---
title: "POST /projects/{projectId}/eval-workbenches/{workbenchId}/runs"
method: POST
path: "/projects/{projectId}/eval-workbenches/{workbenchId}/runs"
---

# POST /projects/{projectId}/eval-workbenches/{workbenchId}/runs

`POST /projects/{projectId}/eval-workbenches/{workbenchId}/runs`

## Request

- Base URL: `https://api.caveman.so/api/v1`
- URL: `https://api.caveman.so/api/v1/projects/{projectId}/eval-workbenches/{workbenchId}/runs`
- Auth: one of:
  - OAuth 2
    - Authorization code: `/api/v1/oauth/authorize`
    - Token URL: `/api/v1/oauth/token` scopes: `experiment:run`
  - HTTP bearer
  - API key in cookie `cave_access`
  - API key in cookie `__Secure-cave_identity.session_token`

POST /api/v1/projects/{projectId}/eval-workbenches/{workbenchId}/runs on the Caveman Cloud control plane. Requires the experiment:run scope. Exposed to autonomous agents as the run_workbench action of the caveman_evals_write tool (write access). Author versioned datasets, reusable evaluators and workbenches; create strict suites, ingest evidence, and run baselines. Evidence ingest supports dry_run — always dry-run first. Ask Caveman to draft quality criteria again within the drafting allowance, propose a criterion for a person to turn on, and hold the pull request that adds the Caveman check to CI. Track a criterion on a slice of traces in one step: from an agent it lands proposed, and a person turns it on. Queue units for review and label them (an agent's label is recorded as an agent's, never as human gold), and make a dataset from recorded traffic.

## Request body

- object
  - `scope` object
    - `kind` 'all' | 'target' | 'row' | 'cell' | 'evaluator', required
    - `target_key` string
    - `row_key` string
    - `evaluator_key` string
    - `source_run_id` string, uuid
    - `trials` integer
    - `split` string
    - `tags` string[]
    - `budget_usd` number: The most the run may spend, at list prices. Required when trials is over 1, and for any run an agent connection or a project key starts.
    - `mode` 'full' | 'partial': partial: the run holds only some of the cases (a quick look at a sample). A partial run is never a baseline and never compared. Left out, the run is full.
    - `commit` object: The change the run was made for, as the caller reports it (a CI job's commit, ref, base and pull request). Returned on the run's scope; nothing grades or compares on it.
      - `sha` string
      - `ref` string
      - `base_sha` string
      - `base_ref` string
      - `pr_number` integer
  - `concurrency` integer
  - `outputs` object[]: Outputs the caller produced, one per trial of a row, for an experiment with an external target. The run holds the rows they name; Caveman grades them and never reads them as its own execution.
    - `row_key` string, required
    - `trial` integer
    - `output` unknown
    - `error` string
    - `trace_id` string
    - `latency_ms` number
    - `cost_usd` number
    - `attempts` integer

## Response `2XX`

Successful response. Exact status depends on synchronous or durable delivery outcome.

## Other responses

- `4XX`: Structured failure. `error.code` is stable; `error.request_id` is echoed in the x-cave-request-id header.
- `5XX`: Structured failure. `error.code` is stable; `error.request_id` is echoed in the x-cave-request-id header.

---

[API](https://skmtc.dev/caveman/apis/caveman-cloud-api.md) - [All operations](https://skmtc.dev/caveman/apis/caveman-cloud-api/llms.txt) - [OpenAPI document](https://skmtc.dev/caveman/apis/caveman-cloud-api/revisions/508312d01cfc?raw)