policy: AI Safety "Tool-Call Recursion & Rate" Limit

[akarlaraytu]

📘 Context

LLM agents can enter recursive tool-call loops (agent calls tool A, tool A
returns a result, agent calls tool A again with slightly different args,
forever). They can also burn API budget by issuing high-frequency calls.

This policy acts as a deterministic safety gate against runaway behavior at
the agent runtime layer, regardless of what the model decides to do.

🎯 The Goal

Write a tool-call recursion and rate-limit policy for any LLM agent runtime.

Variables you can use:

• tool_name (String): Name of the tool being called.
• tool_call_count_session (Integer): Total tool calls in the current agent session.
• tool_call_count_last_minute (Integer): Tool calls in the last 60 seconds.
• same_tool_consecutive_calls (Integer): Consecutive calls to the SAME tool.
• same_args_repeat_count (Integer): Times this exact (tool, args) tuple has been called this session.
• tool_cost_usd (Float): Cost-per-call estimate.
• session_cost_so_far_usd (Float): Cumulative session cost.
• session_budget_usd (Float): Configured session budget.
• recursion_depth (Integer): Current call stack depth (agent → tool → agent → tool ...).

Requirements:

• BLOCK if same_tool_consecutive_calls > 10 (recursion guard).
• BLOCK if same_args_repeat_count > 3 (idempotency loop guard).
• BLOCK if tool_call_count_last_minute > 30 (rate limit).
• BLOCK if session_cost_so_far_usd + tool_cost_usd > session_budget_usd.
• BLOCK if recursion_depth > 8.
• ESCALATE if tool_call_count_session > 200.
• ESCALATE if session_cost_so_far_usd > session_budget_usd * 0.8.

📚 Resources

• Start here: docs/getting-started.md
• Check syntax: quickstart/01_hello_world.csl
• See how examples are structured: examples/community/

✅ Definition of Done

Submit a PR with: examples/community/ai_safety_tool_recursion_limit.csl