Language Models

A Devlish workflow can ask a language model to produce text or structured data when a step needs model judgment. The program decides where that request occurs and what to do with the answer; the model call is an external operation provided by the host runtime.

Use this page after learning ordinary inputs and structured output. A call requires a configured provider and model, credentials where the provider needs them, and a declared model permission. The following example shows the statement forms; it does not configure a provider by itself.

Request a completion

Use with to supply the prompt and as to name the result. Add expecting json for a structured response, or supply a model ID to select a model explicitly. The input variable prompt must already exist.

Permissions:
  Call language models

Ask the model with prompt as draft
Ask the model with prompt expecting json as classification
Ask the model "YOUR_MODEL" with prompt as answer

Configure providers in ~/.devlish/config.toml (see devlish harness init-config). Credentials resolve through the existing .env / --env chain (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.).

CLI:

devlish --run workflow.dvl --provider anthropic --model YOUR_MODEL
devlish --server --bind 127.0.0.1:7420