Reference

Calling

One-shot LLM calls and structured output.

call.call(): Run one CLI-backed LLM call and parse its response.
schema_for(): Serialize a Pydantic model’s JSON schema, with additionalProperties set to false.
resolve_schema_path(): Resolve a JSON schema into the argument form the backend’s CLI expects.
extract_structured(): Return the validated structured_output from a stream-json event list, if present.
parse_structured_output(): Parse raw CLI stdout into text or a validated model.
parse_result_envelope(): Parse a {is_error, result} JSON envelope into its result text.

CLI backends, the specialty registry, and Claude status checks.

LlmBackend: Abstract interface for an LLM CLI backend.
ClaudeCliBackend: LlmBackend for the Anthropic claude CLI.
CodexCliBackend: LlmBackend for the OpenAI codex CLI.
LlmBackends: Registry mapping each specialty to the LlmBackend that serves it.
check_status(): Check whether the claude CLI is installed and authenticated.
ClaudeStatus: Result of check_status: ClaudeReady, ClaudeNotInstalled, or ClaudeNotAuthenticated.
ClaudeReady: The claude CLI is installed and authenticated.
ClaudeNotInstalled: The claude CLI is not on PATH.
ClaudeNotAuthenticated: The claude CLI is installed but not authenticated.

Methods for the ClaudeCliBackend class

ClaudeCliBackend.cc_sentiment(): Build a backend preset for the sentiment/pushback scoring path.
ClaudeCliBackend.build_command(): Build the claude -p argv for one stdin-prompted invocation.
ClaudeCliBackend.parse_response(): Parse claude stdout into text or a validated model.
ClaudeCliBackend.env(): Return no extra environment variables; the claude CLI runs with the inherited environment.
ClaudeCliBackend.build_argv(): Build the inline -p argv for the sentiment/pushback scoring path.
ClaudeCliBackend.parse_result_envelope(): Parse the {is_error, result} JSON envelope from claude -p --output-format json.

Subprocess plumbing shared by the CLI backends.

run_cli(): Run a CLI command to completion and return its stdout.
arun_cli(): Run a CLI command asynchronously and return its stdout.
collect_process(): Drain a subprocess’s stdout and stderr concurrently and wait for it to exit.
map_concurrent(): Map an async function over items with bounded concurrency.

Shared type aliases.

types.TModel: Abstract model tier; each backend maps it to a provider-specific model name.
types.TSpecialty: Task specialty; LlmBackends.for_specialty maps each to its registered backend.

Local Apple-Silicon engine, adapter codec, fuser, and runtime patches.

mlx.MlxEngine: Runs batched MLX inference on a dedicated worker thread.
mlx.AdapterCodec: Compresses and decompresses a homogeneous-dtype safetensors LoRA adapter.
mlx.AdapterFuser: Fuses a shipped LoRA adapter into a base MLX model.
mlx.MLXPatches: Idempotent runtime patches for mlx_lm.