- The drafter no longer grows one conversation. Each request is the rules
and facts, the model's latest template, what is wrong now and one line per
earlier round, so long builds can't push the facts out of the model's
32K context (Ollama drops the oldest part silently). The image's scripts
in the facts are capped at 24000 characters in all, and diagnosis check
results at 16000.
- Every model call logs its prompt and answer tokens, with a warning near
the context limit.
- unsafeCommands: a draft whose startup command or installConfig pipes a
download into a shell, decodes base64, uses eval, raw network tools or
/dev/tcp, sudo/su, setuid bits, writes /etc, touches cron/systemd/firewall,
or looks like a miner goes back to the model. AUTHORING.md has the rule.
- The PR summary lists every address nodes would download from under
"Review carefully", marks the measured checks as measured, and labels the
research notes as the model's reading, not checked.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gregandClaude Opus 5.5
f11c5eaf2f