v2.7.9: Jump to Any Prompt Instantly With the New Prompt Rail, Plus Cleaner Streaming and the Full Effort Scale

Long conversations have always had the same quiet tax: finding something you said an hour ago means scrolling, and scrolling, and scrolling. v2.7.9 fixes that with a Prompt Rail, a set of ChatGPT-style markers down the right edge of the chat that let you see the whole shape of a conversation and jump anywhere in it in one click. It ships alongside two fixes worth their own paragraphs: streaming that no longer scrambles, and the full reasoning-effort scale finally reaching custom models.

The Prompt Rail

Look at the right edge of a chat now and you will see a stack of thin horizontal bars, one for each thing you said. Every prompt gets its own bar, newest at the bottom, and the bar for the message you are currently reading stays lit so you always know where you are in a long thread. It is a map of the conversation you never had before, sitting quietly in space that used to do nothing.

Hover the rail and it opens a scrollable prompt sheet: every prompt you have sent, listed and truncated so you can scan them fast. It fades in against the right edge of the chat, so it never covers your file explorer, and it follows your mouse: stay over it and it stays open, move away and it closes itself. No buttons to hunt for.

The sheet is built for real, sprawling conversations, not toy ones:

• It opens where you are. Instead of dumping you at the top of the list, the sheet scrolls straight to the prompt you are currently reading, so you land exactly where you left off.
• Search built in. A search box at the top filters the list live as you type, narrowing dozens of prompts down to the one you half remember, with a counter showing how many matched.
• Badges that tell you the source. Prompts that came in over a channel wear it: WhatsApp, Telegram, Slack, Teams, Discord, Gmail, Outlook. Prompts from the system carry theirs too (Scheduled task, Manual task, Kanban, Reply, Context). The raw prefixes that used to clutter these messages are stripped and replaced with a clean label, so the list reads like language, not machine noise. Pure system chatter with no real content stays out of the rail entirely.
• Clean entries. Each row has its number, a two-line preview, a relative timestamp, and an "ACTIVE" pill on the one you are reading.

Click to jump, and actually load the conversation

Clicking a prompt jumps the chat straight to it and lights its bar. The important part is what happens underneath. Clicking a prompt now loads the full conversation from that point all the way to the end, not just a small window around it. Click your first message and you get the whole conversation back. Click something from the middle and everything from there down streams in, with a spinner while the tail is in flight and your scroll held steady so nothing jumps under your finger.

That "load everything from here to the end" behavior is the difference between a navigation feature that works and one that only pretends to. You should be able to click a point in your history and trust that the rest of it is really there.

Two guardrails keep it fast and sane on huge threads. The rail itself caps at 60 visible bars so a conversation with thousands of prompts never overflows the column, while the prompt sheet always lists every single prompt, so nothing is ever unreachable. And the lazy loading that keeps big conversations opening quickly is completely untouched underneath.

A pile of scroll fixes

Navigation is only as good as its aim, so a lot of this release went into making jumps land exactly where they should. Prompt jumps no longer overshoot or land in the wrong spot. Jumping to an old, not-yet-loaded message now widens the window and reveals it properly. The lit bar tracks the reply you are actually reading instead of flickering dark between prompts. Auto-scroll no longer yanks you back to the bottom when a reply streams in after you have jumped somewhere. Clicking the very first prompt no longer hangs the chat. And messages that arrived over a channel, like an inbound WhatsApp or Telegram, now get their own bar with clean readable text in the sheet, where before they were silently dropped from the rail altogether.

None of those are glamorous on their own. Together they are the difference between a feature you trust and one you learn to avoid.

Streaming that stops scrambling

If you run custom models, you may have seen replies arrive jumbled under load, where "Je vais corriger" came back as "Je corr vaisiger", or text landed in bursts. That was a race: pieces of the streamed reply were being sent without waiting their turn, so they arrived out of order and the interface dropped the stragglers. The bridge that streams custom-model replies now sends everything through a single ordered chain with a short batched flush, so the words arrive in the order they were written. Clean streams, even when the system is busy.

The full effort scale for custom models

The reasoning-effort selector for custom models used to stop at high, and any level above that was quietly discarded before it reached the provider. Now the selector follows each model's real vocabulary. Models that support it get the full range up through xhigh and max, OpenAI-native models get the exact set they accept, and the interface never offers a level the model would reject. The upshot: when you dial a capable model up to max, it actually runs at max instead of being silently capped.

And a sharper set of eyes

The built-in vision fallback, the local model an employee leans on to read an image when its own model cannot, got an upgrade after a benchmarking pass. It reads small UI text and icons far more accurately, no longer flips light and dark, and answers multi-part questions instead of dodging them. You will not see it directly, but the employees that rely on it now describe what they see with noticeably more precision.

v2.7.9 is live now. Open a long conversation, glance at the right edge, and try jumping around. The scrolling tax is paid off.