Skip to content

Latest commit

 

History

History
63 lines (55 loc) · 2.92 KB

README.md

File metadata and controls

63 lines (55 loc) · 2.92 KB

ScienceClaw gateway plugin

Registers the ScienceClaw engine (packages/scienceclaw) as optional agent tools. The gateway agent acts as the policy of the engine: it edits a typed workflow graph step by step (scienceclaw_canvas), finds scientific tools and pretrained-model wrappers (scienceclaw_tools), inspects or rolls back the versioned Skill/Operator program (scienceclaw_program), and turns verified sessions into gated program updates (scienceclaw_evolve).

Tool Operations
scienceclaw_canvas open, act, render, replay, finish, status, list
scienceclaw_tools search, show, status, weights, setup
scienceclaw_program summary, skills, operators, show, history, rollback
scienceclaw_evolve val_add, val_list, val_remove, propose, gate, run, status, candidates, show

scienceclaw_canvas(operation=open) takes a live task declaration (objective, input files, the schema of the deliverable and its hard constraints); inputs must lie under the configured input roots.

The plugin keeps one long-lived Python process (python -m scienceclaw.rpc, line-delimited JSON) so canvas sessions survive between tool calls. Deployment settings (packageRoot, pythonBin, home, inputRoots, modelRoot, configPath, llm) are plugin configuration and are never accepted as tool parameters. Gateway and provider credentials are not forwarded to the engine; the model used by llm nodes is the OpenAI-compatible endpoint given under llm, or any backend registered through scienceclaw.llm.interface.

{
  plugins: {
    entries: {
      "scienceclaw": {
        enabled: true,
        config: {
          packageRoot: "<checkout>/packages/scienceclaw",
          pythonBin: "<venv>/bin/python",
          home: "<state-dir>",
          inputRoots: ["<workspace>"],
          llm: { baseUrl: "<endpoint>", model: "<model>" },
        },
      },
    },
  },
  agents: {
    list: [{
      id: "main",
      tools: {
        allow: ["scienceclaw_canvas", "scienceclaw_tools", "scienceclaw_program", "scienceclaw_evolve"],
      },
    }],
  },
}

Keep home (session receipts, program store) outside the directories the agent can read freely. An admitted evolution candidate waits as ready until the user promotes it (python -m scienceclaw.cli live promote <id>); autoPromote: true lets the gate promote by itself.

On first start the plugin installs the whole tool library (Python packages, pretrained weights, upstream sources; about 19 GB for the full profile) in the background and the engine refuses tasks until that is done. autoSetup: false turns the automatic start off (run python -m scienceclaw.cli setup yourself), setupProfile: "light" skips the assets larger than 1.5 GB.

The ScienceClaw-Eval benchmark that accompanies the paper keeps its evaluation data on Hugging Face: https://huggingface.co/datasets/beita6969/scienceclaw-eval.