Demonstrate what's now possible with new format

Build a single task_completions entry from all available sources.

Derive task_completions records for an agent directory and (unless dry_run) write them to economic/task_completions.jsonl. Returns the l

Use GPT-5.2 to estimate hours needed for a specific task

Evaluate work artifact(s) using LLM and category-specific criteria Args: task: Task dictionary with occupation, sector,

Extract date from filename like '2027-09-06.log'.

Force-assign a specific task to a date, bypassing the 'used' check. Used in exhaust mode to pre-select a task before run_daily_sessio

Format tool result for logging (handles binary data gracefully)

Format tool result into proper message format

Build task list from task_completions.jsonl (authoritative — no duplicates). task metadata is looked up from tasks.jsonl (first occurrence per ta

Generate a summary report of all generated meta-prompts

Generate a summary report of all estimates

Generate summary statistics for task values

List all agent directories.

Get all task IDs in the filtered task list (for exhaust mode)

Get list of all LiveBench tools Returns: - 4 core tools (decide_activity, submit_work, learn, get_status) - 6 productivity tools (search_

Generate system prompt for LiveBench agent with economic awareness Focus: Work and Learn capabilities only Args: date: Current s

Get current session token cost

Return the active provider (or requested provider before initialization).

Get total work earnings for an agent Args: signature: Agent signature Returns: Total earnings

Extract unique occupations from tasks

Initialize agent components

Check if a terminal log indicates a failed run. Returns failure reason or None.

Load economic/balance.jsonl keyed by task_id. Used as fallback source for wall-clock time.

Load balance history from economic/balance.jsonl

Load already processed task IDs from the output file

Load the gdpval parquet file

Load task_completions.jsonl, summing wall_clock_seconds per date.

Load task hour estimates from JSONL file

Load task values from JSONL file

Load task_id -> task_value_usd mapping and full pool metadata.

Load task_values.jsonl -> {task_id: {task_value_usd, occupation, sector}}

Load task assignments from work/tasks.jsonl

Load task assignments from work/tasks.jsonl. Returns {task_id: record} with the most recent assignment per task_id (handles re-runs where

Parse economic/token_costs.jsonl into two indexes: cost_records : {task_id: record} — records with timestamp_start/timestamp_end incom

Load hourly wage data from CSV file

Main execution function

Main execution function

Main execution function

Main execution function

Run all tests

Main execution function

Use GPT-5.2 to match a GDPVal occupation to a BLS wage occupation

Print formatted table for one agent.

Print summary of corrections

Recalculate balance history by scaling actual payments This preserves evaluation logic (0.6 cliff, etc.) by scaling actual payments rath