[Feature Request] Add more 1v1 text-only negotiation tasks for LLM-only / non-VLM users

[LUTIAN233]

Problem

Currently, the single_buyer_product_seller category has 28 tasks, but the breakdown by modality is highly skewed:

Type	Tasks	Can run with LLM-only?
Text-only, price-only	Task1, Task2, Task3	✅ Yes
Multimodal + contract	Task4–Task28 (25 tasks)	❌ Requires VLM (image input)

Users who run local LLMs via vLLM/SGLang without vision support (e.g., --language-model-only) can only evaluate on 3 out of 28 single-buyer-single-seller tasks. This makes the benchmark nearly unusable for text-model evaluation or RL training with local models.

Request

Add more text-only, 1v1 negotiation tasks that:

1. Do not require image input — no product_info.image_url, runnable with BaseLLM / OpenAILLM / CustomLLM / VLLMLLM
2. Cover more diverse scenarios — the 4 scenario families (E-commerce, Taxi, Food Delivery, Rent House) already have rich text metadata per task (product descriptions, pricing rules, surcharge breakdowns) that work perfectly fine without images
3. Optionally include contract_config — multi-dimensional contracts can also be done purely via text, since all field_descriptions, continuous_bounds, discrete_options are already text-encoded in the prompt

Proposed approach

The simplest way would be to provide "text-only variants" of the existing s1–s25 scenario tasks — same contract_config, same product_info text, same scoring, just dropping the product_info.image_url line. These could be registered under environment IDs like Task_s1_text-v0.

Alternatively, a batch of 10–20 new text-only tasks with varied difficulty levels (close-price, wide-spread, market-aligned, multi-round grinding) would be very helpful for systematic evaluation of text-only models.

Impact

This would significantly lower the barrier for:

• Researchers evaluating open-weight LLMs (Qwen, LLaMA, DeepSeek) without VLM capability
• RL training experiments (e.g., GRPO/PPO on negotiation) that use text-only environments
• Reproducibility studies on commodity GPU setups where VLM inference is prohibitively expensive
  （written by deepseek）

[liuxianyang1]

Thanks for the detailed request! We've added 25 text-only 1v1 negotiation task variants covering all four scenario families (E-commerce, Taxi, Food Delivery, Rent House), i.e. text-only versions of s1–s25.

They live under a new isolated directory so nothing existing is touched:

agenticpay/examples/single_buyer_product_seller/text_only/

Each variant:

• No image input — product_info.image_url removed; runs with CustomLLM (any text LLM, e.g. gpt-3.5-turbo, DeepSeek-R1), no VLM required.
• Same scenarios & scoring — identical contract_config / product_info text metadata and MAUT scoring; only the image is dropped.
• Full multi-dim contracts — field_descriptions, continuous_bounds, discrete_options are text-encoded, so contract negotiation works exactly as before.

Results are written to a separate results/single_buyer_product_seller_text/ directory. A batch runner is included:

cd agenticpay/examples/single_buyer_product_seller/text_only
export OPENAI_API_KEY='your-key'
export OPENAI_BASE_URL='your-url'
bash run_all_text_tasks.sh   # runs all 25 with the default text model