# --- GitHub App ---
# Numeric App ID from https://github.com/settings/apps/<your-app>
GITHUB_APP_ID=
# Path to the App's private key PEM file (downloaded once from the App settings).
# The key is loaded from this file only - never paste it into env values or chat.
GITHUB_PRIVATE_KEY_PATH=./private-key.pem
# Webhook secret you set on the App's settings page (used for HMAC verification)
GITHUB_WEBHOOK_SECRET=

# --- Queue ---
# Redis URL for the arq job queue. Leave UNSET for dev-inline mode
# (webhook handler runs the review in-process instead of queueing).
REDIS_URL=redis://localhost:6379/0

# --- Review backend: vllm | hf_endpoint | api | stub ---
REVIEW_BACKEND=vllm

# Primary: vLLM serving your fine-tuned model (OpenAI-compatible)
VLLM_BASE_URL=http://localhost:8000
VLLM_MODEL=code-review-llama3-lora
VLLM_TIMEOUT_S=120

# Hugging Face Inference Endpoint alternative
HF_ENDPOINT_URL=
HF_TOKEN=
HF_TIMEOUT_S=120

# --- Backup API providers (no self-hosted model needed) ---
# JSON array of OpenAI-compatible providers tried IN ORDER per request.
# If one key dies / quota runs out / the endpoint is down, the next
# provider picks it up automatically. Each entry:
#   {"name": "<label>", "preset": "<agentrouter|openai|deepseek|openrouter>",
#    "api_key": "<key>", "model": "<model-id>", "timeout_s": 120}
# Use "preset" to fill in base_url, or set "base_url" directly for any
# other OpenAI-compatible gateway. Takes precedence over FALLBACK_API_*
# when set. Keys stay in this file only - never commit .env.
PRISM_PROVIDERS=
# Example (replace dsk-... with your real key):
# PRISM_PROVIDERS='[{"name":"agentrouter","preset":"agentrouter","api_key":"dsk-...","model":"deepseek/deepseek-v4-flash"},{"name":"openai","preset":"openai","api_key":"sk-...","model":"gpt-4o-mini"}]

# Stronger fallback model (OpenAI-compatible API) for the verify second pass
FALLBACK_API_BASE=https://api.openai.com/v1
FALLBACK_API_KEY=
FALLBACK_MODEL=gpt-4o

# --- Confidence gate ---
# >= AUTO_POST: post directly. [VERIFY, AUTO_POST): second-pass verify. < VERIFY: drop.
CONFIDENCE_AUTO_POST=0.80
CONFIDENCE_VERIFY=0.55

# --- Limits ---
MAX_FILES_PER_PR=30
MAX_CONCURRENT_FILES=4
MAX_FILE_BYTES=200000
GITHUB_TIMEOUT_S=30

# --- Misc ---
REPO_CONFIG_PATH=config.yaml
FEEDBACK_DIR=feedback_data
LOG_LEVEL=INFO
# Log output format: json (production) | console (local dev, human-readable)
LOG_FORMAT=json

# --- Webhook rate limiting (single-process token bucket) ---
# Max webhook deliveries per minute, per GitHub App installation.
RATE_LIMIT_WEBHOOK_PER_MIN=30

# --- Model serving ---
# Directory holding the fine-tuned LoRA adapter. Mount it into the vLLM
# container (see docker-compose.yml `serving` profile).
PRISM_ADAPTER_DIR=
# Model name sent to vLLM for the adapter (defaults to the adapter dir name)
PRISM_ADAPTER_MODEL=
# MLflow tracking server URI for training/eval runs (optional)
MLFLOW_TRACKING_URI=
# Retries when the model's structured output fails schema validation
STRUCTURED_OUTPUT_MAX_RETRIES=2

# --- Eval gate ---
# CI fails the build if golden-eval F1 drops below this
EVAL_MIN_F1=0.70