v2.7.11: The 'LLM API Error' Flood Is Gone, Sessions Resume Where They Left Off, and Model Switches Finally Stick

There is a specific kind of bug that erodes trust faster than an outage: the one that fills your screen with alarm when nothing is actually on fire. v2.7.11 kills the worst offender, the "LLM API Error" messages that could flood the chat panel in a torrent, and along the way it makes your sessions resume where they left off, close cleanly when idle, and actually keep the model you switch to.

The error flood is gone

When a model call failed and the agent retried in a loop, each failure could post its own "LLM API Error" message into the chat. A handful of retries and your conversation was buried under a wall of identical red alerts. The underlying call might recover a second later, but the panel already looked like a catastrophe.

Two things caused it, and both are fixed:

• One alert, not a hundred. Error alerts are now throttled to at most one every 30 seconds per employee, no matter how many calls fail in between. You still get told something went wrong, once, clearly, instead of being drowned in duplicates. The raw response is still relayed underneath, untouched. Only the noisy chat broadcast is rate-limited.
• No more false alarms. The old detector was too eager. If an agent ran a command that merely mentioned an error phrase, or reasoned out loud about one, that text could trip the alarm and start a self-sustaining loop of alerts about errors that never happened. The detector now only fires on a genuine error signal, a real error prefix or an actual HTTP failure code, so an agent talking about an error is no longer mistaken for one.

The net effect is calm. When something truly breaks you hear about it once. When nothing breaks, you hear nothing.

Sessions resume where they left off

This is the quiet favorite of the release. Until now, if a session ended for a reason you did not choose, the container was swept for being idle, the backend restarted, something crashed, the employee could come back with a blank memory of the conversation, as if you had never spoken. Maddening, especially mid-task.

The rule now is simple and predictable:

• You cleared it, or you switched the model? Fresh start. Pressing Clear session, or changing the employee's intelligence, begins a clean session, exactly as you would expect.
• It died for any other reason? It resumes. Idle sweep, crash, restart, whatever it was, the employee wakes up exactly where you left it, with the conversation intact, and picks up as if nothing happened.

Resume is now the default, and only the two deliberate actions start fresh. So any future hiccup we did not foresee lands in the safe case automatically: your conversation survives.

Idle sessions close cleanly

To free up resources, sessions that have sat untouched for a long stretch get their container stopped. That has always happened, but the way it announced itself was rough: a jarring browser popup, and worse, it used to claim a summary had been saved somewhere that did not exist.

Now it is honest and gentle. When a long-idle session is closed you get a clean, unobtrusive notification that says what actually happened: the container was stopped to free resources, and your conversation history is preserved. No fake promises, no blocking popup. And a bug where a still-open conversation could get stuck in a stop-and-respawn loop is gone, so closing an idle session actually closes it.

Model switches finally stick

If you ever picked a different model for the current session from the model picker and watched it snap right back to the employee's base model, that was a real bug, and it is fixed. A small error in how the switch was saved meant the change was recorded but the live session kept running on the old model, silently. Now the switch takes effect immediately and stays put, and your reasoning-effort setting resets to default on the switch, which avoids sending an effort level to a model that does not accept it.

A tidier engine picker

One small bit of housekeeping: the old OpenCode option in the intelligence picker is gone. It was a legacy entry no active employee was using, and removing it leaves a cleaner list of Claude, Codex, and your custom providers. Nothing you rely on changed.

The theme, again

If you have read a few of these notes, you know the refrain: the flashy features get the headlines, but reliability is what earns the daily trust. An error panel that stays quiet unless something is really wrong, a conversation that survives a crash, a model switch that means what it says. None of it is glamorous. All of it is the difference between an employee you lean on and one you keep double-checking.

v2.7.11 is live now. If your chat ever lit up with a wall of error alerts, or an employee ever came back having forgotten a conversation, those days are over.