Show when Claude's prompt cache expires in the composer

[pBoelck]

When I come back to a Claude thread, I can't tell whether the prompt cache is still warm. If it has expired, the next message re-reads the whole context uncached, which costs time and usage limits. Claude Code writes 1h cache entries on a subscription and 5m entries with an API key, Bedrock or Vertex, so the window is short enough to miss.

T3 Code already knows this. The Claude adapter derives promptCacheTtlMs from the cache_creation buckets of each response, and the web client uses it to offer the resume compaction chip on threads above 100k tokens. The expiry itself is never shown.

I'd like to propose a small opt-in indicator next to the send button:

• Cache 42m while the cache is warm
• yellow in the last fifth of the TTL (at most the last 5 minutes)
• Cache cold in red once it has expired
• hovering shows the exact expiry time and how much of the last request came from the cache

It would be a switch in Settings > General, off by default like the context window indicator. It only appears where the provider reports a TTL, which today means Claude. Codex reports cached input tokens but no TTL, so a countdown there would be a guess.

One server detail: a response that only reads the cache reports no TTL, although it refreshed the cache. The adapter would keep the last known TTL per session so the countdown doesn't disappear after such a response.

I have this working on a branch of my fork and can open a PR if the direction and scope sound right. Would you take this as an opt-in setting, or would you rather fold it into the existing context window indicator?

[pBoelck]

Screenshot from the prototype on a Claude thread. The composer shows Cache 46m, and hovering shows the exact expiry time and that 99% of the last request came from the cache.

This build still shows the details inside the context window popover. On the branch they now have their own hover card, so the indicator also works with the context window indicator turned off.

[ixtli]

I have gotten used to this with ccstatusline in CC.

I think its important to show how long until the model's cache would expire as a countdown or an icon or something. Further, CC has a new-ish feature flag controlled thing where it will auto-compact when you're nearing your cache timeout. I asked claude to search the harness source to figure out what the deal is, and for posterity here is what it emitted:

[it] is controlled by Anthropic through a remote feature flag (tengu_sunny_locket), so the flag was most likely just turned on for your account. [...]

[...] each time a request finishes, it sets a timer. When the timer fires, it compacts only if all of these hold:
- The prompt cache lasts 1 hour,
- The timer is set for 90% of the way to the cache expiring, so about 54 minutes after the last request. The flag can move this anywhere from 50% to 95%.
- The context has at least 200k tokens. The flag can lower this to no less than 100k.
- Auto-compaction is on, and so is a separate idleCompaction setting (on unless you turn it off).
- You haven't touched the session in the last 60s.
- No turn is running and nothing newer has been sent.
- Another process isn't holding the session.
- You aren't near your usage limit.

Anyway this is a pretty important feature for an application like t3 to have.

[ixtli]

Someone has also done something similar in a fork. https://github.com/n0ah37/t3code/pull/3/changes

[ixtli]

There is another idea post for a similar thing: #7590