# Cursor API key used by the Cursor Cloud Agents LLM provider.
# Create one at https://cursor.com/dashboard/api
CURSOR_API_KEY=

# Kimi / Moonshot API key for the OpenAI-compatible provider.
KIMI_API_KEY=

# OpenRouter API key for OpenAI-compatible routed models.
OPENROUTER_API_KEY=

# Groq API key for the OpenAI-compatible provider.
GROQ_API_KEY=

# Google Gemini API key for the Gemini OpenAI-compatible endpoint.
GEMINI_API_KEY=

# Provider selection and model defaults.
# Supported: cursor, kimi, openrouter, groq, gemini
LLM_PROVIDER=cursor
LLM_API_KEY=
# Model override. This Google model is routed through Cursor Cloud Agents.
LLM_MODEL=gemini-3.7-flash
# Model used by the optional LLM summarizer and as the Cursor default when
# LLM_MODEL is unset.
LLM_FAST_MODEL=gemini-3.7-flash
# Ask the Cursor Cloud Agents API to run composer models in fast mode.
CURSOR_FAST_MODE=true
# When true, the orchestrator spends one LLM call (fastest model) summarizing
# each artifact before it is handed downstream. Default OFF: deterministic
# Python digests preserve the same cross-artifact contracts with zero LLM calls
# (saves ~5 calls and several minutes per run).
SUMMARIZE_WITH_LLM=false
LLM_BASE_URL=https://api.cursor.com/v1
KIMI_BASE_URL=https://api.moonshot.cn/v1
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
GROQ_BASE_URL=https://api.groq.com/openai/v1
GEMINI_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai

# Provider behaviour
# LLM_POLL_TIMEOUT_S bounds each agent generation. Keep it tight: the workflow
# runs five sequential agents plus a review, so a long timeout multiplies
# wall-clock time on failures. Lower values fail fast instead of hanging.
LLM_REQUEST_TIMEOUT_S=120
# Must clear the largest artifact an agent emits, or the response is cut off
# mid-object and the run is wasted.
LLM_MAX_TOKENS=8192
LLM_POLL_INTERVAL_S=1.0
LLM_POLL_TIMEOUT_S=300

# Workflow limits
STRUCTURED_OUTPUT_MAX_RETRIES=1
# The reviewer runs at most once; a single targeted regeneration round follows.
MAX_REVIEW_ROUNDS=1
# Each artifact is regenerated at most once per workflow.
MAX_ARTIFACT_REVISIONS=1
# Bounded retries for transient LLM/provider failures.
MAX_LLM_RETRIES=1