Record a single file entry in the manifest. Small text files are stored inline; large or binary files are backed up to *backup_dir* (

Run prediction in micromamba environment.

Structured plan mode: decompose → per-task plan-execute cycle → synthesize. Phase 1 — Decomposition: LLM decomposes the task usin

Search OpenAlex API.

Initialize MCP tools, orchestrator tools, plan generator, and system prompt.

Quote a string for safe use in shell commands.

Check whether *name* matches any ignore pattern (simple glob).

Return the step type key (e.g. 'step', 'input', 'set_variable').

Summarize a raw plan execution result into a compact, structured form. Uses the plan generator model (cheap, e.g. gpt-5) to produce a

Attempt to parse YAML. Returns error message on failure, None on success.

Update a KEY=value line in the env file.

Write Mermaid diagram to file

Backup the source directory to the target directory using rsync for incremental sync. Falls back to shutil.copytree if rsync is not a

Build a complete evaluation summary dict. Args: results: List of result dicts. mode: "baseline" or "agent". model: Model

Clean up dashboard process gracefully. Mimics the cleanup() function in run_agent.sh: - Check if process is still running - Send SIGTERM

Convert existing JSON task to YAML format

Enable tool-output replay: invoke() will not call the server; it will return recorded tool outputs from the provided trace jsonl in o

Load and execute a YAML plan through the Interpreter. Args: plan_path: Path to the verified YAML plan file. model: Op

Run a shell command in the sandbox and return output with exit code. Args: mcp_client: MCP client for invoking sandbox commands c

Execute multiple sub-module calls in parallel and gather results. Uses ThreadPoolExecutor for true parallel execution of submodule calls.

Execute steps or submodule calls in parallel. Supports two execution formats: 1. List of different steps executed concurrently

Execute a while loop with condition checking. Repeatedly executes nested steps while the condition remains true. Has a configurable m

Expand template variables in text using context values.

Extract metrics dict from a ChemBench report dict. Args: report: A single question's report dict (from prompter.report().model_dump()

Format a tool result for display to the LLM. Handles special formatting for specific tools (like shell_run) and provides generic formatting f

Emit a Lean verification file for `ir`.

Detailed summary including full result_summaries for completed tasks. Used for inter-task context passing and final synthesis. Unlike

Check if task data has a config section. Args: task_data: The parsed YAML task data Returns: True i

Host directory mounted as ``/workspace`` in the sandbox container.

Check if a question was answered correctly. Reads all_correct from metrics (computed by score_answer). Args: report: A single questi

Main-flow handle first (next/bottom), then branch handles.

Normalize model name to LiteLLM format (provider/model).

Parse a single log line.

Parse parsed_task.json → WorkflowIR.

Read the top-level ``name:`` key from a workflow YAML without full parse. We need the task name *before* the harness runs YAML env-var expansion

Print a formatted evaluation summary to stdout. Args: eval_summary: Evaluation summary dict from build_eval_summary().

Print a horizontal rule to the console without logging to file.

Interactive REPL mode. Reads tasks from stdin. Special commands: /plan <task> — run task in plan mode (decompose → generate YAML p

Save evaluation summary to output_dir/evals/evaluation_summary.json. Args: eval_summary: Evaluation summary dict. output_dir: Bas

Save a writing result as JSON into the per-instance directory. Args: result: The WritingResult to save. instance_dir: Per-instanc

Save an ELAIPBench result as JSON.

Save an AIME result as JSON.

Score model response against ground truth. Uses ChemBench's regex patterns to extract answers: - MCQ: extracts letter(s) from [ANSWER]A[/ANSW

* Trigger file download

Host-only persistent store for logs, checkpoints, and outputs.

Wrap a Python code block as a bash heredoc so it can be executed via shell_run.

Yield cleaned candidate lines most likely to contain an action.

Add a basic step node - simplified format without HTML

Add a basic step node without auto-connecting from previous node

Add a module call node