Extract text content from message, stripping thinking parts. Use this after parse_response when you only need the text output, ignoring any t

Get available tools from the MCP server. Returns: List of tool schemas in OpenAI format

Get available tools in OpenAI format. Returns: List of tool schemas in OpenAI format

Get available tools from the API server. Returns: List of tool schemas in OpenAI format

RoleColon satisfies the extension property - no content is stripped from history.

Llama3 satisfies the extension property - no content is stripped from history.

Whether this renderer satisfies the sequence extension property. A renderer has the extension property if, for any multi-turn conversation,

Extension property depends on strip_thinking_from_history setting. When strip_thinking_from_history=False, thinking traces are preserved in

Non-thinking mode always satisfies extension - no thinking to strip from history.

Extension property depends on strip_thinking_from_history setting. When strip_thinking_from_history=False, thinking blocks are preserved in

Qwen3 Instruct always satisfies extension - no thinking to strip from history.

Check if there are buffered tokens waiting for more data.

Initialize AlphaVantage server configuration. Returns: StdioServerParameters for AlphaVantage server

Install ART agent skills for Claude Code and OpenAI Codex. Copies bundled SKILL.md files into .claude/skills/ and .agents/skills/ in the targ

Check if a renderer name is registered.

Check if a tokenizer name is registered.

Load only the latest step metrics for each model. This is a convenience function for comparing final model performance. Parameters

Log constant metrics to W&B as horizontal lines across all training steps. Creates a W&B run and logs the same values at every step from 0 t

path.to.module:func_name or func_name (assumes default_module)

Migrate trajectory files from JSONL to Parquet format. This command converts old .jsonl trajectory files to the new .parquet format, whi

Migrate all trajectory files for all models in a project. Args: art_path: Path to the .art directory. project_name: Name of

Migrate existing checkpoints from the old structure to the new structure. Old: .art/{project}/models/{model_name}/{step} New: .art/{proje

Parse response, prepending <think>\\n since the generation prompt prefills it. When sampling with build_generation_prompt, <think>\\n is part

Parse response, prepending <think> since we prefill with it. When sampling with build_generation_prompt, the <think> tag is part of the

Parse response tokens with streaming, yielding incremental deltas. This method enables real-time display of model output by yielding

Create a bar chart visualising *metric_name* for all *models* at the **first** and **last** available training step. For *trained* models (th

Preview the first n scenarios.

Preview generated scenarios.

Pull trajectory checkpoints for *model* from the configured S3 bucket. This is a lightweight helper that mirrors the S3-sync logic used inside

Registers a model with the Backend for logging and/or training. Args: model: An art.Model instance.

Register a LoRA adapter for a specific checkpoint step. This is called when training is skipped but the checkpoint is renamed.

Register a custom renderer factory. Args: name: The renderer name factory: A callable that takes (tokenizer, image_processor) and

Register a custom tokenizer factory. Args: name: The tokenizer name factory: A callable that takes no arguments and returns a Tok

Render message, adding </think> to header when stripping thinking from history. HF's thinking=True template uses </think> at the start of his

Render message in non-thinking mode. For assistant messages (not following tool): - Strip any ThinkingPart from structured content

Reset parser state for reuse.

Resize an image so that its longest side is at most max_size pixels. Preserves aspect ratio and uses LANCZOS resampling for quality. Ret

Retry decorator with exponential backoff for both synchronous and asynchronous functions. Args: max_attempts: Maximum number of retr

Adapter function to fit the async reward logic into GRPOTrainer's expected sync signature. It extracts only the scalar reward value.