cursorrules

ConductGuard Policy

At the start of every conversation:

• Call guard_activity ONCE with a one-line summary of what the user is asking.

Then, for the request itself:

• Call guard_check ONCE per intent — not per file, per command, or per edit.
• Re-check only when scope changes (reads → writes, local → network, or a new task).
• If BLOCKED: stop and explain the policy rule to the user.
• If WARNING: proceed but surface the warning inline.
• If the response is ok or empty: proceed silently. Do NOT narrate it in the chat.

Example:

1. User asks to fix a bug → call guard_activity once with the summary.
2. Before starting the code changes → call guard_check once with the intent (e.g. write_file, bash).
3. Do NOT re-check for every subsequent read or edit in the same task.

Never write secrets to files or output. Never send PII to external endpoints.

Agent Booster — Context-Efficient Reads

Prefer booster MCP tools over native file reads:

• search_context instead of searching files — semantic search across all indexed symbols
• smart_read instead of reading full files — returns only the relevant symbol slices for a task
• get_symbols to survey a file's structure before reading it
• route_model at the start of any non-trivial task to pick the right model tier

Run booster gain to see token savings.