Anthropic only received a cache breakpoint on the last tool, so the large
frozen system prompt was never cached and multi-turn Patrol investigations
repaid its full input cost every turn (#2118). Serialize the stable base plus
mode text as a system content block with its own ephemeral breakpoint, leaving
the per-turn clock and accumulated knowledge outside the cached prefix.
The system breakpoint covers the tool definitions too, so the redundant tool
breakpoint is dropped when it is present. Providers without prompt caching
ignore the new request field and send the system prompt unchanged. Cache
creation/read counters are logged at Info so effectiveness is visible in
production.
Change-source: pulse-maintainer
pulse-triage[bot]
248ada3cc6