config.yaml

# autoskill configuration
#
# LLM backend for skill synthesis. Detection/clustering always runs locally;
# only redacted cluster summaries are sent to the LLM.
#
# Local is the default — your screen content never leaves the machine.
# Cloud backends (claude, foundry) are available for users who explicitly
# opt in; see the backend sections below.
backend: local    # local | claude | foundry

# Per-backend settings. Only the selected backend's block is used.
local:
  # LM Studio exposes an OpenAI-compatible server. Start it from the
  # "Developer" tab; the default port is 1234.
  endpoint: http://localhost:1234/v1
  # Gemma-4-31B-it is the recommended default: strong reasoning at a size
  # most modern workstation GPUs can run. Swap for any LM Studio model ID.
  model: Gemma-4-31B-it

claude:
  model: claude-opus-4-7
  # api_key read from ANTHROPIC_API_KEY env var

foundry:
  endpoint: https://foundry.example.com/anthropic
  model: claude-opus-4-7
  # api_key read from FOUNDRY_API_KEY env var

# Screenpipe HTTP endpoint. For TLS, point this at your local Caddy proxy
# (see references/https-proxy.md).
screenpipe:
  url: http://localhost:3030
  # Screenpipe requires a bearer token for its local API. Either set `token`
  # here, or export SCREENPIPE_TOKEN in your environment (preferred — keeps
  # the token out of version control). Retrieve with: `screenpipe auth token`.
  # token: your-token-here

# Embedding model for matching against existing scientific skills.
# Local only; no API calls.
embeddings:
  model: sentence-transformers/all-MiniLM-L6-v2

# Clustering thresholds.
cluster:
  min_session_minutes: 5        # skip sessions shorter than this
  idle_gap_minutes: 10          # new session after this much inactivity
  min_cluster_size: 2           # need this many similar sessions before proposing

# Content redaction regexes applied before any cluster summary leaves
# the local detection layer. Defense-in-depth on top of screenpipe's
# own app/window filtering.
redaction:
  enabled: true