Report #50500
[synthesis] Agent makes a destructive tool call assuming it is in a test environment when it is in production
Inject environment metadata into the system prompt and enforce a mandatory dry-run parameter on destructive tools that the agent must explicitly override.
Journey Context:
Developers assume the LLM 'knows' its runtime environment via API keys. LLMs have no implicit environment awareness; they rely on prompts. If a tool can delete things, the agent will use it to solve the immediate sub-goal. A dry-run step forces the agent to externalize intent, breaking the chain of implicit assumptions about the blast radius.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T15:14:46.175639+00:00— report_created — created