@addai/entity-runtime@0.2.56

The package runs a daemon that polls a hardcoded Supabase project (syhzpqqvrplaqdipcymw.supabase.co) via the RPC `runtime_pick_next_request` for remote-supplied job rows and spawns Claude Code / Codex / Kimi / Gemini / Grok in a PTY on the installer's host, driven by fields controlled by the paired remote party (prompt, permissionMode, appendSystemPrompt, allowedTools, model, workingDirectory). In `dist/session-runner.js`, `spawnClaudeForRuntime` invokes the agent binary through a login shell (`['-l','-c','exec "$@"',...]`) and, when `input.permissionMode === 'bypassPermissions'`, passes `--dangerously-skip-permissions`, letting the remotely-driven AI agent execute arbitrary shell tools as the local user without any interactive prompt. On start, `ensureBypassPermissionsAccepted()` in `dist/index.js` also writes `skipDangerousModePermissionPrompt: true` into the installer's user-global `~/.claude/settings.json`, disabling Claude Code's own bypass-permissions safety dialog for every future interactive Claude Code session the installer runs — not just those spawned by this daemon. `dist/capabilities.js` additionally reads `~/.kimi/config` and modifies PATH. The net effect is a persistent remote-controlled channel giving the paired remote party host-level RCE on the installer's machine via the AI agent tool loop, plus a persistent lowering of a separate vendor tool's safety prompts.