v1.3.0 — Remote tool execution, Codex tool, agent balance

Highlights

Remote agents can now execute a single named tool directly (no LLM loop) via remote.call / co call, a new codex tool drives Codex through its native app-server, and agents publish their credit balance in the ANNOUNCE profile and /info. Browser input is now humanized (curved pointer paths, real wheel scroll, per-key dwell) to pass fingerprint/CDP detectors.

✨ Features

• Direct remote tool execution — remote.call(...) / co call <addr> <tool> runs one tool with no LLM loop, gated by the host permission whitelist (#225)
• Codex tool — drive Codex via its native app-server: session resume, live streaming, per-action approval (#226)
• Agent balance — published in the ANNOUNCE profile and /info for co/* managed-key agents (#231)
• Humanized browser input — pointer/keyboard/scroll + fingerprint/CDP evasion (#224)
• Gemini 3.6 Flash model support (#220)
• browser-workflow-skill-builder added to useful_skills (#207)
• bash tool's description is now optional (#236)

🔒 Security

• Tightened the default remote-exec whitelist: removed entries that allow arbitrary code execution (pytest, perl, npm test) or dumping arbitrary file/env contents (cat, head, tail, nl, grep, env, printenv). These now require human approval instead of being auto-permitted (#240)

🐛 Fixes

• Windows GBK/cp936 encoding errors resolved (#233)

📚 Docs

• Synced docs with the core refactor and current defaults (#227)

Install

pip install --upgrade connectonion

Full Changelog: v1.2.1...v1.3.0