Evaluate a condition expression, optionally with ``and`` / ``or`` connectives. Supports: - Single comparisons: ``"score >= 0.5"`` -

Expand template variables in text using context

Run the YAML agent with a YAML workflow file. This is the main entry point for executing a workflow. It handles: - Loading and parsin

Unified long-term memory tool for persistent information storage and retrieval. This tool provides episodic memory capabilities that persist

Parse an LLM message and extract all tool calls. Args: message: The message object from LLM response. Expect

Log a structured event for machine parsing.

Compute effective tool configuration for a step. Applies filtering rules to determine which tools and submodules are available for th

Load and parse a YAML task file

Resolve a dotted path like 'A.B.C' through nested dicts in context. Walks each segment of the dot-separated path, descending into nested dict

Run a single task through the agentic loop. Args: task: The user's task description. plan_mode: If True, runs structu

Resolve an operand to its actual value. Args: operand: Operand string (variable name, number, or quoted string) context: Con

Create a NEW EffectiveArgs with YAML config overrides (immutable pattern). This method does not modify the current instance. Instead

Generate next node ID

Tool message whose content JSON contains som_elements.

Execute a step with persistent conversation history. Steps maintain conversation history across iterations, enabling multi-turn inter

Execute a single tool and return the result. Args: tool_name: Name of the tool to execute. tool_args: Arguments for t

Extract a computer-use action from LLM response text. Args: text: Full LLM response (may include thought text before the action line).

Clean text for Mermaid (remove problematic chars, NO HTML escaping here)

Scan *workspace_dir* and return a lightweight manifest dict. Files that are text and smaller than *max_file_size* are stored inline. Larger o

Create EffectiveArgs from an argparse namespace or similar object. Args: args: The original args object (typically from

Make a chat completion request with automatic retry.

Sanitize text fields returned by upstream APIs. In particular, strip NUL bytes ('\\x00') which can cause JSON files to be treated as bina

Check if a code block language indicates shell/bash content. Args: lang: The language specifier from the code block (e.g., "bash"

Return context window size in tokens for a model. Tries LiteLLM's model registry first, then falls back to a local dict for models not covered

Try to parse a string as a dict (JSON or Python literal). Handles LLM output that wraps JSON in markdown code blocks, e.g.:: ```json

Count tokens in a message list for the given model. Uses litellm.token_counter when available — routes to the correct tokenizer per provider

Add user message to history.

Extract {{var}} or {{var.field}} references from a string.

Root directory for all run outputs.

Print info message.

Deterministic RGBA color seeded by sequential element ID.

Check if content looks like a function/tool call. Args: content: The code block content to check. Returns: T

Draw numbered bounding boxes on a screenshot using the two-phase approach. Args: screenshot_path: Path to the input PNG screenshot.

Evaluate a condition string against context values.

Print error message.

Update task fields and persist. Accepts any Task field as kwarg.

Load memory from file, return empty dict if file doesn't exist

Create a temporary PNG screenshot and return its path.

Strip som_elements from tool results beyond the last `max_keep` browser calls. Element IDs are regenerated on each browser_navigate, so old AC tr

Keep only the last `max_images` screenshot messages as actual images. Older screenshot messages have their image_url block stripped, leaving

Add fields to a record only if they have a defined value.

Set a context variable from a value or expression. Args: step_data: Contains 'set_variable' key with: - name: Var

Check if a step ID is marked as completed.

Normalize submodule declarations from YAML data to a standard format.

Recursively check all steps in a workflow or nested block.

Run a git command, return stdout or None on failure.

Execute a single workflow step based on its type. This is the main entry point for step execution. It logs step start/end events and

Get system prompt for YAML agent.

Mark a step as completed and optionally update the checkpoint file. Args: step_id: The step ID to mark completed.

Print success message.

Parse list of steps

Parse list of steps

Save full text only if it differs from display label (i.e., was truncated)

Truncate text to maximum length

Write self.data to disk atomically.