Why does onboarding ask me to accept TRUST_MODEL.md?
HybridClaw enforces explicit trust-model acceptance before runtime starts. Acceptance is recorded in config.json with policy version and timestamp so operators can prove security acknowledgement. Technical runtime controls are documented separately in SECURITY.md.
Do I still configure everything through .env?
No. Runtime behavior now lives in typed config.json (with defaults, validation, and hot reload). Keep .env for secrets like API tokens and credentials.
Is it safe to let the agent run shell commands?
Yes. All tools execute inside ephemeral Docker containers with read-only filesystems, memory caps, and a deny-list of 28+ dangerous command patterns. The host machine is never exposed.
Can browser tools test real login flows?
Yes, when explicitly requested by the user for the intended site. Runtime guidance allows authenticated browser-flow testing, while sensitive credential values are redacted from structured audit tool-argument logs.
How does HybridClaw choose web_fetch vs browser tools?
Default is web_fetch for static/read-only retrieval. HybridClaw escalates to browser automation for JavaScript-heavy apps, auth-gated flows, interaction steps, or when web_fetch reports hints like javascript_required, spa_shell_only, empty_extraction, boilerplate_only, or bot_blocked.
Is the audit trail immutable?
Audit logs are append-only and hash-chained per session, so modifications are tamper-evident. Use hybridclaw audit verify <sessionId> to validate integrity, and hybridclaw audit approvals --denied to review denied actions.
What happens if instruction files are changed?
HybridClaw checks core instruction files before TUI startup. If they changed, you must approve them first. Use hybridclaw audit instructions to review status and hybridclaw audit instructions --approve to approve updates. Approval decisions are audit logged.
Can I use it without Discord?
Absolutely. Run hybridclaw tui for the terminal UI. The same assistant behavior and context model is shared across supported interfaces.
What is the heartbeat system?
A configurable periodic poll (default every 30 minutes) where the agent wakes up, checks its HEARTBEAT.md file for tasks, and proactively reaches out if something needs attention. Silent acknowledgments are discarded to avoid noise.
Can different Discord channels use different bots?
Yes. Each channel can be independently configured with its own HybridAI chatbot, model, and RAG setting using the bot set and model set commands.
How does scheduling work?
Three modes: cron expressions for recurring tasks, ISO-8601 timestamps for one-shot execution, and simple intervals (minimum 10 seconds). All tasks persist in SQLite and one-shot tasks auto-delete after running.
What is HybridAI?
HybridAI is the bot platform that powers HybridClaw. It provides managed chatbots with document knowledge bases, retrieval-augmented generation (RAG), and multi-model support through an OpenAI-compatible API. You create and configure bots on the HybridAI platform, then connect them to HybridClaw.
What AI models does it support?
HybridClaw works with any model available through the HybridAI API. You can list available models with model list and switch per channel at any time.
Does the agent remember things between conversations?
Yes. The agent writes daily memory files (memory/YYYY-MM-DD.md) and curates long-term insights into MEMORY.md. Before session compaction, a dedicated flush turn ensures durable notes are persisted. These files plus SOUL.md are injected into every request, giving the agent continuous, growing context.
Can I use skills from CLAUDE.md or OpenClaw?
Yes. HybridClaw supports OpenClaw/CLAUDE-compatible SKILL.md frontmatter, including always, requires, and metadata fields. Skills can be loaded from workspace, project .agents/skills, personal (~/.codex/~/.claude/~/.agents), bundled, or enterprise extra directories; higher-priority tiers override lower ones. Trust-aware scanning blocks risky personal/community skills while allowing vetted bundled skills.
