preserved-thinking-audit tests Claude agent harnesses

As reasoning models enforce strict prefix-integrity checks on thinking tokens, context compaction can no longer rely on naive client-side truncation. On-demand compaction paired with dedicated harness auditing provides the architectural blueprint developers need to keep long-running agents stateful without blowing token budgets or triggering invalid-signature errors.

• –Prefix integrity enforcement: Claude models like Fable 5.1 strictly validate signatures on preserved thinking blocks, meaning any client-side tampering with past messages or tool definitions triggers API errors or drops thinking.
• –On-demand background compaction: The compact-2026-09-04 beta allows agents to request signed compaction blocks asynchronously while execution continues on full history, eliminating latency spikes during context truncation.
• –Keep-tail preservation: Retaining recent conversation turns uncompacted ensures that an agent's immediate chain of thought remains coherent and valid following a compaction swap.
• –Automated harness verification: The preserved-thinking-audit skill automates 23 separate static and dynamic checks against local recording proxies, catching subtle signature-breaking bugs before agents hit production.