The only tool you need to know what is happening and how to fix it.
Go alertingexception-trackerissue-trackinglogsmetricsmonitoringobservability-platformopen-telemetryotelsession-replaytelemetrytraces
package agent
// Prompt caching. Agent loops resend the whole history every turn, so caching
// is most of the cost saving.
//
// - Anthropic models (caching: anthropic) get explicit cache_control
// breakpoints: one on the system message (the generic prompt plus any
// preloaded skill; tools sit before it in the cached prefix) and one on
// the most recent message, so each turn reads all earlier history from
// cache. Anthropic checks earlier block boundaries for hits on its own, so
// a single moving breakpoint is enough.
// - OpenAI models (caching: automatic) cache any stable prefix. The system
// prompt therefore carries no timestamps and tools keep a fixed order.
// applyCaching returns the messages to send for a caching mode, without
// mutating the conversation itself.
func applyCaching(mode string, msgs []Message) []Message {
if mode != "anthropic" || len(msgs) == 0 {
return msgs
}
out := make([]Message, len(msgs))
copy(out, msgs)
mark := func(i int) {
m := out[i]
text, ok := m.Content.(string)
if !ok || text == "" {
return
}
m.Content = []Part{{Type: "text", Text: text, CacheControl: &CacheControl{Type: "ephemeral"}}}
out[i] = m
}
if out[0].Role == "system" {
mark(0)
}
if last := len(out) - 1; last > 0 && (out[last].Role == "tool" || out[last].Role == "user") {
mark(last)
}
return out
}