Remocode v1.1.3 — AI Supervisor, Auto-Yes & GPT-5 Support

Big update for hands-off AI coding. The AI Supervisor now runs the same battle-tested pipeline as the manual test button — including arrow-key menu navigation. A new Auto-Yes button handles simple approvals without AI. And GPT-5 reasoning models finally work.

AI Supervisor — Full Autonomous Pipeline

The 🧠 brain button now uses the exact same execution logic as the ✍ test button. Previously, the supervisor had a stripped-down decision handler that couldn't navigate Claude Code's arrow-key menus — it would send raw numbers instead of arrow-down presses, which did nothing.

Now it handles everything:

• ●Numbered menus: Converts option numbers to arrow-key presses (e.g. option 3 = two down arrows + Enter)
• ●Yes/No prompts: Sends the appropriate keystroke
• ●Text responses: Types the response and presses Enter
• ●Rejections: Sends Escape
• ●Escalations: Shows in the AI panel without sending any keystroke

The supervisor reads your project brief, respects autonomy settings, and uses your configured AI model — all while running on a 2-second scan loop that auto-detects when the terminal needs input.

Auto-Yes Button

New Y button on each pane's title bar. One click to toggle — no brief, no AI calls, no cost.

How it works: Scans the terminal every 2 seconds for numbered menus where the first option starts with "Yes". When found, it presses Enter to select it. That's it.

When to use it: Perfect for Claude Code sessions where you trust the agent and just want to keep it moving. For smarter decisions (rejecting risky actions, answering questions), use the 🧠 supervisor instead.

• ●Mutual exclusion with supervisor — turning on one turns off the other
• ●Right-click for a log of everything it auto-approved
• ●3-second cooldown between actions to avoid double-firing

AI Panel — Live Supervisor Log

The AI panel now shows every supervisor decision in real-time — not just escalations. When the supervisor detects a prompt, you'll see:

• ●Thinking... (yellow) — AI is processing the terminal content
• ●Done (green) — shows the action taken, reasoning, and what was sent
• ●Error (red) — if the AI call failed

This gives you full visibility into what the supervisor is doing without checking debug logs.

GPT-5 Reasoning Model Support

GPT-5-nano and other reasoning models (gpt-5, o-series) now work correctly as supervisor models.

The problem: Reasoning models allocate a reasoning_tokens budget from max_completion_tokens. With the supervisor's 256-token limit, all tokens went to reasoning — leaving zero for actual output. Every response came back empty.

The fix: Supervisor calls now set reasoning_effort: 'minimal' for reasoning models, which produces zero reasoning tokens and leaves the full budget for the JSON response. All AI providers also now support a configurable maxTokens parameter.

Paste Fix

Cmd+V stopped working in the terminal after a previous update.

Root cause: A paste guard checked if the active element was a TEXTAREA and blocked the paste. But xterm.js uses a hidden TEXTAREA internally (xterm-helper-textarea) — so the guard was blocking xterm's own paste handler.

The fix: The guard now allows xterm's helper textarea through: if the active element has the class xterm-helper-textarea, paste proceeds normally.

Memory Leak Fix

The supervisor's scan loop was writing debug logs via IPC on every 2-second tick — even when nothing changed. Over hours, this accumulated 57,000+ IPC calls and inflated memory usage to 13+ GB.

The fix: All debug logging now happens after the dedup check. If the terminal content hasn't changed, zero IPC calls are made. The scan loop is now effectively silent when idle.

Full Changelog

Added

• ●Auto-Yes (Y) button on each pane — one-click approval for "Yes" menus
• ●AI panel live logging for all supervisor decisions (thinking/done/error)
• ●GPT-5 and reasoning model support (reasoning_effort: 'minimal')
• ●Configurable maxTokens across all AI providers (OpenAI, Anthropic, Google, Groq, Ollama)

Fixed

• ●AI Supervisor now handles arrow-key menus correctly (same as test button)
• ●Cmd+V paste broken in terminal (xterm helper textarea guard)
• ●Memory leak from supervisor debug logging (57K IPC calls, 13GB RAM)
• ●GPT-5-nano returning empty responses (reasoning tokens consuming full budget)
• ●Supervisor detectedPrompt context increased from 15 to 20 lines

Changed

• ●Supervisor scan only logs after dedup (no more per-tick IPC overhead)
• ●Supervisor onDecision callback now includes pane name for UI logging

---

Download: [Remocode-latest.dmg](https://dl.remocode.org/releases/Remocode-latest.dmg)