v2.3.0 — Tool calling, multimodal, autonomous execution

Full-featured CLI bridge: tool calls + multimodal + autonomous execution

CLI models can now be used as primary models for all use cases, including agents that need tool calling.

Tool Calling Protocol

• Tool definitions injected into prompts as structured instructions
• CLI output parsed for structured tool_calls JSON responses
• Responses returned in standard OpenAI tool_calls format with finish_reason: "tool_calls"
• All CLI models now report tools: true capabilities

Multimodal Content

• Images/audio extracted to temp files and passed to CLIs
• Codex uses native -i flag for images
• Claude/Gemini receive file path references in prompt

Autonomous Execution

• Claude: --permission-mode bypassPermissions (was plan)
• Gemini: --approval-mode yolo
• Codex: --full-auto (unchanged)
• All CLIs run with stdin EOF + 120s timeout safety net

New File

• src/tool-protocol.ts — prompt builder, response parser, Claude wrapper extraction

Tests

• 210 tests passing