Crate ag_harness

Expand description

Lightweight, Rust-native LLM harness for application-facing agent workflows.

The crate provides a provider-neutral model client, validated structured output, and private provider backends.

Structs§

Harness: Application-facing harness for one complete model turn.
KimiConfig: Configuration for a Kimi model served through Moonshot AI’s OpenAI-compatible API.
LocalFileSystem: Tokio-backed filesystem implementation for local repositories.
ModelClient: Application-facing client for provider-neutral model requests.
ModelMetadata: Validated provider and model identity used by the shared client lifecycle.
ModelRequest: Provider-neutral input for one model request.
Muse: Muse model configured from the standard Model API environment variables.
MuseConfig: Configuration for a Muse model served through Meta’s Model API.
OutputSchema: A validated, provider-independent JSON Schema for model output.
QwenConfig: Configuration for a Qwen model served through Alibaba Cloud Model Studio’s OpenAI-compatible API.
ReadArguments: Validated arguments for the native read function.
ReadOutput: Bounded text returned by one successful read execution.
ToolCall: Provider-neutral model request for one native tool invocation.
ToolDefinition: Provider-neutral definition of a native model tool.

ModelError: Failure returned while completing a model request.
ModelMetadataError: Invalid identity attributes supplied by a model provider.
ModelResponse: Provider-neutral output from one model request.
MuseError: Failure returned while configuring Muse from the environment.
OutputSchemaError: Failure returned while constructing a structured-output schema.
ReadError: Failure while safely executing one repository-relative read.
Tool: Built-in tool that can be enabled for a harness run.
TurnError: Failure returned by a complete harness turn.

MUSE_SPARK_1_2: Standard Muse Spark 1.2 model whose prompts and completions are not used to train Meta models.
MUSE_SPARK_1_2_CONTRIBUTOR: Discounted Muse Spark 1.2 model that permits Meta to use prompts and completions to train future models.