fabula.config.example.json

{
  "share": "disabled",
  "model": "lmstudio/your-local-model-id",
  // The «Enhance prompt» button uses the model SELECTED in the composer by default, with a 45s timeout (so a
  // slow reasoning model still completes instead of failing). This optional section tunes it PER MODEL: keys
  // are model refs ('provider/model'), '_default' applies to all. Each entry may set request params merged
  // into the /chat/completions body (max_tokens, temperature, reasoning flags), a 'timeout_ms', and — if you
  // want a slow reasoning model to enhance FAST — a 'model' override pointing enhance at a faster model for
  // that selected model. Remove this section to just use the selected model as-is.
  "enhance": {
    "_default": {
      "max_tokens": 1024,
      "timeout_ms": 45000
    },
    "provider/some-slow-reasoning-model": {
      "model": "provider/some-fast-instruct-model"
    }
  },
  // The background checkpoint writer is OFF here (`thresholds: []` — the engine logs "checkpoint schedule
  // empty" and takes none). With ONE local model there is one inference slot, and the writer competes for
  // it with the agent it is summarising: measured, main-agent steps took 317/180/727/319 s while a writer
  // held the slot versus 8-23 s when it was free, writers spent 1.32x the agent's own tokens, and 121 of 121
  // compactions went through the overflow path anyway. If you serve more than one slot, delete the
  // `thresholds` line to get the default schedule back; `fork: true` then lets the writer reuse the agent's
  // prefill cache instead of evicting it.
  "checkpoint": {
    "fork": true,
    "thresholds": []
  },
  // Loaded into EVERY request of EVERY project, so every byte here is a per-turn cost that competes with the
  // task for the model's attention — keep this list small and general. Project-specific rules do NOT belong
  // here: the engine already auto-discovers the nearest project-level agent instructions file for the
  // repository being worked on. Situational, deep material belongs in skills, loaded on demand.
  "instructions": [
    "system-prompt.example.md",
    "prompts/runtime-tools-addendum.md",
    "prompts/agent-rules.md"
  ],
  "provider": {
    "lmstudio": {
      "name": "LM Studio (FABULA-LLM local)",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://localhost:1235/v1"
      },
      "models": {
        "your-local-model-id": {
          "name": "Local model",
          "family": "MLX · 4bit",
          "tools": true,
          "attachment": true,
          "limit": {
            "context": 131072,
            "output": 32768
          },
          "modalities": {
            "input": [
              "text",
              "image"
            ],
            "output": [
              "text"
            ]
          }
        }
      }
    },
    "nvidia": {
      "name": "NVIDIA (FABULA-LLM cloud, optional)",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://integrate.api.nvidia.com/v1",
        "apiKey": "{env:NVIDIA_API_KEY}"
      },
      "models": {
        "z-ai/glm-5.1": {
          "name": "GLM-5.1 (agentic/coding)",
          "tools": true
        },
        "deepseek-ai/deepseek-v4-flash": {
          "name": "DeepSeek V4 Flash (fast)",
          "tools": true
        }
      }
    }
  },
  "mcp": {
    "web-search-internet": {
      "type": "local",
      "command": [
        "npx",
        "-y",
        "mcp-searxng"
      ],
      "environment": {
        "SEARXNG_URL": "http://localhost:8888"
      },
      "enabled": false
    }
  },
  "lsp": true
}