LLM integration

This guide sketches a closed loop: load policy → distil page → LLM plans action → bridge executes → re-distil.

System prompt template

You are an agent for {{site.name}}.

{{wci.md contents}}

Rules:
- Only use node IDs from the distilled view.
- Respect denied scopes: {{denied}}.
- Scopes requiring human confirmation: {{confirm}} — ask the user first.
- On ActionResult.error, read `hint` and adjust; do not retry denied scopes.

Load with:

const ctx = await WciContextLoader.load(origin);
const system = buildPrompt(ctx);

Tool / function schema (example)

{
  "name": "wci_dispatch",
  "description": "Interact with an WCI-annotated page",
  "parameters": {
    "type": "object",
    "properties": {
      "nodeId": { "type": "string" },
      "action": { "enum": ["click", "fill", "select", "check", "submit", "navigate", "focus", "clear"] },
      "value": {}
    },
    "required": ["nodeId", "action"]
  }
}

Handler:

async function handleToolCall(args: ActionRequest) {
  // Policy enforced inside bridge.dispatch when bridge.setPolicy(ctx.policy) is set.
  return bridge.dispatch(args);
}

Context window strategy

1. Static — wci.md + wci.json task flow (small)
2. Per turn — distiller.distilJSON() for current scope only
3. Feedback — append last ActionResult JSON (success, stateChange, sideEffects, or error)

const messages = [
  { role: 'system', content: system },
  { role: 'user', content: distiller.distilJSON(scopeRoot) },
  { role: 'user', content: JSON.stringify(lastResult) },
];

Error handling

error.code	Agent behavior
NODE_NOT_FOUND	Re-distil; page may have navigated
PRECONDITION_UNMET	Complete prerequisite fields first
SCOPE_DENIED	Stop; inform user (policy enforced by bridge)
AUTH_REQUIRED	Run auth flow for scope, then retry
HUMAN_CONFIRMATION_REQUIRED	Ask user, then retry
UNKNOWN_ERROR	Try alternate action or ask user

Rate limits

Parse Rate-Limit-Actions and Rate-Limit-Distil from wci.txt and throttle client-side before calling Bridge or Distiller.

Server-side distillation

For SSR or crawlers without a browser, run pruneDOM + serializeJSON in jsdom (or emit HTML with attributes server-side and distil at the edge). The demo runs client-side only.

See also

• Getting started
• Action protocol