Repository navigation
v0.7.0
Cheaper interactive sessions on Token Factory. Token Factory only credits cached tokens when everything up to the last user message matches an earlier request, so every follow-up used to re-bill the whole conversation. Follow-ups now reach the model as the result of an ask_user tool, which only extends the prompt. A replayed follow-up went from 0% to 98% cached. The model can also call ask_user itself to ask you a question. (#17)
More robust agent loop. Replies cut off at the token limit get a nudge to continue, calls to unknown tools get the list of valid tools, and broken tool arguments get a clear error instead of running with no arguments. (#18)
/fix
- Branches report back through a
run_teststool, so retry rounds stay cached. (#20) ember -p "/fix <issue>"runs the full pipeline non-interactively. Add--yesto apply the winning patch. (#20)
Other changes
/model ultrawarns that Ultra gets few cache hits on Token Factory: it caches in 4,352-token blocks and never counts the last one. (#15)EMBER_NO_WEB=1removes the web tools. (#16)- New SWE-bench Verified runner and tuning-loop tools in
bench/. The baseline with Super resolved 15 of 19 at $0.077 per resolved task. (#16)
Full Changelog: v0.6.0...v0.7.0