cortex.yaml

Rama main · réplica de lectura del repositorio original.

# CORTEX configuration. Every key can be overridden with a CORTEX_<KEY>
# environment variable (lists and nested blocks as JSON), e.g.
#   CORTEX_OLLAMA_MODEL=llama3.1:8b   CORTEX_API_KEY=change-me
# Run `cortex doctor` after editing to check the setup.

# --- Models (served by Ollama) ----------------------------------------------
ollama_base_url: http://localhost:11434
# One tool-capable model for every chat role: a single `ollama pull`.
ollama_model: qwen2.5:7b
reasoning_model: qwen2.5:7b
code_model: qwen2.5:7b
embed_model: nomic-embed-text

# --- Storage (created on first run; wipe with `cortex reset-memory --yes`) --
# Relative paths in this file are relative to this file's folder.
chroma_path: ./.cortex/chroma
db_path: ./.cortex/cortex.db

# --- API ----------------------------------------------------------------------
# Loopback only by default: the agent can run code and read/write workspace
# files. To expose it, set api_host, api_key, allowed_hosts and cors_origins.
api_host: 127.0.0.1
api_port: 8000
log_level: INFO

# --- Observability ------------------------------------------------------------
# Off for native runs (no collector). The Docker stack turns it on for Jaeger.
telemetry_enabled: false
otel_endpoint: http://localhost:4317

# --- Agent and tools ----------------------------------------------------------
max_agent_steps: 10
plugins_dir: ./src/cortex/tools/plugins
# The only folder the agent can read and write (created on start). Keep it
# separate from this project: the agent must not reach cortex.yaml or CORTEX's
# own code. Hidden files (.env, .git, ...) are off-limits wherever it points.
allowed_root: ./workspace
allowed_write_extensions:
  - .txt
  - .md
  - .json
  - .csv
  - .py
tool_timeout_seconds: 30.0
code_exec_timeout_seconds: 10.0
lats:
  enabled: false
  max_depth: 5
  n_branches: 3
  budget: 10
supervisor:
  max_workers: 3
rate_limit:
  enabled: true
  requests_per_minute: 200
  chat_requests_per_minute: 60