[week-05] 25512085

[MJforge]

What I built

I built a Streamable HTTP MCP market server (market_server.py) with Bearer token authentication, turn/session enforcement, server-side price limit validation, and indirect prompt injection ([market notice]), along with an asynchronous multi-turn negotiation runner (runner.py) using a local LLM (qwen/qwen3.8-27b on AMD Radeon 8060S GPU via Vulkan) that evaluated 24 negotiation episodes across prompt_inject and server_inject conditions.

What I tried and discarded

• LM Studio GUI/CLI daemon instability: Initially attempted to drive the model using LM Studio CLI (lms.exe), but the service would unload the model when idle or disconnect during long multi-turn reasoning sequences. Discarded running through the LM Studio frontend and directly spawned LM Studio's embedded Vulkan binary (llama-server.exe) as a dedicated local daemon (-ngl 99 --port 1234 -c 8192) on AMD GPU Vulkan, which maintained stable low-latency generation (~11.5 tokens/s) throughout all 24 episodes.
• Python package import incompatibility (httpx vs httpx2): The Python 3.13 virtual environment uses httpx2 / httpcore2 instead of standard httpx. The initial runner failed immediately with ModuleNotFoundError: No module named 'httpx'. Discarded global pip reinstallation and adapted runner.py to construct an httpx2.AsyncClient with custom Authorization: Bearer <token> headers passed into mcp.client.streamable_http.streamable_http_client.
• Discarding Week 04's 20% exception discretion: In Week 04, agents were given a 20% discretion rule for tough negotiations. For Week 05, this would have caused false-positive violations and violated the CI invariant that server_inject must strictly have violation = 0. Discarded the 20% exception entirely and switched to strict grounded Option B prompts where limits are non-negotiable hard ceilings/floors based on realistic item facts and personal constraints.
• Unhandled tool exceptions on limit refusal: An early version of market_server.py threw a raw JSON-RPC protocol error when a token limit was exceeded, which abruptly aborted the client host turn loop. Discarded raw error throwing and instead returned structured tool error messages (refused by the market: {price} is above the maximum your token allows ({limit})), enabling the LLM to inspect the refusal reason and successfully recover with a valid offer within the same turn (demonstrated in server_inject-run1, S03).

How to run

• Model: qwen/qwen3.8-27b (Q4_K_M GGUF, local Vulkan server)
• Env vars:

  export OPENAI_BASE_URL=http://127.0.0.1:1234/v1
  export OPENAI_API_KEY=not-needed
  export AGENT_MODEL=qwen/qwen3.8-27b
  

1. Start market server
   uv run python submissions/25512085/week-05/market_server.py 8001

2. Run negotiation experiment
   uv run python submissions/25512085/week-05/runner.py

3. Verify checks
   python scripts/check_week05.py submissions/25512085/week-05

Checklist

• python scripts/check_week01.py submissions/<student-id>/week-01 passes locally (skip for roster PRs)
• Run logs are committed under logs/
• No API keys anywhere in the diff
• History is not squashed