Probe results and errors

A probe is one real request to your endpoint, built and read exactly the way a customer's run is. Checking a declaration's shape catches typos; only a probe catches a declaration that is well formed and wrong — a path that answers 404, a header your API does not read, a result path your answer does not have. Nothing goes live without a passing probe.

What the probe sends#

• The model: one model your provider serves that has a descriptor. An openai-compatible provider with none uses the first model its GET {baseUrl}/models lists.
• Your key: the one you declared, sent the way your auth says.
• Settings: stand-in values, of which the model's descriptor keeps only what it declares:

• Input files: a public sample image and 3D model that your API can download. There is no fetchable sample audio yet, so a model that needs an audio input cannot pass a probe.

Outcomes#

{
  "ok": true,
  "outcome": "pass",
  "probedUrl": "https://api.example.com/v2/task"
}

{
  "ok": false,
  "outcome": "fail",
  "error": "YOUR_PROVIDER_KEY: no value at declared output path \"data.output.model\" — the declared path is wrong, or the provider changed its response shape",
  "probedUrl": "https://api.example.com/v2/task"
}

{
  "ok": false,
  "outcome": "inconclusive",
  "error": "YOUR_PROVIDER_KEY: task \"abc123\" had not reached a terminal state yet — …",
  "probedUrl": "https://api.example.com/v2/task",
  "pendingTaskId": "abc123"
}

Answers from before the request is sent#

• { "ok": false, "stage": "validate", "problems": [...] } — the stored declaration does not pass its checks. Declare it again.
• { "ok": false, "stage": "credential", "error": "… no credential configured to probe with …" } — declare your key first. The claim records it as a failed probe.
• 503 with stage credential or resolve-descriptor — Sushibelt could not read your key or the descriptors. Nothing is recorded; try again.
• 400 Declare a transport before probing.
• 404 No claim in verifying for this provider. — the claim is not approved, or has moved on.

Why a probe fails#

The error starts with your provider key. The common reasons, by where they happen:

Building the request#

• No descriptor names a model for your provider, so a sync-rest or submit-poll run has no request to send. A descriptor has to exist first.
• dispatch cannot build its request — … followed by the reason: the transport declares GET (declare POST), or the descriptor's endpoint is on another host than your baseUrl.
• no model to probe with — GET …/models answered HTTP … or … listed none — openai-compatible only.

Your answer to the request#

• rejected the request (HTTP …): … — your API answered with an error status; your message follows.
• response had no "choices" array — not an OpenAI-compatible shape
• submit response was not valid JSON
• could not find a task id in the submit response (tried …) — set taskIdPath, or return the id at one of id, task_id, data.id, data.task_id, result.id, result.task_id.
• A sync-rest answer with no file where result.path points, such as response had no base64 payload at ….

The poll#

• poll rejected (HTTP …): … — the poll URL (baseUrl + pollPath + "/" + taskId) answered with an error.
• poll response was not valid JSON
• task reported failure (status "…"): … — the status was in poll.failure; the message at errorPath follows.
• no value at declared output path … — the declared path is wrong, or the provider changed its response shape

The network#

• probe threw — … — the request did not complete: DNS, TLS, a refused connection or similar.

Declaration problems#

A declaration that breaks a rule is refused with 400 Invalid transport declaration and a problems list. Each line names the field. The lines:

A claim declaration can also be refused for its price or key:

• credits_per_unit must be a number greater than 0 · credits_per_unit must have at most 6 decimal places
• pricing_unit must be one of: run, token, character, second, megapixel, step, usd_micro
• credential must be an object holding the key this provider will be called with

Replacing a key#

Replacing the key of a provider you operate runs a probe with the new key first. The key is saved only on a pass. Otherwise the answer is 422 — The new key did not work, so the old key is still in use: …, or The provider did not finish the test call in time. The old key is still in use — try again. — and nothing changes. See keep your key current.