Reflect & Pivot Recovery

When a loop is detected, simply killing the agent is a wasted opportunity. In many cases, the LLM simply needs an explicit signal that its current strategy is failing.

LongGuard’s Reflect & Pivot mechanism injects targeted guidance into the agent context to force a course correction.

---

Default Pivot Templates

LongGuard formats a context-aware prompt depending on which detector triggered:

Trigger Pattern	Prompt Strategy
tool_repeat	"You have called {tool} multiple times without success. STOP calling this tool. Try a completely different tool, re-scope your goal, or summarize what you have so far."
semantic_oscillation	"Your reasoning is cycling between states. Step back, state what you know with certainty, identify the single missing piece of information, and change direction."
dead_end_drift	"Your last {steps} actions produced no new information. Provide your best answer with current info or abandon this path entirely."
token_velocity	"Your token consumption has spiked dramatically. Summarize concisely in 2-3 sentences and proceed directly to completion."

---

Customizing Pivot Templates

You can override templates globally or per detector:

from longguard import GuardConfig

config = GuardConfig(
    pivot_templates={
        "tool_repeat": (
            "[SYSTEM ALERT] You have repeated {tool} {count} times! "
            "Explain your difficulty to the user instead of trying again."
        )
    }
)

Available Template Variables

• {tool}: Name of the repeating tool.
• {count}: Number of repetitions.
• {window}: Sliding window size.
• {steps}: Consecutive dead-end steps.
• {velocity} / {baseline}: Token consumption metrics.