PROJECT THREE - TOOLS, THE MODEL CONTEXT PROTOCOL (MCP), AND SAFETY

YOUR WRITTEN NOTE The acceptance criteria ask for a short written note about the idempotency bug. Write three or four sentences answering: where did you find it, what happened when it was triggered, and how did you fix it. Save it as stage3_notes.txt in the project folder. Writing down a finding is a core professional habit, because a fix nobody understands will be undone by the next person. EXAMPLES Example 1: MCP integration with a pre-built server. A learner connects the Project Two harness, as a client, to a simple ready-made server such as a local file-system server, and confirms the agent can call its tools through the same registry as native tools. In the lab, mcp_client.py and the mini server do this at a small scale. Example 2: Permission resolution with simulated approval. The learner adds a rule that any write or delete tool call needs a human yes, typed at the keyboard. In the lab, pre_execution_hook asks before any existing report is replaced. Example 3: Defending against a fake web page. The learner feeds the agent a fake page containing a hidden order to reveal a fake secret and verifies the harness's untrusted-content handling prevents compliance. In the lab, the same defense stops the resume_injection_test.txt order. Example 4: A deliberately malicious tool description. The learner writes a harmful-looking description purely as a test and confirms the validation step blocks it. In the lab, setting LAB_POISON to 1 does exactly that.