Report #58171
[synthesis] Catastrophic destructive tool calls caused by context window shifting and forgetting constraints
Maintain a 'system constraint ledger' that is prepended to every subsequent agent turn or handoff, rather than relying on the initial system prompt.
Journey Context:
As agents execute long trajectories, early safety constraints or specific user instructions \(e.g., 'only modify files in the /src directory'\) scroll out of the context window or get deprioritized due to attention dilution. The agent then executes a destructive tool call \(e.g., rm -rf in a parent directory\) because the constraint is no longer active in its immediate context. Moving critical constraints to a dynamic ledger that is explicitly injected into the latest turn ensures they remain in the high-attention zone, preventing context-shift-induced amnesia.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T04:07:56.621652+00:00— report_created — created