Get evaluation history for an agent Args: signature: Agent signature Returns: List of evaluation re

Returns list of Python packages to preinstall in the E2B sandbox. Based on analysis of GDPVal tasks and agent terminal logs showing

Get local paths to task reference files Args: task: Task dictionary Returns: List of file paths (ab

Load the ``agents.clawwork`` section from nanobot's config JSON. Returns a ``ClawWorkConfig`` with defaults for any missing fields.

Load the gdpval parquet file

Load task_completions.jsonl indexed by task_id → entry dict.

Read a Microsoft Word document and extract text. Args: docx_path: Path to DOCX file Returns: Extracted text

Read an image file and return base64-encoded data URL. This format is compatible with OpenAI's vision API format: {"type": "image_ur

使用 SiliconFlow DeepSeek-OCR 对单个 PDF 做 OCR。 NOTE: This is the old OCR-based method. The default is now read_pdf_as_images() which con

Read a plain text file. Args: txt_path: Path to TXT file Returns: File content as string

Read a Microsoft Excel spreadsheet and extract data. Args: xlsx_path: Path to XLSX file Returns: Formatted

Run daily session: decision-making and activity execution Args: date: Date string (YYYY-MM-DD) Returns:

* Send templated email via SES

Track API call with flat-rate pricing (e.g., Tavily search at $0.0008 per call) Args: cost: Flat cost in dollars for thi

Track token usage from a LangChain API response into EconomicTracker. Prefers response_metadata["token_usage"] (raw API dict) over LangChain's

Agent invocation with retry, timeout, and token tracking Args: messages: List of messages to send to the agent

Build the LangGraph workflow

Build multimodal evaluation content with text and images Returns: List of content blocks for OpenAI's multimodal

Construct ClawWorkState from the ``agents.clawwork`` plugin config. Reads clawwork settings directly from the raw JSON in ~/.nanobot/config.j

使用 Qwen VL OCR 对多页 PDF 逐页识别并返回原始结构： { "content": "<text>", "total_pages": ..., "model": "...", "usage": {...}

Create a LangChain Tool from MCP tool definition

Execute a tool by name with given arguments

Extract numerical score from LLM evaluation response Args: evaluation_text: Full evaluation text from LLM Retur

Extract web page content using Tavily Extract API Args: urls: Single URL or comma-separated list of URLs to extract content from

Format single image as multimodal message

Format PDF/PPTX images as multimodal message

Format regular text tool result

Try to match an occupation name, falling back to default.

Get default MCP configuration - Work and Learn only

Get global state from parent module

Get global state from parent module.

Get global state from parent module

Get global state from parent module

Get tools from a specific MCP server

Parse /clawwork <instruction>, classify, assign task, run agent.

Set EVALUATION_* env vars from nanobot's provider config. The LLMEvaluator (livebench/work/llm_evaluator.py:46-55) already checks EVALUATION_

Read task_completions.jsonl to find dates and task IDs already conducted in a previous run. Returns (already_done_dates: set[str], a

Load tasks from inline configuration

Load tasks from JSONL file

Load latest economic state from balance file

Load meta-prompt for a specific occupation category Args: occupation: Occupation name (e.g., "Software_Developers")

Load the occupation-to-wage mapping JSON.

Load tasks from parquet file (existing logic)

Load task_completions.jsonl, summing wall_clock_seconds per date.

Load task values from JSONL file

Log work income to token costs file

Create a LiteLLMProvider from nanobot config (mirrors nanobot CLI).

Copy task reference files to agent sandbox and upload to code sandbox for execute_code. Args: date: Current date

Wrap super()'s processing with start_task / end_task. Intercepts /clawwork commands to classify and assign tasks before handing off t

Read artifact file contents with separate handling for images and text Args: artifact_paths: List of artifact fi

Save consolidated task-level cost record (one line per task)

Search using Jina AI API (alternative - markdown-based results) Args: query: Search query string max_results: Maximum number

Search using Tavily API (recommended - structured results with answers) Args: query: Search query string max_results: Maximu

Select task based on explicit assignment configuration Args: date: Date string available_tasks: List of task

Set up log file path for activity messages

Summarize conversation to extract what the agent was working on Args: messages: List of message dicts with 'role

Track token usage from the API response. Delegates to the shared track_response_tokens() function. Prints the full response_metadata

Analyze the old format to show limitations

Replicate the server.py /api/agents/{sig}/tasks endpoint logic. Returns list of task dicts with: task_id, occupation, sector, task_value_usd,

Return {date: [task_id, task_id, ...]} preserving file order. Multiple entries for the same date come from repeated simulation runs.

Return {task_id: seconds} using logs/info.jsonl. A task may be started more than once if the agent was restarted mid-run. We collect ALL

Build the E2B custom sandbox template with preinstalled packages. Args: alias: Alias name for the template (default: gdpval-work

Calculate value for each task

Check if a category has already been generated

Note backward compatibility considerations

Check if e2b package is installed

Classify an instruction into an occupation with estimated value. Returns: { "occupation": str,

Scan and clean up failed runs for a single agent.

Replicate Dashboard.jsx domainChartData logic exactly.

Compute wall-clock seconds from the best available source. Priority: 1. token_costs.jsonl timestamp_end - timestamp_start 2. bal

Create mapping from date to task_id from tasks.jsonl

Creates a high-energy showreel for Goodsin Studios showcasing their best CGI work. Follows all specifications from the task description.

Create the prompt that asks GPT-5.2 to estimate hours needed for a task

Create the meta-prompt that asks GPT-4o to generate evaluation guidelines for a specific category

Create mappings for all occupations

Create prompt for GPT-5.2 to match occupations

Factory function to create a wrap-up workflow instance Args: llm: Language model instance (if None, creates default) logger:

Demonstrate the new format with actual examples