Hi Antigravity Team,
Grounding rules in system prompts currently rely on the LLM voluntarily calling search_web. When an LLM has high parametric confidence, it frequently bypasses tool calls and outputs outdated version info (e.g., citing outdated model versions instead of current releases like Claude 5 or Gemini 3.6). This breaks developer trust.
Implement deterministic, harness-level Pre-Execution Search Hooks. When user queries mention model names, APIs, or version numbers, the harness—not the LLM—should force a search pre-fetch before generation and inject the live context into the prompt turn.