API reference: tools and calls

Tool dialects

GET /v1/envs/:slug/tools?format=:

native: { name, description, inputSchema, rung, attention, scores } per tool, with what the twin knows about it.
openai: { type: "function", function: { name, description, parameters } } entries, ready for a chat completions or Responses request.
anthropic: { name, description, input_schema } entries, ready for a messages request.

The names are the production tools’ names; your agent calls them as it does in production.

A call

POST /v1/sessions/:sid/call
{ "name": "lookup_order", "arguments": { "id": "o_1" }, "requestor": "assistant" }

requestor is assistant (the model asked) or user (a person or a harness asked); it is recorded, not enforced. The result:

{
  "content": "{\"id\":\"o_1\",\"status\":\"open\",...}",
  "error": false,
  "fidelity": "A_executed",
  "rung": "world",
  "downgraded": false,
  "callId": "c12",
  "ms": 41
}

content: the tool’s answer as text, exactly as the agent should see it.
error: the tool answered with an error (the way production would).
fidelity: the grade of this answer: A_executed (provided or custom code ran), B_twin (the world model), B_replayed (a recorded answer), B_inferred, B_simulated (a model imagined it).
rung: which rung answered.
downgraded: the bound rung could not answer and a lower one did (the tool page says why).
callId: the call’s id in the transcript.

A call the twin cannot route answers 404 with the tool name; a session that expired answers 404; a twin still booting answers 409 with notReady, retry after a moment.

Seeding

Words seed a session on the fly: the twin turns them into rows for the entities they mention. A named scenario is a seed spec the mirror’s config defines (rows per entity, files, a clock), reproducible by name. seed makes generated values deterministic.

​Tool dialects

​A call

​Seeding

Tool dialects

A call

Seeding