batch-api

/batch-api — Agent-prepared Batch API workflow

Hard rule for this whole skill: if any qwen batch … command fails, or the task turns out to be unsuitable, report what happened in a few lines and stop (the one exception, a plan-field error from run, is in §4). Never fall back to doing the transform yourself in this session — the user chose the half-price asynchronous path explicitly, and silently doing the work at full realtime price is exactly what they opted out of. Offering it as a choice ("I can do this realtime instead, at full price") is fine; doing it without being asked is not.

The user explicitly chose async batch mode by typing /batch-api. This mode trades latency for price: the provider bills Batch requests at 50% of the realtime list price (with no context-cache benefit), and a job takes tens of minutes to hours to finish (completion window 24h or more). Your job is to turn the user's task into a small plan file that the deterministic executor (qwen batch run) submits. You never write request JSONL by hand and never call the Batch API yourself.

0. Check readiness before anything else

Run every qwen batch … command with the shell tool as "${QWEN_CODE_CLI:-qwen}" batch … — QWEN_CODE_CLI names the CLI running this session, so a plain qwen on PATH (possibly an older install without these subcommands) is only the fallback. First make sure the CLI you reach has them. Help output never calls a model:

"${QWEN_CODE_CLI:-qwen}" batch --help

If the output does not list batch run <plan>, the qwen it reached is an older install without these subcommands — say so (the session's CLI is not on PATH as qwen) and stop. Do not run any other batch command there: an older CLI treats batch check as a prompt and answers it with a billed model call. Then:

"${QWEN_CODE_CLI:-qwen}" batch check

It proves the credentials, endpoint and Batch route work and shows the model, thinking mode and output limit a run would freeze from the user's current settings — without a billed request. settings.batch.model can select a separate modelProviders entry; its endpoint, envKey and generationConfig are used without changing the conversation model. Do not edit authentication or settings to work around a failed check. If it fails (for example Qwen OAuth, which has no Batch route), relay its message and stop: do not read files or draft a plan the executor cannot submit. Pass its note: lines on to the user.