Bug: arun_society() adds completion tokens to prompt counter — completion_token_count is always 0

[marlonbarreto-git]

Problem

In owl/utils/enhanced_role_playing.py, the async method arun_society() (line ~564) has a copy-paste bug where assistant_response completion tokens are added to overall_prompt_token_count instead of overall_completion_token_count. Additionally, user_response completion tokens are never counted at all.

Code Comparison

run_society (sync, lines ~499-505) — CORRECT:

overall_completion_token_count += assistant_response.info["usage"].get(
    "completion_tokens", 0
) + user_response.info["usage"].get("completion_tokens", 0)

overall_prompt_token_count += assistant_response.info["usage"].get(
    "prompt_tokens", 0
) + user_response.info["usage"].get("prompt_tokens", 0)

arun_society (async, lines ~564-570) — BROKEN:

overall_prompt_token_count += assistant_response.info["usage"].get(
    "completion_tokens", 0          # <-- WRONG: completion tokens go to prompt counter
)

overall_prompt_token_count += assistant_response.info["usage"].get(
    "prompt_tokens", 0
) + user_response.info["usage"].get("prompt_tokens", 0)
# user_response completion tokens are NEVER counted

Impact

• token_info["completion_token_count"] is always 0 when using the async path
• token_info["prompt_token_count"] is inflated by the assistant's completion tokens
• Any cost tracking, billing dashboard, or benchmark reporting that relies on arun_society's token counts is silently wrong
• The async path is the one used by background processing and web app integrations

Proposed Fix

# arun_society — fixed
overall_completion_token_count += assistant_response.info["usage"].get(
    "completion_tokens", 0
) + user_response.info["usage"].get("completion_tokens", 0)

overall_prompt_token_count += assistant_response.info["usage"].get(
    "prompt_tokens", 0
) + user_response.info["usage"].get("prompt_tokens", 0)

This matches the correct logic already present in run_society.

Happy to submit a PR for this fix.