@wasmagent/eliza-rollout-plugin

elizaOS community plugin that captures action runs as rollout-wire/v1 training records (DPO/PPO-ready JSONL) for fine-tuning via evomerge.

elizaOS already isolates agents in a Bun Worker sandbox. This plugin does not add another sandbox layer — its sole purpose is the training data loop: capture → rank → export JSONL.

Install

npm install @wasmagent/eliza-rollout-plugin

Usage

Zero-config (writes to ./rollouts/)

import { createRolloutPlugin } from "@wasmagent/eliza-rollout-plugin";

export default {
  plugins: [createRolloutPlugin()],
};

With evomerge HTTP sink

export default {
  plugins: [createRolloutPlugin({
    sink: { type: "http", url: "https://evomerge.example.com/ingest" },
  })],
};

Multi-branch mode (enables DPO export)

export default {
  plugins: [createRolloutPlugin({
    branches: 3,       // run each action 3 times concurrently, rank, export chosen/rejected pair
    format: "both",    // write both DPO and PPO records
  })],
};

Composing with withCapabilityGovernance (elizaOS ≥ 1.x)

elizaOS core ships withCapabilityGovernance in @elizaos/core/security — a per-call deadline + host/path predicate layer that composes with the existing Bun Worker sandbox. This plugin instruments the action after governance is applied, so you get both:

import { withCapabilityGovernance } from "@elizaos/core";
import { createRolloutPlugin } from "@wasmagent/eliza-rollout-plugin";

// Governance wraps the action first
const governed = withCapabilityGovernance(myAction, {
  allowedHosts: ["api.example.com"],
  cpuMs: 10_000,
});

// The plugin's registerAction hook then instruments the governed action for rollout capture.
// registerAction is called once per action at startup — the plugin sees the governed wrapper.
export default {
  actions: [governed],
  plugins: [createRolloutPlugin()],
};

The two are complementary: governance in @elizaos/core, training data export here.

Options

| Option | Type | Default | Description | |---|---|---|---| | sink | FileSinkOptions \| HttpSinkOptions \| ConsoleSinkOptions | { type: "file", dir: "./rollouts" } | Where to write JSONL records | | format | "dpo" \| "ppo" \| "both" | "ppo" | Export format. DPO requires branches >= 2 | | branches | number | 1 | Independent branches per action run. 1 = single-run mode | | scorer | (answer, task) => number | length heuristic | Heuristic scorer for PPO reward signal (0–1) | | includeActions | string[] | [] (all) | Action names to instrument. Empty = instrument all |

Output format

Records follow the rollout-wire/v1 schema defined in @wasmagent/core/beta.

PPO record

{
  "prompt": "What is 2+2?",
  "completion": "four",
  "reward": 0.8,
  "tool_call_sequence": [],
  "provenance": {
    "source": "wasmagent-rollout",
    "rollout_id": "agent-123-1700000000000",
    "branch_index": 0,
    "objective_score": 1,
    "exported_at_ms": 1700000000000,
    "n_gram_hash": "52cb6b5e4a038af1"
  }
}

DPO record (requires branches >= 2)

{
  "prompt": "Write a sort function",
  "chosen": "function sort(arr) { return [...arr].sort((a,b) => a-b); }",
  "rejected": "",
  "tool_call_sequence": [],
  "provenance": {
    "source": "wasmagent-rollout",
    "rollout_id": "agent-123-1700000000000",
    "chosen_branch": 0,
    "rejected_branch": 2,
    "objective_score": { "chosen": 1, "rejected": 0 },
    "exported_at_ms": 1700000000000,
    "n_gram_hash": "52cb6b5e4a038af1"
  }
}

License

Apache-2.0