Usage page: a Cache writes card. It shows what prompt-cache writes cost and their share of your spend, how many were written for 5 minutes or for 1 hour, and how much of a conversation was written to the cache again after it had already been cached, grouped by the break before it: under 5 minutes (the cache was invalidated or missed), 5 to 60 minutes (the 5-minute cache ran out; a 1-hour cache would have kept it) and over an hour. It also estimates whether a 1-hour cache would have saved money on your own traffic: what it would have saved on those rewrites against what 1-hour writes (2x the input price instead of 1.25x) would have cost. The figures start once the gateway on your machine is updated (hnl update); older gateways keep syncing as before.
Usage page: Gateway savings shows the gateway's cache assist: how many requests HarnessLink added prompt caching to, and what their cache reads saved.
Setting gateway.cache.anthropicTtl (5m by default, or 1h) sets how long the caching the gateway adds lasts. It applies only to requests that arrive without any caching of their own; caching that omp, Claude Code or another harness sets itself is never changed.

The gateway's prompt-cache assist (for requests that arrive with no caching of their own) now uses Anthropic's automatic caching, which moves forward with the conversation, plus one cache point on the system prompt, or on the tool list when there is no system prompt, so a new conversation with the same tools still starts from cache. Before, it marked the system prompt and the last message, which could miss the cache once a long agent turn added more than 20 blocks.
Cost figures price 1-hour prompt-cache writes at Anthropic's 1-hour rate (2x the input price) instead of the 5-minute rate (1.25x): the gateway reads the 5-minute / 1-hour split Anthropic reports, the price list carries a 1-hour rate, and the Usage page, Overview, usage limits and the cache-write advisory all use it. Before, 1-hour writes showed about 37% too cheap.

A synced table that gained columns in a new release kept uploading only its old columns on existing machines, because the change-capture triggers were never rebuilt. They are rebuilt when the synced columns change.

Flag notifications