Imported from mitsuhiko/pi-ds4 (pi-ds4-config/SKILL.md). Install upstream with npx skills add mitsuhiko/pi-ds4 --skill pi-ds4-config. Copyright stays with the author.

pi-ds4 configuration

To reconfigure pi-ds4/ds4, edit ~/.pi/ds4/settings.json (create it if missing). Do not edit other runtime state files in ~/.pi/ds4 unless explicitly debugging.

The file is a JSON object. Environment variables override it. Keys can be env names (DS4_READY_TIMEOUT_MS), camelCase without DS4_ (readyTimeoutMs), or lower snake_case (ready_timeout_ms). Use settings.schema.json / settings.example.json from the pi-ds4 package for validation/examples.

Minimal example:

{
  "$schema": "https://raw.githubusercontent.com/mitsuhiko/pi-ds4/main/settings.schema.json",
  "protocol": "openai-responses",
  "contextTokens": 393216,
  "readyTimeoutMs": 900000
}

Common settings:

• protocol: openai, openai-responses (default), or anthropic
• contextTokens: ds4/Pi context ceiling (default 393216; GLM models use their lower supported limit)
• power: DeepSeek GPU duty-cycle target from 1 to 100 (default 100). Lower values reduce heat and fan noise. GLM 5.2 and DeepSeek V4.1 always run at 100 because upstream does not support throttling them.
• ssdStreaming: auto (default), true, or false. In auto the server gets --ssd-streaming when a model's resident weights exceed 90% of system RAM; only DeepSeek V4.1 declares such a footprint today (152 GiB for Q2, 294 GiB for Q4, with Engram tables staying on disk). Streaming models also cap the context at 131072 tokens unless contextTokens is set explicitly.
• autoUpdate: fast-forward the package-managed ds4 checkout before use (default true; local/external checkouts are untouched)
• readyTimeoutMs: server startup timeout in ms
• runtimeDir: existing antirez/ds4 checkout instead of ~/.pi/ds4/support
• supportRepo / supportBranch: runtime checkout source
• serverBinary / watchdogScript: custom executable/script paths
• apiKey: token Pi sends to the local provider (default dsv4-local)

DS4_GGUF_DIR is an environment-only upstream ds4 option for storing downloaded GGUF files outside the runtime checkout. It is not a settings.json key.

Download progress is measured from the requested growing files in the download directory (including their matching hashed files under the Hugging Face downloader's .cache/huggingface/download directory, which print nothing when piped) and compared against the exact sizes in download_model.sh's artifact_identity() where upstream declares them. If progress looks stuck, check ~/.pi/ds4/log for the raw downloader output.

After editing, run /reload or restart Pi. Models are never downloaded automatically. Use /ds4 to download an explicit DeepSeek V4 Flash, DeepSeek V4.1 Flash, DeepSeek V4 Pro, or GLM 5.2 variant and to view the log or control the server. Existing and completed downloads are discovered and registered automatically in Pi's model catalogue.

DeepSeek V4.1 Flash (dsv41-flash-q2, dsv41-flash-q4) needs 341 GiB and 483 GiB of disk respectively; the Q4 download arrives in two parts that are joined automatically and needs another 37 GiB free while joining. It runs on Metal, and on CUDA for text only. V4 Flash vision encoders and DSpark files are not compatible with it.

Images: download the V4.1 vision encoder from the same /ds4 menu (about 0.9 GiB, target ds41f-vision). When it is present, ds4-server is started with --vision and the V4.1 models advertise image input. The menu entry only appears on macOS because upstream rejects the V4.1 encoder outside Metal. Changing the encoder state restarts the server on the next request.