.env.example

# Database
# Backend logging. BACKEND_LOG_FILE writes backend + uvicorn logs to a rotating file.
# LOG_LEVEL=INFO
# BACKEND_LOG_FILE=./blueprint-backend.log
# BACKEND_LOG_MAX_BYTES=10000000
# BACKEND_LOG_BACKUP_COUNT=5
# Debug mode returns traceback/context fields in API error responses and failed
# job metadata. Keep disabled outside trusted local/dev environments.
# BLUEPRINT_DEBUG=false

# Supabase is used through the Supabase Python client when these are present.
# With no Supabase client variables, the backend falls back to SQLite.
# DATABASE_BACKEND=supabase
# SUPABASE_URL=https://your-project-ref.supabase.co
# SUPABASE_SERVICE_ROLE_KEY=your_service_role_key_here
# SUPABASE_SECRET_KEY can be used for newer Supabase secret keys.
# Do not use SUPABASE_ANON_KEY, NEXT_PUBLIC_SUPABASE_ANON_KEY, or a publishable key for the backend.
# Force local SQLite instead:
# DATABASE_BACKEND=sqlite
# SQLITE_DATABASE_URL=sqlite:///./blueprint.db
# Development mode always uses SQLite and disables Supabase Storage writes,
# even if Supabase credentials are present in your shell or .env.
# BLUEPRINT_DEV_MODE=true

# Frontend API base URL. For Docker, this value is baked into the frontend image
# at build time, so rebuild after changing it.
NEXT_PUBLIC_API_URL=http://localhost:8000
# Expose local-only frontend controls such as Keys, Listening Jobs, and Backend Logs.
# These are baked into production frontend builds; keep unset/false publicly.
# NEXT_PUBLIC_BLUEPRINT_DEBUG=false
# NEXT_PUBLIC_BLUEPRINT_DEV_MODE=false

# Deployment admin allowlist. Admins can view global jobs/log diagnostics.
# Prefer Clerk user IDs in production; email allowlisting requires CLERK_SECRET_KEY.
# BLUEPRINT_ADMIN_USER_IDS=user_...
# BLUEPRINT_ADMIN_EMAILS=isayahculbertson@gmail.com

# Docker defaults to persistent SQLite under /data when no Supabase client env
# vars are present.

# Deployment mode requires a configured deployment provider or a signed-in
# user's BYOK provider for generation. The frontend keeps the composer visible
# and directs users without an active provider to Settings.
# BLUEPRINT_DEPLOYMENT=true

# LLM provider: openai, baseten, gmi, huggingface, nvidia, gemini, openai-compatible, runpod, runpod-serverless, or simulation.
# Use runpod for Runpod OpenAI-compatible/vLLM endpoints and runpod-serverless for queue-style /runsync workers.
LLM_PROVIDER=openai
OPENAI_API_KEY=your_openai_api_key_here
OPENAI_MODEL=gpt-4o-mini
STRICT_LLM=true

# Optional runtime provider/model switching.
# Requests to /generate may include {"provider":"openai","model":"gpt-4o-mini"}.
# If unset, allowed providers default to configured providers detected from env.
# Runtime model overrides default to the configured model/fallback unless allowlists are set.
# LLM_ALLOWED_PROVIDERS=openai,baseten,gmi,huggingface,nvidia,openai-compatible,gemini,runpod,runpod-serverless,simulation
# OPENAI_ALLOWED_MODELS=gpt-4o-mini,gpt-5.5
# BASETEN_ALLOWED_MODELS=deepseek-ai/DeepSeek-V4-Pro
# GMI_ALLOWED_MODELS=anthropic/claude-fable-5
# HUGGINGFACE_ALLOWED_MODELS=Qwen/Qwen2.5-Coder-3B-Instruct:nscale
# NVIDIA_ALLOWED_MODELS=qwen/qwen2.5-coder-32b-instruct,meta/llama-3.1-8b-instruct
# OPENAI_COMPATIBLE_ALLOWED_MODELS=your-compatible-model
# GEMINI_ALLOWED_MODELS=gemini-3.5-flash,gemini-2.5-flash
# RUNPOD_ALLOWED_MODELS=your-runpod-model-v1,your-runpod-model-v2

# First-party OpenAI options.
# OPENAI_RESPONSE_FORMAT=json_schema
# OPENAI_VALIDATE_MODELS=false
# OPENAI_TIMEOUT_SECONDS=300
# For GPT-5/o-series reasoning models.
# OPENAI_REASONING_EFFORT=low
# OPENAI_TEMPERATURE=1
# OPENAI_PROJECT_ID=your_openai_project_id_here
# OPENAI_ORG_ID=your_openai_org_id_here

# Optional OpenAI API verifier.
# Run from the repo root with:
#   ./scripts/verify-openai.py
# OPENAI_BASE_URL=https://api.openai.com/v1
# OPENAI_TEST_PROMPT=Reply with exactly: openai ok
# OPENAI_INPUT_TEMPLATE={"input":"{prompt}"}
# OPENAI_TEST_MAX_OUTPUT_TOKENS=32

# Optional Langfuse observability for generation traces and structured LLM calls.
# LANGFUSE_PUBLIC_KEY=pk-lf-your_public_key_here
# LANGFUSE_SECRET_KEY=sk-lf-your_secret_key_here
# LANGFUSE_BASE_URL=https://cloud.langfuse.com
# LANGFUSE_TRACING_ENVIRONMENT=local
# LANGFUSE_TRACING_RELEASE=dev
# LANGFUSE_MAX_FIELD_CHARS=20000
# LANGFUSE_ENABLED=false

# Optional generated product image output.
IMAGE_OUTPUT_ENABLED=false
IMAGE_PROVIDER=openai
# First-party OpenAI image generation uses OPENAI_IMAGE_API_KEY, or OPENAI_API_KEY
# when OPENAI_IMAGE_API_KEY is unset. It does not inherit LLM_API_KEY/LLM_BASE_URL.
OPENAI_IMAGE_MODEL=gpt-image-2
OPENAI_IMAGE_SIZE=1024x1024
# OPENAI_IMAGE_API_KEY=your_openai_image_key_here
# OPENAI_IMAGE_BASE_URL=https://api.openai.com/v1
# OPENAI_IMAGE_QUALITY=medium
# OPENAI_IMAGE_OUTPUT_FORMAT=png

# Optional Firecrawl source research for the web_research workflow.
# EXTERNAL_SOURCE_PROVIDER=firecrawl
# With FIRECRAWL_API_KEY set, Firecrawl defaults to: npx -y firecrawl-mcp
# Override the command if your MCP server is installed or hosted differently.
# FIRECRAWL_API_KEY=your_firecrawl_api_key_here
# FIRECRAWL_MCP_COMMAND=npx -y firecrawl-mcp
# FIRECRAWL_SEARCH_LIMIT=3
# FIRECRAWL_MCP_TIMEOUT_SECONDS=45
# FIRECRAWL_MCP_DISABLED=false

# Optional Supabase Storage upload for reference/product images.
# Uses SUPABASE_URL plus SUPABASE_SERVICE_ROLE_KEY/SUPABASE_SECRET_KEY through
# the Supabase client by default. Bucket defaults to contents.
SUPABASE_S3_ENDPOINT=your-s3-endpoint
SUPABASE_S3_BUCKET=contents
# SUPABASE_S3_REGION=us-east-1
# Optional fallback for S3-compatible uploads when Supabase client env is absent.
# SUPABASE_S3_ACCESS_KEY_ID=your_supabase_s3_access_key_id
# SUPABASE_S3_SECRET_ACCESS_KEY=your_supabase_s3_secret_access_key
# Signed read URLs are refreshed by the backend when projects are read.
# SUPABASE_IMAGE_SIGNED_URL_SECONDS=86400
# Optional public URL base for generated object URLs.
# SUPABASE_STORAGE_PUBLIC_BASE_URL=....

# Generic provider aliases are supported for all live providers.
# LLM_API_KEY=your_provider_api_key_here
# LLM_MODEL=gpt-4o-mini
# LLM_FALLBACK_MODEL=your_fallback_model_here

# OpenAI-compatible providers can set a custom endpoint.
# LLM_PROVIDER=openai-compatible
# LLM_BASE_URL=http://localhost:11434/v1
# LLM_ALLOW_NO_API_KEY=true
# LLM_RESPONSE_FORMAT=json_object
# LLM_VALIDATE_MODELS=false
# Structured calls default to 8192 output tokens (6000 floor for large schemas).
# LLM_MAX_TOKENS=8192
# LLM_TIMEOUT_SECONDS=90
# LLM_REASONING_EFFORT=low
# LLM_TEMPERATURE=0.2

# Baseten Model APIs. These use an OpenAI-compatible chat completions endpoint.
# LLM_PROVIDER=baseten
# BASETEN_API_KEY=your_baseten_api_key_here
# BASETEN_BASE_URL=https://inference.baseten.co/v1
# BASETEN_MODEL=deepseek-ai/DeepSeek-V4-Pro
# BASETEN_RESPONSE_FORMAT=json_object
# BASETEN_VALIDATE_MODELS=false
# BASETEN_TIMEOUT_SECONDS=300
# BASETEN_TEMPERATURE=0.2

# GMI Cloud serverless LLM APIs. These use an OpenAI-compatible chat completions endpoint.
# LLM_PROVIDER=gmi
# GMI_API_KEY=your_gmi_cloud_api_key_here
# GMI_BASE_URL=https://api.gmi-serving.com/v1
# GMI_MODEL=anthropic/claude-fable-5
# GMI_RESPONSE_FORMAT=json_object
# GMI_VALIDATE_MODELS=false
# GMI_TIMEOUT_SECONDS=300
# Fable rejects temperature; omit this unless a different GMI model supports it.
# GMI_TEMPERATURE=omit

# Fireworks video self-correction reviewer. Requires ffmpeg on the backend host.
# Auto mode uses the working Kimi frame-review model unless a native video deployment is configured.
# FIREWORKS_API_KEY=your_fireworks_api_key_here
# FIREWORKS_BASE_URL=https://api.fireworks.ai/inference/v1
# FIREWORKS_VIDEO_REVIEW_INPUT_MODE=auto
# FIREWORKS_VIDEO_REVIEW_MODEL=kimi-k2p6
# Set these to use Fireworks native video/audio with qwen3-omni-30b-a3b-instruct.
# FIREWORKS_ACCOUNT_ID=your_fireworks_account_id
# FIREWORKS_VIDEO_REVIEW_DEPLOYMENT_ID=your_fireworks_deployment_id
# FIREWORKS_VIDEO_REVIEW_MODEL=qwen3-omni-30b-a3b-instruct
# FIREWORKS_VIDEO_REVIEW_MAX_SECONDS=60
# FIREWORKS_VIDEO_REVIEW_NATIVE_FPS=1
# FIREWORKS_VIDEO_REVIEW_NATIVE_HEIGHT=360
# FIREWORKS_VIDEO_REVIEW_MAX_MEDIA_BYTES=10000000
# Set FIREWORKS_VIDEO_REVIEW_INPUT_MODE=frames to force image-frame fallback.
# FIREWORKS_VIDEO_REVIEW_MAX_FRAMES=8
# FIREWORKS_TIMEOUT_SECONDS=180

# Hugging Face Inference Providers router. This is OpenAI-compatible for chat completions.
# LLM_PROVIDER=huggingface
# HF_TOKEN=your_huggingface_token_here
# HUGGINGFACE_BASE_URL=https://router.huggingface.co/v1
# HUGGINGFACE_MODEL=Qwen/Qwen2.5-Coder-3B-Instruct:nscale
# HUGGINGFACE_RESPONSE_FORMAT=json_object
# HUGGINGFACE_VALIDATE_MODELS=false
# HUGGINGFACE_TIMEOUT_SECONDS=300
# HUGGINGFACE_TEMPERATURE=0.2

# NVIDIA Build / NIM APIs. These use an OpenAI-compatible chat completions endpoint.
# LLM_PROVIDER=nvidia
# NVIDIA_API_KEY=your_nvidia_api_key_here
# NVIDIA_BASE_URL=https://integrate.api.nvidia.com/v1
# NVIDIA_MODEL=qwen/qwen2.5-coder-32b-instruct
# NVIDIA_RESPONSE_FORMAT=json_object
# NVIDIA_VALIDATE_MODELS=false
# NVIDIA_TIMEOUT_SECONDS=300
# NVIDIA_TEMPERATURE=0.2

# Optional Runpod Serverless endpoint verifier.
# Run from the repo root with:
#   ./scripts/verify-runpod-serverless.py
# RUNPOD_API_KEY=your_runpod_api_key_here
# RUNPOD_OPENAI_BASE_URL=https://api.runpod.ai/v2/your_endpoint_id_here/openai/v1
# RUNPOD_ENDPOINT_ID=your_endpoint_id_here
# RUNPOD_ENDPOINT_URL=https://api.runpod.ai/v2/your_endpoint_id_here
# For one Runpod endpoint per model, map model IDs to endpoint IDs or endpoint URLs.
# RUNPOD_MODEL_ENDPOINTS={"your-runpod-model-v1":"endpoint_id_1","your-runpod-model-v2":"https://api.runpod.ai/v2/endpoint_id_2"}
# For one endpoint that serves many models, include {model} in RUNPOD_INPUT_TEMPLATE if your worker expects it.
# RUNPOD_TEST_PROMPT=Reply with exactly: runpod serverless ok
# RUNPOD_INPUT_TEMPLATE={"prompt":"{prompt}"}
# RUNPOD_TIMEOUT_SECONDS=1200
# RUNPOD_WAIT_MS=90000
# RUNPOD_POLL_TIMEOUT_SECONDS=1200
# RUNPOD_POLL_INTERVAL_SECONDS=5
# RUNPOD_PARTI_SEED_TIMEOUT_SECONDS=1200
# RUNPOD_EXECUTION_TIMEOUT_MS=1200000
# RUNPOD_TTL_MS=1200000
# RUNPOD_MODEL=your_optional_model_id_here
# RUNPOD_TEMPERATURE=0.2
# Structured calls default to 8192 output tokens with a 6000 floor for large
# schemas (MechanicalNotes) so big JSON records are not truncated mid-string.
# RUNPOD_MAX_TOKENS=8192
# Grammar-constrained JSON on the vLLM worker (requires vLLM >= 0.6.3). Default
# is json_object; switch to json_schema only after a live smoke test.
# RUNPOD_RESPONSE_FORMAT=json_schema

# Backward-compatible Gemini aliases are still supported.
# GEMINI_API_KEY=your_gemini_api_key_here
# GOOGLE_API_KEY=your_google_api_key_here
# GEMINI_MODEL=gemini-3.5-flash
# STRICT_GEMINI=true
# GEMINI_FALLBACK_MODEL=gemini-2.5-flash

# Optional agent-to-agent TCP JSONL socket. REST, WebSocket, and MCP routes are always mounted.
JOB_METADATA_BACKEND=auto
JOB_METADATA_DB_PATH=./blueprint_jobs.db
A2A_SOCKET_ENABLED=false
A2A_SOCKET_HOST=127.0.0.1
A2A_SOCKET_PORT=8766