[week-05] 26512073

[Tulga0419]

What I built

I built an MCP negotiation market with a buyer and seller using separate bearer tokens. I tested four scenarios three times under each condition: 24 episodes and six logs.

| Condition | Correct | Deal violations | Attempted violations | Refused calls | Mean turns |
| prompt_inject | 0/12 | 0 | 2 | 0 | 8.00 |
| server_inject | 0/12 | 0 | 0 | 0 | 8.00 |

All episodes ended open. The two attempted violations were seller proposals below its reserve. Separate authorization checks confirmed that the server blocks unauthorized and out-of-limit actions.

What I tried and discarded

The SDK initially hid refusal reasons behind generic errors. I changed these to ToolError so agents could read the reason, and preserved the earlier output. I removed the unused lab folder; its committed history remains.

I used AI assistance for coding, debugging, and English wording.

How to run

See REPORT.md for setup and commands.

Host: Python MCP. Model: gpt-4o-mini. Temperature: 0.

Checklist

• python scripts/check_week05.py submissions/26512073/week-05 passes locally
• Six run logs and 24 episode results are committed
• Four authorization checks are recorded
• History is not squashed
• No API keys or bearer-token values appear in the diff