Originally published at recca0120.github.io }}">The first post scanned 95 days of logs and found sub-agent cache TTL silently dropped to 5m. }}">The second tracked it to 17 consecutive days of 100% 5m — conclusion: it's the new default. This time I broke down the model dimension. Scanning March through May 7, I originally wanted to confirm whether the cache TTL had reverted (it hasn't). Instead I found something bigger: the server doesn't just control cache TTL — it silently switched the main agent model three times . Data Source Same as before: ~/.claude/projects/{project-path}/{session-uuid}.jsonl . This time I also checked the message.model field in API responses: { "message" : { "model" : "claude-opus-4-6" , "usage" : { "input_tokens" : 142 , "cache_read_input_tokens" : 892041 , "output_tokens" : 3847 , "cache_creation" : { "ephemeral_5m_input_tokens" : 0 , "ephemeral_1h_input_tokens" : 8234 } } } } Enter fullscreen mode Exit fullscreen mode The model field comes from the server, not the client.…