# Append to ~/.dsh/cordis.yml (or merge into the plugin list of your deployment). # Every key is optional — the values below are the defaults unless commented otherwise. # The second entry lets DSH's skill-filesystem plugin load skills distilled by the dreamer. - name: dsh-somni config: dataDir: ~/.dsh/somni # memory/ skills/ models/ live here python: python3 # interpreter ≥ 3.9 with `somni-memory` importable (bundled under python/) capture: enabled: true maxMessageChars: 4000 inject: # resident system-prompt sections identity: true # about_me + user profiles intentions: true # open intentions / due reminders discipline: true # when to call the memory tools assoc: # System-1 associative hints (needs the vector layer) enabled: true preStep: true # cue = latest user message postExecute: true # cue = tool name + args threshold: 0.52 gap: 0.04 maxItems: 2 tools: enabled: true # recall_memories / search_episodes / … / dream_now dream: enabled: true idleMs: 900000 # 15 min of quiet before a dream may start maxIntervalMs: 21600000 # dream at least every 6 h while logs are pending and agents are idle minPendingLogs: 1 llm: host # host = reuse the harness model via ctx.llm # llm: openai # openai = sidecar calls an OpenAI-compatible endpoint itself # openai: # baseUrl: https://api.openai.com/v1 # apiKeyEnv: OPENAI_API_KEY # model: gpt-4o-mini maxIterations: 60 embed: provider: local # local (ONNX bge-small-zh) | http (OpenAI-compatible /embeddings) | off # modelDir: ~/.dsh/somni/models/bge-small-zh-v1.5 # baseUrl: http://127.0.0.1:8080/v1 logLevel: info - name: '@deepseek-ai/dsh-skill-filesystem' config: customSkillDirs: - ~/.dsh/somni/skills/agent_learned_skills