Test runs
Test runs

Get one run's transcript

Read what happened in one eval run, both what was SAID and what was DONE. Use this to debug a failing run beyond the judge's reasoning: get_test_results tells you WHICH rubric failed; this shows the exchange and the tool calls that failed it. Returns `turns` ([{role, content}]) and `toolCalls` ([{name, args, ok, detail}], compacted, the raw tool payload is not included), each as JSON inside an untrusted-data block once it holds anything, plus the run's `status`, `channel`, and its linked `sessionId` (pass that to get_call_metrics for the metric grid). IMPORTANT, `toolCalls` ABSENT means no tool record was captured for this run, so you cannot conclude anything about whether a tool fired; `toolCalls: []` means the record exists and the agent called nothing, which IS a finding. A phone-mode run copies its turns and tool calls from the call it was linked to; one that was never linked has neither and says so, with the sessionId to read instead when there is one. The run must belong to the current agent.

GET
/workspaces/{workspace}/agents/{agent}/runs/{run}/transcript

Authorization

bearerAuth
AuthorizationBearer <token>

A personal access token from Settings > API tokens.

In: header

Path Parameters

workspace*string

Workspace slug (GET /me lists yours).

agent*string

Agent slug within the workspace.

run*string

Response Body

application/json

application/problem+json

application/problem+json

application/problem+json

application/problem+json

application/problem+json

application/problem+json

application/problem+json

curl -X GET "https://example.com/workspaces/string/agents/string/runs/string/transcript"
{
  "runId": "string",
  "status": "string",
  "property1": null,
  "property2": null
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}
{
  "type": "http://example.com",
  "title": "string",
  "status": 0,
  "detail": "string",
  "code": "string",
  "errors": [
    {
      "path": "string",
      "message": "string"
    }
  ],
  "estimate": {},
  "retryAfterSeconds": 0
}

Cancel a test run POST

Cancel an in-flight test batch (the one run_all_tests returned). Queued runs never start; calls already in flight finish and keep their verdicts; the batch ends as 'cancelled'. Use this immediately when the operator wants a runaway or mis-configured sweep stopped, every queued run that never starts is credits saved. Fails cleanly if the batch already finished.

Get a test run's results GET

Read test (eval) verdicts. Pass the batchId from run_all_tests for that batch, or omit it for the latest GRADED run of every test; pass `test` (a name from list_tests, or an id) to read one test, the shape you want while iterating on one scenario. Lost the batch id? list_batches lists this agent's recent batches. Each row carries the test name, its `runId` (pass a failing one to get_run_transcript to read the exchange behind the verdict: that is the debugging ladder), its `sessionId` (pass it to get_call_bundle), the verdict (passed, failed, scoring, review_required, error, infra_error, running or pending: an operator's override first, then a failed critical metric turning a pass red, then the run's own status), the failed rubrics with their reasoning, and `rubricsNote`, which explains any row whose verdict or empty rubric list needs explaining. A red run with no failed rubric was vetoed by the critical metric that `vetoedByMetrics` names: fix what it measures rather than re-running. `scoring` means the scenario passed but critical metrics are still grading; it counts as in flight. CHECK `complete` BEFORE YOU TRUST THE COUNTS: run rows exist up front and a voice run takes 1-5 minutes, so poll every 30-60 seconds until `complete` is true. Faster polling buys nothing, a batch finishes on its own, and a quiet run is not a stuck one: compare `elapsedSeconds` with `typicalSeconds`, and never cancel a batch for it; `batchStatus: 'cancelling'` is already stopping. `infraFailed` runs and a `truncated` batch tested nothing about the agent: re-run them, and never publish on a truncated batch. A row whose `stale` is a sentence scored an earlier version, and publish accepts only greens earned on the version being shipped (`currentVersion`), so re-run stale tests before publishing. With a batchId, gate on `perTest` (passed, failed, flaky or inconclusive across repeats), not the flat counts. A `fixLadder` on a failing row means prompt-only edits have failed across two or more versions: stop editing the prompt and change the tool's description (set_tool_description), its output (set_tool_response_template) or its inputs (set_tool_params), as `fixLadder.nextRung` says. For the success-rate trend across many runs, use get_runs_overview.