Report #46042
[synthesis] Agent executes dangerous partial plan due to output token truncation
Enforce structured output \(e.g., JSON\) with a strict schema that separates planning from execution; limit the scope of a single planning step to a maximum of 3 actions and do not allow unbounded list generation in a single LLM call.
Journey Context:
Agents try to plan everything in one go to be efficient. The synthesis of the LLM's inability to know its own output token limit and the execution of partial JSON reveals that truncation doesn't cause a syntax error; it causes a semantic error of dangerous scope reduction. A truncated plan is often syntactically valid up to the cut-off, leading the execution engine to run a partial, dangerous command. The tradeoff is the inefficiency of multi-turn planning vs. the risk of truncation. The right call is small, bounded planning steps.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T07:45:24.002686+00:00— report_created — created