crun-task-runnerlisted

# Crun Task Runner Treat this skill as the authority for task creation, authorization, monitoring, recovery, and result delivery. Use `../../runtime/crun_cli.py`; optionally set `CRUN_BASE_URL` for a non-default endpoint. Never print or persist an API key in task files. ## Prepare media and JSON Upload a local image, video, or audio file before putting it in a task input: ```text python <runtime>/crun_cli.py upload <local-file> ``` Put the returned `file_url` in the model input. Never send Base64, data URIs, or local paths. Reuse a prior Crun task's resource URL directly for derivative work. Prefer UTF-8 files to avoid shell-specific JSON quoting: ```text python <runtime>/crun_cli.py task estimate --model <model> --input-file <input.json> ``` Use `--input-file -` to read one JSON object from stdin. Use `--input-json` only when the current shell preserves the JSON string exactly. ## Run a task 1. Describe the selected model when its exact input schema is not already known: ```text python <runtime>/crun_cli.py models describe --model <model> ``` 2. Construct only fields permitted by `input_schema`. 3. Estimate the final model and exact input before creating any task, including a user-specified model. Require `affordable: true`. This matters because `CreateTask` charges credits and is never retried, so submitting an unaffordable or malformed task wastes a real charge. 4. If routing selected the model, report the model, estimate, and relevant settings,