A failed boot went straight back to the model as "fix it", so it guessed:
one draft raised its memory three rounds in a row while the real cause, an
empty setting the server can't run without, stayed untouched.
Now a failed boot is investigated first:
- Evidence: the server's distinct error lines (timestamps and repeats
removed), and a web search on the most telling one, with the top answer
read through the reader.
- Diagnosis: the model states the cause, the evidence for it, its
confidence and the exact fix. If unsure, it asks for up to 3 checks: a
search, a page, the files the server writes (a first boot of the draft),
or a boot test with settings or the startup command changed. Results go
back to it, two rounds at most.
- Fix: the model applies exactly the diagnosed fix and nothing unrelated.
- Attempt log: each cause, fix and outcome. A fix that changed nothing
rules its cause out; if the diagnosis returns to a ruled-out cause, the
build stops instead of guessing, and the PR lists what the builder found.
Boot rounds go up to 5, since rounds now end on a ruled-out cause rather
than a fixed count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
