKENS=128000
#
# TIER_4_PROVIDER=anthropic
# TIER_4_MODEL=claude-opus-4-5
# TIER_4_COST_INPUT=0.005
# TIER_4_COST_OUTPUT=0.025
# TIER_4_CONTEXT_WINDOW=200000
# TIER_4_MAX_TOKENS=64000

# ── Planner ──────────────────────────────────────────────────────────────────
# The lightweight per-turn routing-decision call. Independent of the four
# tiers above — can stay on OpenAI even if every tier is local, or vice versa.
# Only openai-completions and anthropic-messages providers are supported here.
# Required, same as the tiers — no in-code default.
PLANNER_PROVIDER=openai
PLANNER_MODEL=gpt-5-mini
PLANNER_REASONING_EFFORT=low   # openai-completions only; ignored on anthropic-messages

# ── Logging & budgets ─────────────────────────────────────────────────────────
LOG_FILE_PATH=./routing.log.json
# Daily token budgets per tier (reset at local midnight)
MAX_BUDGET_TIER_1=20000
MAX_BUDGET_TIER_2=30000
MAX_BUDGET_TIER_3=40000
MAX_BUDGET_TIER_4=80000

BUDGET_STATE_PATH=./budget-state.json