Attacker brain stalls with zero tool calls (agent-prime via thegrid), even after explicit tool-name steering

[kazani-351]

Environment

• WallBreaker installed from source (pip install -e .), macOS
• Config: brain = agent-prime (thegrid, tool-tagged per Cherry Studio's model picker), target/judge = agent-standard (thegrid), protocol openai, base_url = https://api.thegrid.ai/v1

Repro steps

1. /sysprompt load <file>, /objective <text>, /rounds 10, /auto on
2. Send a kickoff message describing the objective
3. Agent produces zero tool calls for 2 consecutive rounds → Agent stalled twice with no action. Give it direction.

What I tried

• Originally used agent-standard as the brain — confirmed via Cherry Studio's model picker that it's not tool-tagged (it's a reasoning model that burns its token budget on reasoning_content before emitting real output/tool calls — reproduced this independently via a raw curl to /v1/chat/completions with max_tokens: 5, got content: null, finish_reason: "length", reasoning-only output). Switching to agent-prime (tool-tagged) fixed that stall on one occasion.
• On a later run (different challenge, same config), agent-prime stalled again — twice — even after sending an explicit follow-up naming the exact tool to call (query_target) and providing literal text to send as the probe. Still zero tool calls, empty response both times.

Question
Is there a known reliability issue with agent-prime tool-calling via thegrid's OpenAI-compatible endpoint? Does system-prompt/objective length affect tool-call emission? Is there a config flag to force stricter tool-call mode, or a recommended thegrid model for the attacker brain role specifically?