You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Context estimate very incorrect when using with local strata server #96
When using Strata locally, I'm seeing agentty report 65.0k/65.5k while streaming a response, and then sometimes resetting down to the actual 33.0k/65.5k context size after the request completes. I'm often seeing the context size reported by agentty report being above 100% even though the strata web dashboard reports it at a much lower value. This makes the context usage estimate extremely unreliable as it rarely reports anything near what is in actual use.
Strata makes current prompt token usage available in the /metrics endpoint at .live.prompt_tokens, which can be used instead of guessing current context usage.
fixed in 60f3e22 and 1c3ebfd, cached tokens were counted twice on openai-style servers like strata so 33k read as 65k; the gauge now shows the real prompt size, reopen if you still see it off.
Activity
usage: openai-style cached tokens are part of the prompt, not extra (#96
context gauge: right number across threads, compaction and every wire (…
1ay1 commentedon Oct 15, 2026
fixed in 60f3e22 and 1c3ebfd, cached tokens were counted twice on openai-style servers like strata so 33k read as 65k; the gauge now shows the real prompt size, reopen if you still see it off.
web-connect commentedon Oct 15, 2026
I might see off. Maybe verify with author before closing
1ay1 commentedon Oct 15, 2026
hehe, lemme know I will reopen it