Challenge detection previously regex-matched the entire page innerText, so
chat transcripts or console copy containing words like 验证码/安全验证/verify
flagged healthy accounts as challenge_required. Detection is now scoped to
real evidence only (visible captcha vendor widgets, short modal copy, sparse
interstitial pages) via a shared challenge-signals module used by the generic
AI probe, doubao/kimi/qwen auth rules, and the deepseek runtime snapshot.
Platform API error codes (e.g. baijiahao_challenge_required) keep working.
Account health also could not recover after lid-close sleep or a network
drop: offline probes misread cached console pages as expired, and expired
records null out nextProbeAt so they were never re-probed. Probing is now
skipped while offline (network_error backoff instead), and powerMonitor
resume/unlock plus an offline-to-online transition watch schedule jittered
recovery probes for all tracked accounts, including expired ones. Challenge
rechecks wait 8s so transient widgets during page load no longer stick.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
root
abacc1f421