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
Subagent collaboration: subagents can now proactively reach their lead agent mid-task — ask for context (info), request a handoff described as a needed capability (handoff, the lead picks the specialist), or escalate a decision (escalate) — with structured payloads and per-turn rate limits. Blocked children are instructed to ask instead of guessing, and decisions still reach a child that already finished via Agent(resume=...).
Bot mode (/bot): an unattended permission mode where reversible actions (reads, in-workspace edits, read-only commands, goal and collaboration tools) auto-approve while irreversible operations (push / publish / delete / sensitive files / unclassified) are denied and parked for human review — with the denial reason fed back to the model. Web sessions in bot mode keep the server alive so unattended runs finish, and blocked goals write a decision report under unattended/.
Evidence-based goal verification: rejections must name concrete gaps (issue-free verdicts are returned for specifics); evidence gaps self-repair for up to 3 rounds, subjective calls park the goal immediately for a human decision, and disputed verdicts can be arbitrated by a reviewer subagent.
Web search upgrade: new Bing provider (server-rendered HTML parsing with click-tracker decoding) joins a reordered fallback chain — DuckDuckGo → Bing → Sogou → Baidu → 360 — guarded by a chain-level 40s budget and a 15s per-source ceiling so a dead path cannot stack waits. Parsing tolerates alternate markup, undecodable redirect links are dropped, and anti-bot challenges fail fast into the next engine.
Documentation: AGENTS.md gained sections for subagent collaboration, bot mode, and goal verification triage; the README (EN/ZH) reflects the new capabilities and the /bot command.