Conol 0.38.0 — Replies that land where they should

This release is about continuity — making the conversation with your agent hold together across a reply, a long session, or a task running in the background.

Reply to any message on Telegram and the agent keeps the thread

Why. Telegram is a stream of messages, and until now the agent read every incoming message as "the latest thing you said." If the agent had sent you three updates and you replied to the first one with "change this," it had no idea which one you meant. Worse: any message you sent while the agent was mid-question got swallowed as the answer to that question, even when you were replying to something else entirely. And if a background task — a timer, a long job — pinged you on Telegram, replying to it started a brand-new conversation that knew nothing about the task.

How to use it. Just use Telegram's normal reply gesture. Three things now work the way you'd expect:

• Reply to a specific message and the agent sees exactly which one — the quoted context comes through, so "yes, do that" or "fix this part" acts on the message you picked, not the most recent one.
• Reply to a message a background task sent you, and your answer is routed to that task's session — so answering "yes, ship it" reaches the job that asked, instead of a fresh chat.
• Ask about your other work. When several of your sessions are active, the session on your phone can now see what the others have been doing and pass your message along to the right one, rather than answering blind.

Long conversations keep going

Why. Every model has a memory limit, and a long, involved session would eventually fill it and stall — leaving you to start a new chat and lose the thread of what you were doing.

How to use it. Nothing to do. When a conversation starts filling the model's memory, Conol now compacts it automatically: the older turns are summarized and the recent thread is kept intact, so the agent keeps its footing and you keep working in the same session. You'll simply notice that long sessions no longer hit a wall.

Your scheduled timers, right in the composer

Why. You could ask the agent to check on something later or run a task on a schedule, but once set, those timers were tucked away in settings — easy to forget, awkward to cancel.

How to use it. A small chip now sits above the message box whenever you have timers running. It shows how many there are and when the next one fires. Tap it to see the full list and cancel any you no longer need — right where you're already working.

Longer sandbox and coding runs

Why. When the agent clones a private repo or runs the coding tools, that traffic goes through a secure proxy so your credentials never enter the sandbox. On production, that proxy previously ran inside the web app, which capped any single operation at about a minute — long clones and longer agentic coding runs could get cut off partway.

How to use it. Automatic. Production now runs a dedicated gateway for that traffic with no such ceiling, so bigger clones, builds, and coding runs finish cleanly. You'll only notice that things that used to time out now complete.

Fixes

• Tidier chat with tool calls — some models left blank gaps between groups of tool calls (a stray empty bubble), and the plan strip above the composer could linger after a session had finished. The transcript now renders clean, and the strip clears once the agent settles.