Apply YAML config as fallback for unset CLI arguments. Args: args: Parsed CLI arguments yaml_config: Config from YAML file

Build a BM25 index over a list of documents and save it under the given collection name. Args: collection: Name for this index (e.g.

Public method to clean up the client

Convert nested FlowNodeData[] via recursive map, defaulting to [].

Copy fields from src to dst only when explicitly defined (not undefined).

Extract config section from task data with default values. This method extracts the 'config' section from a YAML task file,

Extract the tool name from a tool definition. Handles various tool definition formats including: - OpenAI function format: {'function': {'nam

Load all submodule tools defined in a YAML file. Returns: Tuple of (tools list, registry dict mapping function name to spec)

Parse a submodule's function declaration from its YAML data.

Rebuild workspace state from a previously captured manifest. Returns True if restore was fully successful, False if partial/failed.

Get the next LLM replay event with alignment validation. Retrieves the next event from the replay queue and validates that it matches

Search a BM25 collection by name and return the top-k results with scores. Args: collection: Name of the index to search. qu

Create the runtime directories and export env vars for subprocesses. Idempotent. Safe to call from any entry point before path-dependent work.

Route tool invocations through docker exec. Handles shell_run, fs_read, and fs_write tools.

Copy a file if it exists. Args: src: Source file path dest: Destination file path Returns: True if file was copied,

Mark a sub-step as completed and persist its output. For ``input`` steps, an empty output means the user likely did not provide any input (e.

If *step_id* is already completed in the checkpoint, return its saved output (or "" if none was stored). Otherwise return ``None`` to signal

Recursively estimate pixel height for nested steps (used for container node sizing).

Populate ``AGENTSPEX_RUN_*`` env vars and return ``(run_id, host_run_dir)``. This is the single point where per-run paths are decided. All entry

Load config section from YAML file. Args: workflow_file: Path to the YAML workflow file Returns: Dictionary with config

Replace the current process with a script from scripts/.

Expand template variables

Extract variable names from a condition string.

The core while(True) tool-calling loop.

Load LLM response events from a trace file for replay. Reads all LLM response events from a trace file and stores them for sequential

Normalize a line for comparison (strip whitespace, lowercase).

Process a workflow without connecting from a previous node. Returns the first node ID.

Save memory to file atomically

Convert a real path under `session_root` to a vpath. Raises PermissionError if `real_path` is outside `session_root`.

Write a trace event for durable execution. Thread-safe method that records events to the trace file when tracing is enabled.

Best-effort: capture current MCP session id from transport.

Build a tool definition from a submodule function spec.

Close shell and clean up session directory.

Map a YAML step-type key to the Lean output BaseType. Current YAML workflow language (src/harness/execution/interpreter.py::_STEP_TYPE_KEYS)

Initialize or retrieve conversation history from context. If conversation history doesn't exist, creates it with a system prompt. The system

Execute a task as a fresh, stateless LLM call. Tasks send an instruction to the LLM and capture the response. They are stateless - ea

Execute the multi-step tool calling loop. This is the core execution loop that: 1. Sends messages to the LLM 2. Parses tool c

* Resolve branch steps: prefer inline container steps over edge-connected legacy steps.

If text contains multiple code blocks, truncate to keep only the first one. Returns the original text if there are 0 or 1 code blocks.

Map a YAML step-type key to the Lean StepType constructor name. Historical note: Lean's `step` and `task` meanings were once reversed relativ

Extract the leading whitespace from a line.

Get a configured Slack WebClient.

Inject default cwd and conda_env for shell_run tool calls. Automatically adds working directory and conda environment settings from c

Check if command contains the task completion token. When ``require_submodule_submit`` is set in *context*, direct submission is bloc

Tool message with no som_elements (e.g. a non-browser tool).

Process a single step and return its node ID

Process a workflow and return the last node ID