Runtime operations

The universal RuntimeClient contract every provider implements. Optional methods can be absent. Check both advertised support and method presence before calling; optional chaining alone can hide an unavailable operation.

getRuntimeCapabilities

Signature

client.getRuntimeCapabilities(): Promise<RuntimeCapabilities>

HTTP

n/a (provider metadata)

Capability

always present

Response

Example

const caps = await client.getRuntimeCapabilities();
if (caps.supports.batch) { /* … */ }

startRun

Signature

client.startRun(body: RuntimeRunStartBody): Promise<RuntimeRunStatus>

HTTP

POST /v1/messages (Claude) · POST /v1/responses (Codex) · gateway RPC (Hermes/OpenClaw)

Capability

supports.runs

Request body / Parameters

Response

Example

const run = await client.startRun({
  input: "Summarize the changelog.",
  model,
});
const view = { runId: run.run_id, status: run.status, text: run.output };

getRun

Signature

client.getRun?(runId: string): Promise<RuntimeRunStatus>

HTTP

GET /v1/responses/:id (Codex) · gateway RPC (Hermes/OpenClaw)

Capability

supports.runs + stateful provider

Optional on the type, but implemented by every shipping provider. Synchronous providers (Claude Messages, AGY) serve getRun from a client-side SynchronousRunStore (remembered terminal status; unknown ids degrade without throwing). Codex and gateway providers poll real upstream run handles. Returns the same RuntimeRunStatus shape as startRun.

Example

const status = await client.getRun?.(run.run_id);

cancelRun

Signature

client.cancelRun?(runId: string): Promise<{ status: string }>

HTTP

POST /v1/responses/:id/cancel (Codex) · gateway RPC (Hermes/OpenClaw)

Capability

supports.runs + cancelable provider

Optional. Same availability rules as getRun.

Response

streamRun

Signature

client.streamRun?(body: RuntimeRunStartBody, handlers: RunEventStreamHandlers, options?: { signal?: AbortSignal }): Promise<void>

HTTP

POST /v1/messages?stream=1 (Claude) · POST /v1/responses (SSE, Codex)

Capability

supports.streaming

Optional. Gateways use a subscribe-by-runId model and omit this, exposing a RunEventStreamProvider instead. handlers receive canonical RunStreamEvents.

Example

await client.streamRun?.(
  { input: "Write a haiku.", model },
  { onEvent: (e) => process.stdout.write(e.event) },
);

submitBatch

Signature

client.submitBatch?(requests: RuntimeBatchRequest[]): Promise<RuntimeBatchStatus>

HTTP

POST /v1/messages/batches (Claude) · POST /v1/batches (Codex)

Capability

supports.batch

Implemented by Claude and Codex. Provider-specific batch endpoints and result retrieval differ — see Claude, Codex.

Request body / Parameters

Response

Example

const batch = await client.submitBatch?.([
  { customId: "a", body: { input: "One" } },
  { customId: "b", body: { input: "Two" } },
]);

getBatch

Signature

client.getBatch?(batchId: string): Promise<RuntimeBatchStatus>

HTTP

GET /v1/messages/batches/:id (Claude) · GET /v1/batches/:id (Codex)

Capability

supports.batch

Poll until resultsAvailable is true. Same RuntimeBatchStatus response as submitBatch.

cancelBatch

Signature

client.cancelBatch?(batchId: string): Promise<RuntimeBatchStatus>

HTTP

POST /v1/messages/batches/:id/cancel (Claude) · POST /v1/batches/:id/cancel (Codex)

Capability

supports.batch

Returns the updated RuntimeBatchStatus.

getBatchResults

Signature

client.getBatchResults?(batchId: string): Promise<RuntimeBatchResult[]>

HTTP

GET /v1/messages/batches/:id/results (Claude) · GET /v1/batches/:id then GET /v1/files/:id/content (Codex)

Capability

supports.batch

Throws an EndpointNotFound-class error if the batch has not ended — poll getBatch first. Codex has no single results endpoint: read the batch to obtain its output/error file id, then download the JSONL via the Files API (GET /v1/files/:id/content).

Response