jev-curl

Jev curl invocation

Access the official HTTP API only with curl; for authentication, read directly from the current execution process's TYPESAFE_API_KEY, with no SDK, MCP, or separate configuration file needed.

Applicable and non-applicable scenarios

• Suitable for semantic tasks with sufficient material and clear judgment criteria: classification and triage, selecting from existing candidates, judging whether content meets requirements, and scoring according to given levels. Being expressible as a multiple-choice question does not mean it is suitable for Jev.
• When results can be obtained through deterministic conditions, computation, or database queries, prefer code or queries and do not make an additional model call.
• When information retrieval, multi-step reasoning, or solution design is needed, complete that work first, then hand the local judgments whose material and criteria are already clear to Jev; do not simplify a complete complex decision into a single choice.
• When candidates may not cover the actual situation, add a "none of the above" or "insufficient information" option and clarify what each means. Do not force a business option to be selected when evidence is insufficient.
• Before connecting results to automatic execution, validate judgment effectiveness and error consequences with actual task samples, and determine which results can be executed and which require additional information or human review; do not decide automatic execution based solely on model confidence.

Execution

1. Distinguish between generating commands and actually invoking: output the command only when an example is requested; when execution is requested, check that curl is available and the environment variable is non-empty, report only whether it exists, and do not print the key. If the key is missing, ask the user to configure it in the execution environment; do not search other credential files. The execution process of a desktop Agent may not inherit variables newly set in the terminal.
2. When constructing the request for the first time, read HTTP requests and responses. When new parameters need to be added, interface changes need to be verified, or field errors need to be troubleshooted, then read the official API documentation linked therein; do not load the SDK or unrelated examples.
3. Put the content to be judged into state, the question into instructions, and the candidates or levels into criteria. Use choicefor options, noulfor whether it holds, and scorefor degree scoring. The question ID is only used to match returned results and does not carry prompt meaning.
4. Questions that are independent of each other under the same state can be placed in the same request; questions cannot reference answers that have not yet been returned. External requests contain only content required by the task and authorized to be sent to TypeSafe.
5. Pass the request body using a JSON file or a quoted heredoc, and execute via curl. The key only goes into the authentication header and is not written into the request body, logs, or repository; disable shell tracing during execution to avoid curl -v or trace outputting the authentication header.
6. Check the curl exit code, HTTP status, and the type and result of each question in answers at the same time. A failed call must not be reported as a successful model judgment. Selecting an operation does not mean that operation has been executed.

Returns and failures

• Report the actual model, question results, and necessary probabilities; attach usagewhen needed. Clearly mark request examples when there is no actual call.
• choice must fall within the submitted candidates;noul is the probability of "yes", not a boolean value;score should be interpreted according to the returned legend , and must not be arbitrarily treated as a 100-point scale.
• confidence is not equal to factual accuracy; business thresholds are determined by the task, and a fixed threshold is not uniformly applied.
• 401/403: stop and report the authentication or permission issue; 422: correct the explicit field errors before calling again. 429/529: retry at most twice according to the server's wait hint or with exponential backoff; a timeout or connection interruption means the result is unknown, and retrying may result in duplicate billing.

Source

Based on the TypeSafe official skill 's request design principles and the official HTTP documentation, condensed and rewritten; this skill is a personally maintained curl-specific version. API verification date: 2026-09-21.