Built to run unattended.

jobs submit/list/status/logs/cancel/rerun — SQLite-backed background tasks with attempts and self-healing retry. They survive restarts.

Optional OS-process isolation per agent over Unix sockets: one crash never takes down the fleet, per-process resource limits.

Subprocess hooks like in ZCode: PreToolUse (allow/deny), PostToolUse (extra context), Stop (limited "continue").

Stall detection (warn 450s / kill 1200s), cancellation tokens across the whole agent tree, retry with error classification and Retry-After, prompt-injection defense on web content, doom-loop detector.

stream: true in the agent loop — text deltas hit the TUI live, tool calls are assembled from stream fragments, automatic fallback to non-streaming on failure.

Side-effect tools (save_contacts, git_push by default) wait for approval: y/n in the TUI, POST /sessions/:id/approve in the API.

Agents can ask the operator mid-task; without an operator present they continue autonomously.

run --repeat N collects on a schedule and diffs runs: new emails/phones/persons are printed and alerted via webhook, Telegram or email.

Real BPE token counting (tiktoken cl100k_base), CJK-aware fallback, per-tool truncation with persistence-to-disk, Hermes-style compaction with anti-thrashing cooldown.

PDF, HTML, JSON, DOCX reports plus webhook, email and Telegram notifications on completion.

amoCRM, Bitrix24, HubSpot — dedup by crm_id, no duplicates on repeated pushes.

run --profile hunter|analyst|validator (or your own TOML): ready prompts, model/depth/tool overrides. Per-role models via [agent.role_models].

After fan-out an LLM judge checks the result against the goal and runs up to replan_rounds gap-filling rounds; count-based reflection for lead-gen quotas.

Session token limits, budget-capped sub-agent results, p50/p95 per-tool stats from real sessions.