SuperAPI AI Directory

polpo-integrations

Agents & models

View repository

Updated 10-16 05:44

About this skill

Polpo Integrations

Keep model-visible arguments separate from host-owned identity, authorization, and credentials. The model expresses intent; the runtime resolves trusted context and capabilities.

Workflow

  1. Decide whether the integration is a built-in tool, custom tool, MCP server, Channel handler, or request-scoped client tool.
  2. Define the smallest model-visible JSON schema and separately define trusted bindings and Connection slots.
  3. Resolve user, tenant, resource, scope version, and grants server-side before tool execution.
  4. Package the complete relative source graph and validate it before upload.
  5. Resolve only active Connections with all required scopes and an unambiguous trusted binding.
  6. Execute with bounded time, cancellation, sandbox policy, idempotency, and redacted trace data.
  7. Test denied scope, missing/ambiguous Connection, rotated credentials, timeout, retry, malformed result, and package dependency failures.

Invariants

  • Connection IDs and credentials are not model-controlled tool arguments.
  • ToolInvocationContext is copied, JSON-validated, deeply frozen, and host-owned.
  • serverBindings can read only supported immutable invocation paths.
  • Hidden bindings and Connection capabilities never enter model history or serialized tool arguments.
  • A tool cannot widen its granted scopes, filesystem access, or execution surface.
  • Packaging failure is detected before deployment; runtime 502 is not an acceptable dependency discovery mechanism.

References